Speaker
Description
In recent times, the combination of remote sensing and machine learning applications has led to great perspectives in the context of space economy. A flagship use case "AI algorithms for (satellite) imaging reconstruction" has been therefore established within the Working Group 6 (WP6) "Cross-Domain Initiatives and Space Economy" under Spoke 2, to focus on the analysis of satellite and aerial images.
In this contribution we presents the work accomplished in the segmentation of wildfire-affected areas by means of deep learning techniques, in particular Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) Networks, tested separately and together in a UNet-like architecture with ConvLSTM layers. A custom library, specifically designed to optimize the entire workflow from download to pre-processing and analysis, has been employed to create a dataset of multispectral images from the Sentinel-2 satellites, in combination with the information about wildfires provided by the Copernicus Emergency Management Service. The results are more than encouraging, proving the power of such methodologies in environmental monitoring and disaster response.
Moreover, the analysis of a dataset of satellite images for the early detection of vineyard diseases is currently underway. Such a dataset has been built using the geolocation data and disease impact assessments from a recent airborne survey (8-19 July '24) covering approximately 2000 hectares across the Emilia-Romagna region, performed in the context of the PERBACCO project. This collaboration aims to integrate high-resolution, airborne observations with satellite data to develop more robust, multi-scale monitoring solutions.
