GWSC-SegMamba: Gate Wavelet Spatial Convolution Enhanced State Space Model for Multi-Temporal Agricultural Land Segmentation

The study of multi-temporal satellite data for agricultural land segmentation faces significant computational challenges when processing extended temporal sequences, particularly due to CNNs’ limited receptive fields and Transformers’ quadratic complexity, since convolutional neural networks are constrained by local receptive fields, whereas Transformers experience quadratic complexity in their self-attention mechanisms. These constraints particularly impede the precise identification of spectrally analogous crop varieties and minority classes within agricultural landscapes. We present GWSC-SegMamba, an innovative hybrid architecture that integrates State Space Models (SSMs) with Gate Wavelet Convolution (GWC) and Gate Spatial Convolution (GSC) components to tackle multi-temporal agricultural segmentation issues. The GWC component conducts multi-resolution analysis with discrete wavelet transforms to address spatial resolution constraints for medium-resolution satellite data, whereas the GSC component identifies spatial correlations essential for delineating crop boundaries. Our thorough assessment of three benchmark datasets (Munich, Lombardia, and PASTIS) reveals substantial performance enhancements: a 7.65% increase in mIoU relative to the conventional SegMamba, a 4.43% improvement over the Swin-UNETR baseline, and an impressive 53.13% augmentation in IoU for difficult minority classes, including winter triticale. The design processes temporal sequences with linear computational cost, incurring about 8–10% additional computational overhead relative to the basic SegMamba and demonstrating enhanced scalability properties compared to transformer-based approaches. These findings provide enhanced crop monitoring systems vital for precision agriculture, especially in differentiating spectrally analogous crop varieties needed for yield estimation and sustainable land management decisions.