Crop classification in time-series remote sensing images based on multi-feature extraction and attention connection semantic segmentation model

Time-series multispectral remote sensing imagery provides a dynamic representation of crop variations over time, highlighting the disparities in growth conditions among diverse crop types. This method offers a superior capability to differentiate between crop types compared to single temporal phase imagery. However, when applied to practical issues of crop segmentation, challenges persist, including low segmentation accuracy and underutilization of features. To comprehensively address these challenges, we design a new multidimensional multi-attention semantic segmentation network suitable for crop extraction from time-series multispectral remote sensing images (TSRSnet). Specifically, we propose a multidimensional feature extraction module (MFE) for efficiently mining spatial, temporal and spectral dimensional features in multi-temporal remote sensing images. The module consists of a spatial convolutional layer with temporal modeling capabilities, and a pairwise attention mechanism for joint spatial and channel. Furthermore, we have introduced a difference skip connection module based on a multi-attention mechanism (ADC). This module can accomplish superior performance during the feature fusion phase, contingent on the significance of intra-feature and inter-feature relationships. We experimentally compare the proposed network with a typical semantic segmentation network on the dataset. The results show that the network proposed in this paper has the best segmentation performance and is well suited for the task of crop segmentation in time-series multispectral remote sensing images. The Mean Intersection over Union reached $86.37 \%$ and overall accuracy over 93.24 %. Through the combination of meticulous feature extraction and multi-attention mechanism, our approach improves the segmentation accuracy while better utilizing the multidimensional information of the image, thus realizing the accurate recognition and segmentation of different crops.