Soybeans play a vital role in global food security and sustainable agriculture, particularly in North and South America, which account for over 86% of global production. However, these regions are increasingly threatened by climate change-induced extreme weather events, necessitating advanced monitoring and predictive capabilities to safeguard soybean yields.
This PhD research enhances large-scale soybean yield estimation under varying climatic conditions by leveraging innovative remote sensing (RS) and deep learning (DL) techniques. The dissertation makes three key contributions:
First, this study aims to develop a novel Phenology-guided Bayesian Neural Network (PB-CNN) framework for county-level yield estimation and uncertainty quantification in the US Corn Belt. This framework integrates phenological data with Bayesian neural networks to evaluate yield response to environmental stresses within different growing stages. The developed PB-CNN framework demonstrated improved accuracy and provided valuable uncertainty estimates compared to benchmark models. Feature importance analysis revealed that satellite-based predictors and reproductive growth stages contribute most significantly to yield formation, while soil predictors and early growth stages introduce greater uncertainty.
Second, this study introduces a comprehensive approach to analyzing domain shifts in crop yield prediction and evaluating transfer learning strategies through combined crop model simulations and empirical analysis. It demonstrates that agricultural systems face unique challenges in transfer learning as environmental variations, cultivar adaptations, and management practices create multiple, simultaneous domain shifts. Comparative evaluation showed that Model-Agnostic Meta-Learning (MAML) achieved superior performance across various domain shift types, while Fine-tuning Learning (FTL) provided an efficient solution with moderate amounts of target data.
Third, this study incorporates the Madden-Julian Oscillation (MJO) into the deep learning framework to assess the impact of MJO-driven extreme events on soybean production. By quantifying MJO teleconnections across different phases and ENSO conditions, it revealed regional patterns of temperature stress and soil moisture responses. Integrating projected MJO and ENSO information with within-season environmental variables reduced average prediction error, with significant improvements in major soybean-producing states during critical growth periods.
By enabling accurate soybean yield prediction across the Americas, where the crop covers over 100 million hectares and contributes to $200 billion in annual global trade, this framework will allow rapid responses to potential food crises. The proposed framework supports rapid governmental and humanitarian responses to potential food crises while informing commodity pricing, crop insurance, trade decisions, and economic planning, helping to mitigate the adverse effects of climate change on a critical global food resource.