4.3. Adaptive Identification Results of Potential Landslide Hazards
Before the start of the potential landslide hazard adaptive identification experiment, as a comparison, we first constructed CNNs using 48 × 48, 32 × 32, 16 × 16, or 8 × 8 px fixed windows alone. The results show that the model using the 16 × 16 px window alone has the best identification effect, with 76.51% for precision, 74.38% for recall, and 75.43% for F1 score. Then, we carried out experiments on potential landslide hazard identification with different window sizes adaptively adjusted based on , and the results are as follows.
Using 102 landslide samples compiled from multisource data and 352 randomly selected non-landslide samples, the CNN for the 48 × 48 px window was trained and used to identify 120 samples with significant deformation within the 48 × 48 px window, among which a total of 30 landslide samples with significant deformation were identified (taking the 48 × 48 px window as an example, the corresponding samples and their
values are shown in
Table 4). The quadratic threshold T
48 = 0.013 was set, and only the 7 samples with
≥ 0.013 were retained and expanded to some extent using data enhancement techniques. The selected sample dataset was entered into the model for training and the results (
Table 4) show that the precision was 79.42%, recall was 77.80%, and F1 score was 78.60%.
Using 310 landslide samples compiled from multisource data and 942 randomly selected non-landslide samples, the CNN with a 32 × 32 px window was trained and used to identify 150 samples with significant deformation within the 32 × 32 px window. Among the landslide samples with significant deformation in the 32 × 32 px window, the 32 × 32 px samples at the location of the above 7 retained 48 × 48 px samples were excluded, so 37 32 × 32 px samples were substantially retained, and a trisection threshold T
32 = 0.007 was set, from which only 12 32 × 32 px samples with
≥ 0.007 were further screened. Then, the data enhancement technique was used for some degree of expansion. The selected sample dataset was input into the model for training, and the results (
Table 5) show that the precision was 83.63%, recall was 82.92%, and F1 score was 83.27%.
Using 1040 landslide samples compiled from multisource data and 3168 randomly selected non-landslide samples, the CNN with a 16 × 16 px window was trained and used to identify 300 samples with significant deformation within the 16 × 16 px window. Among the landslide samples with significant deformation in the 16 × 16 px window, the above 12 retained 32 × 32 px samples at the location of the 16 × 16 px samples were excluded, so 69 16 × 16 px samples were substantially retained, and the dichotomous threshold T
16 = 0.0038 was set, from which only 35 16 × 16 px samples with
≥ 0.038 were further filtered. Then, after data enhancement of the 140 samples, the number of samples was further expanded using the oversampling technique. The selected sample dataset was input into the model for training, and the results (
Table 5) show that the precision was 87.56%, recall was 85.06%, and F1 score was 86.29%.
Using 3744 landslide samples compiled from multisource data and 12,672 randomly selected non-landslide samples, the CNN with an 8 × 8 px window was trained and used to identify 500 samples with significant deformation within the 8 × 8 px window. Among the landslide samples with significant deformation in the 8 × 8 px window, the above 35 retained 8 × 8 px samples at the location of the 16 × 16 px samples were excluded, so 122 8 × 8 px samples were substantially retained. Then, after data enhancement, 488 samples were obtained, and the number of samples was further expanded using the oversampling technique. The selected sample dataset was input into the model for training, and the results (
Table 5) show that the precision was 90.57%, recall was 86.35%, and F1 score was 88.41%.
Combining the above adaptive identification results for landslides of different scales, the final identification results were calculated as 85.30% for the precision, 83.03% for the recall, and 84.15% for the F1 score. As shown in
Table 6, this result outperforms the identification effect of the 16 × 16 px preferred fixed-window CNN model (compared to the models with 48 × 48, 32 × 32, 16 × 16, and 8 × 8 px fixed windows, which we have found through previous studies [
59] to be optimal in this study area) and the traditional machine learning model SVM. Thus, we have demonstrated the effectiveness of the adaptive identification method of potential landslide hazards proposed in this study. These results are better than the recognition effect of the fixed-window CNN [
6,
33], which proves the effectiveness of the adaptive identification method for potential landslide hazards proposed in this study. In addition, using the potential landslide hazard data provided by Shandong GEO Surveying and Mapping Institute for the study area (56 potential hazards with deformation, of which 20 were covered by deformation samples) for practical testing, a total of 16 potential landslide hazards were identified using the method in this study (
Figure 7), and the identification rate was 80%.
The information in respect of the potential landslide hazards was further verified through detailed field verification data from Shandong GEO Surveying and Mapping Institute. For example, potential hazard A (99°30′E, 25°14′N) in
Figure 7 is located in Songpo Village, Suyang Town, Yongping County. In terms of topography and geomorphology, the elevation range of this potential hazard is 1783–1925 m, with an elevation difference of 142 m. It is a mountain slope terrain, with a slope gradient of about 20° and a slope direction of about 285°. The slope is transformed into cultivated land, with less vegetation cover, bare surface at the back edge, broken surface at the front edge, a large elevation difference, and heavy shadow. The geomorphology comprises a tectonic denudation geomorphic unit, with dense river valleys and gullies, strong weathering of the rock body, mainly denudation, and multilevel razor planes distributed at different elevations. In terms of lithology and geological structure, it belongs to the medium-to-thick laminated harder rock group sandwiched by the thin laminated softer rock group, with a fine-grained structure dominant, joint fissure development, weak weathering resistance of the joint fissure surface, and a soft rock interlayer prone to landslide formation. In terms of hydrology, the topography of the landslide area has a large slope, with slope surface step-like spreading, which is not conducive to the rapid drainage of rainfall at the surface; rainfall is able to fully infiltrate to recharge the underground pore water, creating down-slope runoff, resulting in pore water pressure. At the same time, groundwater infiltration converges at the bedrock surface, softening the clay layer, reducing its shear strength, and forming a potential weak zone, which directly leads to the formation of landslide deformation and settlement under the action of gravity. The optical image is characterized by an overall irregular shape, with clear front and back edges and perimeter. The left lower side of the landslide body has broken topography and has collapsed, with clear steep canyons; the right lower-image texture is blurred, with deformation of terraces and a creeping phenomenon; the left upper-middle has landslide steps and uneven traps. The InSAR deformation is characterized by the deformation rate exceeding the threshold value of 20 mm/a in both the ascending and descending results. Through field verification (
Figure 8), it was found that the overall slope of the potential landslide hazard is about 20°, the slide direction is about 285°, the terraces in the middle of the slope are locally slippery, the surface texture of the back edge is rough, the front edge is sloped and unsupported, there are many landslide steps, and the main signs of deformation are located in the middle and lower parts. Through the field investigation, the reliability of this study on the adaptive identification of potential landslide hazards based on multisource factor dataset and multiple InSAR monitoring techniques was once again proved. Inevitably, the identification results we obtained are not always accurate. For example, the deformation rate at B (99°40′E, 25°13′N) in
Figure 7 slightly exceeds the threshold value of 20 mm/a. Also, the adaptive modeling results based on the landslide formation context and generation conditions indicate that it can be classified as the potential landslide hazard. However, it was found that this identification result was not accurate enough through field verification. The site did once have a geological hazard potential due to the construction of the road, but the corresponding local hazards have all stabilized at present and the geological environment conditions in its vicinity have improved significantly. Therefore, it is no longer necessary to give priority to its prevention and control as a potential hazard site.