Hybrid CNN-LSTM-attention model for enhanced reference crop evapotranspiration estimation toward optimized irrigation scheduling

This study developed an artificial intelligence-based optimization framework for irrigation management to address the challenges of insufficient integration between reference crop evapotranspiration (ETo) metrics and irrigation decision-making systems. First, a hybrid deep learning architecture combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and an Attention Mechanism was developed. Key climate parameters (average temperature, sunshine duration, and average wind speed) were selected as input features through Grey Relational Analysis (GRA). This approach enabled the development of a high-accuracy ETo estimation model even under limited climate data conditions. Second, an optimized Alternate Wetting and Drying (AWD) irrigation system based on accurate estimation of ETo was developed. An empirical analysis using the 2012–2023 climate dataset from Shaoxing, Zhejiang, China, revealed that compared with the existing irrigation model, the optimized scheduling achieved water savings of 16.5 % and 37.0 % during the 2022–2023 rice growing seasons. The ETo estimation model proposed in this paper can predict both the ETo and the required irrigation water consumption for the next few days using climate forecast data, thereby offering an effective solution to irrigation delays caused by water scarcity in mountainous and arid regions. Overall, this framework offers a scalable and intelligent solution to enhance water-use efficiency and support sustainable agricultural management, particularly in data-scarce and water-sensitive regions. • A CNN-LSTM-Attention model was developed to estimate ETo with high-accuracy using limited available climate parameters. • An optimized irrigation scheduling was then formulated based on accurate estimation of ETo. • The current irrigation scheduling can also predict the required irrigation water consumption using climate forecast data.