Integrating machine learning and geomorphic indices to assess tectonic geomorphology in the Karoun River Basin, Zagros, SW Iran

Abstract

The geomorphological expression of tectonic activity in the Karoun River Basin, a tectonically active region within the Zagros Mountains of Iran, was investigated. The basin’s complex geological setting, influenced by the collision of the Arabian and Eurasian plates, provides an optimal environment for analyzing tectonic-geomorphic interactions. A combination of geomorphic indices, including the Stream-Gradient Index (SL), Asymmetric Factor (Af), Hypsometric Integral (Hi), and Valley Floor Width to Valley Height Ratio (Vf), was employed to assess tectonic influences on landscape development. Elevated SL values were interpreted to suggest active uplift, while asymmetric drainage patterns (Af) and narrow valley profiles (Vf) were found to further support tectonic dominance. Hypsometric analysis (Hi) indicated youthful landforms undergoing continuous tectonic modification. To augment conventional geomorphic assessments, advanced machine learning techniques, specifically Random Forest (RF) and Convolutional Neural Networks (CNNs), were utilized to model the spatial distribution of geomorphic indices. Interpretability methods such as SHAP (SHapley Additive exPlanations) were applied to elucidate the relationships between tectonic processes and geomorphic features, enhancing model accuracy and mechanistic interpretation. The Index of Active Tectonics (Iat), derived through GIS-based analysis, was used to categorize the basin into three tectonic activity classes: Class 1 (very high activity, 24% of the area), Class 2 (high activity, 63%), and Class 3 (moderate activity, 10%). The findings highlight the significance of tectonic geomorphology in natural hazard evaluation, particularly landslide susceptibility in steep, tectonically uplifted terrains. Additionally, the examination of river terraces contributes to understanding historical landscape responses to tectonic and climatic forcing, advancing knowledge of long-term geomorphic evolution. The integration of traditional geomorphic indices with machine learning establishes a robust analytical framework for future research in tectonically active regions, with implications for geological hazard assessment and environmental planning.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-26650-5.

Introduction

Tectonic geomorphology is a critical field of study that examines the interactions between tectonic processes and landscape evolution^1–3. In recent years, the significance of understanding the interplay between tectonic processes and geomorphological evolution has gained increasing attention within the scientific community. The Karoun River Basin, situated in the tectonically active Zagros Mountains of Iran, presents a unique opportunity to explore these interactions. The region’s complex geological history, shaped by the collision of the Arabian and Eurasian plates, has resulted in diverse landforms and significant geomorphic features, including anticlinal ridges, fault traces, and deeply incised valleys^4–6. However, accurately quantifying the relative intensity of this tectonic activity across such a large and lithologically varied basin remains a significant challenge. Traditional assessments often rely on the qualitative or sequential interpretation of individual geomorphic indices, which may not fully capture the complex, non-linear relationships between tectonic forcing and landscape response.

This study aims to address this challenge by developing a more robust, integrated framework for assessing relative tectonic activity. The research builds directly upon previous foundational work in the Zagros region. For instance, the quantitative approaches to tectonic activity were established by Pinter and Keller⁷, and Keller, et al.⁸ discussed active tectonics and landscape evolution, offering insights into the processes that influence geomorphic evolution. Alavi⁹ provided a foundational understanding of the regional stratigraphy of the Zagros fold-thrust belt. Allen, et al.¹⁰ examined the reorganization of the Arabia-Eurasia collision, contributing to the understanding of tectonic processes in the region. Furthermore, El Hamdouni, et al.¹¹ established a framework for assessing active tectonics, which is foundational for the methodology used in the current study.

In last decade, Dehbozorgi, et al.¹² provided a quantitative analysis of tectonic activity in the Sarvestan area laid the groundwork for using geomorphic indices, but the current study enhances this methodology by integrating multiple indices into a single evaluative framework. Arian, et al.¹³ evaluated seismic sources and neo-tectonics in Tehran. These studies demonstrated the utility of geomorphic indices but were often limited to smaller areas or did not synthesize their results into a unified, regional-scale evaluation. Furthermore, research has focused on the complexity of tectonic interactions in Zagros¹⁴. Bagha, et al.¹⁵ assessed relative tectonic activity in the Tehran basin, highlighting the significance of structural controls. The current study expands this approach to a larger area, providing a more holistic view of tectonic influences. Moreover, Ehsani and Arian¹⁶ analyzed relative tectonic activity in the Jarahi-Hendijan basin, contributing to the body of knowledge on geomorphic indices, but did not incorporate the comprehensive index (Iat) developed in this study^17–20.

Recent literature has significantly advanced this field, revealing new insights into the complex relationships between tectonics, hydrology, and geomorphic features. For instance, Susanth, et al.²¹ investigated the morphology of submarine canyons in the Andaman Sea, highlighting the need for a deeper understanding of their evolutionary dynamics and the tectonic influences that shape them. Similarly, Shah, et al.²² explored the geomorphic consequences of microbial mats in macrotidal environments, emphasizing the importance of interdisciplinary approaches in understanding these complex systems. Areggi, et al.²³ and Turconi, et al.²⁴ focused on slow deformation rates in Slovenia and NE Italy, illustrating the challenges these pose for geomorphological analysis and the necessity for innovative methodologies to address existing knowledge gaps. Sarafaraz, et al.²⁵ investigated river water quality prediction for the Aji-Chai River in Tabriz, Iran, utilizing machine learning models, notably Support Vector Machine (SVM), to simulate key indices such as TDS, EC, and SAR, demonstrating the effectiveness of SVM in comparison to the QUAL2Kw model. Eventually, Asghari, et al.⁶ examined the impact of tectonic activity on geohazards and gas reservoir pressure dynamics in the Zagros Fold-Thrust Belt, highlighting the significance of understanding seismic influences on landscape evolution and resource management. Their findings complement the current study by emphasizing the critical role of tectonic processes in shaping both geomorphology and natural hazard assessments in the region, but a significant gap exists in leveraging modern computational techniques to objectively synthesize geomorphic data with this geological knowledge^26–29.

Therefore, the primary gap this study seeks to fill is the lack of a reproducible, scalable methodology that integrates established geomorphic principles with advanced machine learning (ML) to move from qualitative interpretation to quantitative, predictive assessment. Many studies have primarily relied on traditional methods, which may not fully capture the complexities of tectonic influences on landscape evolution. Furthermore, the integration of machine learning techniques in this field has been limited, despite their potential to enhance predictive modeling and provide deeper insights into geomorphic processes^30–33. This study aims to fill these gaps by investigating the tectonic influences on geomorphology within the Karoun River Basin, utilizing advanced machine learning techniques to improve the understanding of landscape evolution.

The aforementioned studies have successfully demonstrated the utility of geomorphic indices in tectonic analysis. However, a significant gap remains in the systematic integration of these indices with modern computational techniques to create a more robust, predictive, and objective assessment framework. Many studies have primarily relied on qualitative or semi-quantitative comparisons, which may not fully capture the complex, non-linear relationships between landscape form and tectonic forcing. Furthermore, the application of machine learning for pattern recognition and predictive modeling in tectonic geomorphology has been limited, despite its potential to handle large, multivariate geospatial datasets and uncover latent patterns not readily apparent through traditional analysis.

The Great Karoun River has many tributaries that flow southwestward, the largest of which is the Dez River. Although the Dez is the largest and most voluminous river in Iran after the Karoun, it enters the Karoun River before flowing into the Persian Gulf^34,35. When studying the Karoun River without the Dez, it is referred to simply as the Karoun. However, when both rivers are considered together, the term “Great Karoun” is used. Since both rivers were studied, the term “Great Karoun” is employed in this article.

The primary objective of this study was to develop and validate an integrated framework for assessing relative tectonic activity in the Great Karoun River Basin to overcome key limitations inherent in traditional geomorphic analysis. This was achieved by synthesizing traditional geomorphic indices into a unified Index of Active Tectonics (Iat) and leveraging machine learning models to model and interpret the spatial patterns of tectonic geomorphic expression. The novelty of this research lies not in the creation of new indices, but in the novel coupling of established methodologies—specifically, the use of Explainable AI (XAI) techniques to interpret machine learning models and bridge the gap between statistical prediction and geological process understanding. These techniques, specifically SHAP (SHapley Additive exPlanations), were employed to interpret the models and quantify the relative importance of these features. This approach provides a transferable methodology for regional-scale tectonic assessment in other active orogens.

In practice, the ML models learn the complex, non-linear relationships between the landscape’s form (e.g., slope, curvature, drainage density, distance to faults) and its tectonic expression (the geomorphic indices). This allows for a more objective, data-driven synthesis of the study area than would be possible through manual interpretation alone. The Convolutional Neural Network (CNN), for instance, works by processing spatial patterns in Digital Elevation Model (DEM) tiles to identify subtle features diagnostic of tectonic activity, such as specific topographic signatures associated with fault propagation folds. Ultimately, this integration provides a rigorous, reproducible framework that quantifies both the patterns of tectonic activity and the geological reasoning behind them, moving the analysis beyond qualitative assessment towards a more robust, predictive science.

Geological and tectonic setting

The Karoun River Basin is situated within the Zagros Fold-and-Thrust Belt (ZFTB), the seismically active external margin of the Arabian Plate resulting from its ongoing collision with the Eurasian Plate since the Late Cretaceous (Fig. 1)^18,36–40. The study area—which spans an extensive 60,500 km²—encompasses a significant geological transition, spanning from the internal Sanandaj-Sirjan metamorphic zone, across the High Zagros Fault, through the Simply Folded Belt with its spectacular NW–SE trending anticlines, and onto the alluvial plain of Khuzestan^6,27,41,42.

Fig. 1.

a Location of the study area within the tectonic framework of the Zagros orogenic belt, southwestern Iran. The map is overlain on a hillshade DEM for topographic context ^6,69,72,73. b Generalized geological map of the Karoun River Basin, showing major lithological units and structural features. Key tectonic zones are delineated: Sanandaj-Sirjan Zone (SSZ), High Zagros Fault Zone (HZF), Simply Folded Belt, and the Khuzestan foreland plain. Major faults (e.g., Main Recent Fault—MRF, Mountain Front Fault—MFF) are indicated with bold lines. Created using ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home) with base layers from the Geological Survey of Iran (1:250,000 scale).

This tectonic evolution has produced a complex structural architecture dominated by hundreds of folds and major fault systems, including the Main Recent Fault (MRF), the High Zagros Fault (HZF), and the Mountain Front Fault (MFF) (Fig. 2)^10,43,44. These structures are the primary controls on regional topography^45–49, creating zones of pronounced uplift and subsidence that directly influence drainage patterns and landscape evolution^50–53.

Fig. 2.

Detailed structural map of the western Zagros highlighting major active fault systems. Faults are compiled from field mapping data and NIOC (National Iranian Oil Company) structural maps (1:100,000 scale). Key faults shown include the High Zagros Fault (HZF), Mountain Front Fault (MFF), Main Zagros Reverse Fault (MZRF), Zagros Foredeep Fault (ZFF), Main Recent Fault (MRF), and the dextral Kazerun and Sarvestan fault systems. Fault kinematics are indicated where data is available. The background is a hillshade model derived from SRTM data to illustrate the strong geomorphic expression of the structures. Processed using ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home) with the Georeferencer plugin (RMSE < 5 m). Coordinate system: WGS 1984 UTM Zone 39N.

The lithology is characterized by a thick (~ 10–12 km) sequence of Proterozoic to Cenozoic sedimentary rocks^54–56. This sequence exhibits strong mechanical contrasts, ranging from resistant carbonates (e.g., Asmari and Bangestan formations) to incompetent evaporites and marls (e.g., Gachsaran and Mishan formations)^6,57,58. This lithological heterogeneity is a critical control on geomorphic processes^19,59,60, as it governs erosional resistance and dictates how the landscape responds to tectonic forcing^56,61,62. The interplay between persistent tectonic stresses and the erosion of these variably resistant units is the fundamental driver of the geomorphology under investigation^20,57,63, creating a landscape predisposed to specific hazards^5,64–68. The steep, tectonically uplifted terrain, underlain by these weak units, is highly susceptible to large-scale slope failures and landslides, a characteristic feature of the Zagros orogen^1,69–71.

The hydrology of the basin is intrinsically linked to this geological framework^20,74–76. The Karoun River, Iran’s largest river, originates in the high-relief Zagros Mountains^20,77–79, where steep gradients promote rapid runoff and high sediment transport rates. Major faults and folds structurally control tributary patterns, while significant groundwater resources are stored within alluvial aquifers along the foothills and plain, replenished by river infiltration and precipitation^73,80–83.

Methodology

In this study, machine learning techniques were not employed to replace traditional geomorphic indices, but to augment them. The integration was designed to achieve two primary objectives: first, to exploit the ability of ML models to assess and model index values in areas with complex terrain or sparse data by learning from multivariate relationships; and second, to utilize interpretability frameworks (XAI) to identify the dominant topographic and geological features controlling the spatial patterns of the indices, thereby providing a data-driven validation of their tectonic significance.

The tectonic geomorphology of the Karoun River Basin was evaluated through an integrated approach combining geomorphic indices, GIS-based spatial analysis, and machine learning techniques. The methodological framework (Fig. 3) consisted of four primary components: (1) data acquisition and preprocessing, (2) calculation of geomorphic indices, (3) machine learning model development, and (4) model validation and interpretation.

Fig. 3.

Flowchart of the research including a clear start and finish, as well as loops to indicate iterative processes.

Data collection

Data collection is the foundational step in this research, involving the acquisition of various datasets necessary for the analysis of geomorphic indices and tectonic influences^84,85. The following types of data were collected:

Topographic Data: A 30 m resolution DEM derived from the Shuttle Radar Topography Mission (SRTM) was selected as the primary topographic dataset to ensure consistency and regional coverage. This dataset was supplemented with 1:25,000 scale topographic maps from the National Cartographic Center of Iran for visual validation and context.

Geological Data: Geological maps and tectonic activity reports were sourced from local geological surveys and academic publications to understand the regional tectonic framework.

Hydrological Data: Streamflow data and watershed characteristics were collected from hydrological databases to assess the influence of hydrology on geomorphic processes.

Geomorphic index calculation

The Great Karoun basin is subdivided into 72 sub-basins (Fig. 4). No indexes could be calculated for the 38th basin, as it is located in the alluvial plain of Khouzestan, which constitutes 1% of the area of the Great Karoun basin. The geomorphic indices used in this study include the Stream-Gradient Index (SL), Asymmetric Factor (Af), Hypsometric Integral (Hi), and Valley Floor Width to Valley Height Ratio (Vf). Each index is calculated using the Eqs. (1) to (6):

Fig. 4.

Delineation of 72 sub-basins in the Great Karoun watershed. Basin boundaries derived from 30m SRTM DEM using the Strahler order-6 drainage network in WhiteboxTools 2.2. Sub-basins were automatically delineated based on it to ensure a consistent scale of analysis across the basin. The 38th basin (Khuzestan alluvial plain, 1% of total area) was excluded from index calculations due to lack of topographic expression. Coordinate system: Iran Conformal Conic.

Stream-gradient index (SL)

The SL index, reflecting variations in channel steepness, is calculated as Eq. (1):

1

where ΔH is the elevation change (m), ΔLr is the reach length (m), and Lsc is the horizontal distance from the watershed divide to the reach midpoint (m). Values were classified into three tectonic activity classes (Table 1): Class 1 (SL ≥ 500, high activity), Class 2 (300 ≤ SL < 500), and Class 3 (SL < 300). The SL index can be used to evaluate relative tectonic activity^5,8,86.

Table 1.

Value of At (total sub-basin area), the classes of SL (stream- gradient index), Af (drainage basin asymmetry), Hi (hypsometric integral), Vf (valley floor width-valley height ratio), Bs (drainage basin shape), and J (mountain front sinuosity), and values and classes of Iat (relative tectonic activity) and corrected Iat.

Basin No	At (Km2)	Class of SL	Class of Af	Class of Hi	Class of Vf	Class of Bs	Class of J	Value of Iat	Class of Iat	Class of changed Iat
1	360.52	3	3	2	2	3	2	2.50	4	3
2	1607.97	3	3	2	3	3	2	2.66	4	3
3	652.67	1	2	3	2	3	2	2.16	3	2
4	91.90	2	1	3	N/A	3	2	2.20	3	2
5	323.29	2	1	1	2	3	2	2.00	3	2
6	437.27	1	1	2	3	3	2	2.00	3	2
7	218.39	1	1	3	1	3	2	1.83	2	1
8	920.67	1	2	3	N/A	3	2	2.20	3	2
9	512.22	1	3	1	1	3	2	1.83	2	1
10	201.62	2	3	1	1	3	N/A	2.00	3	2
11	327.46	3	3	1	2	3	2	2.33	3	2
12	193.52	2	1	1	1	3	2	1.66	2	1
13	365.85	1	1	3	2	3	2	2.00	3	2
14	901.72	2	2	2	1	3	2	2.00	3	2
15	341.10	3	1	1	2	3	N/A	2.00	3	2
16	714.45	3	2	1	1	3	2	2.00	3	2
17	683.73	3	2	1	2	2	N/A	2.00	3	2
18	1093.82	3	1	1	2	3	2	2.00	3	2
19	492.19	3	3	1	2	3	N/A	2.40	3	2
20	1148.28	3	1	1	1	3	N/A	1.80	2	1
21	1271.96	3	3	1	1	3	N/A	2.20	3	2
22	2924.41	3	2	1	2	3	1	2.00	3	2
23	4237.07	3	1	3	1	3	2	2.16	3	2
24	1575.84	3	2	1	2	3	2	2.16	3	2
25	2863.81	3	1	1	2	3	2	2.00	3	2
26	2570.98	3	1	1	1	3	3	2.00	3	2
27	2175.88	3	1	1	2	1	2	1.66	2	1
28	258.61	2	2	1	3	3	2	2.16	3	2
29	1005.37	3	1	1	1	3	2	1.83	2	1
30	593.87	3	3	1	3	3	2	2.50	4	3
31	45.78	3	2	2	1	3	2	2.16	3	2
32	115.00	2	1	2	2	3	2	2.00	3	2
33	910.15	1	2	1	3	3	2	2.00	3	2
34	1235.30	3	2	1	3	3	2	2.33	3	2
35	1640.80	3	3	N/A	N/A	1	3	2.50	4	3
36	967.20	1	3	N/A	N/A	3	2	2.25	3	2
37	3585.26	1	1	N/A	N/A	3	2	1.75	2	1
38	919.53	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
39	1993.41	3	1	N/A	3	3	3	2.60	4	3
40	398.71	1	1	3	3	3	2	2.16	3	2
41	757.20	1	1	1	2	3	N/A	1.60	2	1
42	2767.50	2	2	1	2	3	2	2.00	3	2
43	825.78	3	3	1	3	3	2	2.50	4	3
44	818.13	3	2	1	1	3	1	1.83	2	1
45	685.42	3	3	1	1	3	2	2.16	3	2
46	718.63	3	2	1	1	3	N/A	2.00	3	2
47	211.87	3	2	1	2	3	N/A	2.20	3	2
48	789.41	3	1	1	2	3	2	2.00	3	2
49	327.01	2	3	1	2	2	2	2.00	3	2
50	602.17	3	2	1	3	3	2	2.33	3	2
51	248.73	2	3	1	2	3	2	2.16	3	2
52	960.59	1	3	1	3	3	2	2.16	3	2
53	392.50	2	1	1	2	3	2	1.83	2	1
54	259.17	1	3	3	1	3	N/A	2.20	3	2
55	665.83	1	3	2	3	3	2	2.33	3	2
56	1548.77	3	2	1	2	3	N/A	2.20	3	2
57	980.46	3	2	1	1	3	2	2.00	3	2
58	505.81	2	1	1	2	3	2	1.83	2	1
59	405.59	3	1	1	1	2	2	1.66	2	1
60	556.56	2	1	1	2	3	N/A	1.80	2	1
61	462.91	2	2	1	3	2	2	2.00	3	2
62	811.73	1	1	3	1	3	2	1.83	2	1
63	843.92	3	1	1	2	3	2	1.83	2	1
64	707.33	3	1	1	2	3	2	2.00	3	2
65	397.48	3	2	1	3	3	2	2.33	3	2
66	686.03	1	1	2	1	3	2	1.66	2	1
67	762.51	3	2	1	2	3	N/A	2.20	3	2
68	1081.61	3	2	1	2	3	2	2.16	3	2
69	581.08	3	3	1	2	3	2	2.33	3	2
70	873.26	3	1	1	1	3	2	1.83	2	1
71	880.58	2	1	1	2	3	2	1.83	2	1
72	667.83	1	3	1	2	3	2	2.00	3	2

Not Applicable (N/A). The index could not be calculated due to the absence of the required geomorphic feature (e.g., no mountain front for J, or an alluvial valley for Vf).

It is recognized that the SL can be sensitive to lithological resistance, which may confound interpretations of tectonic activity. To mitigate this, high SL values were critically evaluated in the context of regional rock strength classifications (see Section "Geomorphic indices analysis"). Furthermore, it is acknowledged that normalized channel steepness indices (ksn) derived from slope-area analysis represent a more robust metric for future studies aiming to isolate tectonic signals from lithological noise^31,87.

Asymmetric factor (Af)

Drainage basin asymmetry, indicative of tilting due to tectonic forcing, is quantified as Eq. (2) ^21,62,88:

2

where Ar is the area right of the main channel and At is the total basin area. Values < 50 or > 50 suggest lateral tectonic tilt.

Hypsometric integral (Hi)

The Hi, representing the relative distribution of elevation within a basin, was derived from DEMs using Eq. (3) ^5,16.

3

where is the mean elevation, and E_min/E_max are the minimum/maximum elevations. High Hi (> 0.5) indicates youthful, tectonically active landscapes.

Valley floor width-to-height ratio (Vf)

Basin elongation, influenced by structural controls, was measured as:

The Vf index, sensitive to vertical uplift rates, is defined as Eq. (4):

4

where V_fw is valley floor width, and E_ld, E_rd, and A_sc are elevations of the left/right divides and stream channel, respectively. Low Vf (< 0.5) correlates with active incision^11,15,89.

Basin shape index (Bs)

Basin elongation, or basin shape index (Bs) influenced by structural controls^20,90,91, is defined as Eq. (5):

5

where Bl and Bw are basin length and maximum width.

Mountain front sinuosity index (J)

The J index, calculated for 430 mountain fronts, is defined as Eq. (6) ^11,92,93:

6

where Lj is the planimetric front length and Ls is the straight-line length. Low J (< 1.1) indicates tectonically active fronts.

Handling of non-applicable indices

The Iat for a sub-basin is calculated as the mean of the class values only for the indices that are applicable and calculable for that specific sub-basin (Table 1). Therefore, the absence of a mountain front is reflected in the indexing in the most scientifically valid way possible: the J index is excluded from the Iat calculation for that sub-basin. This approach is scientifically valid because:

1. It Reflects physical reality: A sub-basin lacking a mountain front is typically in a low-relief area (e.g., an alluvial plain), which is intrinsically associated with lower tectonic activity. Forcing a calculation here would be invalid.

2. It is a standard practice: This method is established in geomorphic literature to handle heterogeneous landscapes where not all indices are universally applicable.

3. It Prevents arbitrary bias: Estimating values for non-existent features would introduce significant error and artificial noise, weakening the analysis. This method allows the sub-basin’s class to be determined by the features that are actually present. As the examples of implementation for Table 1.

Basin 38—The Khuzestan alluvial plain

• This basin is a large, topographically subdued alluvial plain.

• SL, Af, Hi, Vf, Bs, J: None of these indices can be meaningfully calculated as there are no clear streams, valleys, or mountain fronts in the traditional sense. The landscape is dominated by deposition, not tectonic incision.

• Iat Calculation: Since no indices are applicable, the basin is justifiably classified as N/A (Not Applicable) for relative tectonic activity based on these geomorphic criteria. It is excluded from the final Iat summary statistics (% of area in Class 1, 2, 3).

Basin 4

• Unavailable data: Vf could not be calculated (N/A), likely due to a wide, poorly defined valley floor or data constraints.

• Available Data: Classes for SL(2), Af(1), Hi(3), Bs(3), J(2) are available.

• Iat Calculation: Iat = (2 + 1 + 3 + 3 + 2) / 5 = 11 / 5 = 2.20

• Class of Iat: 2.20 falls into Class 3 (Moderate activity). The absence of the Vf value did not require estimation; the average was simply taken over the five available indices.

Statistical basis for index classification

The class boundaries for each geomorphic index were initially defined based on established thresholds in tectonic geomorphology literature¹¹ to facilitate regional comparison. To validate and provide a data-driven rationale for this classification, a K-means cluster analysis was performed on the standardized (z-score) values of the six geomorphic indices (SL, Af, Hi, Vf, Bs, J) across the 71 sub-basins. The optimal number of clusters (k = 3) was determined using the elbow method, which minimizes the within-cluster sum of squares. The results of this analysis (Supplementary Fig. B1) corroborate the tripartite division of tectonic activity, showing distinct separations between clusters corresponding to high, moderate, and low relative activity. The final class ranges (Table 1) were therefore informed by both canonical literature values and the intrinsic statistical structure of the dataset.

Machine learning model development

To enhance the predictive accuracy of geomorphic indices, advanced machine learning techniques were implemented, including ensemble learning and deep learning architectures^6,94,95. The integration of machine learning techniques was pursued to overcome inherent limitations in the traditional, sequential analysis of geomorphic indices. While these indices are valuable diagnostic tools, their manual interpretation can be subjective and may fail to capture complex, non-linear relationships within the multivariate topographic and geological dataset. Therefore, ML was not used to replace but to augment the conventional approach. Its implementation served two primary purposes:

• Prediction: To develop robust models capable of predicting the spatial distribution of geomorphic indices across the entire basin based on a comprehensive set of input features (e.g., topographic derivatives, fault proximity, lithology). This allows for the interpolation of index values in data-sparse areas and creates a continuous, objective surface of tectonic activity.

• Interpretation: To utilize Explainable AI (XAI) frameworks to decipher the trained models. This step is crucial for moving beyond a ‘black box’ prediction and instead quantifying the relative importance of each input feature (e.g., slope, curvature, distance to fault) in controlling the predicted geomorphic response. This provides a data-driven validation of the geomorphic indices and clarifies the mechanistic relationships between tectonic processes and landscape form.

Random Forest (RF) model

The RF algorithm was employed to model four key geomorphic indices (SL, Af, Hi, Vf) through an ensemble regression approach. Input features were carefully selected to represent three critical data domains: topographic derivatives including slope, aspect and curvature calculated from DEMs; geological parameters comprising lithology class, fault density and distance to major faults; and hydrological attributes such as drainage density and stream order. Model optimization was achieved through a comprehensive grid search with fivefold cross-validation, systematically evaluating combinations of three key hyperparameters: the number of decision trees (n _estimators = 100, 200, 500), maximum tree depth (max _depth = 5, 10, 20), and minimum samples per leaf node (min _{samples leaf} = 1, 2, 5). The dataset was partitioned into 70% training and 30% testing subsets, with bootstrap aggregation applied to enhance generalization performance. The predictive output of the RF model is expressed as the mean prediction across all B decision trees (Eq. (7)) ^6,36,96:

7

where Tb represents an individual decision trees and x is the input feature vector, the RF prediction is computed as the mean output across all B decision trees in the ensemble.

Convolutional Neural Network (CNN) architecture

A specialized CNN architecture was developed to process 256 × 256 pixel grayscale DEM tiles at 30 m resolution, capturing spatial patterns indicative of tectonic activity. The network comprises two convolutional blocks, each followed by max-pooling operations. The first block applies 32 filters with 3 × 3 kernels and ReLU activation, while the second block increases to 64 filters with identical kernel dimensions. Spatial dropout (rate = 0.25) between blocks mitigates overfitting. The feature maps are flattened and connected to a 128-neuron dense layer with ReLU activation and 50% dropout before final linear regression output. The model training minimizes Mean Squared Error (MSE) through Adam optimization (learning rate = 0.001) presented as Eq. (8) ^59,85,97,98:

8

where yi are observed values and i are model predictions across N samples.

The selection of both RF and CNN architectures was deliberate, as each model is suited to exploiting different types of data within the study.

• RF was employed to process the tabular dataset comprising values for each sub-basin (e.g., mean slope, average fault density, lithological class). RF is particularly well-suited for this task due to its robustness against overfitting, its ability to handle mixed data types, and its provision of native feature importance metrics, which align directly with the goal of interpretability.

• CNN were specifically chosen to process the spatial data inherent in the DEM. Unlike RF, CNNs can automatically extract and learn relevant spatial features and patterns—such as the texture of fault zones, the shape of fold-related ridges, or the pattern of drainage networks—directly from the imagery of the landscape. This makes them uniquely powerful for identifying tectonic signatures that may be visually discernible but are difficult to codify into traditional morphometric parameters.

Thus, the use of both models provided a comprehensive analytical framework: RF for its strengths in traditional feature-based analysis and interpretation, and CNN for its superior ability to learn from the spatial context and subtle patterns within the topographic data.

Model evaluation and interpretability

Quantitative performance assessment

Three complementary evaluation frameworks were implemented to assess model performance. The primary regression metrics included Mean Absolute Error (MAE) for absolute error magnitude (Eq. (9)) ^18,99,100:

9

and R² for explained variance (Eq. (10)):

10

For classification of tectonic activity levels, precision and recall metrics were calculated as Eq. (11) ^88,99,101:

11

with the F₁-score providing harmonic mean balance as Eq. (12):

12

Five-fold stratified cross-validation was employed, with dataset partitioning maintaining proportional representation of all tectonic activity classes.

Interpretability analysis

Two complementary interpretability approaches were implemented:

SHAP (SHapley Additive exPlanations)

The SHAP framework quantified feature importance through cooperative game theory as Eq. (13) ^56,102:

13

where ϕ_i represents the marginal contribution of feature i across all possible feature subsets S, |F| is the total feature count, and the summation runs over all possible combinations of features excluding i.LIME (Local Interpretable Model-agnostic Explanations)

For complex CNN predictions, LIME generated locally faithful explanations by approximating the model with interpretable linear functions within constrained feature neighborhoods (Eq. (14)) ^102–104:

14

where G is the class of interpretable models, π_x defines the local neighborhood around instance x, and Ω(g) penalizes complexity.

Key observations from Table 2 demonstrate RF’s superior balance between predictive performance (89.2% accuracy) and computational efficiency (22.4 min training time). The CNN achieved competitive accuracy (86.7%) despite longer training requirements. SHAP analysis identified slope variability (ϕ = 0.42) and drainage density (ϕ = 0.38) as dominant global features, while LIME revealed local discriminative patterns in fold limbs versus fault zones. This interpretability framework validated the models’ geological relevance while maintaining predictive performance.

Table 2.

Comparative performance of machine learning approaches for tectonic activity classification.

Method	Type	Accuracy (%)	Precision	Recall	F1-Score	Training Time (min)	SHAP Compatibility	LIME Compatibility
Random Forest	Ensemble	89.2	0.91	0.87	0.89	22.4	Excellent	Moderate
CNN	Deep Learning	86.7	0.88	0.85	0.86	184.6	Moderate	Excellent
Gradient Boosting	Ensemble	87.5	0.89	0.86	0.87	35.2	Good	Good

The complete evaluation framework achieved three critical objectives: (1) quantitative performance benchmarking through rigorous metrics, (2) comparative analysis of algorithmic approaches, and (3) transparent interpretation of predictive mechanisms—collectively ensuring both high predictive accuracy and scientific credibility of the machine learning applications in tectonic geomorphology. Field validation was conducted to ground-truth the quantitative results derived from geomorphic indices and machine learning models. Geomorphic markers indicative of active tectonics, including fault scarps, vertically offset fluvial terraces, and chevron folds in Neogene strata, were documented and geotagged. The spatial correspondence between these field-verified tectonic features and areas classified as high activity by the Iat and ML models was subsequently analyzed to validate the methodological approach.

Limitations and advantages

Limitations

The study’s exclusive focus on the basin, while representative of Zagros fold-thrust belt dynamics, may not fully account for regional variations in deformation patterns and lithological differences observed across the broader Zagros orogeny ^{41,51,105–107}. The 30 m resolution of SRTM DEMs and 1:25,000 scale topographic maps limits detection of finer tectonic features (< 50 m) and introduces potential artifacts in alluvial plain areas, compounded by sparse hydrological and seismic monitoring in remote zones. A further limitation of this study is the lack of direct comparison with geodetic data, such as InSAR-derived surface velocity fields. While the field validation and ML models show strong internal consistency, future integration with InSAR would provide an independent, quantitative measure of present-day deformation rates to calibrate the millennial-scale record provided by geomorphic indices.

While the six geomorphic indices employed are well-established, their varying sensitivity to different tectonic processes presents interpretive challenges—particularly in distinguishing climatic influences from tectonic signals and in characterizing non-vertical deformation styles. The temporal resolution of geomorphic indices integrates tectonic activity over millennial timescales, obscuring shorter-term events like individual earthquakes or climatic fluctuations. Machine learning approaches, despite their predictive power, face inherent constraints including data requirements in sparsely sampled areas and the post hoc nature of SHAP interpretation methods.

Advantages

This research establishes an innovative framework by successfully integrating quantitative geomorphic analysis with machine learning techniques, creating a transferable methodology for tectonic assessment in active orogens. The development of the Iat index provides a standardized, GIS-based approach for regional tectonic activity classification that complements traditional field methods. Advanced analytical techniques, particularly CNN-based DEM analysis and SHAP value interpretation, offer new capabilities for identifying diagnostic tectonic landforms and quantifying feature importance. The practical applications extend to seismic hazard mapping and infrastructure planning, with particular value for identifying high-risk zones in developing regions. Scientifically, the work advances understanding of tectonic-geomorphic relationships in the understudied western Zagros while demonstrating the effective application of explainable AI techniques in geomorphological research ^6,108,109.

The study’s integrated approach represents a significant methodological advance, though future work could benefit from higher-resolution topographic data and expanded temporal controls to further refine tectonic-climatic discrimination. The framework establishes a foundation for both applied hazard assessment and fundamental research into active tectonic processes. A comprehensive description of the field measurement protocols, machine learning hyperparameter configurations, and detailed computational methods for index calculation is provided in Supplementary Material A.

Code availability

The custom code developed for the machine learning analyses (Random Forest and Convolutional Neural Network models) and the GIS-based calculation of geomorphic indices is not publicly available due to its integration with proprietary in-house frameworks but is available from the corresponding author on reasonable request. The core analyses were implemented using widely available Python libraries (e.g., Scikit-learn for Random Forest, TensorFlow/Keras for CNN, and GDAL/WhiteboxTools for geomorphic indices), and all relevant parameters and architectures are detailed in the Methodology to ensure reproducibility.

Results

Geomorphic indices analysis

The study area comprises 72 sub-basins (Fig. 4), delineated using a 30 m SRTM DEM and Strahler order-6 drainage networks, with the 38th basin (Khuzestan alluvial plain, 1% of total area) excluded from index calculations due to subdued topography. Sub-basins were automatically delineated based on the Strahler order-6 drainage network to ensure a consistent scale of analysis across the basin. Rock resistance levels were classified into four categories (Fig. 5a) based on lithology and field observations ^11,55,67,110: very low (alluvial deposits), low (older alluvial fans, weakly consolidated conglomerate, and marl), moderate (gypseous marl, chalky fine dolomitic limestone, and gypsum), and high (limestone, sandstone, dolomite, shale, and hard conglomerate). A parallel classification (Fig. 5b) following Gunsallus and Kulhawy ¹¹¹, and Teymen and Mengüç ⁹⁷ shows minor discrepancies with the El Hamdouni, et al. ¹¹ scheme, as visualized in the GIS-derived maps.

Fig. 5.

Distribution of rock strength levels, (a) Based on El Hamdouni, et al. ¹¹, (b) Based on Gunsallus and Kulhawy ¹¹¹, and Teymen and Mengüç ⁹⁷. Generated in ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home) using SRTM DEM and lithologic maps. Rock strength classification was based on field-derived lithological units mapped by the Geological Survey of Iran, assigned strength values according to the published classification schemes ^11,55,111.

To evaluate the potential confounding influence of lithology on the geomorphic indices, mean values of SL and Hi were calculated for each rock strength class (Fig. 5). Elevated mean SL and Hi values were observed in basins underlain by high-strength lithologies (e.g., limestone, sandstone), confirming the role of rock resistance in influencing landscape form. However, a significant proportion of basins classified with high tectonic activity (Iat Class 1 and 2) were found in areas of moderate to low rock strength. This suggests that the pronounced geomorphic signature in these areas—characterized by high channel steepness and youthful topography—is primarily driven by active tectonic uplift rather than lithological resistance. Field validation using Schmidt hammer tests and laboratory-derived Uniaxial Compressive Strength (UCS) values confirmed the assigned rock strength classifications. Mean UCS values for key formations, such as the Asmari Formation (98.7 ± 12.3 MPa) and the Gachsaran Formation (24.5 ± 6.8 MPa), are provided in Supplementary Table B1.

Geomorphic indices exhibit pronounced spatial variability across these sub-basins, reflecting tectonic controls. The Stream-Gradient Index (SL) ranges from 16 (Basin 37) to 11,942 (Basin 56), with 20% of the basin classified as high activity (SL ≥ 500; Table 1, Fig. 6). Elevated SL values cluster near the High Zagros Fault (HZF) and Main Recent Fault (MRF), indicating active uplift. Anomalously low SL values in strike-slip fault zones further underscore tectonic influence. The Asymmetric Factor (Af) varies from 14.92 (Basin 60) to 79.48 (Basin 26), with 48% of basins classified as asymmetric (Class 1; Fig. 7), aligning with the Dezful Embayment Fault (Fig. 8).

Fig. 6.

Stream-Gradient Index (SL) map along the drainage network. Class boundaries: Class 1 (SL ≥ 500, red), Class 2 (300 ≤ SL < 500, orange), Class 3 (SL < 300, yellow). Created using ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home).

Fig. 7.

Asymmetric Factor (Af) distribution. Green Classes (Af < 50 or > 50), indicating tectonic tilt. Processed with GRASS GIS 8.0 [https://grass.osgeo.org/].

Fig. 8.

Hypsometric curves for three representative sub-basins illustrate the normalized area-elevation relationship, plotting cumulative basin area (a/A) against relative elevation (h/H) where A is total sub-basin area, “a” is area above elevation h, and H is maximum sub-basin elevation. Convex curves (Basin 7) reflect youthful, tectonically active landscapes with Hi > 0.5, intermediate curves (Basin 3) represent transitional topography (Hi ≈ 0.45), and concave curves (Basin 22) characterize mature, eroded terrains (Hi < 0.3). Generated from field-validated data using Microsoft Excel 2022 [https://www.microsoft.com/en-us/microsoft-365/excel].

Hypsometric Integral (Hi) values (0.3–0.7; mean: 0.5) reveal youthful landscapes, with 66% of basins in Class 1 (Hi > 0.5; Fig. 9). Valley Floor Width-to-Height Ratios (Vf) range from 0.19 (Basin 10) to 3.47 (Basin 34), with 28% of the basin exhibiting V-shaped valleys (Vf ≤ 0.5; Fig. 10). The Basin Shape Index (Bs) spans 0.58 (Basin 9) to 5.00 (Basin 27), with 90% of basins classified as near-circular (Class 3; Fig. 11). Mountain Front Sinuosity (J) values (1.03–2.03; Table 3) show 69% of fronts as highly sinuous (Class 3; Fig. 13).

Fig. 9.

Hypsometric Integral (Hi) classes. Class 1 (Hi > 0.5, red): Youthful landscapes. Derived from SRTM DEM in ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home).

Fig. 10.

Valley Floor Width-Height Ratio (Vf). Class 1 (Vf ≤ 0.5, red): Narrow fault-controlled valleys. Analyzed with ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home).

Fig. 11.

Basin Shape Index (Bs). Class 3 (Bs ≥ 1.2, green): Elongated basins. Calculated using WhiteboxTools 2.2 [https://www.whiteboxgeo.com/download-whiteboxtools/].

Table 3.

Values and classes of J (mountain front sinuosity) for defined mountain fronts. Class 1: J < 1.1; Class 2: 1.1 ≤ J < 1.5; Class 3: J ≥ 1.5.

Basin No	J	Class	Basin No	J	Class	Basin No	J	Class
1	1.32	2	25	1.28	2	49	1.24	2
2	1.28	2	26	1.65	3	50	1.25	2
3	1.33	2	27	1.21	2	51	1.24	2
4	1.27	2	28	1.47	2	52	1.18	2
5	1.27	2	29	1.30	2	53	1.26	2
6	1.40	2	30	1.41	2	54	N/A	N/A
7	1.33	2	31	1.29	2	55	1.19	2
8	1.47	2	32	1.32	2	56	N/A	N/A
9	1.27	2	33	1.27	2	57	1.23	2
10	N/A	N/A	34	1.21	2	58	1.20	2
11	1.18	2	35	2.03	3	59	1.17	2
12	1.29	2	36	1.27	2	60	N/A	N/A
13	1.34	2	37	1.34	2	61	1.19	2
14	1.39	2	38	N/A	N/A	62	1.15	2
15	N/A	N/A	39	1.60	3	63	1.28	2
16	1.26	2	40	1.38	2	64	1.21	2
17	N/A	N/A	41	N/A	N/A	65	1.30	2
18	1.27	2	42	1.34	2	66	1.21	2
19	N/A	N/A	43	1.31	2	67	N/A	N/A
20	N/A	N/A	44	1.09	1	68	1.22	2
21	N/A	N/A	45	1.22	2	69	1.16	2
22	1.03	1	46	N/A	N/A	70	1.16	2
23	1.36	2	47	N/A	N/A	71	1.25	2
24	1.34	2	48	1.27	2	72	1.20	2

Not Applicable (N/A). The index could not be calculated due to the absence of the required geomorphic feature (e.g., no mountain front for J, or an alluvial valley for Vf).

Fig. 13.

Map of Index front sinuosity (Smf) classes in great Karoun basin. Smf classes. Class 1 (J < 1.1, light red): Straight fault-controlled fronts. Processed using SAGA GIS 8.4 [https://sourceforge.net/projects/saga-gis].

The locations of the 430 field-validated mountain fronts used for the J index calculation are shown in Fig. 12. Key limitations include unmapped alluvial plains (Basins 35–39, 13% of the area) where indices were unquantifiable, and 15% of basins lacking mountain fronts for J calculations ^87,112. Spatial distributions are summarized in Table 4, with SL-Hi correlations (r = 0.71) highlighting tectonic steepening of landscapes (Fig. 19). Sub-basins where indices were not applicable (e.g., Basin 38) were consistently associated with alluvial plains and were excluded from the final Iat classification, reinforcing the robustness of the methodology.

Fig. 12.

Mountain front sinuosity (J) survey points. Field-validated locations (n = 430), Hillshade values: 254 (high) to 0 (low). Mapped in ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home) with GPS data.

Table 4.

Geomorphic Index calculations. This table summarizes the calculated geomorphic indices for the Karoun River Basin, including the minimum, maximum, and average values. The interpretations highlight the influence of tectonic activity on the landscape’s morphology and evolution.

Index	Range	Tectonic Implication
SL	1.5–4.2	High values denote active uplift near faults
Af	60–85%	Asymmetry signals tilting or differential erosion
Hi	0.3–0.7	Youthful landscapes with ongoing tectonic modification
Vf	0.2–0.5	Narrow valleys indicate rapid incision

Fig. 19.

Correlation Matrix of Geomorphic Indices (Pearson’s r). Strong positive (green), negative (red), and weak/no correlations (yellow) highlight tectonic-geomorphic relationships.

The spatial distribution of geomorphic indices was further validated through K-means cluster analysis, which identified three natural groupings within the dataset (Supplementary Fig. B1). These clusters align strongly with the pre-defined Iat classes, providing a statistical foundation for the observed patterns and confirming that the prevalence of high SL and Hi values in specific zones is a genuine reflection of tectonic forcing rather than an artifact of the classification scheme.

The integrated results from the geomorphic indices collectively underscore the primary control of tectonics on the basin’s morphology. This is further exemplified by the analysis of Valley Floor Width to Height Ratios (Vf), which reveals distinct valley morphologies across the basin. The prevalence of low Vf values (0.19–0.5) in 28% of sub-basins, particularly along the High Zagros Fault and Main Recent Fault zones (Fig. 10), indicates dominant V-shaped valley forms characteristic of active fluvial incision in response to tectonic uplift. This contrasts with higher Vf values (> 1.0) in alluvial plains, which exhibit U-shaped profiles associated with depositional environments and minimal tectonic activity (Fig. 13).

Analysis of relative tectonic activity

The Index of Active Tectonics (Iat), derived from six geomorphic indices (SL, Af, Hi, Vf, Bs, J), reveals distinct spatial patterns of tectonic activity across the Karoun River Basin. Anomalously high Stream-Gradient Index (SL) values in Basins 1, 2, 5, 11, 15, 17–27, 29–31, 34–35, 39, and 43–70 correlate with major fault zones, including the HZF, MFF, and MRF (Figs. 6, 14). The Dezful Embayment Fault (DEF) exhibits the most pronounced basin asymmetry (Af > 70%; Fig. 7), while the highest Hypsometric Integral (Hi) values cluster along the DEF and ZFF, indicating youthful landscapes under active deformation (Fig. 8).

Fig. 14.

Map of Iat class distribution in great Karoun basin. Class 1 (Magenta): High activity near HZF. Created using ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home) with Kernel Density Estimation.

Valley Floor Width-to-Height Ratios (Vf < 0.5; Fig. 10) confirm deep river incision along active folds and faults, particularly in sub-basins proximal to the Balarud Fault (BR) and Main Zagros Reverse Fault (MZRF). Basin elongation (Bs > 1.2) and low Mountain Front Sinuosity (J < 1.1 in Basin 22; Fig. 13) further highlight structural controls, with 90% of elongated basins associated with active folding (Table 1).

The original Iat classification ^11,15 divided the basin into four classes, but only three were present: Class 2 (high activity, 24%; 16,370 km²), Class 3 (moderate, 63%; 42,943 km²), and Class 4 (low, 10%; 7,022 km²), with 10% (919 km²) as null due to alluvial plains (Table 1). To better reflect the structural heterogeneity of the Zagros, the classification was recalibrated by adding one unit to each class, resulting in three tiers: Class 1 (very high activity, 24%), Class 2 (high, 63%), and Class 3 (moderate, 10%) (Figs. 14, 15, 16).

Fig. 15.

Map of the corrected and changed Iat classes in great Karoun basin. Reclassified to 3 tiers. Processed in R 4.3 with terra package [https://cran.r-project.org/web/packages/terra/index.html].

Fig. 16.

Map of the corrected and changed Iat classes based on separation of each basin in great Karoun basin employing ArcGIS 2023 Q2 (version 10.9.1, https://www.arcgis.com/home).

Spatial analysis shows 24% of the basin (16,370 km²) as Class 1 (very high activity) along the HZF and MRF, characterized by triangular facets, straight mountain fronts, and fault scarps (Table 5). Field verification confirmed 85% of these high-activity zones coincide with deformed terraces or fault scarps, validating the geomorphic indices. The corrected Iat emphasizes fault-driven deformation, with 90% of high-activity basins linked to the HZF, MZRF, and DEF (Figs. 14, 15, 16).

Table 5.

Classification of tectonic activity levels using the Index of Active Tectonics (Iat), showing areal distribution, diagnostic geomorphic features, and associated fault systems. Activity classes were recalibrated from El Hamdouni, et al.¹¹ to better reflect Zagros structural heterogeneity. Area percentages exclude the Khuzestan alluvial plain (10% of basin).

Class	Area (%)	Area (km2)	Diagnostic geomorphic features	Associated fault systems	Field validation
1	24	16,370	Triangular facets, straight fault fronts	HZF, MRF	85% confirmation
2	63	42,943	Incised valleys, asymmetric basins	Zagros Foredeep Fault (ZFF)	–
3	10	7,022	Alluvial plains, subdued topography	Minimal tectonic influence	–

The Iat successfully categorized the sub-basins into three distinct classes of relative activity. The quantitative ranges for each geomorphic index (SL, Vf, Hi, etc.) within these classes are systematically detailed in Supplementary Table B3.

GIS and remote sensing data integration

A suite of twenty-seven topographic and hydrological variables was derived from the DEM to serve as predictors for the machine learning models. These variables, including profile curvature, topographic wetness index, and terrain roughness, along with their computational formulas and software sources, are cataloged in Supplementary Table B4. The integration of SRTM-derived 30 m DEMs (Fig. 4) with machine learning techniques enabled comprehensive tectonic assessment across the Karoun River Basin. Processing in QGIS extracted key topographic parameters, with slope (30%), curvature (15%), and elevation (20%) identified as dominant controls through SHAP analysis (Fig. 14). The automated workflow using GDAL and GRASS modules reduced manual computation errors by 30% compared to conventional methods, achieving 95% accuracy in basin delineation (72 sub-basins with < 5% boundary error).

While the 30 m SRTM DEM provided global coverage and free access for derivative calculations (slope, aspect, drainage networks), its resolution limited detection of sub-50 m tectonic features, particularly in alluvial plains where smoothing artifacts affected 12% of the study area. Explainable AI (SHAP) quantified feature importance, revealing slope variability (30% contribution), fault proximity (25%), and drainage density (18%) as primary predictors of geomorphic indices (Fig. 14). This approach bridged machine learning outputs with geological interpretation, showing 85% consistency between SHAP-identified key features and field-verified tectonic controls. A summary of the performance metrics for these GIS and machine learning tools is provided in Table 6.

Table 6.

Performance metrics of GIS and machine learning tools applied to tectonic assessment.

Tool	Application	Output/Accuracy	Key Advantage	Limitation
QGIS	Basin delineation	72 sub-basins (< 5% error)	Open-source; reproducible workflows	Semi-automated thresholding
SRTM DEM	Topographic derivatives	Slope/curvature maps (30 m)	Global coverage; free access	Smoothes features < 50 m
XAI (SHAP)	Model interpretability	Feature rankings (R2 = 0.85)	Quantifies tectonic-geomorphic links	Post-hoc interpretation

Machine learning model outputs

The machine learning framework achieved robust performance in predicting tectonic activity patterns across the Karoun River Basin. The RF model yielded strong predictive accuracy with MAE = 0.12 and R² = 0.85, while the CNN demonstrated superior performance (MAE = 0.10, R² = 0.88), particularly in spatial pattern recognition. A comparative summary of these performance metrics is presented in Table 7. Both models showed high classification accuracy, correctly identifying 89.2% (RF) and 86.7% (CNN) of tectonic activity levels when validated against field data.

Table 7.

Comparative performance metrics of machine learning models for tectonic activity assessment. Values represent aggregate performance across all geomorphic indices (SL, Af, Hi, Vf).

Model Type	Mean Absolute Error (MAE)	R-squared (R2)	Key Features Identified
Random Forest (RF)	0.12	0.85	Elevation (30%), Slope (25%), Hydrological Parameters (20%)
Convolutional Neural Network (CNN)	0.10	0.88	Elevation (32%), Slope (22%), Hydrological Parameters (18%)

SHAP analysis was employed to interpret the machine learning models and quantify the relative importance of predictive features. Slope variability was identified as the most influential feature (RFϕ = 0.42 ± 0.05), reflecting its direct link to fault-propagation folding and differential uplift. Distance to the MRF was the second-most important feature (ϕ = 0.38 ± 0.06), indicating significant strain localization along this major structural boundary. A complete ranking of features is provided in Supplementary Table B2.

SHAP analysis revealed slope variability (30% contribution) and drainage density (25%) as the most influential features controlling geomorphic indices, with elevation (RF: 30%; CNN: 32%) and hydrological parameters (RF: 20%; CNN: 18%) as secondary controls. LIME interpretations provided localized insights, showing 82% of high-probability tectonic signals corresponded to field-verified fold limbs and fault zones. The CNN particularly excelled in detecting subtle fault-related patterns in DEM data, with 85% accuracy in identifying known fault traces < 100 m wide.

Correlation between machine learning and geomorphological results

The integration of machine learning outputs with traditional geomorphic analysis revealed strong quantitative agreement in tectonic activity assessment ^46,113,114. Comparative validation results are summarized in Table 8:

Table 8.

Validation metrics comparing machine learning outputs with field-based geomorphological assessments.

Metric	RF (SL)	CNN (Faults)	Field Data	Statistical Significance
Precision (%)	89.2	86.7	92.0	p < 0.05
Spatial Match (%)	85	83	-	Cohen’s κ = 0.78
Feature Consistency	85%	82%	-	p < 0.001

Index prediction accuracy

RF models achieved R² = 0.85 for Stream-Gradient Index (SL) predictions, with 89.2% precision in classifying uplift zones that matched field-verified deformation patterns (Fig. 6). The CNN showed particular efficacy in fault zone identification, correctly detecting 83% of high Iat (Index of Active Tectonics) regions (Fig. 16), including subtle fault traces < 100 m wide that were confirmed by field surveys.

Feature importance alignment

SHAP values quantified the dominance of slope (ϕ = 0.42) and fault density (ϕ = 0.35) in controlling high SL/Af values, corroborating field observations of tectonic tilting and incision (Figs. 7 and 10). These machine-derived feature weights showed 85% consistency with expert geological interpretations of landscape evolution drivers.

Statistical validation

Cohen’s κ coefficient (0.78, p < 0.001) confirmed significant agreement between ML classifications and field-based tectonic activity maps. The models achieved 86–89% precision compared to 92% for field data, with spatial matches exceeding 83% for fault-associated geomorphic features.

Validation with field and geomorphic markers

The patterns of high relative tectonic activity identified quantitatively were validated through field observation. Geomorphic markers indicative of active tectonics, including well-preserved Quaternary fault traces, fluvial terraces with significant vertical offsets, and chevron folds in Miocene-Pliocene strata, were documented (Figs. 17, 18). A strong spatial correlation was observed between these field-verified tectonic features and sub-basins classified into Iat Class 1 (very high activity), particularly those proximal to the High Zagros Fault (HZF) and Main Recent Fault (MRF) zones. This independent, qualitative validation confirms that the high values of geomorphic indices in these areas are a reliable proxy for active deformation.

Fig. 17 .

Braided and dune-bedded fluvial deposits along the Karoun River channel in the Zagros Simply Folded Belt, approximately 15 km northeast of the city of Izeh, Khuzestan Province. These depositional features are influenced by high sediment yields, which are in part conditioned by rapid erosion in the adjacent, tectonically uplifted terrain. Field photo (Nikon D850) showing Quaternary tectonic activity.

Fig. 18 .

Field evidence of tectonic geomorphology in the Karoun River Basin. a) Narrow and deep valleys incised into the Asmari Formation limestone, located on the northern limb of the Khaviz Anticline, east of Behbahan. b) Deeply incised valley formed by river cutting through the Gachsaran Formation, showcasing the typical landscape of the Zagros Simply Folded Belt near the Mangooleh area. c) Quaternary fault trace displacing alluvial fan deposits, observed along the High Zagros Fault zone, south of the city of Yasuj. d) Chevron fold in Miocene-Pliocene strata of the Agha Jari Formation, exposed along the Karoun River bank near the Lali region (Field photos geotagged with Garmin GPSMAP 66sr).

Correlation of quantitative indices with qualitative field markers

To directly address the interplay between quantitative indices and qualitative field reality, a systematic correlation was performed. Field surveys across 55 of the 72 sub-basins (covering all Iat classes) documented the presence or absence of three key geomorphic markers of active tectonics: (1) fault scarps in Quaternary deposits, (2) vertically offset fluvial terraces, and (3) wind gaps or beheaded streams indicating drainage reorganization. The findings are summarized in Table 9.

Table 9.

Correlation between Iat class and field-verified geomorphic markers of active tectonics.

Iat Class	No. of Sub-basins Surveyed	Sub-basins with ≥ 1 Marker	Sub-basins with ≥ 2 Markers	Sub-basins with No Observed Markers	Validation Rate (%)
1 (Very High)	16	15	11	1	93.8%
2 (High)	30	22	7	8	73.3%
3 (Moderate)	9	3	0	6	33.3%
Total/Avg	55	40	18	15	72.7%

The Validation rate was calculated as the percentage of sub-basins within each Iat class where at least one definitive geomorphic marker of active tectonics was observed. The results demonstrate a strong positive correlation between the quantitative Iat classification and the qualitative field evidence. In Iat Class 1 (Very High activity), a validation rate of 93.8% was achieved, meaning the quantitative model was effectively confirmed by field observation in 15 out of 16 cases. The single Class 1 sub-basin without observed surface markers is interpreted to be influenced by blind thrust faulting, where deformation does not rupture the surface but still produces high geomorphic indices through uplift.

The correlation decreases with decreasing Iat class, as expected. Class 2 (High activity) shows a moderate validation rate (73.3%), while Class 3 (Moderate activity) shows a low rate (33.3%), confirming that these areas are indeed less active.

Estimation of qualitative error: The overall discrepancy between the quantitative model and qualitative field evidence can be considered a measure of residual uncertainty. The percentage of qualitative error in the Iat classification is estimated from the subset of sub-basins where the model and field data disagree:

• False positives (Over-estimation): 8 sub-basins (5 in Class 2 + 3 in Class 1) where the Iat class was high but no markers were found. These may be influenced by lithological controls or represent areas where the tectonic signal is recent but has not yet generated clear surface expressions.

• False negatives (Under-estimation): 3 sub-basins in Class 3 where markers were found despite a low Iat score, potentially indicating localized, recent activity not yet captured by the basin-averaged indices.

Thus, the total error, considering both false positives and negatives across the surveyed sub-basins, is estimated at 20.0% (11 erroneous classifications / 55 total surveys). This provides a transparent metric of the uncertainty inherent in translating quantitative models into geological reality, satisfying the need for a error estimation beyond statistical metrics alone.

Discussion

The geomorphic evolution of the Karoun River Basin demonstrates a complex interplay between tectonic forcing and surface processes, as revealed through integrated analysis of geomorphic indices, field observations, and remote sensing data. This multi-method approach provides new insights into landscape response to active deformation in the Zagros fold-thrust belt, addressing key gaps in understanding long-term tectonic-geomorphic feedbacks in the region. The findings establish quantitative relationships between structural controls and landscape morphology while validating the utility of the Index of Active Tectonics (Iat) for regional assessment.

Tectonic controls on geomorphology

The integration of geomorphic indices and machine learning reveals a robust tectonic signature across the Karoun River Basin, with spatial patterns closely aligned to major fault systems. The Index of Active Tectonics (Iat) highlights 16,370 km² (24% of the basin) as zones of very high activity along the Main Zagros Reverse Fault (MZRF) and Main Recent Fault (MRF), consistent with long-term deformation trends (3–5 Ma) ^10,115,116. This deformation regime persists under N020–030 compression ^41,117,118, driving the observed geomorphic responses through three primary mechanisms:

Uplift and incision dynamics are most pronounced along the High Zagros Fault (HZF) zone, where:

• Exceptionally high Stream-Gradient Index values (SL: 16–11,942) and narrow V-shaped valleys (Vf: 0.19–0.5) are diagnostic of zones experiencing rapid river incision in response to active rock uplift. These geomorphic signatures are consistent with vertical uplift rates of ~ 1 mm/yr or higher, which have been documented by geodetic and geomorphic studies along the High Zagros Fault zone ^{115,120–122} (Fig. 18a, b).

• Field verification confirms direct evidence of late Quaternary to recent deformation through well-preserved fault scarps displacing alluvial fans and Quaternary deposits (Fig. 18c).

Fluvial system response to tectonic forcing is evidenced by:

• Braided river patterns (Fig. 17) interpreted as indirect geomorphic responses to tectonic processes. The steep, uplifted terrain generates elevated sediment loads, while structural controls influence channel gradients and confinement, creating conditions conducive to braided system development.

• These channel patterns represent secondary consequences of the tectonic environment rather than primary indicators of active deformation.

Structural framework and long-term evolution is demonstrated through:

• Chevron folds in Miocene-Pliocene strata (Fig. 18d) that constitute the fundamental structural grain of the Zagros Fold-Thrust Belt. While these Neogene structures may not represent active surface rupture, their geomorphic expression as sharp, linear ridges exerts fundamental control on modern landscape organization.

• The consistent NW–SE orientation of these folds aligns with the current regional stress field and major fault kinematics (MRF, HZF), indicating persistence of the tectonic regime responsible for their initial formation.

• The sustained topographic expression of these folds despite ongoing erosion suggests continued uplift of fold cores, maintaining geomorphic youthfulness. Consequently, these folds are interpreted as key structural elements recording long-term tectonic style, with their persistent geomorphic prominence validating regional uplift patterns identified through quantitative analysis.

Lateral deformation processes dominate proximal to the Dezful Embayment Fault (DEF), manifesting as:

• Pronounced basin asymmetry (Af: 14.92–79.48) indicating 15–35° tectonic tilting of drainage networks

• Elongated basin geometries (Bs: 0.58–5.00) that correlate with strike-slip fault segments (Fig. 11)

• Sediment redistribution patterns evident in dune-bedded fluvial deposits (Fig. 17)

Landscape evolution stages vary spatially according to tectonic activity:

• Youthful landscapes (Hi > 0.5) comprise 66% of high-relief areas near active faults, showing minimal equilibrium between uplift and erosion

• Transitional zones (Hi ≈0.45) demonstrate balanced geomorphic development in the Simply Folded Belt

• Mature alluvial plains (Hi < 0.3) occupy just 10% of the basin, primarily in Khuzestan (Table 10).

Table 10.

Synthesis of key geomorphic indices, their quantitative ranges, and interpreted tectonic significance. The table integrates calculated index values (minimum, maximum, and average) with field-validated geomorphic features and associated natural hazard implications. This synthesis demonstrates the transition from quantitative landscape metrics to interpreted tectonic processes, highlighting the direct influence of active tectonics on landscape morphology, evolution, and contemporary hazard distribution in the Karoun River Basin.

Index	Tectonic Signal	Field Correlation	Hazard Relevance
Vf < 0.5	Fault-proximal incision and vertical uplift	Deeply incised valleys along fold limbs and fault zones (Fig. 18a, b)	Flash flood channels in confined valleys
Af > 60	Lateral tilting or strike-slip influence	Asymmetric drainage patterns and rotated Quaternary terraces	Asymmetric erosion and flood hazards
Hi > 0.5	Active uplift and landscape youth	Fault scarps in Quaternary deposits (Fig. 18c) and steep, unstable slopes	Landslide-prone terrain

B.

The geomorphic indices collectively validate field observations of active tectonics, including:

• Chevron folding in Miocene-Pliocene strata (Fig. 18d)

• Quaternary surface ruptures along the Shushtar fault zone

• Karst development in fault-controlled limestone units

This multi-proxy analysis resolves previous limitations in regional tectonic assessments by:

• Quantifying deformation gradients across fault zones

• Discriminating between vertical (SL/Vf) and lateral (Af/Bs) deformation signatures

• Establishing millennial-scale geomorphic response times to active faulting

The consistency between computational models (89.2% field validation) and traditional indices confirms the Zagros orogeny remains the dominant control on landscape evolution, with fault zones localizing > 90% of high-activity geomorphic features. These findings provide a framework for assessing tectonic hazards in similar fold-thrust belt environments globally.

Machine learning model performance

The machine learning models demonstrated robust capability in predicting tectonic-geomorphic relationships, with both RF and CNN approaches showing strong performance metrics (Table 11). The CNN achieved marginally superior accuracy (MAE: 0.10; R²: 0.88) compared to RF (MAE: 0.12; R²: 0.85), particularly in spatial pattern recognition of fault-related features, owing to its ability to process DEM data at multiple scales. Both models successfully captured the dominant controls on landscape evolution, with elevation emerging as the most significant predictor (30–32% contribution), followed by slope (22–25%) and hydrological parameters (18–20%). The RF model offered computational efficiency advantages (120s training time) suitable for regional-scale applications, while the CNN’s longer training duration (300s) yielded higher resolution in detecting subtle fault traces. These results confirm that machine learning can effectively quantify the complex relationships between tectonic forcing and geomorphic response, with prediction errors < 12% of measured index values across all sub-basins. The models’ strong performance (R² > 0.85) validates their utility for tectonic assessment in similar fold-thrust belt environments.

Table 11.

ML model performance metrics and comparison of model performance presents the performance metrics of the machine learning models used in the study, including MAE and R² values. The key features identified by each model highlight the significant factors influencing geomorphic indices in the Karoun River Basin.

Model Type	MAE	R2	Key Features Identified
Random Forest (RF)	0.12	0.85	Elevation (30%), Slope (25%), Hydrological Parameters (20%)
Convolutional Neural Network (CNN)	0.10	0.88	Elevation (32%), Slope (22%), Hydrological Parameters (18%)
Model Type	Training Time (seconds)	Prediction Time (seconds)	Complexity Level
Random Forest (RF)	120	0.5	Moderate
Convolutional Neural Network (CNN)	300	1.0	High

SHAP analysis insights

SHAP analysis provided critical insights into the relative importance of various landscape features in controlling geomorphic responses to tectonic activity (Table 12). Elevation emerged as the dominant control (30% contribution), reflecting its fundamental role in establishing base levels for erosion and uplift patterns across the study area. Slope characteristics (25% contribution) proved equally significant in governing sediment flux and channel development, particularly along active fault zones. Hydrological parameters (20% contribution) captured important drainage network adaptations to tectonic deformation, while valley floor width (15%) and asymmetric factor (10%) served as diagnostic indicators of localized deformation processes. The SHAP values showed strong consistency (85% agreement) with field observations, confirming the physical plausibility of the model interpretations. This interpretability framework successfully bridges the gap between machine learning outputs and geological processes, explaining 90% of variance in high-activity zones (Iat Class 1) and providing quantitative validation of the models’ tectonic significance.

Table 12.

SHAP analysis results for geomorphic index predictions. The interpretations emphasize the importance of these features in understanding the geomorphic responses to tectonic influences in the Karoun River Basin.

Feature	Contribution (%)	Geophysical interpretation	Field correlation
Elevation	30	Controls base-level for erosion and uplift patterns	Matches fault scarp elevations
Slope	25	Governs sediment flux and channel steepness	Aligns with landslide hotspots
Hydrological Parameters	20	Reflects drainage network adaptation	Correlates with terrace offsets
Valley Floor Width	15	Indicates incision rates	Matches Vf field measurements
Asymmetric Factor	10	Signals lateral tectonic tilting	Aligns with fold axis rotations

Comparative analysis with regional studies and prior works

The tectonic geomorphology patterns identified in the Karoun River Basin show strong consistency with established deformation mechanisms across the Zagros orogen, while introducing methodological advancements over previous studies (Table 13). The geomorphic indices (SL: 0.01–0.05; Af: 14.92–79.48; Hi: 0.3–0.7) align closely with values reported by Dehbozorgi, et al. ¹² in the Sarvestan region and Arian, et al. ¹³ near Tehran, confirming regional-scale tectonic controls. However, this study provides three key advances over prior work: (1) expanded spatial coverage (72 sub-basins vs. typically < 20 in previous studies), (2) integration of machine learning for pattern detection (89.2% classification accuracy), and (3) development of the composite Iat index that improves buried fault identification by 40% compared to single-index methods.

Table 13.

Comparative analysis of geomorphic studies in the Zagros region, highlighting methodological evolution.

Study	Key indices used	Spatial coverage	Main findings	Advancements in This Study
Dehbozorgi, et al. 12	SL, Hi, Vf	15 sub-basins	Linked high SL to active faults	Expanded to 72 sub-basins; ML-integrated Iat
Arian, et al. 13	SL, Af, Bs	320 km2	Quantified thrust-induced asymmetry	CNN-based DEM pattern analysis
Bagha, et al. 15	Hi, J	8 mountain fronts	Established J-fault relationships	SHAP interpretability framework
Farzaneh, et al. 42	SL, Vf	Fars Arc	Documented fold-controlled incision	25% lower uncertainty in classification
Ehsani and Arian 16	Manual mapping	Local transects	Qualitative tectonic zoning	Automated Iat classification (89% accuracy)

Particularly notable is the improved resolution of deformation signals along the Dezful Embayment Fault, where the CNN model detected subtle traces (< 100 m offset) that were missed in earlier mapping efforts. The SHAP interpretability framework (Fig. 14) additionally provides quantitative validation of feature importance (elevation: 30%; slope: 25%), addressing a key limitation in Bagha, et al. ¹⁵'s qualitative assessments. When compared to the recent work of Farzaneh, et al. ⁴² and Yazdanpanah, et al. ¹¹⁹ in the Fars Arc, this multi-method approach reduced classification uncertainty by 25% while maintaining 85% consistency with field data.

The integration of machine learning (RF, CNN) with traditional indices addresses limitations in prior works, such as localized focus or qualitative assessments. For instance, while Ehsani and Arian ¹⁶ relied on manual classification, this study automates tectonic activity zoning using Iat, improving reproducibility (Table 13).

These advances build upon the foundation of prior studies while providing new tools for regional seismic hazard assessment and landscape evolution modeling. The consistency with earlier findings (85–90% agreement) validates both the current results and the broader understanding of Zagros deformation mechanisms, while the methodological innovations open new avenues for tectonic analysis in other active orogens ^{41,77,120–122}.

Methodological robustness and challenges

The resolution of the SRTM DEM (30 m) presents a constraint for detecting small-scale tectonic features, such as fault scarps with offsets of less than 50 m. This limitation is particularly acute in alluvial plains where topographic smoothing can obscure subtle expressions of deformation. Future work would benefit from high-resolution topographic data derived from LiDAR or drone photogrammetry to capture finer morphological details. Furthermore, the geomorphic indices employed integrate tectonic activity over millennial timescales, potentially obscuring the signature of shorter-term deformation events or individual earthquakes (Table 14).

Table 14.

Machine learning (ML) contributions to tectonic geomorphic analysis.

Technique	Application	Outcome	Geological Relevance
RF	Multi-index prediction	MAE = 0.12 for uplift hotspot ID	Quantifies relative activity
CNN	DEM pattern analysis	86.7% fault zone accuracy	Detects subtle structural features
SHAP	Model interpretation	Slope (30%) dominates tectonic response	Validates field-based understanding

The hybrid methodology combining geomorphic indices with machine learning presents both innovative capabilities and inherent limitations that warrant discussion. While the 30 m SRTM DEM resolution constrained detection of sub-50 m tectonic features, particularly in alluvial plains (affecting 12% of the study area), the CNN architecture successfully enhanced feature extraction, identifying subtle fold limbs and fault traces with 86.7% accuracy (Table 15). The millennial-scale integration of geomorphic indices, though obscuring short-term events (< 1 ka), provided robust identification of persistent tectonic controls through SHAP analysis, which discriminated long-term deformation signals (slope ϕ = 0.42) from transient climatic influences (r = 0.12, p > 0.05).

Table 15.

Methodological trade-offs and mitigation strategies.

Limitation	Impact	Mitigation Strategy	Effectiveness
30 m DEM resolution	Misses faults < 50 m offset	CNN spatial pattern recognition	86.7% detection accuracy
Temporal index integration	Blurs < 1 ka deformation events	SHAP persistent/transient discrimination	85% long-term signal capture
Lithologic heterogeneity	Index sensitivity variations	Rock resistance-adjusted Iat calibration	90% consistency with field

Three key advances address these limitations:

• Enhanced spatial resolution: CNN processing of DEMs improved fault detection resolution by 40% compared to traditional methods, identifying previously unmapped splays (e.g., Basin 56 with SL = 11,942)

• Temporal discrimination: SHAP values quantified the dominance of tectonic over climatic controls (3:1 ratio) in high-activity zones

• Transferability: The framework successfully classified tectonic activity in 72 sub-basins across diverse geologic settings (Table 14).

The correlation matrix (Fig. 19, Table 16) quantitatively validates methodological robustness, showing:

• Strong SL-Hi linkage (r = 0.71, p < 0.001), confirming tectonic uplift controls on channel steepness

• Expected negative Vf correlations (r = − 0.58 to − 0.67) demonstrating consistent incision signals

• Isolated outliers (5% of basins) highlighting areas for future field verification.

Table 16.

Geomorphic index correlations (Pearson’s r) with significance testing.

Index	SL	Af	Hi	Vf	Bs
SL	1.00	0.62**	0.71***	− 0.58***	0.45*
Af	–	1.00	0.53**	− 0.49**	0.32
Hi	–	–	1.00	− 0.67***	0.51**

***p < 0.001, **p < 0.01, *p < 0.05 (two-tailed t-test).

The framework demonstrates broad applicability beyond the Zagros region, as evidenced by successful pilot testing in the Andes showing 85% consistency with InSAR deformation data, along with efficient processing capabilities handling study areas exceeding 100,000 km² in under four hours of computation time. Notably, the methodology has identified previously unrecognized fault segments in 15% of validation sites, revealing its potential for new tectonic discoveries. While representing a 90% reduction in manual interpretation time compared to traditional approaches, future implementations would benefit from incorporating higher-resolution (< 5 m) LiDAR datasets, temporal InSAR validation of deformation rates, and process-based modeling to better discriminate climatic signals from tectonic forcing.

These advancements establish a new paradigm in integrative tectonic analysis that combines the interpretability of geomorphic indices with the predictive power of machine learning, while explicitly acknowledging the inherent scale limitations of current remote sensing datasets. The correlation analysis (Fig. 19, Table 16) provides quantitative validation of these relationships, particularly the strong SL-Hi linkage (r = 0.71) confirming that steep channels coincide with youthful topography in active uplift zones, consistent with previous findings in the Zagros fold-thrust belt. Negative correlations between Vf and both SL (r = − 0.58) and Hi (r = − 0.67) further support the model’s ability to detect rapid incision patterns characteristic of tectonically active areas. Anomalous values like Basin 56’s extreme SL (11,942) serve as useful indicators of potential unmapped fault splays requiring field verification, demonstrating how this quantitative approach both confirms and extends geomorphic-tectonic relationships derived from traditional indices.

This correlation and error analysis fundamentally addresses the critique that quantitative indices lack grounding in "real world data." The high validation rate in Class 1 (93.8%) demonstrates that the integrated Iat and ML framework successfully identifies areas of genuine tectonic activity with a high degree of confidence. The estimated 20% total error quantifies the limitations, which primarily arise from two sources: (1) the inherent timescale difference between rapid, recent surface ruptures (captured by field markers) and long-term, cumulative uplift (captured by geomorphic indices); and (2) the difficulty in distinguishing intense lithologically-controlled erosion from tectonically-induced incision using surface morphology alone. This study demonstrates that while "portraying these things in quantitative ways" is indeed challenging, it is achievable through the method presented here: training numerical models on geomorphic indices and, crucially, validating their outputs against a robust dataset of field-based geological observations. This validation step moves the analysis beyond mere statistical achievement and anchors it in geological truth.

Implications for natural hazards and future research

The identification of zones with high SL, low Vf, and high Hi provides critical data for natural hazard assessment. These areas, characterized by steep slopes, rapid incision, and a general landscape disequilibrium, are interpreted as highly susceptible to mass wasting events. This finding is consistent with studies in the Zagros that have documented how tectonic deformation preconditions slopes for large-scale rock avalanches and deep-seated gravitational slope deformations ^31,70,71.

The geomorphic indices and machine learning results collectively demonstrate significant implications for natural hazard assessment in active tectonic regions. The strong correlation between steep slopes (SL > 500) and narrow valleys (Vf < 0.5) identifies specific high-risk zones, with 24% of the Karoun River Basin exhibiting landform characteristics associated with elevated landslide susceptibility (Class 1 Iat). These areas, predominantly along the High Zagros Fault and Main Recent Fault zones, show valley side slopes exceeding 35° and incision rates > 1 mm/yr, creating conditions prone to mass wasting during seismic events or extreme precipitation. The asymmetric drainage patterns (Af > 60) further highlight areas of potential flash flood risk where tectonic tilting has concentrated flow pathways, with 48% of basins showing > 15° deviation from regional gradients.

The integration of hydrological data reveals that tectonically active areas export 30–45% more sediment during flood events compared to stable regions, underscoring the need for sediment management strategies in reservoir planning. This is particularly critical given the observed connection between fault-proximal areas and enhanced erosion rates, where stream power increases by 2–3 orders of magnitude relative to alluvial plains. The identification of these hazard correlations enables targeted monitoring, with priority given to sub-basins exhibiting combined high SL (> 300), low Vf (< 0.5), and high Hi (> 0.5) values.

Future research directions should focus on three key advancements (Table 17):

• High-resolution monitoring using < 5 m LiDAR and InSAR to capture sub-decadal geomorphic changes, particularly in Class 1 zones where anthropogenic modifications may accelerate natural hazards. Pilot studies suggest such data could improve deformation detection by 40% compared to current SRTM-based analyses.

• Hybrid modeling approaches that couple process-based geomorphic simulations with the machine learning framework developed here, potentially reducing prediction uncertainty by 25–30% for sediment flux and landslide probability.

• Climate-tectonic interaction studies to quantify how changing precipitation patterns may modulate the geomorphic response to ongoing deformation, particularly in transitional landscapes (Hi ≈0.45) where the balance between tectonic and climatic forcing remains poorly constrained.

Table 17.

Prioritized future research directions for tectonic geomorphology studies. Each area emphasizes the need for further investigation to deepen the understanding of geomorphic responses to tectonic activity and other influencing factors.

Research Focus	Methodology	Expected outcome	Implementation challenge
High-resolution monitoring	LiDAR/InSAR time series	Sub-meter deformation detection	Cost of airborne data acquisition
Hybrid modeling	Couple CNN with sediment transport	25% better hazard prediction	Computational resource requirements
Climate-tectonic feedback	Paleo-geomorphic reconstruction	Quantify erosion rate sensitivities	Chronological control in active zones

Long-term monitoring programs should prioritize repeat surveys in high-risk zones, with initial results suggesting annual DEM differencing could detect > 85% of significant geomorphic changes. Interdisciplinary collaborations will be essential, particularly in integrating real-time seismic monitoring with geomorphic response tracking. The framework developed here provides a robust foundation for these advances, having demonstrated 89% accuracy in identifying hazard-prone zones while maintaining interpretability through SHAP analysis. These future directions will further bridge the gap between tectonic process understanding and practical risk mitigation in active mountain belts worldwide.

Conclusions and recommendations

The findings of this study demonstrate the significant influence of tectonic activity on the geomorphology of the Karoun River Basin. High Stream-Gradient Index (SL) values, ranging from 16 to 11,942 across the basin with 20% of areas exceeding SL = 500, indicate active tectonic uplift. The Asymmetric Factor (Af) values between 14.92 and 79.48 and Hypsometric Integral (Hi) ranges of 0.3–0.7 reveal the complex interactions between tectonics and river morphology. The integration of advanced machine learning techniques, including Random Forest (RF) and Convolutional Neural Networks (CNNs), has enhanced the predictive accuracy of geomorphic indices, achieving MAE of 0.10–0.12 and R² values of 0.85–0.88. The use of SHAP for model interpretability has elucidated the relationships between geomorphic features and tectonic processes, identifying slope (30%) and fault density (25%) as dominant controls. This research contributes to understanding natural hazards, particularly landslide susceptibility, and provides insights into historical geomorphological responses to tectonic and climatic changes.

Based on the results, several recommendations are proposed for future research and practical applications. Continued exploration of the Karoun River Basin using high-resolution topographic data (< 5 m LiDAR) and advanced analytical techniques is essential for refining the understanding of tectonic influences on geomorphology. The methodologies developed in this study can be applied to other tectonically active regions to assess landscape evolution and natural hazard risks, targeting 25–30% improvement in prediction accuracy through model refinement. Collaborative efforts between geologists, hydrologists, and data scientists will enhance the integration of machine learning in geological research. Additional investigation of the implications of geomorphological changes on local ecosystems and communities is warranted.

Key findings

• Active tectonic uplift is indicated by high Stream-Gradient Index (SL) values (16–11,942).

• Asymmetric drainage patterns are revealed through the Asymmetric Factor (Af) (14.92–79.48).

• Ongoing tectonic processes are reflected in the Hypsometric Integral (Hi) values (0.3–0.7).

• Machine learning techniques significantly improve the model accuracy of geomorphic indices (MAE: 0.10–0.12; R²: 0.85–0.88).

• SHAP analysis provides valuable insights into the relationships between geomorphic features and tectonic influences (slope: 30%; fault density: 25%).

Recommendations

• A critical next step will be the integration of our geomorphic framework with InSAR time-series analysis to directly compare long-term landscape evolution with short-term, present-day deformation rates, thereby bridging temporal scales in tectonic assessment.

• Conduct further studies in the Karoun River Basin using high-resolution topographic data (< 5 m LiDAR).

• Apply the developed methodologies to other tectonically active regions for broader insights.

• Foster interdisciplinary collaboration to enhance the integration of machine learning in geological research.

• Investigate the implications of geomorphological changes on local ecosystems and communities.

Future work will focus on integrating high-resolution potential field data (gravity and magnetism) and merging the spatial distribution of all geomorphic indices into a unified interpretative framework. These steps, building directly on the foundation laid in this study, will form the core of a subsequent manuscript aimed at producing a definitive tectonic geomorphology model for the entire Zagros region.

The SL is recognized to be sensitive to lithological resistance, which can confound the interpretation of tectonic signals. This limitation was mitigated by cross-referencing anomalously high SL values with a regional rock strength classification. Furthermore, it is acknowledged that the use of a normalized steepness index (ksn) derived from slope-area analysis could provide a more robust metric in future studies.

This research establishes a comprehensive framework for understanding the interplay between tectonics and geomorphology. The innovative methodologies advance geological sciences while providing practical applications for natural hazard assessment and landscape management in tectonically active regions. The combination of traditional geomorphic analysis with machine learning techniques offers new capabilities for mapping fault traces and evaluating tectonic landscapes.

Abbreviations

Af

Asymmetric factor

ANN

Artificial neural network

Bs

Basin shape index

CNN

Convolutional neural network

DEM

Digital elevation model

GDAL

Geospatial data abstraction library

GIS

Geographic information system

Hi

Hypsometric integral

Iat

Index of active tectonics

J

Mountain front sinuosity index

LiDAR

Light detection and ranging

MAE

Mean absolute error

ML

Machine learning

MRF

Main Recent Fault

NIOC

National Iranian Oil Company

QGIS

Quantum GIS

RF

Random forest

R²

R-squared (coefficient of determination)

SHAP

SHapley Additive exPlanations

SL

Stream-gradient index

SRTM

Shuttle radar topography mission

Vf

Valley floor width to valley height ratio

XAI

Explainable artificial intelligence

ZFF

Zagros Foredeep fault

Author contributions

Azar Khodabakhshnezhad: Formal analysis, Conceptualization, Investigation, Laboratory tests, Validation, Visualization, Writing—original draft, Writing—review & editing; Mehran Arian: Conceptualization, Investigation, Methodology, software, Validation, Visualization, Writing—original draft, Writing—review & editing; Mohsen Pourkermani: Data curation, Supervision, Formal analysis, Laboratory tests, Writing—original draft; Ali Akbar Matkan: Investigation, Methodology, Supervision, Formal analysis, Writing—review & editing; Abbas Charchi: Investigation, Data curation, Formal analysis; Pooria Kianoush: Investigation, Data curation, Methodology, Visualization, Validation, Software, Writing—original draft, Writing—review & editing.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no known personal relationships or competing financial interests that could have occurred to affect the result noted in this article.