Sentinel-2 영상과 기상모델 자료를 이용한 기계학습 기반의 벼논 토양염분 추정

Abstract

Seawater intrusion and soil salinity in rice paddies are factors that contribute to crop yield reduction in South Korea; however, current soil salinity measurements are limited to specific points, which restricts spatial analysis. To address this issue, this study estimated soil salinity in a spatially continuous raster map using machine learning; integrating Sentinel-2 satellite imagery for the Normalized Difference Vegetation Index (NDVI) and salinity indices (SI), along with meteorological data from the Local Data Assimilation and Prediction System (LDAPS). Electrical conductivity (EC), a proxy indicator of salinity, was used as the target variable, and a machine learning model was trained to estimate EC and then evaluated. This study focused on Hwaseong-si in Gyeonggi-do and, considering the different characteristics of coastal and inland paddy fields, developed and evaluated separate models for each region. In addition, to account for the variation in reflectance caused by vegetation masking the land cover, we trained and evaluarted separate models for growing and non-growing seasons. The inland model achieved the lowest mean absolute error (MAE) of 0.808 dS/m, while the seasonal models (0.893 and 0.895 dS/m) outperformed the model trained on the entire dataset. Feature importance analysis revealed that SI and NDVI contributed more during Growing season than non-growing season, indicating vegetation masking effects in satellite imagery were considered. These findings suggest that grouping datasets by geographical location and season improves prediction accuracy. Overall, this study demonstrates the potential of integrating satellite imagery and climate data with machine learning to analyze large-scale soil salinity for rice paddies.