Speaker
Description
Forests are of paramount importance for the Earth’s global ecosystem. They act as effective carbon sinks, reducing the concentration of greenhouse gas in the atmosphere, and help mitigating climate change effects. This delicate ecosystem is currently threatened and degraded by anthropogenic activities and natural hazards, such as deforestation, agricultural activities, farming, fires, floods, winds and soil erosion. The availability of reliable, up-to-date measurements of forest resources and evolution is therefore of great importance for environmental preservation and climate change mitigation. In this scenario, Synthetic Aperture Radar (SAR) systems, thanks to their capability to operate in presence of clouds, represent an attractive alternative to optical sensors for remote sensing surveys over forested areas, such as tropical and boreal forests, which are hidden by clouds for most of the year.
In this work, we will present the activities currently being carried out at DLR for mapping forests at large scale and high resolution, with the aim to regularly monitor vegetated areas and to detect long-term changing trends. Focusing on the Amazon rainforest basin, we will investigate the added-value of SAR interferometry (InSAR) in combination with artificial intelligence algorithms, having to cope with the lack of large datasets of labelled data for training deep learning algorithms. In particular, we will concentrate on the differences between bistatic and repeat-pass InSAR configurations, using the examples of the TanDEM-X and Sentinel-1 missions, respectively, proposing different deep learning-based frameworks which exploit the peculiarity of each mission.
Regarding TanDEM-X bistatic data, we will show how we can rely on self-supervised learning approaches and AI-based InSAR denoising algorithms for generating accurate forest maps with a ground resolution down to only 6 meters. This resolution outperforms all currently available spaceborne-based land cover products, setting the groundwork for an effective detection of forest degradation phenomena and selective logging activities. The absence of temporal decorrelation in TanDEM-X bistatic InSAR data allows us to generate highly accurate forest maps using only one single acquisition, but due to the limited acquisition capabilities it is not possible to perform a regular revisit of large-scale areas. For this reason, spaceborne SAR missions which can provide regular revisit, such as Sentinel-1, become of great interest for the development of an operational framework for monitoring of extended vegetated areas. In this case, the challenge to overcome is how to deal with the presence of temporal decorrelation in repeat-pass data. To this aim, we assume a reasonable stationarity of the illuminated scene during a short time period of about one month and we compute both backscatter and multi-temporal coherence stacks (computed at different temporal baselines) from Sentinel-1 short time series. These are used as input features to a convolutional neural network (CNN) for performing a multi-layer semantic segmentation task. Early results on a dedicated test site in Rondonia, Brazil, showed a very promising performance by combining backscatter information with coherence stacks at a temporal baseline of only 12 and 24 days, and that seasonal effects can be mitigated by properly designing the network training strategy. Here, the final goal is the development of a pre-operational framework, running on high performance computing facilities, for the regular monitoring of the Amazon basin, the detection of changes and the setup of a reliable, cloud cover independent early warning system for deforestation activities and forest degradation phenomena.
Finally, we will compare pros and cons of single-pass (bistatic) versus repeat-pass InSAR, discussing their main peculiarities and limitations. In particular, we will concentrate on the analysis of the interferometric coherence and on the relationship between volume and temporal decorrelation with respect to forest mapping, demonstrating the importance of combining multiple sensors for enhancing the overall capability to monitor forests dynamics at large-scale.
