Effects of Digital Elevation Model Resolution on Unmanned Aerial Vehicle-Based Topographic Change Detection in Human-Altered Landscapes

Abstract

UAV-based topographic change detection is widely used in geoscience communities. The change detection involves comparison of two digital elevation models (DEMs) produced by UAV surveys, which are affected by the DEM resolution. Coarse resolution DEMs introduce errors in change detection, but the DEM resolution effect remains poorly understood. Moreover, effective strategies for mitigating the resolution effect have yet to be investigated. This study generated UAV-based DEMs at resolutions ranging from 0.1 m to 10 m with various resampling methods. The impact of DEM resolution on topographic change detection was then evaluated by analyzing the difference of DEM (DoD) and volume budget errors with indices such as mean error (ME), standard deviation (STD), and Moran’s I. The results from two human-altered landscapes showed that the random errors of DoD increase rapidly with the DEM resolution coarsening, while DoD systematic errors (spatial distribution of errors) become stable after 4 m resolution. The volume budget errors also increase with DEM coarsening. Coarser resolution DEMs tend to underestimate the volume budget (gross erosion, gross deposition, and net changes). Moreover, selecting an appropriate method for generating DEM is beneficial in decreasing the errors caused by the resolution effect. Among the seven methods (MAX, MIN, MEAN, BIL, NEAR, NEB, and TIN), the BIL method is optimum for mitigating both DoD and volume errors. The NEAR, NEB, and TIN methods are equivalent, and they are superior to the aggregation methods (MAX, MIN, MEAN). The slope of DoD (SDoD) should be considered when selecting a resolution for change detection. Large errors tend to appear in areas with large SDoD and vice versa. Coarse resolution DEMs are tolerable in areas with low SDoD, while high resolution DEMs are necessary in areas with large SDoD.

1. Introduction

Unmanned Aerial Vehicles (UAVs) combined with Structure-from-Motion (SfM) and Multi-View Stereoscopic (MVS) photogrammetry, which provided flexibility for topographic change detection, are widely used in geoscience communities [1]. The UAV-based change detection involves comparison of two DEMs produced by UAV surveys, namely, DEM of Difference (DoD). DoD obtained by subtraction of two DEMs is one of the most direct and commonly used methods for topographic change detection [2]. Compared with traditional point-based geodetic methods [3] (such as total station and GPS survey), the DEM-based methods can directly obtain spatially continuous topographic changes. As high spatial resolution terrain data become more and more accessible [4], DEM-based topographic change detection is widely used in various fields of earth sciences, such as gully erosion monitoring [5,6], fluvial transport [7], landslides [8], inferring sediment transport in rivers [9,10], and landslide evaluation and prediction [11].

Previous studies have shown that DEM resolution directly affects the results of terrain analysis and geographic modeling [12,13,14]. For example, studies have explored the effect of DEM resolution on the calculation of terrain factors [15], such as slope, aspect [16], and topographic wetness index [17], and their results showed that the terrain factors have distinct differences under different spatial resolution DEMs, while some functional relationships were also found between terrain factors and DEM resolution [18,19]. Scholars have also analyzed the influence of DEM resolution on various types of geographic analysis models, such as Revised Universal Soil Loss Equation (RUSLE) [20], Soil and Water Assessment Tool (SWAT) [21], Water Erosion Prediction Project (WEPP) [22], and flood inundation simulation [13], and reported that DEM resolution is a crucial factor for geographic analysis [23].

Selecting an appropriate resolution is important for DEM-based change detection. Very high-resolution DEMs (e.g., 0.1 m resolution) are not only difficult to acquire [24], but also increase the requirements of computing resources and processing time geometrically, while low-resolution DEMs (e.g., 30 m resolution) cause some losses in terrain information, resulting in reduced accuracy in terrain analysis [19,25].

Apparently, topographic change detection involves the comparison of multi-period DEMs, and hence the resolution effects of multi-period DEMs, which are different from previous studies (the resolution effect of single-period DEM). A few studies have focused on this effect. Kasprak et al. (2019) addressed the effect of survey technique and data resolution on topographic change detection and found that the magnitude of detected topographic change was dampened at coarser pixel resolutions. However, this study was limited to 0.05 to 1 m resolution [26]. Sankey et al. (2022) also looked at topographic changes at varying resolutions compared to sediment transport measures for a Mars analog study. Their results showed large differences between the remote sensing and in-site method results when the remote sensing data were resampled at the 0.1–10 m pixel resolutions [27].

However, topographic change detection includes both elevation and volume change detection. With coarse-resolution DEM, there would be errors in both elevation and volume change detection. The relationship between the errors and the DEM resolution remains poorly understood, especially the spatial structure of the errors. Moreover, how to mitigate the resolution effect is yet to be investigated.

The DEM resolution effect on topographic change detection could involve many factors. First, DEM generation rules, such as the nearest neighbor resampling [28], bilinear interpolation [29], and cubic convolution method [30], can yield DEMs with different details [31]. Previous research showed DEM generation methods have various effects on elevation errors or uncertainty [32,33]. Although these studies revealed the impact of different DEM generation methods on the quality of DEM, the impact of the resampling technique on its subsequent effects on topographic changes is less well understood. Second, the terrain characteristics, such as the terrain slope and roughness, were reported to affect the selection of DEM resolution for terrain analysis [14]. Resolution has less impact on flatter terrain but a greater effect on rugged terrain [34].

Accordingly, this study tries to explore the effects of DEM resolution on UAV-based topographic change detection in two human-altered landscapes. The specific aims are to: (1) evaluate the DoD and volume errors caused by the DEM resolution effect and model the relationships between the errors and the DEM resolutions; and (2) investigate the influencing factors of the resolution effect, such as DEM resampling methods and the topographic characteristics.

2. Methodology

2.1. Overview

To investigate the multiple effects of DEM resolution on topographic change detection, we employed the following workflow (Figure 1). First, we used UAV photogrammetry to collect terrain data (detailed in Section 2.3). Second, employed seven methods (i.e., 3 aggregating point-to-raster methods and 4 spatial resampling methods) to generate DEMs with different resolutions from point cloud data (detailed in Section 2.4). Third, DoDs and volume budgets at different resolutions were obtained by subtracting the two-period DEMs (detailed in Section 2.5). Then, we assessed the errors of DoD and volume budget at different resolutions, respectively, and then analyzed the relations between the errors and resolution (detailed in Section 2.6). Finally, we analyzed the influencing factors of the resolution effects (detailed in Section 2.7).

2.2. Study Areas

We selected two study areas (T1 and T2) with observed topographic changes for investigating the DEM resolution effect. The T1 area (Qiaogou), located in Suide County, Shaanxi Province, China, used to be a natural catchment, while new terraces were constructed on its hillslope in 2020. The hillslopes were cut and filled into terraces with an average width of 15 m. The land use was previously grassland and now is farmland. The T2 area (Tianshan) is an open mining pit, located in Tianshan District, Urumqi City, Xinjiang Province, China. During the survey period, there are significant terrain changes in downward excavation and partial accumulation (Figure 2), but the downward excavation (cut or erosion) is dominant. Compared with T2 area, both erosion (cut) and deposition (fill) alternate spatially, and their magnitudes are equivalent in T1 area. Detailed information about the study areas is shown in

Table 1. Introduction to the study areas.

Study Area	Name	Survey Period	Location	Area (km2)	Topographic Changes
T1	Qiaogou	2019–2021	37°34′2″ N 110°16′48″ E	0.15	New terraces were constructed on their hillslope in 2020
T2	Tianshan	2017–2019	85°3′48″ N 43°18′60″ E	0.15	Significant terrain changes in downward excavation and partial accumulation

.

2.3. Topographic Data Collection

The terrain data were collected by UAV photogrammetry. The DJI Phantom 4 Pro drone, equipped with a 13.2 × 8.8 mm CMOS sensor with an 8.8 mm focal length, was used for image collection. The precision of the on-board GPS of the DJI 4 Pro was ±1.5 m (horizontal) and ±0.5 m (vertical). During the data collection, the Pix4d capture was used for designing the flight and controlling the UAV. The flight comprised two orthogonal blocks (double grid), with an 80% overlap both along strip and cross-strip. The UAV flights started halfway up the mountain in each study area to keep the average flight height (

Table 2. Information on the UAV survey.

Study Area	Period	Flight Height (m)	Number of Images	GSD (m)	Mean Error (m)	Standard Deviation Error (m)
T1	2019	100	~580	0.027	0.001	0.035
	2021	100	~580	0.027	−0.001	0.041
T2	2017	120	~470	0.038	0.001	0.053
	2019	120	~470	0.038	0.001	0.046

). During flight, the camera angle was set to 20° with fixed infinity focus. Due to the low on-board GPS precision, we used 20 ground control points (GCPs) to improve photogrammetry accuracy and used 10 GCPs to evaluate the DEM accuracy (mean error of GCPs) and precision (standard deviation error of GCPs) (Table 2). The GCPs were collected by a Topcon Hiper SR GNSS-RTK. The horizontal and vertical precision of the Topcon Hiper SR GNSS-RTK was ±0.010 and ±0.015 m, respectively. The GCPs were collected at the same time as the images.

The Agisoft PhotoScan Pro 1.5 was used to produce dense point cloud data and orthophoto images. Moreover, 54 and 49 million points were acquired in the T1 and T2 areas with an average density of 360 and 327 points/m², respectively. The dense point cloud data were processed using the CloudCompare 2.12 alpha Stereo. In the T1 area, we used Cloth Simulation Filter filtering to classify ground points and vegetation in 2019, while the T1 area in 2021 and the T2 area had no vegetation, so filtering was unnecessary. Given that 30 GCPs were set in the two study areas, we used the GCPs to register the point cloud in CloudCompare manually. The registration errors of the two-period point cloud are less than 0.01 m in the two study areas. This ensured that the coordinate systems of the terrain data collected in different periods were completely consistent and could be used for change detection.

2.4. Generation of a Series of DEMs

To explore the impact of DEM resolution on topographic change detection, a series of DEMs at varying resolutions was generated from point clouds using different methods. The spatial resolution was set to range from 0.1 to 10 m. When the resolution range is 0.1–1 m and 1–10 m, the spacing between DEM resolutions is set to 0.1 m and 1 m, respectively. DEMs with resolutions exceeding 10 m were excluded because they are hard to capture topographic changes at this resolution [14].

Point cloud datasets provide flexibility for producing a series of DEMs at varying resolutions. Users can set the spatial resolution size to obtain the required resolution DEMs from the same point cloud. When generating grid DEMs from point clouds, multiple points may exist within the range of a pixel (window). Thus, three aggregation rules are widely used for generating DEMs: maximum (MAX), minimum (MIN), and mean (MEAN) aggregation [35]. The MAX, MIN, and MEAN methods take the highest, lowest, and mean value of the elevation points in the pixel, respectively. The three methods were implemented in the ClouldCompare 2.12 alpha Stereo.

In addition to MAX, MIN, and MEAN, many spatial resampling methods can generate DEMs from point features. The natural neighbors (NEB) and Triangulated Irregular Network with linear interpolation (TIN) are standard in the LiDAR-related industry. The nearest neighbor method (NEAR) and bilinear interpolation method (BIL) are also commonly used for DEM resampling in terrain analysis. Hence, we implemented the four resampling methods in ArcGIS 10.7 to generate DEMs at varying resolutions.

2.5. Topographic Change Detection

In this study, the topographic change detection includes two aspects: elevation changes in DEM and the volume budget. The first period DEM was subtracted from the second period DEM for each measurement period to obtain the elevation changes (i.e., DoD), where a positive DoD value represents deposition (also artificial fill) and a negative DoD value represents erosion (also artificial cut). By accumulating the negative and positive DoD values separately or together and multiplying them by the pixel area, we can obtain the volume budget, namely, the gross erosion, gross deposition, and net changes (volume) of the study areas.

2.6. Error Assessment

To assess the errors of DoD and volume, the reference data were generated from the original point cloud due to its high density and accuracy. We used the M3C2 method [36] in the ClouldCompare 2.12 alpha Stereo to calculate the elevation differences (DoD) of the two-period point cloud and then converted the M3C2 results into a raster (0.1 m resolution). Then, the error maps are obtained by subtracting the reference DoD raster from the observed DoD (Equation (1)) using ArcGIS 10.7:

$${Merror}_i={DoD}_i-{DoD}_{0.1}$$

(1)

where ${Merror}_i$ is the error map; i is the corresponding resolution; and ${DoD}_{0.1}$ is the reference DoD generated by M3C2.

Then, the mean value (ME) and standard deviation (STD) of the error grids were calculated. The ME represents the bias (systematic error) of the calculated DoD compared to the reference DoD. STD quantified the random error of the calculated DoD compared to the reference DoD. These two standard metrics have been widely used for data quality analysis.

However, the DoD error could be spatially distributed, which means although the ME and STD may be the same, the spatial structure could be completely different. For this reason, we also quantified the extent to which there is a spatial structure to the error. We used Moran’s I [37] for this purpose (Equation (2)). The Moran’s I ranges between −1 and 1. The closer its value to 1 or −1, the more positive or negative the spatial autocorrelation of errors, respectively. Its value of 0 indicates a random distribution of error.

$$Moran’sI={\displaystyle \frac{n}{S_o}}·{\displaystyle \frac{\sum _{i=1}^n\sum _{j=1}^n{w}_{i,j}{z}_i{z}_j}{\sum _{i=1}^n{z}_i^2}}$$

(2)

where $n$ is the number of spatial units indexed by $i$ and $j$; $x$ is the variable of interest; $\overline{x}$ is the mean of $x$; $w_{i,j}$ is a matrix of spatial weights with zeroes on the diagonal; and $S_o$ is the sum of all $w_{i,j}$.

We also calculated the volume errors. First, we calculated the reference volume budget from the reference DoD. Then, we quantified the relative errors of the volume budget:

$${error}_i={(Volume}_i-{Volume}_{0.1})/{Volume}_{0.1}$$

(3)

where ${error}_i$ represents the relative errors; $i$ is the corresponding resolution; and ${Volume}_{0.1}$ is the reference volume budget.

2.7. Influencing Factor Analysis

In addition to conducting an error assessment, the error indices were regressed against the DEM resolution. Then, the relationship between the DEM resolution and the errors was analyzed. However, the errors caused by the DEM resolution effect may not be constant but may vary in space. This means that there are some factors influencing the resolution effects. We assumed that this spatial variation is related to the characteristics of topography or topographic changes. Then, we used the average slope (SL) of the two-period DEMs, the difference of the SL (DoSL), DoD, and the slope of DoD (SDoD) as the metrics for quantifying the characteristics of topography and topographic changes. Then, the four metrics are regressed against the errors to validate whether a correlation exists between them.

3. Results

3.1. Topographic Changes

The change detection results were based on point cloud, used as reference data, obtained by the M3C2 method in the CloudCompare software (version 2.12). The topographic changes detected in the two study areas differed distinctly (Figure 3). The red and blue patches show deposition (Fill) and erosion (Cut), respectively. Due to the construction of terraces on the slope, the spatial distribution of topographic changes shows the alternating distribution of erosion/sedimentation in the T1 area (Figure 3). The gross erosion, gross deposition, and net changes are 79,337 m³, 514,530 m³, and 435,193 m³, respectively, in the T1 area. Meanwhile, the main changes in the T2 area are artificial erosion (excavation) caused by downward excavation of the mineral bed, especially at the bottoms of the pit. The gross erosion, gross deposition, and net changes are 781,430 m³, 76,398 m³, and 705,032 m³, respectively, in the T2 area.

3.2. The Resolution Effects on DoD Errors

3.2.1. The Effects on Magnitude of Errors

The DoD error changes with DEM resolution (Figure 4). However, this effect is related to the DEM generation method. With the aggregation methods (MAX, MIN, and MEAN), both the ME and STD increase as the DEM resolution becomes coarser. ME and STD show linear trends (r > 0.9, p < 0.01) and quadratic radical trends (r > 0.9, p < 0.01) with the DEM resolution, respectively. Obviously, not only do random errors (STD) increase, but systematic errors (ME) do as well (Figure 4).

Among the aggregation methods, the ME with the MAX and MIN method in T1 areas is positively and negatively correlated with DEM resolution, respectively, indicating that coarser DEM leads to more misdetection of deposition (if positive correlation) or erosion (if negative correlation). Meanwhile, the trends of ME with the MAX and MIN methods in the T2 area are opposite to those in the T1 area. This result could make sense because the dominant change type in T1 area (deposition/fill dominates) is opposite to that in T2 area (erosion/cut dominates). The STDs of the three methods (MAX, MIN, and MEAN) show a similar trend with DEM resolution. However, the MEAN method shows the lowest slope in both ME and STD, which means that it is optimal among the aggregation methods for DEM generation.

With the spatial resampling methods (BIL, NEAR, NEB, and TIN), the ME is always close to 0, indicating that they mitigate the systematic errors caused by the resolution effect. The correlation between ME and resolution decreasing to less than 0.45 also supports this finding. STD still increases with DEM resolution. However, the slope of the trend lines is much lower than that of aggregation methods (Figure 4c,d), showing the spatial resampling methods also decrease the random error caused by the resolution effect. The four spatial resampling methods present a very marginal difference in ME and STD (almost overlapped in Figure 4). The four methods are usable for DEM generation for change detection.

3.2.2. The Effects on Spatial Structure of Errors

We selected the MEAN method as a representative of the aggregation methods in the next comparison, given its better performance in ME and STD (Figure 4). Besides the error magnitude, we found that the spatial patterns of DoD errors vary with DEM resolution (Figure 5). At the 0.2 m resolution, the DoD errors are small and tend to show no spatial patterns in flat and gently sloping areas, except for some relatively higher errors at the edges of slopes. With the resolution increasing, both the negative and positive errors become clustered and alternated in space (Figure 5). This means that systematic errors still exist, although the ME is small when using the spatial resampling methods.

To quantify the spatial structure of the DoD errors, we calculated the Moran’s I (Figure 6). The Moran’s I of the methods show a similar trend with DEM resolution, that is, the Moran’s I increase with the increase of resolution in the two study areas (quadratic radical trends). The Moran’s I increase rapidly between 0.1 and 4 m resolution, but there are just slight changes after 4 m resolution (Figure 6). This result means that the spatial autocorrelation of DoD errors increases with increasing DEM resolution. When the DEM resolution is coarser than 1 m, the systematic errors could exist (Moran’s I > 0.5) (also supported by Figure 5). However, this effect becomes limited after 4 m resolution. Among the methods, the Moran’s I with the BIL and TIN methods are relatively lower than that with other methods. According to this point, the BIL and TIN methods are suitable for DEM generation for change detection.

3.3. The Resolution Effects on Volume Errors

Figure 7 shows the relationship between volume errors and DEM resolution when using different interpolation methods (BIL, NEAR, TIN, and NEB). As the resampling methods generally outperform aggregation methods (Figure 4 and Figure 6), we focus on the former in this analysis. The results show that as the DEM resolution coarsens, the errors in gross deposition, gross erosion, and net volume change all exhibit an increasing trend. However, the magnitude of these errors varies among the interpolation methods.

NEAR, TIN, and NEB methods all display substantial volume errors, demonstrating a negative linear correlation with resolution (r > 0.9, p < 0.01). This suggests that these three methods tend to underestimate gross deposition, gross erosion, and net volume change. Conversely, the BIL method consistently yields volume errors around 0%, although the fluctuation range slightly widens with increasing resolution. This indicates that BIL effectively minimizes systematic errors. With the BIL method, the positive and negative errors appear to be relatively random, effectively canceling each other out during accumulation. Therefore, from a volumetric error perspective, the BIL method emerges as the most suitable approach for generating DEMs for change detection.

3.4. The Influencing Factors

The spatial distribution of DoD errors is not constant but varies in space (Figure 5). We investigated whether this effect is related to characteristics of topography and topographic changes or not. Figure 8 shows the correlation coefficient (r) between SL, DoSL, DoD, and SDoD and DoD errors at different resolutions.

The correlation coefficient between SL and DoD errors initially starts high, exceeding 0.3, but gradually decreases with coarser resolutions, eventually falling below 0.1. This indicates that when using high-resolution DEMs for change detection, areas with larger average slopes tend to have larger DoD errors. However, as the resolution coarsens, the average slope becomes less influential.

Compared to SL, both DoSL and SDoD exhibit more stable correlations with DoD errors. Regardless of the resolution changes, the correlation coefficients between both DoSL and SDoD with DoD errors remain above 0.2. This suggests that these two metrics, to a certain extent, influence the errors caused by resolution. Areas with larger DoSL and SDoD values generally correspond to larger errors caused by resolution. Therefore, during topographic change detection, DoSL and SDoD can be utilized to predict areas more susceptible to the resolution effects.

Interestingly, the magnitude of DoD shows minimal correlation with DoD errors (Figure 8). This implies that the magnitude of topographic change, whether drastic or subtle, does not impact the DoD errors. We expected that areas with drastic terrain changes would be less affected by DEM resolution; namely, coarser resolution DEM could be used in these areas. However, the experimental results underscore the importance of considering the terrain data resolution during topographic change detection, even in drastic change areas.

4. Discussion

4.1. The Effects of DEM Resolution in UAV-Based Change Detection

DEM resolution is a key factor for topographic change detection. The DoD errors show positive correlations with DEM resolution. However, the quadratic radical trends in STD and Moran’s I show that the DoD errors caused by the resolution effect could have only slight changes if the resolution becomes even coarser (Figure 4 and Figure 6). This result is expected mathematically. Previous studies showed that terrain surfaces become smoother with coarser DEM resolution [38], that is, terrain slopes will decrease in steep areas and increase in gentle slope areas (also named “peak load shifting”) [34]. Thus, as DEM resolution becomes coarser, the DoD errors increase. However, if the DEM resolution is much coarser, all terrains tend to be flat, and thus the DoD could change slightly with the resolution.

Notably, we did not consider the propagation of survey errors in this study. Generally, survey errors are propagated through modeled terrains [6,33,39,40,41]. Studies have shown the resolution effects on DEM errors [42,43]. Coarser resolution data generally have higher survey errors. The errors of DoD at different resolutions are more complex than DEM errors because both surveys have measurement errors that are propagated through both DEMs and therefore within the DoD. Previous studies have used a minimum detection threshold to filter detection results [7,44]. However, finding the best confidence level for thresholding is difficult [41,45] and thresholding could lead to observation loss [46]. This study directly investigated the resolution effects on DEM-based change detection, which is beneficial for revealing the resolution effects in composite analysis with multiple resolution DEMs and for pre-evaluating the magnitude of errors.

There are some recommendations and limits that should be noted. First, this paper focused on the resolutions ranging from 0.1 to 10 m. The resolution smaller than 0.1 m is not meaningful because the resolution of the original SfM-MVS-produced DEM is 0.08 m in this study. Further study could extend the range of the resolution. However, this is related to the magnitude of the topographic changes; coarse resolution DEM could be hard to capture the changes if the change is too small. Second, the point cloud was generated from UAV photogrammetry. The top-down measurements of the UAV are different from the ground-based lidar measurements (oblique perspective). The change detection results and resolution effects could vary when the data acquisition method changes. Third, we used two human-altered landscapes: a mining pit and a terraced area. In these areas, anthropogenic disturbance is the dominant factor in the topographic changes. The topographic changes, namely DoD, have obvious mutations (high slope of DoD) in space (Figure 3). In a natural catchment, the slope of DoD could be small. We proposed that considerations should be more focused on the slope of DoD instead of the magnitude of DoD when evaluating errors in different landforms.

4.2. The UAV-Based DEM Generation Methods

The resolution effects on DoD and volume budget errors are related to the DEM generation method (Figure 4, Figure 6 and Figure 7). The selection of an appropriate DEM generation method can mitigate the DoD and volume budget errors. The spatial resampling methods (BIL, NEAR, NEB, and TIN) are superior to the aggregation methods (MAX, MIN, and MEAN). The aggregation methods exhibit not only large random errors but also systematic errors (supported by the ME in Figure 4). The MAX and MIN methods have the biggest errors. This is because sampling of the maximum and minimum value in these two methods leads to systematic bias, although they best preserve valleys (MIN) and peaks (MAX).

Among the spatial resampling methods, the results of ME, STD, and Moran’s I suggest that the BIL method is less sensitive to DEM resolution, and it is optimum for resampling DEM in topographic change detection (Figure 4, Figure 6 and Figure 7). Generally, the sensitivity of the resampling methods to DEM resolution could differ, and they can generate DEM with various surface details. The BIL methods produce a smoother surface, whereas the NEAR, NEB, and TIN methods generate a rougher surface [47]. Studies have shown that the BIL method is more robust for calculating terrain parameters [48,49]. Our result is in accordance with the literature.

However, the differences between the four resampling methods (BIL, NEAR, NEB, and TIN) are small (Figure 4, Figure 6 and Figure 7). All four methods are acceptable for DEM generation for change detection. In addition to the DEM resolution effects, terrain complexity and the specific application field should be considered when selecting an appropriate resampling (or interpolation) method [42,50,51].

4.3. The Characteristics of Topographic Changes

Previous studies have proposed that the potential of a DEM for characterizing a terrain feature (e.g., valley lines and ridge lines) relies on not only DEM resolution but also the terrain feature itself [18,52]. If the size of a terrain feature is large, coarse-resolution DEM is also tolerable. Topographic changes can also be considered a type of terrain feature. However, our results argue that regardless of the size of topographic changes, high resolution is necessary.

Because the topographic change detection involves two-period DEMs, the DEM resolution effect is different from that of single-period applications. In applications of single-period DEM, such as gully extraction [53,54], the size of the object could be more important than DEM resolution. Even coarse-resolution DEMs could capture the object if the object is very large. In topographic change detection, coarse-resolution DEMs may capture the changes when the magnitude of changes is large, but the differences in the two DEM surfaces caused by the smoothing effect of DEM resolution could be complex and induce non-ignorable errors. However, in areas with uniform topographic changes (DoD) (e.g., flat terrain before and after the change), coarser resolution DEMs are tolerable regardless of whether the magnitude of changes is small or large. Then, we proposed that the slope of DoD should be a crucial consideration for determining resolution. High-resolution is critical for change detection in areas with large variation of terrain changes.

5. Conclusions

This paper generated UAV-based DEMs at 0.1–10 m resolutions by various resampling methods. Then, the DEM resolution effects on topographic change detection were evaluated by analyzing the DoD and volume budget errors with indices such as ME, STD, and Moran’s I. The results from two human-altered landscapes showed that both the magnitude and spatial correlation of DoD errors increase with DEM resolution coarsening. The random errors of DoD increase rapidly (supported by STD) with the DEM resolution coarsening, while systematic errors of DoD become stable after 4 m resolution (supported by Moran’s I and ME). The volume budget errors also increase with DEM coarsening. Coarser resolution DEMs tend to underestimate volume budgets (all gross erosion, gross deposition, and net changes).

Moreover, the resolution effects on DoD and volume budget errors are related to the DEM generation method and characteristics of topographic changes. Among the seven methods (MAX, MIN, MEAN, BIL, NEAR, NEB, and TIN), the BIL method is optimum for mitigating both DoD and volume errors. With the BIL method, the volume error is close to zero due to minimized systematic errors and mutual cancelation of positive and negative random errors. The NEAR, NEB, and TIN methods are equivalent, and they are superior to the aggregation methods (MAX, MIN, and MEAN). The characteristics of topographic changes are crucial for selecting a resolution for change detection. We proposed that the slope of DoD (SDoD) can be used for quantifying the topographic change characteristics. In areas with low SDoD (i.e., flat terrain before and after the change), coarse-resolution DEMs are tolerable. However, resolution is critical for change detection in areas with large SDoD (i.e., a flat terrain before the change and a high, steep slope after the change).

These findings provide reference and guidance for selecting an appropriate DEM resolution and DEM generation rule for topographic change detection or for pre-evaluating the magnitude of errors in change detection.