Self-organising maps for change detection and monitoring of human activity in satellite imagery
The increasing number and resolution of satellite sensors to be launched in the coming years will dramatically increase the need for efficient image archiving and processing. Current Earth Observation archiving systems typically support queries by sensor type, acquisition date, imagery coverage or a combination of them. Concurrently, security-concerned applications relying on satellite imagery often demand repeated or continuous monitoring, and intelligent access to the extracted information. There is therefore a growing interest in the remote sensing community to access databases directly by the information contained in images. Content-Based Image Retrieval (CBIR) allows an efficient management of large image archives, as well as satellite image annotation and interpretation. We extended and exploited the potential of PicSOM, a CBIR system based on Self-Organising Maps (SOMs), for remote sensing image analysis, with a particular focus on detection of man-made structures and changes. The same framework allows for supervised and unsupervised change detection, with promising applications in long-term monitoring of strategic sites or content-driven novelty detection.