Geological and machine learning based landslide risk assessment in the context of community knowledge and perception

Abstract

Aizawl, located in Mizoram, northeast India, faces persistent landslide risk due to fragile geology, steep terrain, and unregulated urban expansion. This study assesses the correspondence between community perceived landslide drivers and scientifically derived factors. Research included questionnaire survey, field investigation, and Random Forest based susceptibility modelling (LSM). The survey indicated 66.9% of respondents had experienced landslides. Heavy rainfall (86.2%) and poor drainage (69.0%) were the dominant perceived triggers. Risk perception was statistically uniform across demographic groups (all p > 0.05). Causal knowledge showed a significant positive association with preparedness behaviour (Spearman ρ = 0.479, p < 0.001). The LSM classified 30.5% of the municipal area as very high susceptibility, capturing 78.8% of events (ROC–AUC = 0.726; F1 = 0.793); all failures coincided with settlement zones (38.3%). Geotechnical investigation showed moderate to high rock mass weathering. Out of 15 sites, 13 were classified as Zone IV (LHZ) and rated Bad (SMR). Planar, wedge, and toppling failures were confirmed. Cross scheme analysis demonstrated perfect rank agreement between LHZ and LSM hazard drivers (Spearman ρ = 1.000, p < 0.001), with directional alignment between community perceived causes and LSM factors (ρ = 0.600). These results establish a quantitative basis for community integrated landslide assessment in Aizawl; broader applicability requires validation under site specific geological and socio-cultural conditions.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-52262-8.

Introduction

Landslides are among the most significant and recurring geo-environmental hazards in India, particularly affecting the Himalayan belt, the Western Ghats, and the Northeastern hill regions^1,2. The northeastern states are especially vulnerable due to fragile geology, high rainfall intensity, and increasing anthropogenic interventions³. Aizawl, the capital city of Mizoram, located within the Assam–Arakan fold belt, has been consistently identified as a critically landslide prone urban centre^4,5. Severity of landslides is due to its steep terrain, weak sedimentary lithology, and rapid, largely unplanned urban expansion⁶. The increasing frequency of landslides in Aizawl is closely associated with slope modification, road construction, drainage disruption, and expansion of settlements onto unstable slopes, particularly during the monsoon season⁷. Landslide research has advanced considerably through geotechnical investigation, hazard zonation, and machine learning-based susceptibility modelling, improving both the identification of hazard-prone areas and the quantification of controlling factors⁷. However, these approaches remain predominantly focused on physical and statistical parameters, and are rarely integrated with assessments of how affected communities perceive and respond to landslide risk. The role of local communities in understanding and responding to such hazards remains underrepresented. Residents often possess experiential knowledge of early warning indicators, such as surface cracks, water seepage, and slope deformation, which form an important component of local disaster awareness^8,9. Landslide risk perception is shaped by lived experiences, cultural context, and continuous environmental interaction⁸.

Global disaster studies have shown that community based knowledge can significantly influence survival and adaptive behaviour⁸. During the 2004 Indian Ocean tsunami, indigenous population with culturally embedded hazard awareness demonstrated higher survival rates, compared to migrants and tourists who lacked local knowledge⁹. Similar community responses have been reported during the Mount Merapi eruption and the 2006 Central Java earthquake¹⁰. Although perception based approaches are important in landslide research, several methodological limitations and inconsistencies remain. Empirical studies indicate significant divergence between expert assessments and community perceptions, with non-experts often attributing landslide occurrence primarily to natural factors, whereas experts emphasise anthropogenic controls¹¹. However, perception does not necessarily translate into preparedness or adaptive behaviour; communities residing in high risk zones may still exhibit low perceived risk and limited willingness to adopt mitigation measures¹². Cognitive biases, risk denial, and socio-economic constraints further contribute to underestimation or misinterpretation of hazard risk, thereby reducing the effectiveness of risk communication and mitigation strategies¹³. Recent studies also highlight a persistent disconnect between scientific assessments and community level understanding, which can undermine preparedness and response effectiveness¹⁴. In Aizawl, the collective experiences of repeated landslides have shaped a distinct social perception of hazardous zones and responses. It is crucial to explore how communities perceive landslide’s causative factors and how these perceptions align from scientifically derived susceptibility models. Despite this recognised need, a critical gap persists. Perception based landslide studies in India remain largely descriptive, confined to frequency distributions and qualitative narratives, and are rarely integrated with geotechnical or model based assessments¹⁵.

Therefore, comprehensive understanding of landslide risk necessitates an integrated framework that incorporates geotechnical, environmental, and socio-behavioural dimensions. This gap is particularly acute in the rapidly urbanising hill cities of northeast India, where physical vulnerability and limited institutional risk communication capacity together create urgent demand for evidence based, community informed hazard management. Scientific assessments must ultimately translate into actionable knowledge that is understandable and usable at the community level¹⁵. To address this gap, the present study adopts an integrated framework combining structured community perception analysis, Random Forest based landslide susceptibility modelling, and geotechnical field investigation in Aizawl. The study hypothesises that the primary landslide drivers identified by community respondents will correspond with the conditioning factors ranked as dominant by the Random Forest model and with field documented slope instability mechanisms. The specific objectives are to: (i) analyse community perceptions of landslide causes, risk exposure, preparedness, and institutional trust; (ii) develop a landslide susceptibility model and identify dominant conditioning factors through variable importance analysis; (iii) characterise slope instability through geotechnical field investigation; and (iv) evaluate the degree of correspondence between perception based and scientifically derived assessments. The findings are intended to strengthen hazard communication and support more effective, locally adapted landslide risk management in Aizawl and comparable hill cities of northeast India.

Study area

Aizawl, the state capital of Mizoram in northeast India, is situated in a zone of rugged, steeply dissected terrain characterised by narrow ridges, deep valleys and slopes predominantly exceeding 20°. The region has a humid tropical climate with heavy rainfall (~ 2500 mm annually) and temperatures ranging from ~ 19 °C to ~ 35 °C¹⁶. Geomorphological analysis shows a fairly uniform aspect distribution, with 38% of slopes facing 0°–124°, 30% facing 125°–253°, and 32% facing 254°–360°. Slope distribution indicates that 45% of the area lies within moderately steep slopes (19°–30°), followed by 32% gentle (0°–18°) and 23% steep slopes (31°–66°). Elevation ranges from 131 to 1379 m, with 45% of the area concentrated at higher elevations (902–1379 m), as shown in Fig. 1. Aizawl has undergone rapid and largely unregulated urban expansion over the past three decades, with settlements increasingly extending onto steep and geologically sensitive slopes. The city comprises 19 administrative wards and recorded a population of 291,822 (city) and 400,309 (district) as per the 2011 Census [Census of India, 2011], with recent estimates exceeding ~ 403,000 by 2023¹⁷. LULC analysis by Lawmchullova et al.¹⁷ indicates a substantial increase in built-up area from 5.98% in 1991 to 25.8% in 2021 at expense of agricultural land and vegetation.

Fig. 1.

Location and elevation of the Aizawl Municipal area: (A) Mizoram within India, (B) Aizawl ward boundary, and (C) spatial distribution and frequency of survey respondents. Black hexagons are respondent locations, letters ‘a’ to ‘j’ denote increasing respondent frequency from 1 to 10, respectively. (Software: QGIS, Version: 3.4.4, URL: https://qgis.org/download/).

The underlying geology is characterized by repetitive sequences of sandstone, shale, mudstone, silty sandstone, and silty shale, which are interbedded across much of the district. The Surma Group, comprising Bhuban and Bokabil formations¹⁸, dominates the stratigraphy of the area, with Aizawl city itself underlain by the Bhuban beds. These lithologies consist of alternating layers of compact, competent sandstones and relatively weaker, brittle shale and mudstone beds^19,20, promoting differential weathering, discontinuity development, and reduced shear strength under saturated conditions, making slopes highly susceptible to failure. This inherent geological vulnerability is further exacerbated by rapid and unplanned urban expansion driven by population growth and migration, resulting in widespread hill cutting, slope modification, and intensified land use. Concurrently, increasing vehicle pressure has enhanced road induced slope disturbances. Together, these natural and anthropogenic factors contribute to drainage disruption, environmental degradation, and heightened landslide susceptibility, with failures and population density concentrated in higher elevation zones.

Methodology

The present study adopts an integrated approach combining (i) community based perception analysis, (ii) data-driven landslide susceptibility modelling, and (iii) geotechnical field based assessment. These components are systematically linked through statistical validation to evaluate convergence between societal perception and scientific assessment of landslide hazard. The survey instrument comprised 44 questions organised into six thematic sections: respondent demographics (Section A, 6 questions), landslide experience (Section B, 8 questions), beliefs and risk awareness (Section C, 6 questions), perception of risk exposure (Section D, 2 questions), preparedness and early warning strategies (Section E, 12 questions), and risk communication and institutional trust (Section F, 10 questions). The complete questionnaire is provided as Supplementary Material S1.

The survey adopted a stratified purposive sampling approach to capture diverse socio-economic, occupational, and residential groups within the Aizawl municipal area. Stratification was based on occupation, age, and proximity to landslide-prone locations, with respondents drawn from nineteen localities covering both high-risk elevated zones and relatively stable areas. The questionnaire was prepared in English and administered using a hybrid approach. A pilot survey was conducted online to test clarity and structure, followed by an in person field survey. A total of 145 responses were collected. Based on Cochran’s sample size formulation²¹, the ideal sample size for large populations at 95% confidence and 5% margin of error is approximately 384; however, the present sample corresponds to an estimated margin error of ~ 8%, which remains acceptable for exploratory perception based studies. This limitation is acknowledged in the interpretation of results. The relatively smaller sample size is attributed to accessibility constraints and the rugged, difficult terrain of the study area. Potential knowledge bias was minimised as most respondents were general residents and undergraduate students with limited technical background, while a small proportion of technically informed respondents was included to provide contextual insights.

Landslide susceptibility modelling was carried out using a Random Forest approach based on a verified landslide inventory and selected conditioning factors representing topographic, geological, hydrological, and anthropogenic controls. Geotechnical assessment was performed through Landslide Hazard Zonation (LHZ) and Slope Mass Rating (SMR) to evaluate terrain level and site specific stability conditions, respectively^22,23. Statistical analysis was conducted using appropriate parametric and non-parametric methods to examine relationships within and across datasets. Cross-scheme validation was performed to compare outputs from LHZ, SMR, LSM, and questionnaire data, enabling assessment of convergence between geotechnical, modelling, and community based perspectives.

All participants were briefed on the purpose and voluntary nature of the study prior to participation. Informed consent was obtained from all participants prior to their inclusion in the study. The survey was anonymous and aimed at understanding community perceptions of landslide hazards. No personal, medical, or sensitive information was collected. No personally identifiable information was recorded, and confidentiality was maintained throughout the study. All procedures were conducted in accordance with relevant academic ethical standards and guidelines for social science research involving human participants established by Banaras Hindu University (BHU). It included principles of informed consent, data protection, and respondent confidentiality. The study protocol was reviewed and approved at interdepartmental level including Dr. Ashutosh Kainthola from Department of Geology, Institute of Science, Dr. Priyanka Jha from Department of Political Science, Dr. Ajay Kumar Yadav from Malaviya Centre for Peace Research and Dr. Mahaveer Singh Panwar from Department of Science and Technology - Centre for Interdisciplinary Mathematical Sciences, BHU. The results of this community perception survey serve as a foundational component of the study’s integrated framework, which also includes geotechnical field observations and machine learning based susceptibility modelling, aiming to bridge the gap between scientific knowledge systems and the everyday realities of those living with the risk of landslides in Aizawl.

Results and discussion

Understanding the coherence between scientific assessments and community perceptions of landslides is essential for developing effective, locally adapted risk reduction strategies. Accordingly, this study analysed responses from a structured questionnaire designed to capture lived experiences, local knowledge, awareness, and preparedness behaviours in landslide prone areas.

Respondent profile

A total of 145 responses were collected. The sample shows a slight female predominance (55.2%) and is largely young, with ~ 61% aged ≤ 25 years. Respondents are relatively well educated, with over 80% having graduate or higher qualifications, and include a mix of occupational groups, with a substantial proportion categorized as ‘other/unemployed’ (46.9%) as shown in Table 1. A majority (59.3%) have resided in the area for 6–20 years, while 57.2% live in rented accommodation, reflecting the transient and rapidly urbanizing nature of Aizawl.

Table 1.

Socio-demographic profile of survey respondents (n = 145).

Demographic variable	Category	No of respondents	(%)
Gender	Female	80	55.2
	Male	65	44.8
Age group	≤ 25 years	88	60.7
	26–35 years	5	3.4
	> 35 years	4	2.8
	Not reported	48	33.1
Education	Non-Graduate	27	18.6
	Graduate	66	45.5
	Postgraduate/Doctorate	52	35.9
Employment	Government	26	17.9
	Self-employed	22	15.2
	Private sector	16	11
	Student	13	9
	Other/Unemployed	68	46.9
Residence duration	< 5 years	42	29
	6–20 years	86	59.3
	> 20 years	17	11.7
House ownership	Owner	62	42.8
	Non-owner	83	57.2

Lived experience with landslides

Field observations indicate a high level of awareness across all demographic groups about landslide. Participants from a wide age distribution demonstrated a general understanding of what landslides are, their triggering factors, and associated impacts. This widespread baseline knowledge strengthens the reliability of subsequent responses, as they reflect informed perceptions rather than speculation. Eight questions assessed direct and indirect landslide exposure among respondents. Of the 145 respondents, 66.9% reported having personally experienced a landslide, while 33.1% had not, indicating that over two-thirds possessed first hand exposure. Among those in affected localities, 16.3% had experienced at least one evacuation event, with incidents concentrated in recent years and peaking in 2024. Repeated evacuations were reported by a smaller share, with responses ranging from one to more than five events, reflecting wide variation in individual exposure intensity. When asked about types of losses suffered, residential property damage was most frequently reported (22 respondents), followed by impacts on livestock, crops, and friends or neighbours (16 respondents each), while 32 respondents reported no material loss (Fig. 2). In parallel, 45.9% reported landslide exposure outside their home locality, affecting routine travel and daily activities. A further 48.3% indicated that close friends or family members reside in landslide prone areas, reflecting substantial indirect exposure through social networks. Media engagement was also notable: 62.8% had encountered landslide related news through various channels, underscoring a broader awareness that extends beyond direct personal experience.

Fig. 2.

Distribution of different types of losses reported by respondents due to landslide events in Aizawl.

Beliefs, knowledge (direct/indirect), and landslide risk awareness

To examine community understanding of landslide causes and perceived vulnerability across Aizawl, six targeted questions were included. The first assessed perceptions of primary landslide triggers using predefined options along with an open-ended category. Heavy rainfall was the most frequently cited trigger (125 respondents, 86.2%), followed by poor drainage (100, 69.0%), loose or unstable soil (64, 44.1%), and ongoing construction (61, 42.1%) (Fig. 3). Other cited factors included deforestation (40), earthquakes (22), improper agricultural practices (18), and heavy vehicles (14). The strong emphasis on hydrological and anthropogenic drivers closely aligns with scientific understanding of landslide causation in the region.

Fig. 3.

Distribution of community responses on perceived causes of landslides in Aizawl.

In Aizawl’s multi-hazard setting, 68.3% of respondents reported having experienced at least one earthquake or cloudburst event. Awareness of major past landslides was broadly distributed, with the Laipuitlang event (2013) most widely recognised (73.1%), followed by Lunglei road (39.3%), Bungkawn-Khatla (33.8%), and the Lengpui airport failure (31.0%) and multiples failures at Hlimen Quarry, indicating strong place-based collective memory. To assess perceived current risk, an overwhelming 97.9% of respondents confirmed that Aizawl is at high landslide risk. Perceptions of what is at risk were explored next. Most of the respondents (79.3%) recognised threats to both people and infrastructure, while smaller proportions emphasised infrastructure alone (29.3%) or primarily people (13.6%). Institutional awareness was partial: 52.1% were familiar with government agencies involved in landslide management, while 47.9% were not, highlighting gaps in public outreach.

Perception of risk exposure

Preliminary findings from the survey revealed that residential houses are the most affected during landslide events, a trend also corroborated by previous studies conducted in the region^25,26. This prompted an assessment of residents’ perceptions of location-based housing susceptibility in Aizawl. The survey presented seven common residential settings such as houses near slopes, along rivers, within the city centre using a multiple choice grid format. Respondents rated each housing type across four risk levels (very low, low, moderate, high) (Fig. 4). Houses on slope edges attracted the greatest concern, with the majority rating them as high or moderate risk (58 and 38 respondents respectively). Houses at the foot of slopes showed similar consensus (51 high, 50 moderate). In contrast, city centre housing was perceived as relatively safe (47 low risk, 26 very low), as was government-built social housing (63 low risk), likely reflecting perceived compliance with building standards. Housing on reinforced slopes was most commonly rated as moderate risk (76 respondents). These perceptions reflect a coherent community understanding of topographic controls, with slope proximity viewed as the primary determinant of landslide risk. These perceptions reflect strong community awareness of topographic controls on landslide processes, with slope instability viewed as a more immediate threat than river proximity.

Fig. 4.

Distribution of community responses on perceived landslide risk for different housing types in Aizawl.

Preparedness, risk management and early warning strategies

The first question assessed how frequently residents encountered information on key landslide preparedness indicators using a multiple choice grid format. Exposure to key preparedness information was generally low across all categories (Fig. 5). Shelter locations and evacuation routes were least communicated, with the majority of respondents reporting “very few times” or “never”. Awareness of early warning systems and at-risk areas showed marginally better reach, though still dominated by infrequent exposure. Guidance on food and water preparation was similarly limited. Overall, these patterns indicate a significant shortfall in routine preparedness communication, which may constrain timely evacuation and emergency response²⁷.

Fig. 5.

Frequency of respondent’s exposure to essential preparedness information, including evacuation routes, shelters, early warning systems, risk prone areas, and emergency supplies.

To strengthen the communication among residents and authorities, it’s important to identify dominant channels used. Using a multiple choice grid format, respondents rated the frequency of information received across seven communication modes (Fig. 6). Social media and television were the most effective information channels, with digital networks reaching respondents frequently or sometimes (59 and 51 respondents respectively) and television showing comparable reach (53 frequently, 52 sometimes) (Fig. 6). Radio was least effective, with 72 respondents reporting no exposure. Public talks had moderate reach, while textbooks and short courses were rarely used. Strengthening multi-channel communication, particularly for populations with limited digital access, remains a priority.

Fig. 6.

Frequency of responses indicating sources of information on landslide risks in Aizawl.

The survey examined resident reported preparedness actions for landslide risk mitigation. The most commonly reported preparedness actions were awareness of hazard warnings (94 respondents) and identification of safe exit routes (92), reflecting priority on life safety (Fig. 7). Knowledge of shelter locations (88) and protection of important documents (70) also ranked highly. Fewer respondents prioritised protecting personal belongings (55) or engaging in community activities (52), suggesting that individual and evacuation focused preparedness is more developed than collective participation.

Fig. 7.

Community reported actions and suggestions for landslide preparedness and response in Aizawl.

Understanding the presence and effectiveness of landslide early warning systems (EWS) in locality is critical. Most respondents (128) reported the absence of any formal EWS in their locality, while only 16 acknowledged its presence and two were unsure. Among those aware of an existing system, satisfaction ratings were generally low, indicating limited public confidence. Given this perceived ineffectiveness, respondents were asked about municipal management during alert phases. While 68.3% felt alerts were handled appropriately, a substantial 31.7% expressed dissatisfaction. Similarly, only 60.6% found public instructions during alerts, clear and actionable, whereas 39.4% reported confusion. It is important to assess whether instances of false alerts occurred, the overwhelming majority 131 respondents (90.3%) reported no such instances, while 14 respondents (9.7%) acknowledged having faced false alarms. False alerts were mostly in 2024, specifically in July-September.

Formal landslide EWS in Aizawl remain limited, while community perceptions strongly underscore their importance, with 71% of respondents considered EWS useful. Although, 82.4% people felt that existing arrangements have limited coverage, delayed alerts, and doubts about reliability. The open-ended responses revealed a range of priorities summarized in Fig. 8. Proposed improvements included faster real time alerts, better drainage and road infrastructure, broader information dissemination, and stronger community training. Respondents also highlighted improved inter agency coordination and investment in technology and manpower.

Fig. 8.

Distribution of community responses highlighting proposed improvements to the existing landslide early warning system.

When asked about anxiety during heavy rainfall without formal warnings, 22.1% of respondents reported heightened concern, largely associated with recent events, particularly 2024, and with vulnerable localities such as the Mizoram University area. The remaining 77.9% reported no such anxiety. In response to a clear landslide alert, 47.4% indicated they would evacuate immediately, 17.5% would follow official instructions, and 15.8% would prioritise securing family and essential belongings. Smaller proportions would remain at home if they felt safe (8.8%), react emotionally (7.0%), or were uncertain (3.5%). Evacuation was the dominant intended response, though variation in secondary actions securing belongings, following instructions, or staying calm reflects differences in preparedness levels and trust in official guidance. Willingness to volunteer in EWS management was notably high: 44.1% expressed willingness and 48.3% indicated potential interest, with only 7.6% declining. This strong latent capacity, if supported by structured training and institutional backing, could be effectively leveraged to build a community volunteer network for early warning in Aizawl.

Risk communication and trust

Most respondents (78.9%) agreed that proper land management can prevent landslides, while 39.1% viewed events as consequences of failing to respect nature and 21.8% perceived them as uncontrollable. Only 6% regarded landslide research as unnecessary. These findings indicate that while scientific and managerial approaches are broadly valued, fatalistic and traditional views persist and should be accommodated in risk communication strategies^28,29.

Infrastructure vulnerability was rated as high or very high by 44% of respondents and moderate by 28%, with open-ended responses citing poor design of older structures, recent highway construction, and non-compliance with building guidelines as major concerns. Regarding future risk without mitigation, 25% of respondents perceived it as high to very high, 15.6% as conditional, and 25% specifically proposed mitigation measures including plantation, drainage improvement, and slope stabilisation, closely aligned with established scientific recommendations. On land-use restrictions in landslide prone areas, the largest share (31.3%) reported no clear knowledge, while 18.8% criticised weak planning enforcement. A smaller group (6.3%) highlighted drainage maintenance issues and the role of blocked channels in increasing risk.

Internet and social networks were the most preferred information channels (78.1%), reflecting widespread digital adoption, though television (58.6%), newspapers (50.8%), and siren or helpline systems also featured prominently (Fig. 9). Institutional trust was highest in the state government (62.2%), followed by municipal (47.2%) and central governments (46.5%), with scientists ranking fourth (43.3%) (Fig. 10). Civil protection agencies and local police received moderate trust. These findings point to the need for multi-channel, government-anchored communication strategies that combine digital immediacy with traditional and emergency-specific channels for inclusive reach.

Fig. 9.

Distribution of preferred communication methods for receiving landslide alert.

Fig. 10.

Variation in public trust across institutions for obtaining landslide risk and response information.

Landslide susceptibility modelling

Landslide susceptibility modelling (LSM) was carried out using a Random Forest (RF) model to spatially evaluate slope instability and validate community-identified landslide drivers. RF was selected due to its ability to handle non-linear relationships and its proven reliability in geospatial susceptibility modelling^30–33. A total of 85 landslide locations within the Aizawl municipal boundary were sourced from the Geological Survey of India (GSI) landslide repository, accessed via the Bhukosh portal (https://bhukosh.gsi.gov.in), and were visually verified using Google Earth imagery prior to use. An equal number of non-landslide points were generated from stable terrain areas within the study boundary to improve model reliability. The inventory was binarily classified, assigning landslide points as 1 and non-landslide points as 0. For model development, 70% of the data (59 locations) were used for training, while the remaining 30% (26 locations) were reserved for validation. The final dataset was converted into a 30 m resolution raster for analysis.

For susceptibility modelling, conditioning factors are guided by their relevance to slope instability, measurability, and ability to capture the geo-environmental complexity of the region²⁴. In this study, nine predisposing factors are considered. Previous studies have consistently identified topographic (slope, elevation, aspect), hydrological (drainage, rainfall), geological (lithology, tectonic proximity), and anthropogenic (road influence, land use) factors as key controls on slope instability^34–36. Similar combinations of conditioning factors have also been widely adopted in landslide susceptibility studies across northeast India and comparable tectonically active regions^37–40, supporting their applicability in the present study. The selected nine factors are particularly relevant for the Aizawl region, which is characterised by steep terrain, intense precipitation, active tectonics, and rapid urban expansion. The Variable Importance (VI) analysis further validates this selection, with slope, elevation, drainage, and rainfall emerging as dominant contributors. All predictors were normalized to a maximum score of 100. The results clearly indicate that slope is the most influential factor, contributing the highest score (100), reaffirming its primary role in slope instability. Elevation follows with a relatively strong influence (~ 55), reflecting its control over relief and climatic interactions⁴¹. Drainage (~ 40) and rainfall (~ 30–35) also rank high, highlighting the combined effect of fluvial undercutting, pore pressure build up, and intense precipitation on landslide initiation. Among the remaining factors, distance to thrusts, distance to roads, and peak ground acceleration (PGA) show modest but noteworthy contributions (~ 10–15).

The final Landslide Susceptibility model (LSM) classified the Aizawl Municipal area into four distinct susceptibility zones: Very High (30.5%), High (11.4%), Moderate (10.8%), and Low (47.6%). Spatial validation of the model revealed that out of 85 recorded landslide locations, 67 (78.8%) fell within the very high susceptibility zone, 7 (8.2%) within the high zone, 2 (2.3%) within the moderate zone, and only 9 (10.5%) within the low susceptibility region as shown in Fig. 11. The results highlight that nearly two-fifths of the municipal area is under high to very high susceptibility, underscoring the need for targeted risk mitigation and stricter slope management practices. Complementing the susceptibility assessment, the Land Use/Land Cover (LULC) analysis revealed that within the municipal boundary, settlements occupy 38.3%, vegetation 35.3%, barren land 24.9%, and water bodies 1.5% of the total area. All recorded landslide incidences were concentrated within the settlement regions, indicating strong anthropogenic influence on slope instability. The spatial overlap between landslide occurrences and built-up zones suggests that unregulated urban expansion, road cutting, and inadequate drainage infrastructure play a major role in triggering slope failures⁴². Roads constructed along steep slopes for connectivity within the hilly terrain further exacerbate instability, as most landslide scars are observed along or adjacent to road alignments.

Fig. 11.

LULC map showing distribution of different land cover with landslide locations (A), and Landslide susceptibility map classified into Low, Moderate, High, and Very High zones (B). (Software: QGIS, Version: 3.4.4, URL: https://qgis.org/download/).

To further evaluate the predictive performance of the model, validation metrics were analysed. The model achieved an ROC–AUC value of 0.726, indicating moderate predictive capability in distinguishing landslide from non-landslide locations as shown in Fig. 12a. The Precision–Recall analysis yielded a PR–AUC value of 0.605, demonstrating a reasonable balance between precision and recall, which is particularly important for spatial binary classification problems. The optimal classification threshold was identified at 0.397 using F1-score maximization, resulting in an F1-score of 0.793, indicating satisfactory classification performance with an effective balance between omission and commission errors. The ROC, Precision–Recall, and F1-threshold curves shown in Fig. 12a–c respectively, collectively illustrate the predictive performance of the model and the robustness of the threshold optimisation approach.

Fig. 12.

Performance of the Random Forest landslide susceptibility model.

Geotechnical characteristics of road cut slopes in Aizawl

A detailed geological and geotechnical investigation was carried out within the Aizawl municipality, assessing slope stability and identification of critical landslide prone zones. Total fifteen locations were assessed in detail. Site selection was guided by both local knowledge and expert judgment to ensure a representative coverage of varying slope conditions. Slope geometry and discontinuity orientations were measured using Brunton compass. Joint spacing, persistence, aperture, infilling material, Schmidt Hammer Rebound Number (SHRN), and overall discontinuity conditions were systematically recorded. The general slope trend varied from northeast to southwest. Typically, three prominent joint sets were observed, with a fourth occasionally appearing in certain areas. Discontinuities commonly terminated abruptly, with joint persistence generally ranging from 1 to 2 m, extending up to 10 m in some exposures. The spacing between joints varied widely from less than 10 cm to more than 200 cm, but was most frequently between 15 and 50 cm. Aperture widths ranged from less than 1 mm to about 10 mm, occasionally exceeding 20 mm. The infilling materials consisted mainly of clay and silt, indicating significant weathering and alteration. The rock mass exhibited moderate to high degrees of weathering, with SHRN values ranging mostly between 25 and 35, though locally varying from 16 to 44. The Geological Strength Index (GSI) primarily ranged from 40 to 60, representing blocky rock masses with fair surface conditions; however, values as low as 20 and as high as 70 were also recorded at certain sites. Water seepage was commonly observed along the slopes, especially in wet sections, reflecting a strong hydrological influence on slope stability.

The rock mass is composed of intact blocks separated by natural discontinuities such as joints and bedding planes. These structural features exert a profound influence on the overall mechanical strength and stability of the rock slopes^1,43. The orientation of these discontinuities, when combined with slope angle and aspect, governs potential block movements. Identifying the type of block movement induced by unfavourably oriented discontinuities beforehand is crucial for minimizing potential damage and ensuring slope safety. Consequently, kinematic analysis has become a standard approach for evaluating such block movement hazards⁴⁴. Signs of planar, wedge, and toppling failures were observed across multiple sites, often marked by loose, disjointed rock blocks precariously resting on steep slopes. Instability was further aggravated by unplanned excavation and inadequate surface drainage. Joint Roughness Coefficient (JRC) was evaluated at each location. Roughness varied widely across sites, some joints displayed smooth surfaces with low JRC values, whereas others exhibited distinctly rough surfaces, indicating variable shear strength characteristics within the rock mass.

In Durtlang 1, Vaivakawn, and Bungkawn, the slopes have lost their vegetative cover due to excavation or natural degradation, exposing the soil and regolith to direct rainfall. Residents, drawing from lived experience, have locally adopted the practice of covering these exposed slopes with polythene sheets to reduce infiltration and saturation during heavy rainfall as shown in Fig. 13. This measure reflects practical community adaptation, aimed at limiting pore water pressure buildup and reducing failure risks. In Durtlang 2, Chaltlang, and adjacent areas, rock slopes are highly jointed, with failures primarily controlled by discontinuity orientation relative to the slope. At Durtlang, toppling failures were common, involving pebble to cobble sized blocks detaching from steep slopes. In Chaltlang, toppling of larger boulder sized blocks occurred due to alignment of steeply dipping discontinuities parallel to the slope surface. In Vaivakawn, Edenthar, and Hlimen, slope instability is strongly linked to anthropogenic undercutting, during construction. Field conditions of different locations are shown in Fig. 13. Excavation at the slope toe without adequate retaining structures has left the upper portions unsupported, causing repetitive failures, especially during monsoons. Water infiltration during intense rainfall events raises pore water pressure, reduces cohesion and angle of internal friction, promoting sliding. In Edenthar, inadequate drainage further exacerbates erosion; runoff flowing directly down the slope cuts channels into the soil and regolith, washing debris onto roadways after major rainfall. Preliminary rock mass assessment identified discontinuity orientation relative to slope geometry, intense rainfall, and inefficient drainage as the primary controls on slope failure. Kinematic analyses across all sites revealed three dominant failure modes planar, wedge, and toppling. At Hlimen, planar failures prevail due to discontinuities dipping parallel to the slope, with water seepage along joints further reducing shear strength. The Durtlang region exhibits all three failure modes, reflecting its complex structural fabric and steep relief, while Bungkawn is characterized by wedge failures at upper slopes formed by the intersection of major joint sets.

Fig. 13.

Field photographs showing different types of slope failures and local mitigation measures observed across Aizawl.

Based on field investigation, the prominent causative factors for landslides are, Slope, Strength of rock mass, Orientation and condition of discontinuity, rainfall, drainage, unscientific construction. To view the causative factors for the landslides in Aizawl from geotechnical perspectives, authors used Landslide Hazard Zonation (LHZ)²² and Slope mass rating classification method of slopes²³.

Landslide hazard zonation (LHZ)

Landslide Hazard Zonation (LHZ) is a method used to classify terrain based on its susceptibility to landslides by considering key geological, geotechnical and environmental factors²². In this study, The Landslide Hazard Evaluation Factor (LHEF) approach provides a quantitative evaluation of slope instability through the combined effect of multiple causative parameters, helping to identify high influencing causative factors. LHEF method is applied to all fifteen sites to obtain a quantitative hazard classification. Six causative parameters; lithology, structure, slope morphometry, relative relief, land use/land cover, and groundwater condition were each rated against prescribed maximum values from field observations. Their sum gives the Total Estimated Hazard Rating (TEHR), which classifies each site into one of five Landslide Hazard Zones (LHZ): Zone I (TEHR < 3.5, Very Low), Zone II (3.5–5.0, Low), Zone III (5.1–6.0, Moderate), Zone IV (6.1–7.5, High), and Zone V (> 7.5, Very High). TEHR values across the fifteen sites ranged from 6.10 to 7.80 (Table 2). Two sites Durtlang 2 (TEHR = 7.80) and Durtlang 3 (TEHR = 7.65) fell in Zone V (Very High Hazard), while the remaining thirteen sites were classified as Zone IV (High Hazard). No location was classified below Zone IV, reflecting the uniformly high hazard conditions across the investigated slopes in Aizawl (Table 2).

Table 2.

LHEF parameter scores and LHZ classification at fifteen sites in Aizawl. GW = groundwater condition; LULC = land use/land cover. Maximum allowable ratings: Lithology 2, Structure 2, Slope Morphometry 2, Relative Relief 1, LULC 2, Groundwater 1; Total maximum = 10. Zone classification after Gupta and Anbalagan²².

Site	Lithology	Structure	Slope morphometry	Relative relief	LULC	GW	TEHR	LHZ zone
Sakawrtuichhun 1	1.5	1.2	2.0	0.6	0.8	0.0	6.10	IV – High Hazard
Sakawrtuichhun 2	1.5	1.75	2.0	0.6	0.8	0.0	6.65	IV – High Hazard
Vaivakawn 1	1.5	0.9	2.0	0.6	1.2	0.0	6.20	IV – High Hazard
Vaivakawn 2	1.5	0.9	2.0	0.6	1.2	0.0	6.20	IV – High Hazard
Durtlang 1	1.25	1.55	2.0	0.6	1.2	0.0	6.60	IV – High Hazard
Durtlang 2	2.0	1.6	2.0	1.0	1.2	0.0	7.80	V – Very High Hazard
Durtlang 3	1.95	1.9	2.0	0.6	1.2	0.0	7.65	V – Very High Hazard
Hlimen 1	1.25	1.55	2.0	1.0	0.8	0.8	7.40	IV – High Hazard
Hlimen 2	1.5	1.4	1.7	1.0	0.8	0.0	6.40	IV – High Hazard
Hlimen 3	1.25	1.5	2.0	1.0	0.8	0.0	6.55	IV – High Hazard
Hlimen 4	1.25	1.5	1.7	1.0	0.8	0.0	6.25	IV – High Hazard
Hlimen Quarry	1.0	1.3	1.7	1.0	1.2	0.0	6.20	IV – High Hazard
Bungkawn	1.25	1.5	2.0	1.0	1.2	0.0	6.95	IV – High Hazard
Chaltlang	1.63	1.7	2.0	0.6	1.2	0.0	7.13	IV – High Hazard
Edenthar	1.5	1.2	2.0	1.0	0.8	0.0	6.50	IV – High Hazard

To identify which parameters most actively drive the observed hazard, a utilization rate was calculated for each parameter from Eq. (1), as the mean observed score expressed as a percentage of its maximum allowable score (Table 3). Slope morphometry recorded the highest utilization (97.0%), scoring at or near its maximum (2.0) in twelve of fifteen sites, confirming that terrain steepness is the most persistently active hazard driver. Relative relief followed at 81.3%. Lithology (72.7%) and structure (71.5%) showed comparable contributions, reflecting the weak weathered rock and unfavourably oriented discontinuities common to the Bhuban Formation. Land use/land cover contributed at 50.7%, with residual vegetation providing partial mitigation. Groundwater condition showed the lowest utilization (5.3%) because most surveys were conducted outside the active monsoon period; its contribution during intense rainfall events is expected to be substantially higher.

1

Table 3.

LHEF parameter utilization rates across fifteen sites in Aizawl, ranked from most to least active contribution to landslide hazard.

Rank	Parameter	Mean score	Max rating	Utilization (%)	Sites at Max (n = 15)
1	Slope Morphometry	1.94	2.0	97.0	12 of 15
2	Relative Relief	0.81	1.0	81.3	8 of 15
3	Lithology	1.46	2.0	72.7	1 of 15
4	Structure	1.43	2.0	71.5	0 of 15
5	Land Use/Cover	1.01	2.0	50.7	0 of 15
6	Groundwater	0.05	1.0	5.3	1 of 15

Slope mass rating (SMR)

Slope Mass Rating (SMR) is a widely used geomechanical classification system that evaluates the stability of rock slopes by integrating rock mass quality with the orientation of discontinuities²³. It provides a more realistic assessment of slope behaviour by incorporating structural controls that directly influence failure mechanisms. In this study, SMR is applied to quantify slope stability conditions across the investigated sites. SMR extends the Rock Mass Rating (RMRbasic)⁴⁵ by incorporating three orientation correction factors (F¹, F², F³) that account for the geometric relationship between the dominant discontinuity set and the slope face, and an excavation factor F⁴, as shown in Eq. (2).

2

RMR basic was derived from intact rock strength (SHRN), Rock Quality Designation (RQD), joint spacing, discontinuity conditions, and groundwater ratings. F⁴ was set to zero for all mechanically excavated slopes. SMR values range from 0 to 100, with higher values indicating greater stability: Very Bad (0–20), Bad (21–40), Normal (41–60), Good (61–80), Very Good (81–100). Computed SMR values ranged from 20.0 to 48.5 across the fifteen sites (Table 4). Thirteen sites were classified as Bad (unstable to highly unstable) and only two sites Durtlang 1 (SMR = 48.5) and Bungkawn (SMR = 41.5) reached Normal (partially unstable). No site achieved Good or above. This near universal instability is consistent with the field observations of active planar, wedge, and toppling failures, and reinforces the LHZ finding that all sites carry high to very high hazard.

Table 4.

SMR parameter values and stability classification at fifteen sites in Aizawl.

Site	RMR basic	F1	F2	F3	F1×F2×F3	SMR	SMR class
Sakawrtuichhun 1	46	0.15	1.00	− 50	− 7.5	38.5	Bad (Unstable)
Sakawrtuichhun 2	37	0.18	1.00	− 50	− 9.0	28.0	Bad (Unstable)
Vaivakawn 1	34	1.00	1.00	− 6	− 6.0	28.0	Bad (Unstable)
Vaivakawn 2	45	1.00	1.00	− 25	− 25.0	20.0	Bad (Unstable)
Durtlang 1	61	0.25	1.00	− 50	− 12.5	48.5	Normal (Partially Unstable)
Durtlang 2	43	0.80	1.00	− 25	− 20.0	23.0	Bad (Unstable)
Durtlang 3	55	0.70	1.00	− 25	− 17.5	37.5	Bad (Unstable)
Hlimen 1	41	0.40	0.70	− 60	− 16.8	24.2	Bad (Unstable)
Hlimen 2	41	0.70	0.85	− 25	− 14.9	26.1	Bad (Unstable)
Hlimen 3	62	1.00	0.70	− 60	− 42.0	20.0	Bad (Unstable)
Hlimen 4	58	1.00	0.70	− 50	− 35.0	23.0	Bad (Unstable)
Hlimen Quarry	57	0.85	0.85	− 50	− 36.1	20.9	Bad (Unstable)
Bungkawn	55	0.54	1.00	− 25	− 13.5	41.5	Normal (Partially Unstable)
Chaltlang	52	0.25	1.00	− 60	− 15.0	37.0	Bad (Unstable)
Edenthar	41	0.15	1.00	− 25	− 3.75	37.3	Bad (Unstable)

Contribution of each SMR parameter to slope instability was quantified using utilization metric. For stabilizing parameters from Eq. (3),

3

For the orientation factor, the contribution is expressed in Eq. (4),

4

Both express the same concept: the fraction of a parameter’s maximum damage potential that is realised across the study area (Table 5). UCS recorded the highest contribution (83.1%), reflecting the advanced weathering and low intact strength of the Bhuban Formation. Discontinuity condition (57.1%) and joint spacing (54.0%) ranked next, driven by clay-silt infilling and closely spaced joints respectively. The orientation factor contributed 30.5% of its maximum possible penalty, consistent with unfavourably oriented discontinuity sets observed at several sites. RQD (39.3%) and groundwater (21.3%) were comparatively lower, the latter partly reflecting dry season survey conditions.

Table 5.

SMR parameter contribution to slope instability across fifteen sites in Aizawl.

Rank	Parameter	Mean score	Max rating	Contribution to instability (%)
1	UCS (Intact Rock Strength)	2.53	15	83.1
2	Condition of Discontinuity	12.87	30	57.1
3	Spacing of Discontinuity	9.20	20	54.0
4	RQD	12.13	20	39.3
5	Orientation (F₁×F₂×F₃)	− 18.3	− 60	30.5
6	Groundwater	11.80	15	21.3

Statistical analysis

The statistical analysis includes (i) inferential evaluation of survey responses to examine demographic influences on exposure, risk perception, and preparedness, along with intra-survey comparisons, and (ii) cross-scheme validation to assess consistency across all frameworks. Prior to test selection, the Shapiro–Wilk test⁴⁶ confirmed that all continuous dependent variables significantly violated normality (all p < 0.05), justifying the use of non-parametric methods. Accordingly, standard statistical methods were applied based on data characteristics. Chi-square (χ²) tests⁴⁷ were used to assess associations between categorical variables, with Cramer’s V⁴⁸ as the effect size measure. Group comparisons were performed using Mann–Whitney U⁴⁹ and Kruskal–Wallis H tests⁵⁰. Relationships between variables were evaluated using Spearman’s ρ⁵¹, supported by Kendall’s τ⁵² as a robustness check, while Pearson’s r was applied for linear relationships where appropriate. Agreement between classification schemes was assessed using Cohen’s κ⁴⁸. All tests were conducted at a significance level of α = 0.05 using SciPy⁵³, with exploratory results (Yes [E]) interpreted cautiously due to small subgroup sizes.

Demographic and experiential predictors of risk perception

A mean risk perception score was computed by averaging each respondent’s ordinal ratings of six common housing typologies (scale 1 to 4; sample mean = 2.77, SD = 0.48). Seven independent tests comparing risk perception scores across groups defined by gender, witnessed experience, regional experience, education level, residence duration, continuous age, and employment type, yielded non-significant results (p ranging from 0.226 to 0.657). This consistent pattern across seven tests indicates that risk awareness in Aizawl operates as a community wide phenomenon⁵⁴, not stratified by demographic background or personal exposure history. Two borderline exploratory findings were noted: age group (H = 8.154, df = 3, p = 0.043) and employment type (H = 11.256, df = 4, p = 0.047), interpreted as directional trends only due to small sub group sizes. At the typology level, female respondents rated foot-of-slope housing as significantly higher risk than males (Mann-Whitney U = 1940, p = 0.015), consistent with gender differentiated hazard sensitivity documented elsewhere⁵⁵. Respondents with greater causal knowledge also rated slope-edge housing higher in risk (Spearman ρ = 0.208, p = 0.013).

Casual knowledge and preparedness behaviour for landslide

The strongest inferential finding is the significant positive Spearman correlation between the number of landslide causes identified by each respondent and the number of preparedness actions endorsed (ρ = 0.479, p < 0.001, n = 145). This was independently cross-validated using Kendall’s τ (τ = 0.383, p < 0.001), confirming robustness. The finding is consistent with the cognitive-motivational model of protective behaviour, in which causal knowledge functions as a necessary antecedent to preparedness action. Neither direct experience (p = 0.478) nor gender (p = 0.074) significantly predicted overall preparedness, indicating that knowledge acquisition through education and media exposure rather than personal encounter is the primary driver of preparedness engagement (Table 6).

Table 6.

Complete survey inferential test results. Yes [E] = exploratory result, V = Cramer’s V; r = rank-biserial r; ρ = Spearman rho; τ = Kendall’s tau.

S. No.	Independent variable (Question/Factor)	Dependent variable (Question/Outcome)	Test applied	Statistic	p value	Sig.	Effect size
1	Gender (male/female)	Witnessed a landslide (yes/no)	Chi-Square	χ2=4.56, df = 1	0.033	Yes	V = 0.18
2	Education level (3 groups)	Witnessed a landslide (yes/no)	Chi-Square	χ2=9.50, df = 2	0.009	Yes	V = 0.26
3	Residence duration (3 groups)	Witnessed a landslide (yes/no)	Chi-Square	χ2=1.28, df = 2	0.527	No	V = 0.09
4	House ownership (owner/non-owner)	Witnessed a landslide (yes/no)	Chi-Square	χ2=0.00, df = 1	0.993	No	V < 0.01
5	Gender (male/female)	External landslide exposure	Chi-Square	χ2=6.86, df = 1	0.009	Yes	V = 0.23
6	Whether witnessed a landslide (yes/no)	Mean risk perception score (of 6 housing typology)	Mann-Whitney U	U = 2316	0.637	No	r = − 0.05
7	Whether experienced landslide elsewhere	Mean risk perception score	Mann-Whitney U	U = 2536	0.226	No	r = 0.09
8	Gender (male/female)	Mean risk perception score	Mann-Whitney U	U = 2210	0.228	No	r = 0.08
9	Education level (3 groups)	Mean risk perception score	Kruskal-Wallis H	H = 1.055, df = 2	0.590	No
10	Residence duration (3 groups)	Mean risk perception score	Kruskal-Wallis H	H = 1.331, df = 2	0.514	No
11	Age group (4 groups: ≤18, 19–25, 26–35, > 35)	Mean risk perception score	Kruskal-Wallis H	H = 8.154, df = 3	0.043	Yes [E]
12	Employment type (5 groups)	Mean risk perception score	Kruskal-Wallis H	H = 11.256, df = 4	0.047	Yes [E]
13	Age as continuous variable	Mean risk perception score	Spearman ρ	ρ = 0.050	0.657	No
14	Gender (male/female)	Risk rating given to foot of slope housing	Mann-Whitney U	U = 1940	0.015	Yes
15	Number of landslide causes identified	Risk rating given to slope-edge housing	Spearman ρ	ρ = 0.208	0.013	Yes	ρ = 0.208
16	Whether witnessed a landslide (yes/no)	Preparedness actions selected or endorsed	Mann-Whitney U	U = 2496	0.478	No	r = 0.05
17	Gender (male/female)	Preparedness actions selected or endorsed	Mann-Whitney U	U = 3046	0.074	No	r = 0.10
18	Number of landslide causes identified	Preparedness actions selected or endorsed	Spearman ρ	ρ = 0.479	< 0.001	Yes	ρ = 0.479
19	Communication preference	Information exposure channel	Spearman ρ	ρ = 1.000	< 0.001	Yes	ρ = 1.000
20	Perceived cause ranking	Preparedness action ranking	Spearman ρ	ρ = 0.988	< 0.001	Yes	ρ = 0.988

Communication channel preference vs. actual information exposure

Spearman rank correlation between preferred communication channels and actual information exposure showed perfect alignment (ρ = 1.000, p < 0.001, n = 4). Social media and television emerged as the dominant channels in both preference and exposure, while radio remained the least utilised. This indicates that current information dissemination in Aizawl broadly aligns with community preferences (Table 6).

Perceived causes of landslide vs. endorsed coping strategies

Respondents identified multiple landslide causes (mean = 3.06), with rainfall and drainage emerging as dominant factors, alongside corresponding preparedness actions focused on warning awareness, evacuation routes, and shelter knowledge. Spearman rank correlation showed near-perfect alignment between perceived causes and coping strategies (ρ = 0.988, p < 0.001, n = 8), indicating that community responses are well aligned with perceived risk drivers (Table 6).

To evaluate convergence across the three methodological components, Spearman rank correlation⁵¹, Pearson correlation, and Cohen’s κ⁴⁸ were applied between LHZ, SMR, LSM, and survey outputs.

Spearman correlation between LHZ parameter ranks and LSM Variable Importance showed perfect agreement (ρ = 1.000, p < 0.001), indicating strong consistency between field-based LHEF classification and the Random Forest model. Slope morphometry (97.0% utilization) corresponds directly to slope angle (VI = 100), confirming it as the dominant controlling factor. Comparison between community perceived causes and matched LSM factors showed moderate but non-significant correlation (ρ = 0.600, p = 0.285; n = 5). This reflects directional agreement, where rainfall (86.2%) aligns with hydrological factors and construction related impacts (42.1%) align with anthropogenic controls. Site level correlation between TEHR and SMR scores was weak and non-significant (ρ = 0.190; r = 0.132), reflecting the difference in scale and methodology. LHZ represents terrain level hazard, whereas SMR captures site specific geomechanical stability. Cohen’s κ (≈ 0.00) indicates limited categorical agreement; however, both methods consistently classify most sites as highly unstable, demonstrating overall agreement in hazard identification.

Conclusion

Mizoram records the highest concentration of landslide hotspots, across all Indian states^4,5, and Aizawl has been consistently identified as a critically affected urban centre. This is due to its fragile sedimentary geology, steep dissected terrain, intense monsoonal rainfall, and unregulated urban expansion⁶. The present study integrates a three tier framework, community perception through questionnaire based surveys, geotechnical investigations, and scientific landslide susceptibility modelling. Across all 15 geotechnically investigated sites, weathering was moderate to high. SHRN was predominantly in the range of 25–35 indicating low to moderate rock strength. GSI ranged from 40 to 60 at most sites, representing blocky rock masses with fair surface conditions. Three to four prominent joint sets were documented at most locations, with spacing predominantly between 15 and 50 cm. Aperture widths varied from less than 1 mm to approximately 10 mm, occasionally exceeding 20 mm, and discontinuities were largely filled with clay and silt, reducing shear resistance considerably. Kinematic analysis confirmed planar, wedge, and toppling failure modes, with anthropogenic toe cutting, inadequate drainage, and clay-silt discontinuity infill as the dominant site level failure controls. For LHZ analysis, TEHR values ranged from 6.10 to 7.80 with no site classified below Zone IV (High Hazard) under LHZ. SMR values ranged from 20.0 to 48.5 with 13 of 15 sites classified as Bad (unstable) confirming uniformly high geomechanical susceptibility throughout the investigated localities.

The Random Forest model shows moderate accuracy (ROC–AUC = 0.726; PR–AUC = 0.605; F1 = 0.793). It classifies 30.5% of the area as very high susceptibility zone, High (11.4%), Moderate (10.8%), and Low (47.6%). About 78.8% of landslides fall in very high susceptibility zone, and 8.2% fall in the high zone. All landslides occur within settlement areas (38.3% of the study area). This shows that landslide risk is strongly linked to human activities and urban expansion. The questionnaire survey revealed that risk perception is demographically uniform. Seven independent non-parametric tests found no significant association between perceived risk and gender, education, residence duration, or direct landslide experience (all p > 0.05). It is consistent with community wide hazard awareness documented in analogous risk environments. The strongest behavioural finding is a significant positive association between causal knowledge breadth and preparedness behaviour (Spearman ρ = 0.479, p < 0.001; Kendall τ = 0.383, p < 0.001). It indicates knowledge acquisition, rather than personal exposure, governs adaptive response, a pattern consistent with the cognitive motivational model of disaster preparedness²⁹. A majority of respondents reported direct landslide experience (66.9%) and recognised Aizawl as landslide prone (97.9%). Critically, the rank ordering of LHZ driving parameter utilisation rates corresponds perfectly with the rank ordering of LSM Variable Importance scores (Spearman ρ = 1.000, p < 0.001), demonstrating that the empirical field based hazard classification and the independently trained machine learning model identify identical hazard driver hierarchies. Community perceived cause ranks show directional but non-significant alignment with LSM factor ranks (ρ = 0.600, p = 0.285; n = 5 matched pairs), a pattern attributable to community emphasis on directly observable triggers rainfall and drainage over morphometric predictors such as slope angle, which requires spatial analysis to quantify.

• Preferred communication channels match actual information sources (ρ = 1.000, p < 0.001). Social media and television are the most used.

• Perceived causes and coping strategies show strong agreement (ρ = 0.988, p < 0.001). This indicates that community understanding guides preparedness actions.

• Heavy rainfall (86.2%) and poor drainage (69.0%) are the main perceived triggers. Unstable soil (44.1%) and construction activities (42.1%) also contribute significantly.

• Slope is the most influential factor in the random forest model (VI = 100), followed by Elevation (~ 55), drainage (~ 40), and rainfall (~ 30–35) follow, while thrusts, roads, and PGA have smaller effects (~ 10–15).

• Slope instability is mainly controlled by low rock strength (83.1%), discontinuity condition (57.1%), and joint spacing (54.0%) based on SMR.

• Slope morphometry shows the highest LHEF utilisation (97.0%), followed by relative relief (81.3%) and lithology (72.7%), while groundwater condition is the lowest (5.3%) based on LHZ.

Based on these findings, several targeted interventions are recommended for landslide risk reduction in Aizawl. Remedial measures such as drainage improvement, slope regrading, and retaining structures should be prioritised in Very High susceptibility zones (30.5%), particularly in Durtlang, Hlimen, and Vaivakawn. Strict regulation of slope modification is required, with mandatory geotechnical assessments before road cutting, excavation, or construction on slopes. Community based early warning systems should be developed using rainfall thresholds and locally observed indicators such as seepage and cracks, and disseminated through widely used channels like social media and television. Additionally, awareness programmes should focus on improving understanding of landslide causes and mechanisms, as knowledge has a significant influence on preparedness behaviour (ρ = 0.479). This study shows that community-identified landslide drivers in Aizawl broadly align with geotechnical and model-based factors, supporting integration of local knowledge into risk management. However, limitations include moderate sample size (n = 145) and moderate model performance (ROC–AUC = 0.726; PR–AUC = 0.605; F1 = 0.793). The framework should be applied cautiously to other regions. Future work should focus on improving datasets, model performance, survey coverage.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

All the authors contributed equally.

Funding

The research was funded by TDR-IOE, BHU scheme.

Data availability

The data will be made available on a reasonable request to the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.