Applicability of machine learning techniques in predicting wheat yield based on remote sensing and climate data in Pakistan, South Asia

Abstract

Machine learning (ML) algorithms perform better than classical statistical approaches to explore hidden nonlinear relationships. In this context, the goal of this research is to predict wheat yield utilizing remote sensing and climatic data in southern part of Pakistan. Four remote sensing indices, viz.., Green Normalized Difference Vegetation Index (GNDVI), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI) are integrated with five climatic variables, i.e., Maximum Temperature (Tmax), Minimum Temperature (Tmin), Rainfall (R), Relative humidity (RH) and windspeed (WS) and one drought index, i.e., Standardized Precipitation Evapotranspiration Index (SPEI). Eight model combinations are built within two scenarios of wheat season, i.e., Whole Seasonal mean (WSM) (SC1), and Peak of Seasonal Mean (POSM) (SC2). Two nonlinear ML algorithms, i.e., Random Forest (RF), and Support Vector Machines (SVM), and one linear model, i.e., LASSO is being employed for wheat yield prediction to find the best combination and ML algorithm in two scenarios. Results revealed that in SC1, RF regression for the model combination (GNDVI +Tmax+ Tmin + R + RH + WS) outperformed other models (R2 = 0.71, RMSE = 2.365). Similarly, in SC2 RF regression outperformed SVM with model combination (GNDVI + Tmax+ Tmin + R + RH + WS) performed highest with R2 = 0.78, and lowest RMSE = 2.07, followed by (GNDVI + SPEI + RH + WS; R2 = 0.75). Interestingly, linear LASSSO also performed equally with RF with R2 = 0.77–0.73 in both scenarios. However, the output of this research recommends using SC2 for yield prediction in ML models. Overall, this research reveals the significance and potential of ML techniques for timely prediction of crop yield in different stages of crop growth that provide a solid foundation for food security in the region.