Multisource data and improved machine learning algorithm were used to estimate maize yield at county level in Xinjiang

Accurate and timely prediction of maize yield in Xinjiang, which accounts for 60% of the annual local grain production, is important for maize management, food security and sustainable agricultural development. However, combining different datasets to predict maize yield at the county scale in Xinjiang simply and accurately is still relatively rare. In addition, in the relevant studies in recent years, the optimization of models for estimating yields has been underestimated, limiting the ability of the models to estimate yields. To explore these challenges, we improved the optimization algorithm by combining satellite, meteorological and soil property data and applying it to machine learning yield estimation models (LASSO, RF, LightGBM) for maize yield estimation at the county scale. The results showed that the multi-source dataset significantly improved the prediction accuracy of maize yield in Xinjiang compared with the single-source dataset. Meanwhile, the improved optimization algorithm provides the greatest performance improvement in the prediction model, with an improvement of about 6.3%, and the improved prediction model explains about 76% of the variation in yield. In terms of seasonal performance, the estimation model rises sharply for a few months and then levels off, approaching a maximum in August, and the optimized estimation model performs well with an R² of 0.78, which allows the model to achieve a better prediction of maize yields about one month in advance.