Winter wheat yield prediction using linear and nonlinear machine learning algorithms based on climatological and remote sensing data

Abstract

• Integration of RS indices and climatic variables enhances wheat yield prediction accuracy. • RF algorithms incorporating novel indices like GNDVI, SAVI, and ARVI. • Specific RF model combinations with indices predicting wheat yield across seasonal scenarios. • The LASSO model outperforms RF, indicating its potential as a viable substitute for RF in wheat yield prediction. • Significantly contributes to food security by providing more accurate and timely crop yield predictions. In the pursuit of enhancing agricultural forecasting in Pakistan, this research integrates remote sensing indices and climatic variables through advanced machine learning algorithms. By meticulously examining ten model combinations within different wheat season scenarios, the study employs nonlinear models, such as Random Forest (RF) and Support Vector Machines (SVM), and linear models, like Least Absolute Shrinkage and Selection Operator (LASSO) and Ridge. This research aims to predict wheat yield in Pakistan by integrating five remote sensing indices, including the Green Normalized Difference Vegetation Index (GNDVI), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), Atmospherically Resistant Vegetation Index (ARVI) with five climatic variables: Maximum Temperature ( T max ), Minimum Temperature ( T min ), Rainfall (R), Soil Moisture (SM), and Windspeed (WS) alongside the drought index and standardized Precipitation Evapotranspiration Index (SPEI). Ten model combinations were created within two wheat season scenarios: Full Seasonal Mean Scenario 1 (FSM) (SC1) and Peak Seasonal Mean Scenario 2 (PSM) (SC2). Two nonlinear ML algorithms, RF and SVM, and two linear models, LASSO and Ridge, were employed in both scenarios. Results indicated that in SC1, the RF model combination (GNDVI + SPEI + WS + SM) outperformed other models (R 2 = 0.75, RMSE = 2.40, MAE = 1.98). Similarly, in SC2, the RF regression surpassed SVM, with the model combination (GNDVI + SPEI + WS + SM) demonstrating the highest performance, achieving R 2 = 0.78, RMSE = 2.25, and MAE = 1.88, followed by (NDVI + T max + T min + PPT + PET + WS + SM; R 2 = 0.75). The linear LASSO model also performed similarly to RF, achieving R 2 = 0.74–0.69 in both scenarios. The findings advocate for utilizing SC2 for yield prediction in ML models. Overall, this study underscores the significance and potential of ML methodologies in timely crop yield prediction across various crop growth stages, thereby establishing a robust foundation for ensuring regional food security.

Topics & Concepts