AUTHOR=Yu Yaran , Wang Zhiyong , Li Zhenjin , Ye Kaile , Li Hao , Wang Zihao 

TITLE=A Lightweight Anchor-Free Subsidence Basin Detection Model With Adaptive Sample Assignment in Interferometric Synthetic Aperture Radar Interferogram

JOURNAL=Frontiers in Ecology and Evolution

VOLUME=Volume 10 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/ecology-and-evolution/articles/10.3389/fevo.2022.840464

DOI=10.3389/fevo.2022.840464

ISSN=2296-701X

ABSTRACT=Excessive exploitation of coal resources has caused serious land subsidence, which seriously threatens the lives of residents and the ecological environment in coal mining area. Therefore, it is of great significance to precisely monitor and analyze the land subsidence in the mining area. To automatically recognize and detect subsidence basins in the mining area from the interferometric synthetic aperture radar (InSAR) interferograms with wide swath, a lightweight model for detecting the subsidence basins from InSAR interferogram with anchor-free and adaptive sample assignment based on YOLO V5 network, named Light YOLO-Basin model, is proposed in this paper. First, depth and width scaling of convolution layers and depthwise separable convolution are used to make the model lightweight to reduce the memory consumption of CSPDarknet53 backbone network. Furthermore, the anchor-free detection box encoding method is used to deal with the inapplicability of anchor box parameters, and an OTA (Optimal Transport Assignment) adaptive sample assignment method is introduced to solve the difficulty of optimizing model caused by abandoning the anchor box. To verify the accuracy and reliability of the proposed model, we acquired 64 Sentinel-1A images over Jining and Huaibei coalfield (China) for training model and experimental verification. Contrasted with the original YOLO V5 model, the mAP value of the Light YOLO-Basin model increases from 45.92% to 55.12%. The lightweight modules of the model speed up the calculation with the GFLOPs (one billion Floating Point Operations) from 32.81 to 10.07, and reduce the parameters from 207.10 MB to 40.39 MB. The Light YOLO-Basin model proposed in this paper can effectively recognize and detect subsidence basins in mining areas from the InSAR interferograms.