District-Level Rice Yield Prediction in Karnataka: Evaluating Machine Learning Models Under Different Validation Strategies

Abstract Rice is an important food crop in Karnataka, and its yield is affected by different factors such as rainfall, temperature and irrigation. Machine learning models can be used to study these factors and predict rice yield, but the accuracy of these models may change depending on how the data is divided for training and testing. This study uses district-level rice yield and climate data from Karnataka for the period 2000–2023, with 636 observations from 27 harmonised districts. Rainfall data were collected from CHIRPS, temperature data from NASA POWER and rice area, production and yield data from the ICAR-IIRR Rice APY Dashboard. Three machine learning models, Ridge Regression, Random Forest and XGBoost, were evaluated using random train-test splitting and chronological walk-forward validation, and compared with two simple baselines. The study also looks at whether irrigation information and a PMFBY policy-era indicator improve prediction results. The models achieved reasonably good scores under random splitting (average $ R ^ 2 $ of 0.693–0.726), but their $ R ^ 2 $ was lower under chronological validation (0.643–0.664), which gave a more realistic picture of how the models perform when predicting later years. A simple previous-year yield baseline also had a lower mean absolute error than every model. The additional irrigation experiment showed mixed improvements, while interestingly, the PMFBY policy-era indicator improved prediction performance across all three models in the separate repeated random-split experiment, with $ R ^ 2 $ improvements of approximately 0.019 to 0.036. A variable importance analysis of XGBoost showed that district identity was the most important input in every experiment that used the full dataset, followed by rainfall and then temperature, while the irrigation group was the most important input in the irrigation experiment. These results need to be interpreted carefully because of data limitationsf and do not establish a direct effect of the policy. The study more so highlights the importance of choosing suitable validation methods and conditions before concluding that a machine learning model is accurate.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11712090/v1
Primary Topic
Smart Agriculture and AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

District-Level Rice Yield Prediction in Karnataka: Evaluating Machine Learning Models Under Different Validation Strategies

Vikram Ramavarapu, Dhaksha Gowda
5 links
Research Square
Smart Agriculture and AI
preprint

District-Level Rice Yield Prediction in Karnataka: Evaluating Machine Learning Models Under Different Validation Strategies

Vikram Ramavarapu, Dhaksha Gowda
preprint en

Abstract

Abstract Rice is an important food crop in Karnataka, and its yield is affected by different factors such as rainfall, temperature and irrigation. Machine learning models can be used to study these factors and predict rice yield, but the accuracy of these models may change depending on how the data is divided for training and testing. This study uses district-level rice yield and climate data from Karnataka for the period 2000–2023, with 636 observations from 27 harmonised districts. Rainfall data were collected from CHIRPS, temperature data from NASA POWER and rice area, production and yield data from the ICAR-IIRR Rice APY Dashboard. Three machine learning models, Ridge Regression, Random Forest and XGBoost, were evaluated using random train-test splitting and chronological walk-forward validation, and compared with two simple baselines. The study also looks at whether irrigation information and a PMFBY policy-era indicator improve prediction results. The models achieved reasonably good scores under random splitting (average $ R ^ 2 $ of 0.693–0.726), but their $ R ^ 2 $ was lower under chronological validation (0.643–0.664), which gave a more realistic picture of how the models perform when predicting later years. A simple previous-year yield baseline also had a lower mean absolute error than every model. The additional irrigation experiment showed mixed improvements, while interestingly, the PMFBY policy-era indicator improved prediction performance across all three models in the separate repeated random-split experiment, with $ R ^ 2 $ improvements of approximately 0.019 to 0.036. A variable importance analysis of XGBoost showed that district identity was the most important input in every experiment that used the full dataset, followed by rainfall and then temperature, while the irrigation group was the most important input in the irrigation experiment. These results need to be interpreted carefully because of data limitationsf and do not establish a direct effect of the policy. The study more so highlights the importance of choosing suitable validation methods and conditions before concluding that a machine learning model is accurate.

Research Square
University of Illinois Urbana-Champaign (US)
Zero hunger, Industry, innovation and infrastructure
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon