A dual-branch weakly supervised learning based network for accurate mapping of woody vegetation from remote sensing images

Abstract

Mapping woody vegetation from aerial images is an important task bluein environment monitoring and management. A few studies have shown that semantic segmentation methods involving deep learning achieve significantly better performance in mapping than methods involving field-based measurement and handcrafted features. However, current deep networks used for mapping vegetation require labour-intensive pixel-level annotations. Thus, this paper proposes the use of image-level annotations and a weakly supervised semantic segmentation (WSSS) network for mapping woody vegetation based on Unmanned Aerial Vehicle (UAV) imagery. The network comprises a Localization Branch (LB) and an Attention Relocation Branch (ARB). The LB is trained in stage 1 of the mapping to identify regions with the most discriminative vegetation, while the ARB is introduced to better mine semantic information, which enhances the ability of the class activation maps (CAMs) to represent useful information. The ARB inherits the weights from the LB in stage 2 and uses a Multi-layer Attention Refocus Structure (MARS) into the network to expand the receptive field to enable the model to process global features. Thus, same-category regions that are located farther apart are better captured. Finally, the region focused by the dual branches are integrated to more accurately cover the areas to be segmented. Using UAV imagery datasets, namely UOPNOA and MiniFrance, along with quantitative metrics and qualitative results, the network demonstrates performance better than existing state-of-the-art related methods. The effectiveness and generalization of each module of the network are validated by ablation experiments. The code for implementing the network will be accessible on https://github.com/Mr-catc/DWSLNet.

Keywords

• Vegetation remote sensing
• Weakly supervised learning
• Semantic segmentation
• Environment monitoring