Machine Learning Based Rice Yield Prediction and Rice Growth Stages Identification Using Sentinel-1 SAR and Auxiliary Data in Mwea Irrigation Scheme, Kenya

Abstract

Rice is considered the third most important staple food in Kenya after maize and wheat making its productivity of significant concern. Additionally, conventional preharvest crop yield estimation techniques are dependent on data collection from ground-based field visits which are often subjective, costly and prone to huge errors resulting to poor crop estimates. Satellite remote sensing has been widely accepted as a solution to overcoming these challenges. Therefore, the primary objective of this research was to identify rice growth stages and develop a rice yield-prediction model using Sentinel-1 SAR data and climatic auxiliary data with machine learning (ML) in Mwea Irrigation Scheme (MIS), Kenya. Hence, this study focused on the characterization of rice growth stages in MIS using spectral data derived from Sentinel satellites on Google Earth Engine (GEE) cloud. Thereafter, the SAR backscatter data, coupled with rainfall, and temperature data covering the period 2021-2024 were used to predict rice yield applying four ML models including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM) and Linear Regression (LR) using python programming language on the web-based Jupyter Notebook platform. Finally, evaluation of the performance of RF, XGBoost, SVM models in predicting rice yield for the 2024 season was done using LR as a benchmark. The proposed method produced a rice extent map having an overall accuracy of 83% with a kappa coefficient of 0.53 and effectively identified the stages of rice growth during the 2024 main rice growing season in the study area. From the identified rice clusters, it can be shown that transplanting takes place in July and August while the maturity phase is attained in October and November, giving an indication that the main cropping season was approximately 120 days/ 4 months. The findings from this study illustrate that the RF model outperforms the other models with lower error values, achieving a RMSE of 246.91 kg/acre, MAE of 202.90 kg/acre and MAPE of 7.67% for rice yield predictions utilizing both SAR backscatter and auxiliary climatic datasets. On the other hand, LR model shows suboptimal performance of the four models as indicated by its higher error values attaining a RMSE of 384.09 kg/acre, MAE of 290.86 kg/acre, and MAPE of 11.06%. Furthermore, in predicting yields for the upcoming 2024 season using datasets from previous seasons, the RF model emerged with the highest accuracy obtaining the lowest RMSE of 288.52 kg/acre, MAE of 195.13 kg/acre and MAPE of 8.43%, followed by XGBoost with a RMSE of 290.44 kg/acre, MAE of 201.45 kg/acre and MAPE of 8.56% and finally the SVM model with a RMSE of 493.66 kg/acre, MAE of 395.21 kg/acre and MAPE of 16.73%. The findings show that integration of Sentinel SAR backscatter and optical time series data applying a phenology-driven framework on the GEE platform provides an effective approach for rice mapping and growth stage identification. The study demonstrates that ML models calibrated using previous seasons’ datasets can reasonably predict rice yield for subsequent growing seasons. RF provided the most accurate predictions one month prior to harvesting, followed by XGBoost, and SVM models. The LR model failed to generalize to an unseen growing season, indicating that linear relationships are inadequate for capturing seasonal variability in rice production. Machine learning-based rice yield prediction, particularly using the RF model, should be incorporated into agricultural monitoring frameworks to provide early estimates of rice production before harvest. Reliable predictions would support government and state agencies in evidence-based decision-making on grain procurement, import and export planning, food reserve management and market stabilization. Availability of quality weed control, fertilization and pest management data for rice production was a major limitation in this study which future studies could incorporate for rice yield predictions.