Deep learning for check dam area extraction with optical images and digital elevation model: A case study in the hilly and gully regions of the Loess Plateau, China

Abstract

Check dams are widely used on the Loess Plateau of China to control soil and water loss, develop agricultural land and improve watershed ecology. Detailed information on the spatial distribution of check dams and the area of dam land is critical for quantitatively evaluating hydrological and ecological effects, planning the construction of new dams and repairing damaged dams. Therefore, this research presents a method that integrates deep learning and geospatial analysis to facilitate the extraction of check dam areas in broad areas from high‐resolution Gaofen‐2 (GF‐2) multispectral imageries, including red (R), green (G), blue (B) and near‐infrared (NIR) bands and digital elevation model (DEM). First, we generated three datasets with different band combinations (RGB, RGB + NIR and RGB + DEM) using GF‐2 remote sensing images combined with Advanced Land Observing Satellite—the Phased Array type L‐band Synthetic Aperture Radar DEM data to determine the optimal data combination for dam area extraction. Next, four widely used semantic segmentation networks—Fully Convolutional Network (FCN), U‐Net, PSPNet and DeepLabv3+—were modified to support the arbitrary number of input channels and evaluated for check dam area extraction. Finally, the check dam candidate areas in the Yan River basin were extracted from DEM using geospatial analysis to optimize the dam area extraction results. The results showed that all deep learning (DL) models could extract dam areas quickly and accurately with mean intersection over union and overall accuracy values >85% and >98%, respectively. PSPNet had the best performance for testing datasets with different band combinations. We also found that the DL models in the RGB + DEM images had the best remote sensing image segmentation results, avoiding many miss‐classified pixels. The F1 scores in the RGB + DEM test dataset for FCN U‐Net, PSPNet and DeepLabv3+ reached up to 92%, or 3.3%, 2.3%, 1.9% and 3.1% higher than the corresponding values for RGB images. Potential check dam candidate regions (~3969 km2) were obtained for application analysis, which reduced the original area by 50%. The DL models for the RGB + DEM images were applied in the dam candidate regions, extracting 91.9 km2 of agricultural production dam land and 10.6 km2 of runoff and sediment silted dam land. The extraction results will facilitate quantitative analyses of check dams, improve the management of these structures and promote the efficiency of controlling soil losses.

Intelligent classification and analysis of regional landforms based on automatic feature selection

“…However, such multiple pooling operations are liable to cause spatial information loss, resulting in the gradient features being mistaken, the segmentation boundaries being ambiguous and the fine accuracy being insufficient among landform types (van der Meij et al, 2022). To address the existing questions, the UNet model has been put forward to break the deadlock, which not only applies contextual semantic information to predict the type of each pixel but also uses the jump structure connection to combine deep and shallow features (Sun et al, 2023).…”

“…The smaller the batch size is, and the larger the initial weight of the learning rate is, the easier the overfitting occurs. Given the overall computer performance, the network parameters are set as Table 4 shows, including dropout (Srivastava et al, 2014). The loss curves of different regions exhibit significant jitter peaks (the waveform displays a series of random and periodic amplitude patterns in the histogram due to the occurrence of extreme values).…”