Metode hibride de inteligență artificială pentru extragerea informațiilor latente din imagini de teledetecție: analiza la bord a hazardelor naturale / Hybrid artificial intelligence methods for latent information extraction from remote sensing images: on-board natural hazard analysis

The continuous change of our planet requires the full-time collection of very large volumes of data, especially in the context of an increasingly warming climate, which contributes to more unstable weather patterns. This thesis aimed to value the information available in satellite data for assessment and near real-time detection of natural disasters. The research directions employ Deep Convolutional Neural Networks (DCNN) based on the ResNet backbone in supervised, unsupervised and self-supervised learning methodologies to evaluate and quantify the disastrous effects of floods and wildfires. In the case of flood effects assessment, a parameter-free model based on data compression (NCD) was used on bi-temporal Synthetic Aperture Radar (SAR) images to estimate the changed areas. To estimate the impact of forest fires in the absence of a time series, features extracted exclusively from post-event multispectral imagery using deep learning extractors were statistically modelled to highlight the affected areas. Space technologies contributes to deployment and exploitation of space-based assets, e.g., on-board AI, system miniaturization, improved power storage and efficiency, in a unique operational environment with micro-gravity and extreme conditions. The adoption of new technologies in the hardware components that control the data flow in the on-board acquisition phase is still hampered by reliability and availability requirements. Thus, another line of research focused on developing a deep learning algorithm optimized to operate on embedded platforms with Field Programmable Gate Array (FPGA) logic. This algorithm was designed to detect wildfires in post-event multispectral imagery acquired on-board satellite. SAR images were obtained using the TerraSAR-X satellite and multispectral images via the Sentinel-2 satellite