Description
The application of neural networks to remote sensing has already brought a transformative advancement in environmental monitoring, offering cutting-edge tools for the extraction of complex spatio-temporal patterns essential for informed environmental decision-making and rapid response to emerging crises.
This work leverages satellite observations to mainly address post-disaster assessment challenges, utilizing multi-sensor, multi-resolution, and multi-temporal datasets for precise analysis through deep learning models tailored for heterogeneous inputs. Our goal is specifically the segmentation of burn surfaces from over 100 wildfire events in the Mediterranean region, exploiting the multispectral Sentinel-2 imagery and using as ground truth reference the geolocated impact assessments from the Copernicus Emergency Management Service. The coverage of diverse geographical regions is expected to boost the model robustness and generalizability for impact mapping in different settings. Moreover, this dataset is expanded with complementary Sentinel-1 and Sentinel-3 observations, enabling a more comprehensive multi-sensor characterization of wildfire effects also in terms of revisit time and spatial resolution.
A dedicated Python library employing the SentinelHub API was developed to streamline downloading, preprocessing, and seamless integration of heterogeneous satellite data. Temporal series from pre- to post-event observations are processed using a UNet architecture with Convolutional LSTM layers, to effectively capture spatial and temporal patterns in rapidly transforming landscapes. The results demonstrate that our approach yields accurate segmentation of wildfire-affected areas, providing actionable inputs for emergency response and recovery.
This work has been carried out within the Spoke 2 as part of the activities in the Working Package 6 (“Cross-Domain Initiatives and Space Economy”) under the flagship use-case “AI algorithms for (satellite) imaging reconstruction”.
