FL-DDQN: A federated double deep Q-learning framework for enhancing irrigation management across family multi-farms environments

Abstract

In this paper, we propose a federated deep reinforcement learning framework, named FL-DDQN, designed to enhance irrigation management in smart farming environments. The proposed framework addresses key challenges posed by fragmented data collected from small-scale farming devices, as well as the complexities associated with varying climate and soil patterns. Specifically, our framework optimizes the selection of clients in the federated learning process through deep reinforcement learning. The evaluation of the proposed framework is conducted across different multifarm configurations, showcasing its scalability and adaptability. Furthermore, the FL-DDQN framework facilitates collaborative training of a weather and climate forecasting model using a hybrid convolutional neural network (CNN)–long short-term memory architecture. This model is enhanced by incorporating regulation and normalization techniques, which help mitigate the effects of distribution shifts and pattern changes in dynamic farming environments. The results demonstrate high accuracy, achieving a low mean absolute error for soil moisture (0.0118), temperature (1.0200), and relative humidity (4.3958). Additionally, we integrate autoencoders based on CNNs to detect anomalies in irrigation system sensor readings by evaluating the reconstruction error. The proposed framework achieves a significantly lower reconstruction error compared to recent state-of-the-art methods, confirming its robustness in anomaly detection.