Enhancing Crop Yield Estimation Through Iterative Querying and Bayesian-Optimized Gated Networks

Abstract

The accurate prediction of crop yield is essential not only for sustainable agriculture but also for ensuring global food security. In recent times, deep learning (DL) techniques have made significant strides in improving prediction accuracy by leveraging complex and advanced architectures. However, despite these advancements, the existing methods often struggle in modeling temporal dependencies efficiently, especially when dealing with limited data (a common challenge in crop yield prediction). To address this, an innovative iterative querying (IQ) strategy based on the principles of active learning (AL) to enhance model performance has been proposed. The aim of the IQ strategy is to maximize performance by introducing the model to a batch of uncertain instances in each iteration. The overall prediction framework consists of two key components: first, a Bayesian-optimized gated recurrent unit (GRU) method to capture the complex temporal relationships between crop variables and target yield; and second, the novel IQ strategy, which utilizes an uncertainty-driven query mechanism to refine predictions by focusing on the most challenging and uncertain data points. A comprehensive multisource data, comprising remotely sensed variables, climatic, soil, and corresponding crop yield values from the US Corn Belt region, are used to train and evaluate the proposed IQ-GRU method. Experimental results demonstrate the effectiveness of the proposed IQ-GRU framework in improving yield estimation for both in-season and end-of-season predictions over conventional methods.