CASST_Net

we develop a spatio-spectral-temporal deep learning regression model , termed CASST-Net, which leverages a cosine attention mechanism to enhance feature representation to improve FVC estimation accuracy, can also dynamically adjusted the weighting values of each spectral band in the satellite data based on changes in vegetation physiological characteristics and canopy structure. This end-to-end network model comprises a feature extraction module, which leverages the Swin Transformer and a cosine attention mechanism to capture spatial-spectral-temporal information from multi-temporal unmanned aerial systems (UAS) and satellite data, and a fusion module that serves as a regressor for accurate FVC estimation.

本研究构建了一款命名为CASST-Net的时空谱深度学习回归模型。该模型通过余弦注意力机制强化特征表达，以提升植被覆盖度（FVC）的估算精度；同时可依据植被生理特性与冠层结构的变化，动态调节卫星数据中各光谱波段的权重值。这款端到端网络模型包含特征提取模块与融合模块：其中特征提取模块采用Swin Transformer与余弦注意力机制，从多时序无人机系统（UAS）及卫星数据中捕获时空谱信息；融合模块则作为回归器，用于实现高精度的植被覆盖度估算。