2026/03/17 by Zhixian Lin, Kaiyu Guan, Sheng Wang +5 · 1 voice
Environmental Science · Agricultural and Biological Sciences · Engineering · #Remote Sensing in Agriculture #Smart Agriculture and AI #Remote-Sensing Image Classification
paper · doi:10.1016/j.jag.2026.105231
openalex publication_date 2026/03/17 · openalex created_date 2026/03/18 · openalex updated_date 2026/07/23
• Multi-task temporal fusion (MTTF) framework for triple-season rice yield estimation. • MTTF captures shared and season-specific patterns across Vietnamese rice systems. • Phenology alignment enhances temporal consistency across heterogeneous calendars. • Multi-modal fusion with vegetation, climate, and soil data boosts model accuracy. Accurate seasonal rice yield estimation across Southeast Asia’s intensive cropping systems remains challenging due to complex phenological patterns and heterogeneous environmental conditions. This study develops a phenology-aligned multi-task temporal fusion (MTTF) framework for satellite-based seasonal rice yield estimation in Vietnam’s triple-cropping systems from 2001 to 2020. The multi-task learning treats each cropping season (winter–spring, summer–autumn, monsoon) as related but distinct tasks, enabling knowledge sharing while preserving season-specific characteristics. The framework integrates multi-source time-series data, including climate variables (e.g., temperature, precipitation), satellite-based vegetation indices (e.g., NDVI, EVI, NIRv, GCVI, LSWI), productivity indicators (e.g., SIF, GPP), and static soil properties (e.g., clay content, organic carbon, bulk density) through parallel Transformer encoders and late fusion strategies. To address temporal misalignment across heterogeneous cropping calendars, we developed an automated phenology-based crop season detection method that synchronizes time-series inputs to key growth stages rather than calendar dates. MTTF achieved high performance (R 2 = 0.75, RMSE = 0.63 Mg·ha −1 , and rRMSE = 12.0%), outperforming baseline models including Transformer, AtBiLSTM, ANN, XGBoost, and Random Forest. The multi-task learning approach outperformed both global models (single predictor for all seasons) and local models (separate predictors for each season), demonstrating particular benefits for data-scarce seasons like monsoon rice. Phenology alignment enhanced temporal consistency across all models. Multi-modal data fusion significantly improved performance, with satellite-based vegetation measurements contributing more significantly than climate variables according to SHAP analysis. The proposed framework provides a robust approach for operational rice yield monitoring across intensive cropping systems, with implications for assessing food security and agricultural policy in monsoon regions.