Integrating attention mechanism into CNN-LSTM for spring maize crop coefficient prediction

The crop coefficient (K c ) serves as a critical parameter linking reference crop evapotranspiration to actual crop water requirements. Its accurate prediction holds significant importance for efficient agricultural water resource management and precision irrigation practices. This study selected six representative sites in Northeast China -Changchun, Harbin, Dalian, Shenyang, Tongliao, and Hohhot—and utilized long-term meteorological time series data from 1950 to 2020. Combining feature engineering with long-sequence modeling,we integrated an attention mechanism to fuse and improve convolutional neural networks (CNN) and long short-term memory networks (LSTM), four models for spring maize K c prediction were developed: LSTM, LSTM-Attention, CNN-LSTM, and CNN-LSTM-Attention . Additionally, Extreme Gradient Boosting (XGBoost) was introduced as a traditional machine learning benchmark model for comparison. These models were validated against FAO-56 recommended values. Research findings indicate: (1)Spatiotemporally, the long-term evolution of meteorological variables profoundly influenced the K c variation patterns. Spatially, K c values exhibited significant regional heterogeneity, with high values primarily concentrated in inland arid regions (e.g., Tongliao and Hohhot), while low values occurred in coastal and semi-humid zones modulated by the ocean (e.g., Shenyang and Dalian). The K c values across growth stages displayed a typical unimodal pattern, increasing steadily from the early growth stage, peaking during the mid-growth stage, and significantly declining in the late growth stage (2) The CNN-LSTM-Attention model exhibited optimal prediction performance, showing high consistency with FAO-56 calculated values throughout the entire growing season, and its prediction accuracy and robustness significantly outperformed traditional machine learning and standard deep learning models. Among these, the Harbin site achieved the highest prediction accuracy (R² = 0.963), while others sites such as Changchun and Hohhot also demonstrated good stability and precision. (3) The prediction accuracy varied across different growth stages, in the following descending order: rapid growth stage > late growth stage > mid-growth stage > early growth stage. The hybrid CNN-LSTM model with the integrated attention mechanism developed in this study effectively estimated daily crop coefficients for spring maize, providing a robust methodological framework for optimizing regional-scale maize irrigation scheduling and implementing precision water management.