Machine Learning-Based Regional Assessment of Second-Crop Corn Yield Using Climatic and Water-Balance Variables

Second-crop corn is an important component of Brazilian grain production and is highly exposed to climatic variability because crop development occurs during the transition from the rainy to the drier season. This study assessed the performance of machine learning algorithms for estimating second-crop corn yield within the 2012–2023 historical data domain in Mato Grosso do Sul, Brazil. Municipal yield data from 22 municipalities and daily climatic data covering the period from 2012 to 2023 were analyzed. Yield data were obtained from Aprosoja/MS, whereas climatic variables, including solar radiation, air temperature, evapotranspiration, relative humidity, wind speed, and precipitation, were obtained from NASA/POWER. Pearson correlation analysis was used as an initial screening procedure to identify climatic variables associated with yield variability. Multiple linear regression (MLR), a multilayer perceptron (MLP), a support vector machine (SVM), Random Forest (RF), XGBoost, and a gradient boosting regressor (GBR) were evaluated. Under the observation-level random partition adopted in this study, RF and XGBoost showed the highest performance in the randomly held-out testing subset, both reaching R2 values of approximately 0.80. However, XGBoost exhibited near-perfect calibration performance followed by a substantial decrease during testing (MAPE = 14.05%; R2 = 0.80), indicating overfitting. Random Forest achieved similar testing performance with a smaller calibration-to-testing gap. Therefore, under the specific random partition used in this study, RF was considered the preferred model. In contrast, GBR failed to provide reliable predictive performance under the present dataset and configuration, with a testing MAPE of 547.78%, and was therefore considered too unreliable for interpretation in this study. Importantly, because the partition did not preserve temporal or spatial independence, these results represent generalization within the 2012–2023 historical data domain and should not be interpreted as evidence of future-year or spatially independent predictive performance. Because the dataset was partitioned randomly at the observation level without preserving temporal independence, the reported performance represents generalization within the 2012–2023 historical data domain and should not be interpreted as future-year forecasting performance.