A Multi-Scale Coordinate Attention Feature Fusion Network for Province-scale Aquatic Vegetation Mapping from High-resolution Remote Sensing Imagery

Yuzhe Wang,Yunfei Sun,Peng Zhang,Yanlan Wu,Hui Yang

DOI: https://doi.org/10.1109/jstars.2024.3507023

IF: 4.715

2024-01-01

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

Abstract:Aquatic vegetation plays a crucial role in the sustainable development of water resources. Although high-resolution remote sensing imagery has been used to develop methods for the automatic extraction of aquatic vegetation, these methods are usually limited to specific areas(such as a lake or a river). Extracting aquatic vegetation on a large scale faces numerous challenges, including high environmental complexity, significant seasonal variations, and poor cross-regional adaptability. Currently, relying solely on either prior knowledge-based methods or classifier methods makes it difficult to comprehensively capture the complex characteristics and variations of aquatic vegetation. To address this issue, we propose an expert feature-embedded deep semantic segmentation model. This model uses an attention mechanism to fuse expert features with deep features, dynamically adjusting and optimizing the weight allocation between them. This enables the model to better capture the subtle differences among various types of aquatic vegetation and environmental changes. Specifically, we propose a feature fusion module based on the squeeze-and-excitation coordinate attention mechanism. This module initially employs the squeeze-and-excitation mechanism to compress spectral and index features, generating adjusted weights $\alpha$ and $\beta$ . Subsequently, it dynamically adjusts the combined weights of these features through the coordinate attention mechanism. This enhancement enables the model to more effectively recognize aquatic vegetation in complex environments and bolsters its adaptability to cross-regional and seasonal variations. We selected eight regions in Anhui Province for evaluation. Compared to HRNet, our method improved Precision and IoU by 10.91% and 20.72%, respectively. Finally, using Anhui Province as the study area, we produced a province-scale aquatic vegetation density map. The results indicate that there are approximately 25.78 $\mathrm{km}^{2}$ of aquatic vegetation in Anhui Province.