Lightweight land cover classification via semantic segmentation of remote sensing imagery and analysis of influencing factors

• DOI: 10.3389/fenvs.2024.1329517
• Publication Date: 2024-02-15T04:44:54Z
• Authors: Guoying Wang; Jiahao Chen; Lufeng Mo; Peng Wu; Xiaomei Yi
• Keywords: Land Cover

Land cover classification is of great value and can be widely used in many fields. Earlier land methods traditional image segmentation techniques, which cannot fully comprehensively extract the ground information remote sensing images. Therefore, it necessary to integrate advanced techniques deep learning into study semantic However, most current high-resolution networks have disadvantages such as large parameters high network training cost. In view problems above, a lightweight model via segmentation, DeepGDLE, proposed this paper. The DeepGDLE designed on basis DeeplabV3+ utilizes GhostNet instead backbone feature extraction encoder. Using Depthwise Separable Convolution (DSC) dilation convolution. This reduces number increases computational speed model. By optimizing rate parallel convolution ASPP module, “grid effect” avoided. ECANet channel attention mechanism added after module pyramid pooling focus important weights Finally, loss function Focal Loss utilized solve problem category imbalance dataset. As result, effectively And extensive experiments compared with several existing algorithms DeeplabV3+, UNet, SegNet, etc. show that improves quality efficiency segmentation. other networks, more applied classification. addition, order investigate effects different factors performance images verify robustness model, new dataset, FRSID, constructed dataset takes account influences than public experimental results WHDLD metrics mIoU, mPA, mRecall are 62.29%, 72.85%, 72.46%, respectively. On FRSID 65.89%, 74.43%, 74.08%, For future scope research field, may fusion multi-source data intelligent interpretation