Research on influencing factors and spatial heterogeneity of farmland net primary productivity based on the mobile-UNet model: a case study of Shanghai

Urbanization threatens smallholder agriculture, vital for global food security. This study uses Shanghai as a case, applying 2-m images and an improved Mobile-UNet with data augmentation to enhance the recognition of fragmented farmland. The model achieves 90.38% accuracy (mIoU=0.712, F1=0.831), outperforming logistic regression, random forest and gradient boosting by 6%–12% in the mIoU and 4%–9% in F1-score. Net primary productivity (NPP) data and 10 factors were analyzed through stepwise regression and geographically weighted regression (GWR). Stepwise regression explains 48% of the NPP variation, identifying potential evapotranspiration, evapotranspiration, land surface temperature, solar radiation and the digital elevation model as the 5 most important factors. GWR revealed that the strength and direction of relationships between factors and NPP vary significantly. The results informed hierarchical management maps and strategies for targeted resource allocation and farmland productivity policies, advancing sustainable urban agriculture through improved small-scale farmland regulation.