GMamba: State space model with convolution for Grape leaf disease segmentation

Plant leaf diseases severely impair crop quality and productivity. Accurate segmentation of diseases facilitates the understanding of disease distribution and is a critical step in achieving precise diagnosis and identification of diseases. However, grape leaf diseases suffer from complex backgrounds, small diseases, and high similarity between diseases, leading to limited segmentation precision. To this end, we propose a novel Mamba segmentation model for the grape leaf disease, termed GMamba, which aims to efficiently extract both finegrained and coarse-grained disease features. Specifically, we leverage the UNet hierarchical architecture to deliver multiscale hierarchical disease information from complex backgrounds. Subsequently, we design a Co-SSM module characterized by a formulation of SSM and convolution. The former efficiently models longrange dependencies, which enables the mining of spatial global features of small diseases via multi-directional scanning. The latter focuses on capturing the local detailed features of leaf diseases. Besides, we incorporated SAB and CAB modules to diminish the loss of detail information during decoder fusion and to enhance the richness of feature information scales. Different from Transformer, GMamba globally models the leaf disease without self-attention operation. Extensive experiments have demonstrated that GMamba has competitive segmentation performance over current CNN, Transformer and CNN-Transformer combination architectures on Field-PV, Syn-PV and Plant Village datasets. GMamba yields a 3.28% IoU and 1.65% Precision gain in Black rot over Topformer. To the best of our knowledge, this is the first segmented Mamba tailored for grape leaf diseases. This study can provide an efficient and accurate method for the task of disease segmentation in grapevine leaves, which forms the basis for precise disease analysis.