【Objective】Accurate assessment of phosphorus nutrition is essential for optimizing rice growth and improving fertilizer management. This study proposes a lightweight image classification model, MobileNetV3_U-Net, that integrates a MobileNetV3_large encoder with an improved U-Net decoder for efficient diagnosis of rice leaf phosphorus status.【Method】The encoder leverages pretrained weights to accelerate convergence and extracts multi-scale semantic representations, while the decoder adopts an asymmetric multi-scale upsampling strategy with skip connections to progressively fuse shallow and deep features. Unlike the conventional symmetric U-Net, the redesigned decoder adjusts channel width and spatial resolution in a compact manner, enhancing fine-grained feature characterization and sensitivity to phosphorus deficiency symptoms. The classification is completed by adaptive average pooling and a fully connected layer.【Result】On test sets from the rice tillering and jointing stages, MobileNetV3_U-Net achieved accuracies of 93.33% and 87.40%, respectively, surpassing representative lightweight models including MobileNetV3_large, GhostNet, ShuffleNetV2, and EfficientNet_b1. Furthermore, experiments on the Plant Village dataset using grape leaf images confirmed its cross-domain generalization capability.【Conclusion】MobileNetV3_U-Net demonstrates a favorable balance between lightweight design and diagnostic accuracy. By enhancing multi-scale feature fusion through a tailored decoder structure, it provides an efficient, robust, and transferable framework for intelligent diagnosis of rice phosphorus nutrition.