HI-FIVE: High-Resolution Forest Coverage with InSAR & Deforestation Surveillance

Living Planet Fellowship research project carried out by Francescopaolo Sica.

Forests are of paramount importance for the Earth’s ecosystem, since they play a key-role in reducing the concentration of carbon dioxide in the atmosphere and in controlling climate changes. The study of deforestation and development of global forest coverage and biomass is fundamental for assessing forests’ impact on the ecosystem. Remote sensing represents a powerful tool for a constant monitoring at a global scale of vegetated areas. In particular, given the daylight independence and the capability to penetrate clouds, space-borne synthetic aperture radar (SAR) systems represent a unique solution for the mapping and monitoring of forests. Sentinel-1, with its large coverage and short revisit-time, is a breakthrough technology, ideal for the generation of a constantly updated forest coverage map product and for the rapid monitoring of large-scale areas, aiming at detecting ongoing deforestation activities and forest disturbance.
The main objective of the proposed research project is to develop and implement advanced image-processing methods and strategies for the generation of high-resolution maps of forest coverage and deforestation from Sentinel-1 interferometric SAR data. Even though the detected SAR backscatter already provides useful information on forest coverage and structure, the use of SAR interferometry adds valuable and reliable information to the classification method. In particular, the temporal dynamic of the interferometric coherence, with a sampling period of 6 or 12 days, is investigated and modeled for different types of land cover.
The accurate estimation of InSAR parameters is of fundamental importance for approaching this analysis. The proposed methodology exploits nonlocal estimation methods to retrieve reliable information about InSAR parameters of the full-resolution SAR image. Different classification approaches are compared, from classical pixel/region-based classifier, to more recent machine learning approaches, such as Deep Convolutional Neural Networks.
Furthermore, since both temporal and volume decorrelation phenomena affect the coherence measurement in repeat-pass systems, such as Sentinel-1, I further propose the use of TanDEM-X bistatic data (simultaneous acquisitions) as high-resolution reference to support the modeling of Sentinel-1 backscatter information and its InSAR coherence temporal dynamic. Indeed, the combined use of single- and repeat-pass data allows for the isolation of volume and temporal decorrelation and for a more suitable use of the coherence at the aim of land classification.
Eventually, Landsat, Sentinel-2, TanDEM-X, and possibly laser data, together with ground truth references, will be used for training and validation.

The project’s main contribution has been to provide scientific evidence about interferometric coherence being a crucial parameter to classify land cover from SAR data accurately. Additionally, different frameworks have been proposed to exploit Sentinel-1 time series jointly with machine learning algorithms.

The provided methodologies have been successfully applied to map forest coverage in Europe and Brazil, specifically over the Amazon rainforest. The investigations have proved the ability of coherence time series to accurately follow forest coverage changes and therefore track deforestation phenomena promptly, down to a monthly timescale resolution. Deep Learning approaches were also developed to increase the temporal and spatial resolution. Specifically, the Phi-Net has been proposed to estimate the interferometric coherence accurately at the highest possible resolution, i.e., close to Sentinel-1 Single Look Complex (SLC) original resolution. Similarly, the temporal resolution has been targeted. An image segmentation algorithm based on the U-Net architecture has been developed to improve classification performance and set the groundwork for single image land cover classification, allowing change detection lower than a monthly scale.