Remote sensing-based framework to predict and assess the interannual variability of maize yields in Pakistan using Landsat imagery

Abstract

Predicting crop yields and their spatio-temporal variability under a changing climate is a challenging but essential undertaking for crop management and policymaking purposes. The availability of information on risks associated with effects of climatic variability on agricultural activity outcomes is critical for stakeholders ranging from individual landowners to national economists alike. This research was conducted as a pilot study to (1) develop satellite remote sensing based estimates of maize acreage in a typical Maize growing region in Pakistan, (2) to develop a statistical-empirical model for prediction of maize yields, and finally, (3) to assess the influence of temperature on inter-annual variability in maize yields across a decade. A total of eight machine learning algorithms were tested for identifying maize growing operations in the Faisalabad district of Pakistan using Landsat 8 imagery. Classification models were evaluated via 200 randomly selected ground-verified points across the study region. Results of the maize mapping exercise were used to estimate interannual maize yields using Landsat-derived multi-temporal normalized difference vegetation index (NDVI) and land surface temperature (LST) data as predictors. Predictors for the yield forecasting model were selected via principal component screening and were fed into a least absolute shrinkage and selection (LASSO) regression model. The yield model thus developed was applied to 10 years of past data (2006-2017) and validated against data recorded by government sources. Finally, predictions spanning the ten years were tested for effects of temperature variability to find evidence of influence of ambient temperature on maize yields. Results indicate that support vector machine classifiers work the best in this landscape (accuracies >90%) and reveal that maize cropping area may be underestimated in government sources by as much as 14%. The LASSO regression models also showed very good fits (validation R2 = 0.95) and were fairly accurate in tracking interannual variations in maize yields (R2 = 0.78.) Results also indicate that the maximum temperature has significant negative influence (R2 = 0.76, P < 0.0001) on maize yields in Faisalabad district. Methods presented in this study should be of use to policymakers for better formulating export-import policies and decisions governing food security issues in the larger region.