Iintegrating deep learning and multi-source datasets for drought-resilient winter rapeseed yield prediction in the Yangtze River Basin

Global climate change has led to more frequent and severe droughts in the middle and lower reaches of the Yangtze River, intensifying the spatiotemporal variability of crop yields in this region. Winter rapeseed, a major oilseed crop in China, is particularly vulnerable to these drought conditions, which now pose greater risks to local food security. Accurate and timely regional yield predictions are increasingly important for effective agricultural management and disaster response. However, predicting rapeseed yield at the city level is challenging due to complex climate patterns and the strengthened impact of drought. Addressing these challenges requires the integration of multi-source data, including both remote sensing and weather data, to capture the full range of environmental influences on crop growth. Traditional statistical and machine learning methods have often proven inadequate for robust, transferable yield prediction across different regions and years.

This study presents a deep learning–based yield prediction framework that integrates multi-temporal remote sensing indicators and meteorological variables to estimate winter rapeseed yield under both normal and drought conditions. Using data from 2014 to 2023 for the middle and lower reaches of the Yangtze River, an Attention–Long Short-Term Memory (Attention-LSTM) model was developed by jointly incorporating time-series remote sensing indices, meteorological factors, and statistical yield records. Key phenological periods for yield estimation were identified through multi-temporal and multi-variable combinations, and input configurations were systematically optimized. The proposed framework outperformed LSTM, Random Forest, and Support Vector Regression models, achieving an R² of 0.81 and RMSE of 306.73 kg/ha on the validation dataset. Spatiotemporal yield dynamics and regional applicability were further analyzed, and the model’s robustness and adaptability were assessed under drought conditions. Under drought scenarios, the model maintained high accuracy, with an R² of 0.76 and RMSE of 358.32 kg/ha. These results indicate the framework’s potential for drought-resilient yield prediction and its value for agricultural management and drought assessment under future climate change.