Multi-scale spatial-temporal remote sensing fusion for phenology identification in rice germplasm resources.

Crop phenology is a critical determinant for yield prediction and germplasm evaluation. However, precise phenological monitoring in large-scale rice breeding trials faces significant challenges due to the inherent phenological asynchrony among hundreds of cultivars and the trade-off between spatial resolution and temporal continuity in unmanned aerial vehicle (UAV) remote sensing. To address these issues, this study proposes a multi-scale temporal deep learning framework that integrates high-frequency medium-resolution (MR) images as temporal anchor with sparse high-resolution (HR) images as spatial enhancement. We introduce a Missing Aware Gated Fusion (MAGF) mechanism to dynamically integrate multi-resolution features on non-aligned timelines, enabling robust modeling under irregular sampling conditions. Validated on a massive dataset covering approximately 500 rice cultivars and over 100,000 images across 2023 and 2024 growing seasons, the proposed method significantly outperformed single-temporal-scale baselines despite multiple growth stages coexisting within the same dates. The integration of multi-spatial-scale fusion with LSTM temporal modeling yielded superior performance considering efficiency, achieving an Overall Accuracy (OA) and F1-score of 0.873, with a Kappa coefficient of 0.84. A hybrid sampling strategy (daily MR image combined with weekly HR image) demonstrates that weekly flight time can be reduced from 28 h to approximately 6 h while maintaining high accuracy. Notably, even when HR acquisition was reduced to a once every 14 days frequency, the fusion performance remained significantly superior to that of daily MR monitoring alone. The model exhibited strong generalization capabilities. When directly applying the model trained on 2024 data to the 2023 dataset, it maintained an OA of 0.774 and an F1-score of 0.738 under a 3-day error tolerance, with recall for the maturity stage consistently exceeding 0.96. This framework offers a flexible, scalable, and cost-effective solution for high-throughput phenotyping in precision breeding.