Research

Landslides, floods and other slope and river hazards leave a clear footprint in satellite data, but turning that footprint into maps that responders and planners can use has traditionally meant weeks of manual interpretation. My research builds methods that shorten this path with artificial intelligence: deep-learning models that map hazards directly from optical and radar imagery, Earth-observation foundation models pre-trained on millions of unlabelled images, machine-learning models that say where the next failures are likely, and open datasets and weights that let these models be tested outside the places they were trained. Alongside the AI work, I use satellite radar, from Sentinel-1 to the new NISAR mission, to measure how slopes move before and after they fail.

NATURA: natural and anthropogenic drivers of landslide risk

NATURA is the project I lead at OGS, funded by a Starting Grant of the Italian Science Fund (FIS 3, MUR) from April 2026 to March 2031. It brings physically-based slope modelling, AI-driven analysis of satellite data (Sentinel-2, PlanetScope, Copernicus DEM) and socio-economic risk analysis into one framework. The aim is to capture the feedbacks between landslides and human-driven landscape change: NATURA uses CORDEX climate projections and Shared Socioeconomic Pathways to ask how urban expansion, economic change and climate change will shift landslide susceptibility, and which communities end up most exposed. Planned outputs are high-resolution risk maps, socio-economic vulnerability assessments and scenario-based risk projections for disaster risk reduction and climate adaptation.

Foundation models for Earth observation

Labelled landslide maps are slow and expensive to make, while aerial and satellite archives hold decades of unlabelled imagery. I work on self-supervised foundation models that learn from those archives first and need few labels afterwards. In the ChronoSat study with Xiaochuan Tang and Filippo Catani, we trained a masked autoencoder (MAE) and a joint-embedding predictive variant (JEPA) from scratch on 4.85 million unlabelled 0.2 m aerial patches of Emilia-Romagna (1976–2023) and compared them for landslide segmentation under an identical frozen-encoder protocol. MAE outperformed JEPA at every label fraction from 1 % to 100 % under that protocol, even though JEPA-style objectives often win on natural-image benchmarks, which suggests the fine morphology of scarps and deposits is better kept by masked reconstruction. An end-to-end fine-tuned supervised SegFormer-B2 remains the stronger reference. A revised version of the paper is in preparation. Preprint · open MAE and JEPA weights (867 M parameters each).

Rapid landslide mapping with deep learning

After an earthquake or an extreme rainfall event, the first question is where the slopes failed. I develop and benchmark convolutional and attention-based networks that delineate landslides from very-high-resolution optical and SAR imagery within hours of image acquisition, and study how well they transfer to new regions and sensors. Examples: rapid mapping in the Western Ghats, SAR-based mapping with attention U-Net, and the globally distributed training set HR-GLDD with its multi-region, multi-sensor benchmark.

Landslide susceptibility and risk

Mapping past landslides is only half of the problem. I work on susceptibility models that combine terrain, geology, land cover and triggering factors, with attention to which inputs actually matter and how the choice of mapping unit and training samples changes the answer (Belluno, NE Italy; hybrid machine-learning classifiers).

Slope motion from satellite radar and NISAR

Before many slopes fail, they creep. I use InSAR and radar offset tracking to measure this motion and to test what current missions can and cannot resolve, from displacement forecasting with deep learning to the first NISAR analyses of a Himalayan failure. For the 26 August 2026 Bhote Koshi ice–rock avalanche in Nepal, I processed all five NISAR L-band pixel-offset pairs spanning the event, with ASF's official offsets and an independent GAMMA chain. The source was moving before it failed, about 0.96 m in the last 24-day pre-event pair, but no interval stands out statistically against size-matched regions nearby and the rate does not accelerate. The honest answer to "can NISAR see it coming?" is therefore: it measures the motion, but that motion alone does not single out the slope that will fail. Brief communication (EGUsphere) · full study (Research Square).

Floods, exposure and cascading hazards

Mountain hazards rarely come alone. I also work on flood susceptibility, building-exposure mapping and the downstream effects of mass movements, for example the characterisation of buildings for flood exposure and the changes in water quality after the Chamoli disaster.

The full list of papers is on the Publications page.