HFENet: Hierarchical Feature Extraction Network for Accurate Landcover Classification

Landcover classification is an important application in remote sensing, but it always a challenge to distinguish different features with similar characteristics or large-scale differences. Some deep learning networks, such as UperNet, PSPNet, and DANet, use pyramid pooling attention mechanisms improve their abilities multi-scale extraction. However, due the neglect of low-level contained underlying network information differences between feature maps, difficult identify small-scale objects. Thus, we propose novel image segmentation network, named HFENet, for mining multi-level semantic information. Like HFENet adopts top-down horizontal connection architecture while includes two improved modules, HFE MFF. According levels information, module reconstructs extraction part by introducing mechanism fully mine With help channel mechanism, MFF up-samples re-weights maps fuse them enhance expression ability features. Ablation studies comparative experiments seven state-of-the-art models (U-Net, DeepLabv3+, FCN, DANet SegNet) are conducted self-labeled GF-2 sensing dataset (MZData) open datasets landcover.ai WHU building dataset. The results show that on three six evaluation metrics (mIoU, FWIoU, PA, mP, mRecall mF1) better than other mIoU 7.41–10.60% MZData, 1.17–11.57% 0.93–4.31% landcover.ai. can perform task refining images.