A Machine Learning Approach to Spatial Analysis of Paddy Field Conversion Using Multispectral Sentinel-2A Imagery

The expanse of rice fields is a critical metric as it is intimately linked to agricultural productivity in a given locale. This study investigates the application of satellite imagery to quantify trice fields' acreage and temporal variations. The data utilized was acquired by the Sentinel-2A multispectral satellite. The variables employed are the image's baseband and spectral index. The research area encompasses the Sukamakmur sub-district in Bogor Regency, Indonesia. The types of machine learning models include Extreme Gradient Boosting (XGBoost), Multi-Layer Perceptron (MLP), and k-Nearest Neighbor (kNN). The simulation of class numbers is conducted to achieve the most stable and precise evaluation metric values. The XGBoost algorithm is used for the overall classification process of the region based on the optimal metric score. The model's accuracy, precision, recall, and F1-score are 92.37%, 92.3%, 92.38%, and 92.33%, respectively, indicating a very good performance. The model successfully captures a decline in rice field area between 2020 and 2023. Using the Modified Moran’s Index (MMI), the study reveals a positive spatial autocorrelation, indicating a clustered pattern in land-use change. Regions that experience either substantial or minor changes in land use are commonly situated near areas exhibiting similar characteristics. This study presents a spatially aware machine learning framework that enables the effective monitoring of agricultural land-use dynamics. In the future, this framework can be enhanced by integrating time-series forecasting and socio-economic data, supporting more informed decision-making in food security planning and agricultural policy development.