LAUSR

Abstract Estimating leaf wetness duration (LWD) is fundamentally important for crop disease monitoring and early warning. This study developed and evaluated a novel framework based on machine learning techniques with climate reanalysis data to estimate leaf wetness. We compared the Random Forest, kernelized Support Vector Machine, Feed-forward Neural Network, and Classification and Regression Tree models with hourly surface and near-surface meteorological inputs from state-of-the-art climate reanalysis datasets, ERA5 and MERRA2. The models were trained and cross-validated using three-year hourly leaf wetness observation data at nine agricultural sites in Alabama and six sites in California. The models were intercompared with each other and with a simple model based on observed relative humidity (with 90% used as threshold, referred as RH). The results indicated that the machine learning models based on ERA5 showed better performance than the MERRA2-based models as well as the observation-based RH model. In particular, the ERA5-based Random Forest model showed the best performance for estimating hourly leaf wetness occurrence (Accuracy = 83%, Precision = 78.2%) and daily LWD (MAE =2.8 h, RMSE = 4 h) than the other approaches. In comparison with the RH method, the ERA5-driven Random Forest model showed at least 2 h smaller MAE and RMSE than the other methods for estimating daily LWD, suggesting a great potential for practical applications. The results also showed that the ERA5-driven Random Forest model accurately predicted the leaf wetness occurrence in space (with 70 to 87% accuracy), suggesting a potential for large-scale crop disease surveillance and climate impact assessment, particularly for regions with sparse in situ observation data.