Estimating Rice Production using Machine Learning Models on Multitemporal Landsat-8 Satellite Images (Case Study: Ngawi Regency, East Java, Indonesia)

Abstract

To enhance sustainable food security, the cost-efficient data collection technology for estimating rice production in a major agriculture nation such as Indonesia is undoubtedly vital to support the existing official data collection. The current official data collection is still facing great challenges in terms of its high cost and laborious nature. This study aims to build machine learning-based models for rice production estimation by utilizing multitemporal Normalized Difference Vegetation Index (NDVI) data obtained from Landsat-8 remote sensing satellite imagery focusing on Ngawi Regency, East Java, Indonesia as a case study area. Our investigation reveals the quarterly changes in vegetation conditions of the rice fields can be captured through the NDVI value. Four different machine learning models are constructed and evaluated to process the satellite data. Support vector regression (SVR) was shown to obtain the best performance from 10-folds cross-validation with the average root mean square error (RMSE) of 6952.89 tons and has a quite high coefficient of determination (R2) score which is up to 0.9. The current estimation results provide an incentive to use satellite imagery data and machine learning models to support agricultural monitoring and decision-making.