A mixed convolution and distance covariance matrix network for fine classification of corn straw cover types with fused hyperspectral and multispectral data

Effective management of corn straw and stubble is critical in conservation tillage, as it impacts soil health and productivity. However, accurate classification of different types of straw cover has been hindered by their similar spectral and spatial characteristics and the low spatial resolution of hyperspectral satellite imagery. Moreover, traditional convolution neural network (CNN)-based methods, which rely on first-order statistics for feature extraction, often struggle to extract distinguishable features of highly similar objects effectively, thereby reducing classification accuracy. In this study, a second-order statistical-feature extraction algorithm based on CNN that uses fused multispectral and hyperspectral data was tested for its ability to classify types of straw cover. In the first step, coupled non-negative matrix factorization (CNMF) was used to fuse hyperspectral and multispectral images effectively, thereby enhancing the spatial resolution of the hyperspectral data. In this study, we integrated pointwise convolution (PWC), depthwise convolution (DWC), and a distance covariance matrix (DCM) to form a mixed convolution and DCM (MCDCM) network; we used this to extract and integrate deep spectral–spatial features of the hyperspectral images. Our experimental results show that the MCDCM network significantly improved classification accuracy compared to traditional methods, with accuracy rates for the different straw-cover types exceeding 90% and overall accuracy reaching 98.26%. The fused image also exhibited better preservation of feature edges and contours. The accurate identification of corn-straw-cover types achieved with the proposed MCDCM method is a major step in optimizing conservation-farming practices, improving soil fertility and farm productivity, and supporting sustainable ecological development.