The Evolution of Weather-Based Deep Learning in Smart Irrigation: A Systematic Review of Sustainable Approaches and Perspectives

This paper presents a systematic literature review of 191 peer-reviewed studies that link short-term weather information with learning based forecasting and control in irrigation or related applications, focusing on 191 peer-reviewed studies published between January 2020 and early 2025, with four foundational studies published prior to 2020 included via backward citation tracking. The review follows a PRISMA-inspired protocol, with database searches in Scopus, IEEE Xplore, and Web of Science, clear inclusion and exclusion criteria, and structured data extraction on the application domain, sensing and IoT architecture, forecasting models, reinforcement learning algorithms, and reported performance metrics. The results show that deep learning models, especially CNN, LSTM, and their hybrids, are frequently used for short-term environmental prediction and typically outperform classical machine learning baselines. Almost 50 studies employ reinforcement learning or deep reinforcement learning, but only five (≈2.6% of the full corpus) apply these methods directly to irrigation control, while most DRL applications appear in energy and smart-grid management. Around a quarter of the corpus explicitly implements IoT architectures, yet very few systems integrate IoT with reinforcement learning in a closed loop at the edge or fog. Sustainability-related outcomes, such as water use, energy savings, costs, and emissions, are mentioned, but they are not consistently quantified using comparable metrics. The review provides a structured mapping of methods and architectures, clarifies how existing work is fragmented across domains, and highlights open opportunities for developing weather-aware, IoT-enabled, and sustainability-oriented reinforcement-learning frameworks for smart irrigation.