<p>Climate change poses a critical threat to agricultural sustainability, with direct implications for the global food supply. Rice, a staple crop throughout Asia, is particularly vulnerable to variations in temperature and rainfall, making it essential to understand how it responds to changing climatic conditions. This study integrates historical climate records, rice yield data, and projections from Global Climate Models (GCMs; CMIP3) to assess the potential effects of climate change on rice production in Punjab, Pakistan. We employed multiple machine learning approaches, including Multiple Linear Regression (MLR), Boosted Tree Regression (BTR), Probabilistic Neural Network (PNN), Generalized Feed-Forward (GFF) Neural Network, Linear Regression (LR), and a Multilayer Perceptron (MLP) Artificial Neural Network. The models were trained and validated using observed climate and yield data from 1990 to 2020. Future yields were projected under three IPCC emission scenarios (SR-A2, SR-A1B, SR-B1) through the year 2050. Model evaluation showed that the Multilayer Perceptron (MLP) achieved the highest predictive performance (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> = 0.791, R = 0.868, MAE = 0.215, MSE = 0.0869, NMSE = 0.3681), followed by Boosted Tree Regression (BTR; <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> = 0.779, R = 0.845, MAE = 0.334, MSE = 0.1308). The Probabilistic Neural Network (PNN) and Generalized Feed-Forward (GFF) model also performed respectably (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> = 0.745, R = 0.811, MAE = 0.176, MSE = 0.380 and <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> = 0.643, R = 0.825, MAE = 0.398, MSE = 0.178, respectively). In contrast, Multiple Linear Regression (MLR) and Linear Regression (LR) performed poorly, with low <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> values (0.535), underscoring their inability to capture the non-linear relationships between climate variables and yield. Our analysis identifies maximum temperature as the primary climatic driver of yield loss. Based on the projections, we estimate an average yield decline of 0.12% by 2050. This study demonstrates that non-linear machine learning models, particularly the MLP, are essential for reliable agricultural forecasting under climate change. The results highlight the growing vulnerability of rice production to rising temperatures and provide a robust evidence base for designing adaptation strategies, such as developing heat-tolerant rice varieties, to enhance food security in vulnerable regions.</p>

Predicting rice yield and impact of climate change on rice production using machine learning models

期刊论文
出版模式：

开放获取
发表日期：

2025年11月20日

查看全文

Khawaja T. Tasneem,
Muhammad Umair Shahzad,
Javed Rashid,
Kamal M. Othman,
Tania Zafar,
Muhammad Faheem

摘要

Climate change poses a critical threat to agricultural sustainability, with direct implications for the global food supply. Rice, a staple crop throughout Asia, is particularly vulnerable to variations in temperature and rainfall, making it essential to understand how it responds to changing climatic conditions. This study integrates historical climate records, rice yield data, and projections from Global Climate Models (GCMs; CMIP3) to assess the potential effects of climate change on rice production in Punjab, Pakistan. We employed multiple machine learning approaches, including Multiple Linear Regression (MLR), Boosted Tree Regression (BTR), Probabilistic Neural Network (PNN), Generalized Feed-Forward (GFF) Neural Network, Linear Regression (LR), and a Multilayer Perceptron (MLP) Artificial Neural Network. The models were trained and validated using observed climate and yield data from 1990 to 2020. Future yields were projected under three IPCC emission scenarios (SR-A2, SR-A1B, SR-B1) through the year 2050. Model evaluation showed that the Multilayer Perceptron (MLP) achieved the highest predictive performance ( \(R^2\) = 0.791, R = 0.868, MAE = 0.215, MSE = 0.0869, NMSE = 0.3681), followed by Boosted Tree Regression (BTR; \(R^2\) = 0.779, R = 0.845, MAE = 0.334, MSE = 0.1308). The Probabilistic Neural Network (PNN) and Generalized Feed-Forward (GFF) model also performed respectably ( \(R^2\) = 0.745, R = 0.811, MAE = 0.176, MSE = 0.380 and \(R^2\) = 0.643, R = 0.825, MAE = 0.398, MSE = 0.178, respectively). In contrast, Multiple Linear Regression (MLR) and Linear Regression (LR) performed poorly, with low \(R^2\) values (0.535), underscoring their inability to capture the non-linear relationships between climate variables and yield. Our analysis identifies maximum temperature as the primary climatic driver of yield loss. Based on the projections, we estimate an average yield decline of 0.12% by 2050. This study demonstrates that non-linear machine learning models, particularly the MLP, are essential for reliable agricultural forecasting under climate change. The results highlight the growing vulnerability of rice production to rising temperatures and provide a robust evidence base for designing adaptation strategies, such as developing heat-tolerant rice varieties, to enhance food security in vulnerable regions.