Context:Using sophisticated data-driven technologies, precision agriculture has emerged as a crucial strategy for enhancing agricultural production, sustainability, and resource management. In centralized learning approaches, farmers need to share their farm data with the central server . It raises issues of data ownership, security, and privacy in distributed agricultural settings.
Objective: In this paper, our primary objective is to develop a privacy-preserving intelligent framework for soil fertility prediction in precision agriculture. This study's goal is to compare the predictive performance of several intelligent models and to assess Federated Learning's ability to enable safe, decentralized, and collaborative agricultural analytics.
Materials and Methods: This study utilized 10 ML models and 5 DL models and evaluated them using performance metrics such as accuracy, precision, recall, F1-score, and Cohen’s Kappa. We also developed a privacy-preserving intelligent framework, adopting a federated learning mechanism with five clients and 20 communication rounds, using FedAvg (federated averaging aggregation). The dataset was divided into different farms to \simulate real-world distributed agriculture systems.
Results: From our experimental observations, the Extra Tree (ET) classifier obtained the highest accuracy of 93.94%, F1-Score of 93.99, and Kappa value of 0.90. Similarly, the Shallow MLP model from DL obtained an accuracy of 93.18% and a Kappa value of 0.8977. It has been reported that the FL framework achieved 84.47% accuracy and a Kappa score of 0.76. Apart from these, our framework preserves farm-level data privacy and removes the need for centralized data sharing. The experimental work demonstrates that distributed learning is a secure and collaborative option for smart agriculture. Several Multi-Layer Perceptron (MLP) architectures were constructed and compared to conduct deep learning analysis. The Shallow MLP model with a 128-64 neuron architecture achieved better performance.