ESTIMATION OF NDVI FOR CLOUDY PIXELS USING MACHINE LEARNING

Abstract

The Normalized Difference Vegetation Index (NDVI) is a useful index for vegetation monitoring. However, due to cloud cover the observations of NDVI are discrete and vary in the intensity. Therefore, there is a need to estimate the NDVI during cloud cover using alternative sources of satellite observations. The main objective of this study is to estimate NDVI during cloudy conditions using moderate resolution multi-spectral and synthetic aperture radar (SAR) observations. Two approaches were identified: 1) pixel replacement and 2) machine learning based regression analysis to estimate cloud free NDVI. Moderate Resolution Imaging Spectroradiometer (MODIS) 8-day NDVI composite, Sentinel-1 SAR and cloud masked Sentinel-2 multi-spectral observations were collected for entire cropping season. The satellite observations were selected only for agricultural areas by applying the agriculture, non-agriculture land use land cover mask. Machine learning algorithms such as Linear Regression (LR), Random Forest Regression (RFR), and Support Vector Regression (SVR) were used for NDVI estimation. Regression analysis was performed using Sentinel-2 NDVI as an independent variable and VV, VH, Cross Ratio (i.e., VV/VH), and MODIS NDVI as dependent variables. NDVI of the cloudy pixel was estimated using the trained regression models over the agriculture areas. A regression model was trained and applied to each Sentinel-2 tile that covers an area of 100 km × 100 km. The RFR and SVR showed the highest R2 of 0.73 and a RMSE of 0.12. A visual comparison of time series graphs showed good alignment between actual (Sentinel-2) and predicted NDVI and usual crop growth trend.