Machine learning methods for estimating the impact of climate change on Indian crop yields

Abstract:

Crop productivity is sensitive to both short-term weather variability and long-term climate change. Over the past few decades, predicting crop yields as a function of climate variability and change has been a topic of extensive research, both at a regional and global scale. However, most of these studies employ linear regression techniques that predict yields as a function of seasonally-averaged climate. This is in spite of the substantive literature that shows that crop physiological responses to weather are non-linear, and that intra-seasonal climate variability can have substantial effects on crop yields. There is a lack of consensus among researchers analyzing Indian agriculture’s sensitivity to climatic variability, and studies have reported both a significant and non-significant impact of intra-seasonal climatic variations on crop yields.

Our study addresses this uncertainty by employing two different statistical techniques to model crop yields as a function of climate variability. We build a linear regression model with crop yields as the dependent variable, and various climate parameters like temperature, precipitation, soil moisture availability etc. as the independent variables. However, linear regression models are not flexible enough to detect the aforementioned non-linear relationship between intra-seasonal climate variability and agricultural productivity, so we also build crop models using Boosted Regression Trees (BRTs): a non-parametric machine learning method to analyze the relationship between independent and dependent variables by combining several simpler partial models instead of fitting one “true” model. We then discuss the advantages and disadvantages of linear regression models versus BRTs in terms of their bias, accuracy, interpretability and predictive power. Our study thus improves our understanding of Indian agriculture’s vulnerability to a changing climate, and our results can be used to design adaptation and mitigation strategies by policymakers.