Network-regularized sparse logistic regression for crop identification in remote sensing images

Employing remote sensing imagery for crop classification is an efficient means of rapidly generating crop distribution maps that can assist with crop management and resource utilization. However, the high dimensionality and complexity of remote sensing satellite data often lead to the development of intricate models that are challenging to interpret. The study explores a network-regularized sparse logistic regression (Net-SLR) model for identifying crop distribution in remote sensing images, aiming for efficient and accurate crop mapping. By introducing a Laplacian prior network, we investigated the effectiveness of spectral features in crop classification for spectral feature extraction. Both multispectral satellite data and hyperspectral imagery data were employed to identify the planting distributions of winter wheat, winter canola, and spring corn, and thus, the robustness of the model under different remote sensing data scenarios was investigated. Furthermore, the Net-SLR model was compared with conventional classification methods, and the classification performances were comprehensively assessed. The experimental results demonstrated that the Net-SLR model had outstanding performance in crop type classification. It achieved overall accuracy (OA) of 0.982 and 0.972 for identifying winter wheat and winter canola in multispectral satellite data, respectively. Correspondingly, the area under the curve (AUC) values were 0.984 and 0.95. In hyperspectral imagery data, the OA was 0.9, and the AUC was 0.975 for identifying the spring corn. The proposed Net-SLR model can efficiently achieve crop identification tasks and feature selection simultaneously on remote sensing images to accurately identify crop distribution.