Federated Learning and Transformer Model‐Based Hybrid AI Model for Predicting Crop Yields

Abstract

Timely and efficient management of agricultural practices and subsequently predicting crop yields are fundamental elements in the quest to mitigate global food insecurity. Conventional sampling and secondary historical records, which are usually relied upon, tend to be inaccurate and inflexible to current trends. This paper proposes a hybrid artificial intelligence model that aims to achieve higher prediction accuracy with the aid of USDA Crop Scape data. We rely on a Federated Learning (FL) framework, which allows for different nodes in the network to work together to construct a common model without the need to transfer any training data to the nodes. After federated training, Transformer-based NLP technologies were used to process and fuse unstructured data (weather reports) and structured data (soil properties) typologically into one format, taking into account long-range contextual relations and enhancing the information content of the model. This multicenter multimedia architecture presented a significant advantage over the traditional approaches by integrating spatial and text data in an optimal way. Afterwards, a Multi-Objective Genetic Algorithm (MOGA) was employed to optimize the model parameters, resulting in an increase in forecast accuracy from 93.67% to 96.15%. On the one hand, in order to explain and visualize important parameters of the model better, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied as a means of guaranteeing clarity and transparency in the model predictive analyses. This hybrid AI approach is indicative of a cutting-edge development in agricultural precision practices, considering its adaptability, accuracy, and interpretability in crop monitoring and enhancing food safety.