A STUDY ON RETRIEVING SPATIAL LAND SURFACE TEMPERATURE DATA OF IRRIGATION-COMMANDED AREA BY SENTINEL-2 SATTELITE USING MACHINE LEARNING APPROACH

Abstract

“Per Drop, More Crop” is the plan of action of the Indian Government to achieve efficiency in the expenditure of irrigation water. A Support system that can provide high spatial-temporal resolution soil moisture data maps will bring enormous changes in the water management of Irrigation projects. Remote sensing-based support systems for Irrigation water management are very suitable because they are tractable and inexpensive. In most research studies focused on predicting high spatial and temporal resolution soil moisture (SM) data using remote sensing and machine learning, land surface temperature (LST) and the normalized difference vegetation index (NDVI) are commonly used as covariate data. Currently, the Landsat 8 and 9 satellites provide land surface temperature (LST) data with a combined revisit period of every 8 days. In response, this study aims to improve the temporal resolution of LST data from 8 days to 5 days by sentinel-2 utilizing a machine learning approach. The XGBoost algorithm was used for ML modeling. The ML model building was conducted using input variables as transformed optical sensor data like NDVI, Red, NIR, Shortwave transformed reflectance and IMD maximum air temperature data. The target variable was the Landsat 8 Collection 2 Level 2 LST. After training and validating separate models for the Rabi and Kharif seasons, they were tested with the 2023 Kharif and 2023-24 Rabi datasets. The test results for the Kharif season indicated that the Root Mean Square Error (RMSE) ranged from 1.812 to 2.966, and the Mean Absolute Error (MAE) ranged from 1.246 to 2.473. For the Rabi season, the model's test results showed an RMSE ranging from 0.874 to 3.866 and an MAE ranging from 0.681 to 3.513. To improve the temporal resolution of LST data, input variables generated by Sentinel-2 were used to predict LST, which can be supplied every 5 days.