Vegetation index for crop monitoring from multi temporal SAR data

Democratization of technologies and dedicated space programs developed in the last decades have immensely advanced the field of remote imagery and sensing to a mature state where it can enhance the quality of life on earth. The Copernicus Sentinel program, directed by European Space Agency, has provided in the last years a continuous and stable flow of data coming from space which has become indispensable for environmental monitoring, resource management, and disaster response. Data coming from Sentinel-1 and Sentinel-2 satellites have immensely boosted the research in the agricultural field providing new tools to tackle new challenges and develop technologies for general purpose. These sources cover the field of optical and near-optical with Sentinel-2 data, while Sentinel-1 provides radar data, which has the unique ability to capture information day and night, regardless of atmospheric conditions, including cloud cover. The aim of this work is to explore advanced methodologies for estimating the Normalized Difference Vegetation Index (NDVI), a key indicator used to monitor vegetation health in agriculture and environmental management, under cloud-covered conditions. Since clouds can obscure Sentinel-2 optical data, this study leverages multi-temporal Sentinel-1 radar data to estimate NDVI, offering a solution when optical data is unavailable. The approach addresses critical challenges in monitoring vegetation and enhances the reliability of assessments in agricultural contexts. An introduction to the field of Remote Sensing is presented, including a general overview of the currently available technologies that are available today. It then addressed the most common problems associated with remote sensing data. A review of existing cloud removal and radar vegetation index techniques follows, highlighting their advantages and limitations. Building on this foundation, a multi-temporal data framework is presented, designed to exploit the spatial, temporal and spectral information of imagery from Sentinel-1. The data used for the preliminary experiments comes from the public available SEN12TP dataset. Taking that work as a reference a novel dataset Sensed_TS90 has been generated to address specific challenges of using time series data as input to deep learning models to estimate NDVI in agricultural areas. The thesis concludes with a presentation of xperimental results and a discussion of potential improvements. The use of the UTAE model, a neural network designed to handle time series data, provides promising evidence that incorporating the temporal dimension offers a distinct advantage in NDVI estimation compared to single acquisitions from Sentinel-1. The results look promising, and they lay the groundwork for future investigation. This thesis advances the field of remote sensing applications by introducing a novel dataset that facilitates future research, offering a comprehensive comparison of state-of-the-art methods for NDVI estimation. Finally, it also highlights the potential benefits of using temporal information in analyzing Sentinel-1 data with cutting-edge neural networks for time series imagery.