Downscaling the spatial resolution of satellite imagery based on morphometric parameters to estimate the Topographic Wetness Index using GIS tools

Abstract

Digital elevation models (DEMs) play a key role in extracting morphometric factors like fill sink, flow accumulation, profile, flow width, slope, plan curvature, aspect, and total catchment to estimate the Topographic Wetness Index (TWI) that provides key information for modelling and predicting hazards related to mass movement or landsliding. The range and accuracy of information, including the topographic feature, can influence the quality of results, significantly impacting the severity and likelihood of occurrence of mass movement. Therefore, it plays a crucial role, especially in mountainous regions. This research aims to investigate, evaluate, and identify the optimal downscaling methodology for DEMs and assess its impact on DEMs at different spatial resolutions. Morphometric factors derived from the DEM were examined using six distinct methodologies: kriging, nearest neighbor, majority, bilinear, bi-cubic, and the Hopfield Neural Network (HNN). Three geospatial databases containing a 20 m, 12.50 m, and 1.50 m resolution DEM were used for analysis. Six downscaled topographic maps were generated: kriging, nearest neighbor, majority, bilinear, bi-cubic, and the HNN. Results validated from field elevation survey points have an accuracy of 1.50 m from total stations and the Global Positioning System (GPS). By predicted actual values at corresponding locations, field survey points named GPS, and total station data having vertical accuracy of 1.50 m, were used as a standalone reference source to validate the downscaled results. Consequently, we strongly endorse downscaling DEM employing the HNN technique to obtain the most precise morphometric parameters for fill sink, flow accumulation, profile, flow width slope, plan curvature, aspect, and total catchment to assess TWI maps. The result shows RMSE accuracy improved accuracy approximately 25% to 75% and 93%, respectively, when progressing from low to medium resolution (30, 20, and 12.50 m) DEMs. It is concluded that all the mentioned techniques, Bi-cubic, HNN, Nearest Neighbor, Majority, Bilinear, and Kriging, can improve the overall accuracy of DEM. However, Bi-cubic and HNN methods demonstrate superior accuracy relative to the Bilinear, Nearest Neighbor, Majority, and Kriging methods. This study attempts to address a gap in past research by evaluating and selecting the most effective downscaling approaches for DEMs and testing their accuracy to enhance topographical features at various spatial resolutions in mountainous regions.

Introduction

The Jhelum Basin (study area, Fig. 1) encompasses Azad Jammu and Kashmir, which falls in a mountainous region, with the most significant portions located in the Abbottabad, Muzaffarabad, and Mansehra districts, demarcating its boundaries. This mountainous region, characterized by very rough terrain and flash floods with sudden mass movement, is a high rainfall region. In terrain analysis, hydrological modelling, and the evaluation of natural hazards, Digital Elevation Models (DEMs) and their morphometric derivatives play a crucial role in downscaling processes. There is considerable uncertainty in the accuracy of other relative morphometric parameters when moving from low to high spatial resolutions. However, increasing DEM resolution using resampling and downscaling techniques increases the tolerable resolution. Additionally, resampling and downscaling techniques can be applied to enhance the accuracy and resolution of grid DEMs. These methods can also be performed on different source datasets covering the same areas.

Fig. 1.

Base map of the study area: a Index map of Pakistan, b Satellite imagery, c Elevation map (ArcGIS 10.7.1; https://www.esri.com/software/arcgis).

Digital Elevation Models are essential to digitally represent Earth’s surface geomorphology and hydrological attributes^1,2. Higher DEM resolution will improve the efficiency of the results for accurate morphometric analysis and offer advantages over conventional topographical maps^2,3. Therefore, DEM downscaling has been developed recently and seeks to create finer-scale digital terrain models to fully capture geomorphology and hydrological variation in the Earth’s surface^1,4. DEMs data into GIS, enabling efficient data management, analysis, and visualization. DEMs provide detailed and accurate topographic information for morphometric analysis². They have broad applications in geoscientific research, but the reliability and utility of these applications depend heavily on DEM accuracy.

Digital elevation models have been widely used in a variety of earth-surface applications such as hydrogeology, geology, civil engineering, cartography, and landsliding design⁵. Use of DEM analysis increased in popularity over the preceding years for accurately predicting and then forecasting natural hazards such as mass movement or landslides, topsoil loss, flooding, and rising ocean levels. In this sense, “precise forecasting of natural hazards” mentions having the capacity to accurately forecast the likelihood, magnitude, and exact location of natural hazards^6,7. Effective reservoir water level prediction is one of the more fundamental tasks in this discipline to better manage these valuable resources⁸. With little observational data, the structured framework can be applied to the watershed⁹. The most credible model was subsequently determined by comparing the outcomes from global climate models using statistical standards, and the chosen model was then used to generate subsequent information¹⁰. With the growing availability of high-resolution DEMs, the TWI can now be computed, even in remote regions, although it is intuitive, straightforward, and can perform effectively across diverse applications¹¹.

It is essential to overcome these restrictions and provide trustworthy, high-resolution morphometric outputs using different downscaling methods, post-processing techniques, or smoothing strategies. Our study offers a new comparative analysis of six methods: HNN, Nearest Neighbors, Majority, Kriging, Bicubic, and Bilinear interpolation applied across progressive resolution gains. These methods were combined with seven morphometric factors: sink, flow accumulation, profile, flow width, slope, plan curvature, and aspect. In contrast to other studies, we emphasize the transmission of advances in DEM accuracy into morphometric variables during low-to-high resolution downscaling. We aim to bridge a significant gap in the trustworthiness of DEM-derived morphometrics by emphasizing the preservation of fine-scale terrain signatures and providing recommendations for choosing suitable downscaling techniques for high-resolution disaster modelling.

This discrepancy is crucial in related derivatives, which significantly influence models for erosion, flood susceptibility, landslides, and mass movements, and are extremely sensitive to even minor topographic changes. Although techniques are available that significantly improve the morphometrics of the first derivatives during low-to-medium resolution upscaling¹². They show moderate resolution improvement and, in some cases, slightly degrade second derivatives (i.e., curvatures and TWI) at finer scales. Previous research showed that methods of modeling and data sources used, including DEM plus its generated topographic or morphometric changes variables, including the aspect, slope, plan curvature profile, and TWI have an essential influence on the precision of natural risks prediction^13,14. Generally, high-resolution DEMs are thought to provide more accurate outcomes when modelling natural hazards. Results deliver more accurate statistics to derive topographical characteristics for aspect, profile curvature, plan curvature, slope, and TWI¹⁵. In simpler terms, improved, accurate input of DEMs can deliver further detailed statistics regarding topographical aspects and enhance the precision of these solicitations. DEM can be obtained using several approaches, including field surveying, photography, optical and radar satellite imagery, and light sensing and ranging (LiDAR), mainly in a grid raster format. As a result, depending on the data collection methods, these statistics are accessible at varying resolutions, besides degrees of accuracy^16,17.

For applied climate modelling, increased TWI estimate accuracy has significant practical ramifications. More accurate TWI values enhance the determination of vulnerable to saturation regions and the representation of spatial soil moisture variability in hydrological modeling, resulting in more accurate runoff production and flood response simulations. Improved TWI accuracy reduces misclassification of unstable slopes and increases the durability of susceptibility algorithms for landslide vulnerability assessment by enabling improved identification of areas with high soil moisture and porous water pressure. Parallel to this, enhanced TWI enables more accurate interpretation of the relationship between topography, hydrology, and slope stability in mass movement and slope process evaluation, making it easier to identify regions vulnerable to failure or transport of sediment. The practical importance of enhanced TWI accuracy for method based landscape evaluation, land use planning, and risk evaluation is highlighted by the explicit connection to these applications.

Several approaches and tools were applied to increase the accuracy of DEM for natural hazard forecasting^18,19. Many depend on downsizing DEM to different resolutions²⁰, which they then utilize to establish the initial parameters for natural hazards modelling in various geographic areas. Resampling, also known as downscaling techniques, can therefore be applied to increase grid DEM accuracy and resolution¹². Kriging, Bilinear, and Bi-cubic resampling are the most used techniques for downsizing. The key advantages of these approaches are their computational effectiveness and the absence of dependency on previous DEM statistics at altered resolutions aimed at training. However, accuracy may not be adequate in all cases. The key advantages of these approaches are their computational efficiency and independence from previous DEM data at various training resolutions. Some advanced deep learning artificial intelligence models have recently been proposed for integration learning^21,22.

Most of these techniques are based on methods described for high-resolution pictures^21,23. The results show that these strategies enhanced DEM accuracy, reducing RMSE and improving alignment with the reference dataset. Even though these strategies showed promise for improving the resolution of super-resolved DEMs, they required many training datasets and costly hardware to develop sophisticated neural network models. With positive findings, a downscaling strategy aimed at DEM accuracy for Hopfield Neural Networks (HNN) was recently anticipated. In addition to improving DEM accuracy, this approach is computationally efficient and cost-effective for desktop computers. It is still unknown how DEM downscaling methods may increase the precision of resulting morphometric parameters, like fill sink, flow accumulation, profile, flow width, slope, plan curvature, aspect, total catchment, to calculate the TWI.

This study aims to fill a gap in the previous research through downscaling and linking Bilinear, Nearest Neighbor, Majority, Kriging, Bi-cubic, and Hopfield Neural Network (HNN) techniques for enhancing accuracy for the seven morphometric parameters. This research included three DEMs with varying resolutions: a spatial resolution of 30 m, 20 m, and 12.50 m downscaled using six methods. Downscaled topographic maps of 20 m, 12.50 m, and 1.50 m of vertical resolution were validated from the DEM produced by plotting points of GPS and total station, having 1.50 m resolution. This work addresses a key gap in DEM downscaling research by investigating how topographical factors, namely morphometric derivatives and the Topographic Wetness Index (TWI), impact downscaling precision. The impact of local topographic constraints on downscaling reliability gets little focus, despite the fact that past research has mostly focused on developing innovative downscaling approaches or assessing their success using general accuracy criteria. This study introduces a new paradigm for future understanding of the spatial variability of DEM downscaling performance and increases the physical foundation for accuracy evaluation by linking downscaling errors to morphometric features and TWI.

Data used

This study utilized three DEM samples (Table 1) and survey points collected from the field in the region of Azad Jammu and Kashmir and its nearby areas of various resolutions.

Table 1.

DEM data source of different spatial resolutions for the study area.

DEM Resolution (m)	Source	Vertical accuracy(m)	Resolution sampling before(m)	Zoom Factor
30	SRTM, ALOS PALSAR	20	90	3
20	SRTM, ALOS PALSAR	12.50	60	4
12.50	ALOS PALSAR	1.50	30	4
1.50	Total Station/GPS	0.50	20	4

The SRTM produced the 30 m DEM in dataset 1, which was acquired from USGS Earth Explorer (https://earthexplorer.usgs.gov and https://www.earthdata.nasa.gov). This dataset is situated in the territory of Azad Jammu and Kashmir and covers an area of approximately 100 km x 100 km. This location was chosen because it has exceptionally minimal vegetation cover, which has little influence on the elevation accuracy of the SRTM data. Dataset 1 was used to study the impacts of downscaling and resampling the DEM, having a resolution lower than 30 m. To achieve a lower resolution, the original 30 m DEM was upscaled to 90 m. This was done to determine the impacts of downscaling and resampling.

DEM for dataset 2 was collected at the exact location as the previous dataset. That DEM was produced by integrating contours from 1:10,000 scale topographic maps using Kriging (ordinary Kriging). The original DEM was upscaled from 20 m to 60 m resolution by averaging the 20 m pixel values within each 60 m grid cell. Dataset 3 was for DEM collected through the same procedure mentioned above. That DEM was upscaled to 30 m pixels of the original DEM by kriging methods. Ground surveys collected dataset 4. The test field was in the study area, which measures approximately 200 m by 200 m. One hundred fifty points were collected and interpolated with Kriging methods to provide a gridded 1.50 m DEM dataset for reference. To ensure the accuracy of the reference DEM, the RMSE was calculated for validation points. The total station captures elevation points with a resolution of 1.50 m (Table 1).

Methodology

This section outlines the proposed methodology for downscaling the spatial resolution of satellite imagery based on morphometric parameters to estimate the TWI. The process begins with an open-source DEM data source of different spatial resolutions for the study area, which was acquired from the USGS Earth Explorer. Next, this study performed downscaling and linking of Bilinear, Nearest Neighbor, Majority, Kriging, Bi-cubic, and HNN techniques for enhancing accuracy for the seven morphometric parameters. The workflow adopted for DEM resampling, downscaling, and calculating the TWI is shown in Fig. 2.

Fig. 2.

Workflow adopted for DEM resampling, downscaling, and calculating the topographic wetness index.

Hopfield neural network

In 1982, John Hopfield developed HNN, a recurring artificial neural network²⁴. This strategy maximizes the spatial dependence or resemblance between nearby pixels in an image. This work considers spatial dependence as an essential function for downscaling DEM in hazard risk areas^5,25,26. The workflow adopted for this study is shown in Fig. 2.

1

2

where is the correction value of the elevation pixel, = average computed values around pixel i, j. x, y means the elevation of pixel value (x, y) for the original resolution. = output value for the p, q subpixel within the newly created resolution of pixel and means zoom factor. Pixel elevation value are restored if the average value elevation for all the sub-pixels inside its footprint is greater than the elevation for . The neuron’s resulting value is deducted if the mean value for the elevation of all the subpixels in one pixel are greater than the elevation . On the other hand, the HNN network iteratively operates till the energy value E is as low as possible⁵.

3

Bilinear interpolation

Bilinear interpolation is a typical technique for integrating pixel values using a weighted mean for neighboring pixels²⁷. That approach computes the intensity of each pixel using a specific formula in Eq. 4, assuming it fluctuates linearly across the image:

4

where f ₁ (0,0), f ₂= f (1,0), f ₃ = f (0,1), f ₄ = f (1,1), here f ₁, f ₂, f ₃ and f ₄ are four surrounding pixels of given formula and f (x, y) shows the elevation of the (x, y) pixel for resampling the case of DEM.

Bi-cubic interpolation

This method uses the 16 nearest surrounding pixels (4 × 4) values to estimate the location (x, y). Below are Eqs. 5 and 6 in matrix form, as derived by^27,28.

(5)

6

Here, the elevation value of a pixel represents ĝ (x, y); g (x_i, y_j), value of the surrounding pixel (i) of the original image, and bi weight value calculated by means of the distance from x, y to i pixels from the original image.

Kriging interpolation

Kriging is used as an interpolation technique centred on the spatial connection amongst sampled points²⁹. This technique involves two distinct steps. Fitting a variogram and figuring out the sampled points spatial covariance structure is the first step. The covariance construction obtained from the first step was utilized to produce a weight matrix, which was then used to compute the value for the elevation of pixels x and y. Kriging’s f (x, y) computation is comparable to bi-cubic resampling in the following ways, Eq. 6.

Nearest neighbor technique

The ratio of the observed mean distance to the expected mean distance is known as the Nearest Neighbor Index. The average distance between neighbors in a fictitious random distribution is known as the anticipated distance. If the index is lower than 1, the pattern indicates clustering; if it is higher, a trend toward dispersion or competition. In ArcGIS, there is no direct formula; it uses Voronoi (V_i) diagrams following geometric principles for analysis, calculating the mean distance between every attribute and its closest adjacent feature to compute a nearest neighbor index. Within polygon Vi, the attribute estimates were at unstamped points. According to Webster and Oliver³⁰, the estimated values (z) at the closest single sampled data point x_i were equal to zˆ (x₀) = z(x_i).

Majority technique

The last option used in this research is the Majority technique.is which applies a majority algorithm to decide the new value of the cell^30,31. Similar to the nearest neighbor method, it is mostly employed with discrete data; the majority approach (Eq. 6) often yields a smoother outcome than the nearest.

7

where C_out is the output cell value, is the input value, C_in is the input cell value.

The most common value inside a zone is found in ArcGIS using the “Majority” option in the Zonal Statistics tool. In essence, it determines the value that occurs most frequently inside the zone’s designated region.

Morphometric factors

The current investigation includes seven morphometric variables used to calculate the TWI: fill sink, flow accumulation, profile, flow width, slope, plan curvature, aspect, and total catchment.

Aspect and slope

The most prevalent approach for determining the aspect and slope of a DEM is to use a 3 × 3 grid cell neighbourhood. The 3 by 3 cell for slope and aspect may be calculated using the maximum slope technique, the neighbourhood technique, the Quadratic Surface technique, and the maximum Downhill Slope technique³². Jones³³ demonstrated the Horn methodology’s suitability for computing aspect and slope from DEM by evaluating two approaches using a typical Morison’s surface. This approach calculates the slope and aspect in this study and is part of the features in the ArcGIS software.

Fill sink, flow accumulation, flow width, plan curvature, and profile curvature

Fill sink, flow accumulation, flow width, plan, and profile curvature were used as provisional influences in several mass movement or landslide assessments, hydrogeological and physical modelling, and landform grouping. These parameters were the most used^34,35. The curvatures represent the topography’s second-order products³⁶. Zevenbergen³⁷ proposed a technique for calculating using DEM and implemented it as a feature tool in ArcGIS.

Topographic Wetness Index

The Topographic Wetness Index, based on local topography, indicates the chance of water collecting in a specific location. The morphometric properties of a DEM were used to compute TWI, a soil moisture measurement. A gridded DEM was commonly used to calculate TWI, and the outcome of TWI for a given location is determined from the following Eq. 9^38,39.

9

β is the gradient in radians for the local region, and a is the uphill slope of the area per unit length contour. The calculation of tanβ algorithms affects the TWI value^5,39.

Accuracy assessment

The RMSE is commonly used to assess the accuracy of aspect, fill sink, flow accumulation, slope, flow width, plan curvature, and profile curvature for TWI calculation using an altered DEM^40,41. The RMSE of numerous DEMs can be used to examine differences in topographic factors. Additional statistics were used in addition to the RMSE. The spectrum of values and any patterns or trends in the slope were evaluated using data such as Mean Absolute Error (MAE), which displays results in terms of minimum, maximum, and mean standards. The aspect was assessed using the RMSE and aspect value difference categories, which describe changes in slope direction and may have a substantial influence on hazard and natural process models. Plan and profile curvatures were assessed based on the percentage of correctly identified curvature features, including concave shapes and zero-curvature regions. In addition to the previously mentioned characteristics, histograms for topographical variables were calculated for assessment. Histograms can graphically illustrate value trends and distributions, as well as contrasts or similarities in slope distributions across different resolutions and accuracy levels.

Results and discussions

Six downscaling methods (Kriging, Nearest Neighbor, Majority, Bilinear, Bi-cubic, and HNN) were applied to the DEM test dataset to examine their impact on derived morphometric parameters. The factors, also named as zoom factor, were used to analyse the implications of up-sampling and down-sampling techniques for accessing coarse resolution to moderate resolution, e.g., 30 m to 20 m, 20 m to 12.50, and 12.50 m to 1.50 m DEMs resolution. Table 1 displays the zoom influences that were utilized. Visual and quantitative analyses determine how DEM sampling and downscaling impact topographic characteristics.

Six resampling strategies utilised in the study, HNN and cubic, clearly outperform kriging, nearest neighbor, majority, and bilinear resampling, as seen by a visual evaluation of Table 2. As shown in Table 2, the resampling of bilinear, bi-cubic, and HNN produced comparable results closer to the slope estimated from the referenced DEM. However, the histograms for bilinear resampling digital elevation models are often remarkably similar to those generated from the original low resolutions.

Table 2.

Summarizes the RMSE differences between the resampled DEMs and their reference versions.

Datasets	Method of Resampling	RMSE	Datasets	Method of Resampling	RMSE
30-meter DEM	90-meter resampled	6.26	12.5-meter DEM	20-meter resampled	1.55
	Nearest	6.12		Nearest	1.55
	Bilinear	3.36		Bilinear	1.43
	Bi cubic	2.12		Bi-cubic	1.09
	Majority	6.01		Majority	1.43
	Kriging	6.56		Kriging	1.67
	HNN	1.94		HNN	1.03
20-meter DEM	60-meter resampled	4.22	1.5-meter Ground Survey	12.5 m resampled	1.49
	Nearest	5.30		Nearest	1.44
	Bilinear	1.88		Bilinear	1.36
	Bi cubic	0.94		Bi-cubic	1.05
	Majority	5.12		Majority	1.33
	Kriging	5.47		Kriging	1.55
	HNN	1.76		HNN	1.01

Slopes created from downscaled and resampled DEMs were compared, and it was discovered that both procedures might improve slope accuracy. Slope computations based on low-resolution DEMs exhibit persistent underestimation relative to high-resolution DEMs due to the smoothing effect. Figure 3 depicts this trend in slope histograms obtained from referenced sources and resampling of DEMs from various approaches. This result of utilizing resampling techniques to improve DEM resolution, the number of places having larger slopes improved and became more similar to those shown for the referenced DEM. The visual study contrasted histograms of slope estimates obtained from downscaled and resampled DEMs, indicating that both improve slope accuracy.

Fig. 3.

Slope histogram generated from datasets 1, 2, and 3.

Figure 4c shows that three-dimensional resampling methods do not considerably improve slope values for higher resolution DEMs compared to moderate resolution DEMs. The levelling consequence on geography might explain why this is less noticeable whenever the original DEMs possess a spatial resolution of twenty or thirty meters rather than sixty or ninety meters. The slope histograms of DEM at intermediate resolution in Dataset 1 demonstrate an increasing slope angle after downscaling and resampling procedures, with values exceeding 10°. Kriging interpolation, HNN downscaling, and bi-cubic resampling resulted in more pixels with slope angles between 10°and 40°compared to low-resolution DEMs.

Fig. 4.

Showing the resolution quality of DEMs Downscaling a 30 m, b 20 m, c 12.50 m using six methods, while Table 2 summarizes the RMSE differences between the resampled DEMs.

Table 3 quantitative analysis shows how sampling influences affect slope angles. The slope maximums and averages in the first dataset range from 30 to 20 m DEMs. The second data set (20 m to 12.50 m DEMs) demonstrates increased slope angles at medium resolutions. Resampling techniques considerably improved the statistics, even though the maximum and average slope values.

Table 3.

RMSE components of recognition accuracy data for all three testing datasets.

Datasets	Method of resampling	Slope %		Aspect %		Profile %		Curvature %		Fill %		Flow accumulation %		Flow Width %
		Inaccuracy	Accuracy	Inaccuracy	Accuracy	Inaccuracy	Accuracy	Inaccuracy	Accuracy	Inaccuracy	Accuracy	Inaccuracy	Accuracy	Inaccuracy	Accuracy
DEM 30 m	No resample (90 m)	68.1	32	61.3	39	58	42	59	40.60	58.7	41	58.7	41	58.7	41
	Bilinear	63.9	36	57.6	42	57	43	55	45.00	55.9	44	55	45	58.7	41
	Kriging	57.3	43	56.9	43	54	46	53	47.30	55.2	45	53.1	47	55	45
	Nearest	56.9	43	53.9	46	54	46	52	47.60	53.9	46	51.3	49	54.8	45
	Majority	55.5	44	53.8	46	53	47	52	47.70	51.6	48	51.2	49	51.3	49
	Bi Cubic	54.6	45	53.6	46	53	48	52	47.80	51.5	49	51.1	49	51.2	49
	HNN	54.2	46	51.7	48	52	48	51	49.50	51.4	49	51	49	51.1	49
DEM 20 m	No resample (60 m)	53.9	46	51.1	49	51	49	50	49.60	50.9	49	50.5	50	51	49
	Bilinear	53.8	46	50.3	50	51	49	50	49.70	50.8	49	50.4	50	50.5	50
	Kriging	53.7	46	47.12	53	47	53	49	51.00	50.7	49	50.3	50	50.4	50
	Nearest	53.6	46	47.1	53	46	54	49	51.10	49.2	51	47.8	52	50.3	50
	Majority	53.1	47	47.1	53	46	54	49	51.20	49.1	51	47.7	52	49.6	50
	Bicubic	51.3	49	46.3	54	46	54	49	51.30	49	51	47.7	52	47.8	52
	HNN	50.3	50	43.7	56	46	54	49	51.30	44.1	56	47.6	52	47.7	52
DEM 12.50 m	No resample (20 m)	25.3	75	17.4	83	22	78	21	78.70	20.5	80	24.3	76	24.6	75
	Bilinear	25.1	75	15.12	85	20	80	20	79.80	20.3	80	17.6	82	24.3	76
	Kriging	25.1	75	12.52	87	20	80	18	82.00	18.6	81	17.6	82	22.7	77
	Nearest	23.9	76	11.2	89	19	81	18	82.40	17.6	82	15.7	84	21.6	78
	Majority	20.7	79	10.92	89	19	81	16	84.00	16.5	84	15	85	15	85
	Bicubic	17.4	83	10.1	90	17	83	15	85.00	15.8	84	8	92	8	92
	HNN	8.88	91	8.32	92	16	84	7.3	92.70	9	91	7.1	92.9	7	93

For all utilised resampling practices, except kriging interpolation, increase the maximum slope values of the original 30-meter and 20-meter DEMs (24.40°and 32.91°, respectively), bringing them closer to the finer 20-meter and 30-meter for reference DEMs (44.49° and 57.57°, respectively). Similarly, in medium-resolution datasets for all cases, the average values from the resampled slope statistics are only around 2.1° less when compared with the target slopes having high resolution and increase by roughly 3.1° from 4.2° when linked with the original low resolution of slope data.

When the greatest and mean slopes have been considered, the impacts of resampling a DEM from moderate to high resolution are comparable to those at medium resolution. Except for the bilinear and kriging approaches in the 5-meter dataset 3, all three high-resolution DEM datasets shown in Table 2 exhibited an increase in both mean and maximum slope values. Although mean slopes increased by 0.5° to 1.5° and maximum slopes increased by roughly 5° to 7°, the rise was not as substantial when compared to medium resolution data.

HNN downscaling reduces RMSE for 7.3° and 6.91° for coarser resolution slope information to approximately 2° and 3° of HNN downscale slope information, correspondingly. This is the most significant improvement. Bi-cubic resampling results 4.5° for 30 m DEM for RMSE and 5.23° for 20-meter DEM, which can improve slope accuracy. Kriging, nearest, majority, and bilinear are not especially noticeable compared to these approaches. Bilinear resampling decreases RMSE by over 35% for 30 m to 20 m and 20 m to 12.50 m, reaching 4.10° and 4.86° respectively. In the 30-meter to 20-meter case, however, kriging indicated a larger increase in RMSE values.

RMSE and MAE were utilized to assess the effects of DEM resampling on slope accuracy. The data in Table 2 shows that slopes produced for the coarser DEMs may be substantially similar to those of the fine reference DEM through resampling techniques. When comparing with DEMs that were downscaled from medium to high resolution, downscaled resampling between coarser resolution DEMs to medium resolution results in a larger gain in slope accuracy. A mean reduction of almost 45% in RMSE is seen for both instances of downscale resampling from 30 m to 20 m and 20 m to 12.50 m.

Also, MAE-based results evaluation is almost identical to those of the RMSE. Using bi-cubic resolution reduces the MAE in slope calculations from 4.3° to 3.6° and 1.9° to 2.8° in HNN downscaled. After down-sampling from 30 m to 20 m, the MAE for the slope reduced by nearly 45%, with values almost identical to RMSE, respectively. However, regarding MAE, bilinear, nearest neighbor, majority, and kriging resampling are less effective than bi-cubic and HNN downscaling. Bilinear resampling lowers MAE in both datasets, although kriging does not reduce MAE when down-sampling from 30 m to 20 m.

Compared to the coarser to medium downscaling, the statistics for the three datasets that underwent high to medium downscaling are less impressive. When comparing the resampled slopes from the original low-resolution data, the RMSE and MAE reductions are only between 6% and 40%. HNN and bi-cubic downscaling remain the finest downscaling methods. HNN downscaling outperforms bi-cubic reprocessing in datasets 1, 2, and 3, with RMSEs of 7°, 1°, and 2°, respectively. The MAE statistics values of 6.13°, 1.95°, and 2.13° for dataset 1, respectively. The MAE values for the bicubic resampling are 6.88°, 1.99°, and 2.44° for dataset 2, respectively.

In dataset 3, bilinear resampling produced the poorest results for medium to high resolutions, with RMSE and MAE of 2.13° and 1.21°, respectively. Dataset 3 has an MAE of 2.97° and an RMSE of 2.92°, respectively, indicating much poorer results than those of the low-resolution slope. This suggests that, in many cases, the nearest, bilinear, majority, and kriging interpolations of DEM are ineffective strategies to improve slope data values. Using bi-cubic and HNN downscaling approaches, the slope generated by low-resolution DEMs may be modified to match that of higher-resolution DEMs.

Aspects of downscaled DEM

The aspect refers to a slope’s position concerning the sun, which determines how much solar energy the region receives. Aspect runs from 0 to 360 degrees and is calculated by measuring clockwise from north. Different values can be grouped into different group attributes based on the intervals used. At 90-degree intervals, slope directions may well have been divided into four topographical orders: south, east, north, and west. Alternatively, slope orientations can be divided into eight groups using 45-degree intervals: Northwest, Northeast, West, Southwest, East, Southeast, South, and Northwest. The dimensions and resolution related to referenced DEMs determine these various degrees of aspect changes. The assessment from this aspect is based on the proportions in collections having an aspect assessment alteration within intervals like 0–10°, 10–20°, 20–45°, 45–90°, and 90–180°, as well as statistics like RMSE.

Aspect values derived from the original and resampled DEMs were measured against the high-resolution reference DEM. The resulting differences were categorized; for example, a pixel showing 40° in the 20 m DEM and 55° in the reference was assigned to the 10–15° difference class. This reflects a 15° aspect difference. The proportion is estimated as the proportion of the entire number of pixels in a set to the total number of pixels. Table 3 shows detailed results for each category of aspect differences.

The aspect and slope difference characteristics provided in Table 3 were studied to determine the implications for aspect resampling values and, ultimately, the precision of simulation mass movement or catastrophe forecast. The aspect distributions derived from low-resolution DEMs (90 m, 60 m) differ substantially from those obtained using higher-resolution DEMs (30 m, 20 m), as evidenced by the percentage variations in Table 3. The accuracy of dataset 1 and dataset 2 aspect values from 90 to 60 m falls within the range of 50%, which is a significant difference compared to the aspect and slope values resulting from dataset 3 higher resolution DEMs (approximately 30% to 5% correspondingly). The 45–90° group has an approx. 13% share of the aspect of 30 m DEM, compared to the 90–180◦ group’s approximately 4%. The percentages from the 60 m DEM are approximately 19% and 22%, respectively. If two pixels have facets that are 90 degrees apart, they are facing opposite directions. As a result, these aspects and slope directions are calculated from the low-resolution DEM for the pixels in these groups that deviate significantly from the similar value determined by the higher-resolution DEM.

The aspect values derived from medium-resolution DEMs (20 m, 30 m) show strong agreement with higher-resolution benchmarks (12.50 m, 20 m) in datasets 2 and 3, unlike results from datasets 1 and 2. Across all datasets, over 55% of medium-resolution aspect values fall within 10° of their high-resolution (1.50 m/12.50 m) counterparts at matching locations. Resampling techniques can reduce the gap between aspects calculated from different DEMs. Resampling approaches (HNN, bi-cubic) improve the fidelity of low-resolution DEMs to match high-resolution references (Table 3), effectively enhancing the reliability of derived slope and aspect parameters.

The RMSE results in Table 3 show how resampled methods enhanced aspect accuracy by downscaling the original DEM from low to intermediate resolution. The low-resolution digital elevation models at 20 m for dataset 2 and 30 m for dataset 1 have RMSE ranges between 40° and 34°, respectively, indicating considerable deviations from the reference DEM for 20 m and 30 m. At 30 m resolution, both (bi-cubic and HNN) methods have RMSE values 18°. For dataset 2, aspect analysis showed further improvement with RMSE values of 18° (bi-cubic) and 16° (HNN).

Even though they are less successful than bi-cubic and HNN, added resampling algorithms, like bilinear and Kriging, have improved spatial accuracy. The RMSEs of aspects for 30-meter bilinear DEM resampled and interpolated through kriging are 23° to 32°, respectively, whereas for 20-meter resampling from bilinear and kriging interpolation, they are 25° to 29°. These are still much cheaper than the originals.

In datasets 2 and 3, the impact of DEM resampling from high to medium resolution on aspect accuracy improvements is less noticeable than in datasets 1 and 2. The RMSE enhancement is less than 10◦ across all three datasets. The gain in aspect accuracy is smaller than 20, which can be insufficient for more or less applications, such as mass movement or landslide analysis, to generate an overall accuracy improvement, especially because the DEMs in datasets 2 and 3 have been downscaled by more than four times.

When resampled, techniques were used for DEMs, the proportion from groups 0–10° was about threefold, reaching 49% for dataset (1) In contrast, the ratios of groups 35–100° and 100–180° declined dramatically from 23% to 10% to 5% within dataset 1 and from 18% to 7% to 8% for the 60 m dataset (2) The aspects and slope accuracy are also much improved by the HNN and bi-cubic downscaling for datasets 2 and (3) In dataset 2, the proportion of the 0–10° group increased from 45%, 51%, and 65% to 77%, 87%, and 69%, respectively, using HNN and bicubic downscaling. For dataset 3, HNN and bi-cubic processing reduced the frequency of 35–100° and 100–180° slope/aspect values by a minimum of 33%, potentially exceeding this threshold in some instances.

Plan curvatures, fill sink, flow accumulation, flow width, and profile curvatures from downscaling

In general, plan curvatures, fill sink, flow accumulation, flow width, and profile curvatures, the elevation surface’s second components, describe the slope variation rate along directions parallel to the contour lines and perpendicular to the profile lines. A DEM may be used to calculate curvatures of the plan and profile using m − 1 as the unit of measurement. Identifying these structures (curvature characteristics) is critical because the amount of movement in various soil material movement events, like mass movement/landslides and debris flows, corresponds to landforms like concave and convex shapes.

Concave and convex structures must be precisely identified to predict mud and soil movement. When a concave shape is inaccurately identified as a convex piece, the position of water flow and collection may be inaccurately anticipated. Consequently, the accuracy of curvature feature identification is employed in this work to evaluate curvature computation, along with statistics such as RMSE. The proportions of accurate and unfitting curvature topographies obtained after a DEM resampling, in contrast to curvature topographies generated by the benchmark DEM from higher resolution, serve as statistical indicators for this accuracy assessment. Table 3 displays quantitative results for RMSE and curvature feature identification accuracy across the five experimental datasets.

According to the findings, using DEM reprocessing at low resolutions to moderate resolution, such as from DEM 30 m to DEM 20 m or from DEM 20 m to 12.50 m, may somewhat boost the correctness of plan and profile curvatures, created on both RMSEs and curvature feature assessments. Employing HNN downscaling, the analysis shows that RMSE values, derived from plan, fill, flow accumulation, flow width, and profile curvature, decreased significantly. The percentages declined significantly, reaching approximately 20% for dataset 1 and 7% for dataset 2, compared to the original low-resolution DEM, which had RMSEs ranging from roughly 70% to 25%.

Furthermore, utilizing these parameters reduced the number of incorrectly detected curvature topographies by around 10%. In datasets 2 and 3, these parameter accuracy ratios were reorganized and reduced, despite a slight increase in RMSE from HNN downscaling. For instance, compared to “no resampling” data, the percentage of accurate profile curvature feature identification rose by 25%.

HNN downscaling demonstrates markedly better performance than alternative methods (bi-cubic, nearest neighbor, majority, bilinear) in all test datasets, as evidenced by RMSE (Table 3) and curvature metrics. Kriging methods fail for curve downscaling applications. Although resampling often improved curvature precision, the findings for all three datasets in Table 3 did not consistently surpass “no resampling”, specifically in the five investigations reported in Table 3.

TWI of the resampled and DEMs

As a result, imprecise TWI measurements may yield incorrect evaluations of soil erosion or safe and dangerous landslide zones. As indicated in Fig. 5, TWI was developed in this study area using dataset 3. Figure 5 shows the resulting TWI of kriging, majority, nearest neighbor, bi-cubic, bilinear, and HNN DEMs downscaling, demonstrating an enhancement in the details of these features. The resultant TWI for downscaled Kriging, bilinear, bi-cubic, and HNN DEMs. Approximately specific topographies in TWI images found as of resampled DEMs, mainly those as of bicubic and HNN-downscaled dataset exhibit significantly reduced distortion and pixelation artefacts compared to non-resampled DEMs, despite the fact that the degree of particulars of TWI remains meaningfully lesser than from the reference image. When comparing these resampling algorithms (Fig. 5) to the original DEM. HNN is created from ground survey points, bi-cubic performs equally across resampling approaches, and HNN somewhat outperforms bi-cubic. Performance for HNN is equivalent, with bi-cubic by a tiny margin. The Kriging and bilinear techniques yield significantly inferior results to the other methods. After comparing RMSE and an MAE analysis (Fig. 6), which revealed a higher level of accuracy. TWIs derived from resampled DEMs are compared to those calculated with reference DEMs. The RMSE and MAE values are somewhat higher, indicating that they are more accurate than those computed with “no resampling.” In dataset 3, the bi-cubic and HNN algorithms only considerably reduced RMSE and MAE by 90%.

Fig. 5.

TWI calculated with six methods, kriging, majority, nearest neighbor, bi-cubic, bilinear, and HNN method, using a DEM resolution of 12.50 m (ArcGIS 10.7.1; https://www.esri.com/software/arcgis).

Fig. 6.

RMSE and MAE validation with survey points.

Accuracy assessment for improved downscaled DEM and topographic factors

Likening the accuracy enhancement of topographic factors and DEM reveals that DEM accuracy relates to the accuracy improvement of DEM’s first by products. Resampling methods like bi-cubic and HNN downscaling meaningfully condensed the MAE and RMSE of all the approaches, as shown in Table 2. This is evident because these techniques enhanced DEM accuracy across experimental datasets. Similarly, Table 3 indicates a considerable increase in the precision for the first derived components, such as slope and aspect. The improved DEM accuracy resulted in a significant gain of slope and aspect precision for all three datasets, as seen by the RMSE and MAE statistical markers.

The enhancement in accuracy for DEM, including the subsequent second byproducts of TWI, did not follow the same pattern as the first byproduct. The results of all types of DEMs reveal that increasing resampled DEM accuracy does not necessarily improve second derivative accuracy. The decrease in the RMSE and MAE for the curves and TWI in dataset 1 (30 m to 20 m downscaling) and dataset 2 (20 m to 12.50 m downscaling) indicates that DEM improvement only affected low to medium resolutions. These derivatives improved in other scenarios, even though DEM accuracy augmentation enhanced the precision of all morphometric factors and TWI for the high-resolution corrected DEM. As a result, it is not suggested that these second derivatives be computed using resampled DEMs from medium to high resolution analysis, like downscaling from 30 m to 12.50 m.

Conclusions and future recommendations

The study’s findings show that DEM downscaling accuracy is strongly connected to morphometric derivatives and that of the Topographic Wetness Index (TWI), highlighting the relevance of terrain-specific elements in modelling downscaling efficiency. This research goes further than traditional accuracy assessments by indicating where and why mistakes are more likely to occur in certain topographic environments. By demonstrating this relationship, the work provides an entirely new viewpoint that improves the understanding and application of downscaled DEMs, notably in hydrological and geomorphological applications, while also emphasizing the uniqueness and relevance of the suggested method.

This research presented different methods of statistical downscaling and resampling for 30, 20, and 12.50 m DEMs. Six interpolation techniques, including bilinear, nearest neighbor, majority, HNN, bi-cubic, and kriging, were used to downscale the DEM into 20, 12.50, and 1.50 m vertical elevation values to estimate seven morphometric factors like aspect, curvature, slope, profile curvature, fill sink, and flow accumulation flow width to generate downscaled topographic wetness index maps. HNN downscaling and bi-cubic outcomes outperform all other resampling algorithms compared with survey points. Further, comparative analysis indicates HNN demonstrates marginally superior performance to bi-cubic resampling across evaluated metrics. The resampling approach used for the DEMs effectively improves the value of their derivatives; the lessening of RMSEs and MAE metrics results in significant gains between 49% and 08% for all the scenarios. Techniques for shifting from low to medium resolution (30 and 20 m) and high resolution (12.50 m) DEM yield accuracy percentage assessment by applying RMSE of approximately 25% to 75% and 93%, respectively. These findings show that these topographic characteristics may be more precise by using resampling and downscaling approaches for dataset 3 of high resolution compared to low to medium-resolution DEMs. Lower RMSEs indicate that the 12.50 m DEM’s greater precision corresponds to better precision for the high-resolution DEMs dataset. Generated TWI maps from 12.50 m show a coarser kriging and finer resolution for bi-cubic and HNN methods.

Future studies must look at the benefits of different downscaling approaches in this mountainous data limited region, which have recently been found to influence DEM accuracy positively. To discover the best strategies to increase their accuracy, topographic variables and TWI should be immediately downscaled from lower to higher resolution to compare this dataset with actual survey data for validation. So, more advanced investigation is required to validate the effect of low-resolution resampling and downscaling high to medium-resolution DEMs.

Author contributions

Haider Shabbir, Muhsan Ehsan, and Danish Raza: Conceptualization, Methodology, Software, Formal analysis, Visualization and Writing - Original Draft; Muhsan Ehsan: Methodology, Formal analysis, Writing - Review & Editing; Muhammad Tayyab Sohail, Hamad Alqahtani, Ramzi Almutairi, Yasir Almutairi: Methodology, Writing - Review & Editing; Qadeer Ahmed: Software, Visualization, Writing- original draft preparation. Muhsan Ehsan: Conceptualization, Writing - Review & Editing and supervision, Project administration and supervision.

Funding

No funding is available for current research.

Data availability

The data will be shared with the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

All other co-authors have approved the manuscript and agree with its submission without any conflict.

Consent to publish

The authors agree with the publication of the manuscript.