APPLY MACHINE LEARNING TO LAND USE TRANSITION INVENTORY USING SENTINEL-1 AND SENTINEL-2

Abstract

The Normalized Difference Vegetation Index (NDVI) is a key metric for assessing vegetation health, derived from the red and near-infrared bands of optical satellite imagery. However, clouds frequently obstruct optical observations, preventing the calculation of NDVI in cloud-covered areas. Radar data, such as the VV (vertical transmit, vertical receive) and VH (vertical transmit, horizontal receive) polarizations from Sentinel-1, can penetrate cloud cover, offering valuable surface information that can be used to estimate NDVI in these obstructed regions. This project addresses the challenge of estimating NDVI for cloud-covered pixels in satellite imagery using machine learning regression techniques. By combining radar backscatter data (VV, VH) from Sentinel-1 with optical NDVI from cloud-free Sentinel-2 imagery, various regression models: Multiple Linear Regression, Gradient Boosting, Random Forest Regressor, KNN Regressor are applied to predict NDVI for cloud-obscured areas. The project leverages satellite image data from the Open Data Cube (ODC) platform under the EASI CSIRO Asia project, taking advantage of multi-sensor data integration. To ensure accurate predictions, the VV and VH radar data are resampled and aligned with NDVI values calculated from cloud-free Sentinel-2 imagery, forming the training dataset for the machine learning models.