This paper presents an effective approach for accurate soil humidity prediction using federated learning, which ensures data privacy by sharing only model gradients rather than raw data. In the agriculture sector, accurate prediction of soil humidity is crucial for effective irrigation planning and crop yield optimization. However, collecting and sharing soil data from various farms poses significant data privacy concerns. For an instance, the farmers are usually unwilling to share personal details like crop types, land conditions, or production data with centralized data storage systems. The proposed method overcomes privacy concerns associated with traditional centralized data storage systems, particularly in agricultural applications. In this, we employ Long Short-Term Memory (LSTM) networks to analyze soil humidity data, capitalizing on their ability to capture complex temporal dependencies. The proposed federated learning model enables multiple clients to collaborate on training without disclosing sensitive data. Experimental results show that the model achieves \(R^2\) values of 0.90 and 0.95 for two clients using FedAvg method and \(R^2\) values of 0.91 and 0.97 for same two clients using Q FedAvg method, demonstrating both its accuracy and potential for real-world applications in soil humidity prediction. This work highlights how federated learning can be used to build strong and reliable models for agricultural monitoring while keeping farmers data private.