Abstract
Abstract The MIRAGE project is motivated by the need for rapid, wide-area monitoring of slow mass movements—such as landslides and rock glaciers—to mitigate hydrogeological risks heightened by climate change. By leveraging deep learning, the project aims to unlock the full potential of "raw" wrapped DInSAR (Differential Synthetic Aperture Radar) data to automate the detection and classification of these processes. This approach addresses the technical complexity and massive data volumes that have previously hindered large-scale monitoring, ultimately supporting land-use planning and disaster risk reduction The research initiative is based on two study areas in the Central Alps and Northern Apennines where an extensive expert-annotated dataset was created. A novel YOLO-based convolutional neural network is then used to emulate expert human interpretation and identify displacement patterns within raw, wrapped interferograms. Progress has been documented through specialized work packages that integrate geomorphological constraints with sophisticated artificial intelligence to improve the accuracy of environmental hazard assessments.
Results achieved
The MIRAGE project has achieved significant milestones in the automation of landslide detection, merging expert geomorphological knowledge with advanced artificial intelligence. Below is an expanded overview of the methodological achievements, the technical development of the neural network, and the project’s broader impact. Expert-Driven Methodological Framework A cornerstone of the project was the development of a robust methodology for mapping and labeling DInSAR (Differential Synthetic Aperture Radar) signals. Unlike traditional methods that often filter out complex data, MIRAGE utilized "raw" wrapped interferograms to capture a wider range of movement rates. Experts followed a rigorous four-step process to create high-quality training data: selecting optimal interferograms, identifying phase fringes, and verifying both the morphological consistency (using high-resolution terrain models) and the temporal recurrence of the signals. This effort resulted in the MIRAGE dataset, a publicly available library of 4,910 expert-annotated signals covering hundreds of individual mass movements across the Alps and Apennines. CNN Training and Performance The project developed YOLO-MIRAGE, a Convolutional Neural Network (CNN) specifically tailored for geohazard detection. To emulate human interpretation, the model was trained using a unique three-layer input stack: 1. The "raw" wrapped DInSAR phase containing the movement signal. 2. An InSAR Sensitivity Map (ISM) that weights the reliability of the signal based on satellite geometry and terrain. 3. A composite geomorphometric layer derived from Principal Component Analysis (PCA) to help the model distinguish unstable slopes from stable ground. The network was refined through iterative tuning, utilizing strategies such as 16-bit image handling to preserve data precision and data augmentation to improve its ability to generalize to new areas. In the Northern Apennines, the model achieved a mean Average Precision (mAP) of 0.804 across four process classes. In the more topographically complex Central Alps, the model maintained a solid AP of 0.765 for general mass movement detection. Validation and Scientific Discovery Validation went beyond standard metrics; geomorphologists manually inspected model predictions against field evidence. This process revealed that the model was occasionally more sensitive than human experts, identifying weak or small displacement signals that had been missed during the initial manual mapping stage. Furthermore, the project delivered a QGIS plugin, providing practitioners with a tangible tool to assess InSAR terrain sensitivity. Potential Impact and Future Outlook The results of MIRAGE demonstrate that routinely available satellite data can be rapidly "unlocked" to provide wide-area monitoring without the need for time-consuming traditional processing. The potential impact is twofold: • For Civil Protection and Land-Use Planning: The framework supports the rapid update of landslide inventories and the early identification of active areas, which is critical for disaster risk reduction in a changing climate. • For the Scientific Community: By sharing the expert-constrained MIRAGE dataset, the project provides a foundation for the global research community to develop even more sophisticated physics-informed AI models. Ultimately, MIRAGE transitions satellite monitoring from a specialized, manual task into a fast, automated operational tool for long-term geohazard management. The project's impact was disseminated through high-impact scientific publications and a final workshop involving over 40 stakeholders from government agencies and industry.Dettagli del progetto
Responsabile scientifico: Alessandro Simoni
Strutture Unibo coinvolte:
Dipartimento di Scienze Biologiche, Geologiche e Ambientali
Coordinatore:
Università degli Studi di MILANO-BICOCCA(Italy)
Contributo totale Unibo: Euro (EUR) 78.861,00
Durata del progetto in mesi: 24
Data di inizio
28/09/2023
Data di fine:
28/02/2026
