Rice Ripening Stage Recognition Using Multi-Source UAV Remote Sensing Data and a Modified MobileNetV3-Small Network

To enhance fine-grained recognition of adjacent rice ripening phases, this study employs a rice ripening-stage identification method based on multi-source UAV remote sensing data fusion and a modified MobileNetV3-Small network. RGB and multispectral images were acquired using a DJI Mavic 3 Multispectral UAV at Houbai Improved Seed Farm, Jurong City, Jiangsu Province, China, and were combined with vegetation indices to construct a fused 15-channel input. The original three-channel input layer of MobileNetV3-Small was modified to accommodate RGB, multispectral, and vegetation-index features. Based on the agronomic characteristics of late reproductive rice growth, the ripening process was divided into five phenological stages: filling, milky ripening, early waxy ripening, late waxy ripening, and full ripening. In the field-scale patch assessment using the HB31 dataset, the model correctly identified 4147 of 4185 valid patches, achieving an overall accuracy of 99.09%, with macro-averaged precision, recall, and F1-scores of 99.06%, 99.12%, and 99.08%, respectively. To reduce the influence of within-field spatial correlation, an independent field-holdout test was further conducted using the HB32 field, which was excluded from training, hyperparameter tuning, and model selection. Among 2046 valid HB32 patches, 1897 were correctly classified, yielding an overall accuracy of 92.72% and a macro-averaged F11-score of 92.60%. These results indicate that fusing RGB, multispectral, and vegetation-index features can characterize canopy color, spectral responses, and vegetation-index variations during rice maturation, providing reference information for rice ripening-stage identification, field-scale ripening mapping, and harvest scheduling.