70 / 2023-03-31 18:22:56

Multilevel differential network for change detection from high resolution remote sensing images：a case study in open-pit mine

deep Learning,open-pit mine,change detection,remote sensing

      The monitoring of land use/land cover changes in mining areas is an important guarantee for comprehensive regulation and effective remediation of the mining environment (Xiang Yang et al., 2019). With the continuous development of remote sensing, change detection in mining areas based on remote sensing has become a research focus in the field of mining ecology.

      Existing deep-learning-based change detection studies have mainly extracted change information from two phases of images by both early-differential and late-differential approaches [Zhang C et al. 2020,]. The early difference network directly superimposes or takes the absolute difference between two period images to meet the single input requirement, it is able to focus on the discovery of change regions throughout, but the change detection results are prone to boundary fragmentation and internal missing, because the early layer of the network cannot provide the deep features of the single period image to help image reconstruction [Zhang C et al. 2020]. The late differential network can receive two phases of images separately, and its early layer is responsible for extracting the deep features of the two phases, while the late layer is responsible for extracting the changes. Although the early layer can provide the deep features of a single phase image to help image reconstruction, it is prone to error accumulation [Peng D et al, 2019].

      To solve the appeal problem, this study proposes a multilevel difference network (MDNet), which introduces both early and late difference strategies to compensate for the shortcomings of each. MDNet uses an "encoder-decoder" architecture. The encoder consists of two parts: the siamese network and the semantic segmentation network. The siamese network is responsible for receiving the two phases of the images and extracting the deep features of the images separately, and then extracting the change features. The semantic segmentation network is responsible for receiving the absolute difference between the two phases and extracting the change features directly. In order to fuse the two widely different change features, the multi-level change features fusion module (MCFFM) proposed in this study is used in the decoder to effectively fuse the two in a weighted manner.

      This study applied on the team's home-made mine change detection dataset, show that the Precision , Recall and F1-score of MDNet applied to open-pit mine change detection reach 87.4%, 91.6% and 89.2%, respectively, which is a significant improvement compared to the current state-of-the-art change detection networks, such as FCCDN (Chen et al., 2022), SNUNet (Fang et al., 2022), Siam-U-Net (Chen et al., 2022) and SMCDNet (Li et al., 2022), and are important for intelligent change monitoring in open-pit mines.