A Machine Learning approach to reconstruct cloudy affected vegetation indices imagery via data fusion from Sentinel-1 and Landsat 8

A way to reconstruct optical sensor-derived images allowing cloud-free vegetation monitoring is proposed in this paper. The motivation is the influence that clouds have on optical remote sensing of tropical regions, which hinders Earth observation systems because their presence reduces imaging frequency. To circumvent that problem, a machine learning model-based integration methodology for the fusion of Landsat 8 and Sentinel-1 data is proposed herein. Sentinel-1 constellation has mounted Synthetic aperture radar (SAR) sensors are used because the imaging is not affected by clouds due to microwave spectrum characteristics. To study the problem and predict both the normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI), three algorithms were selected: multivariate linear regression, multivariate adaptive regression splines, and random forest (RF). Two testing strategies were also chosen: k-Fold cross-validation for hyperparameter tuning of the model and holdout testing to assess the generalization ability of the model. The SAR covariables were employed to feed the algorithms, including selected SAR vegetation indices; in addition, environmental data, such as land use and land cover (LULC), the date, and position of the samples were used. The predictions from the NDVI and EVI produced good results, namely, similar Willmott's agreement index (d) values that ranged from similar to 0.64 to 0.96. The best-fitted model was the RF, which was used to reconstruct the NDVI images and produced good results that agreed well with the predictions (d index from 0.58 to 0.87) and spatial patterns. The results obtained show that the integration of radar and environmental covariables with optical vegetation indices can allow vegetation monitoring that is free of gaps due to clouds.