TandemNet: A Multi-Scale Multiple-Instance Learning Framework for Early-Season Rice Yield Prediction

Abstract

Early-season crop yield prediction is critical for food security assessment and timely agricultural decision-making, yet large-scale remote-sensing applications are often constrained by a mismatch between pixel-level observations and county-level yield labels. This mismatch limits the use of fine-grained spatial heterogeneity, especially under partial-season observations. Single-scale approaches are also limited in capturing complementary pixel-level and county-level information, restricting representation of yield formation processes. To address this, we propose TandemNet, a multi-scale multiple-instance learning framework for early-season prediction of japonica rice yield in Northeast China. TandemNet treats each county–year as a bag of rice pixels and adopts a dual-branch architecture to jointly learn pixel-level growth trajectories and county-level statistical responses. A phenology-conditioned cross-attention module fuses the two scales under varying growing-season windows. Using Sentinel-1, Sentinel-2, and MODIS data from 2018 to 2023, leave-one-year-out validation shows that TandemNet outperforms Random Forest, XGBoost, LSTM, and Transformer baselines across most phenological stages. It achieves reliable prediction at the tillering stage, approximately 2–3 months before harvest, with an R2 of 0.69 and an RMSE of 655.98 kg/ha. Ablation and attention analyses further indicate a stage-dependent shift from local heterogeneity to county-level consistency. These results demonstrate that modeling cross-scale interactions improves the timeliness, accuracy, and interpretability of rice yield prediction.