Project Summary

● Implementation of segmentation algorithms for coldspots (clouds, waterpoints) and hotspots (wildfires) for Sentinel-3 data

● Evaluation against FRP (Fire Radiative Products) from Sentinel-3 and MODIS

● Journal publication

Development Tools

● I downloaded data from Sentinel-3 using the phi-down platform developed by the Φ-lab.

● I used ERA5 wind data.

Development Outputs

● Upcoming: Journal paper.

● Most of my research will be included as the last chapter of my PhD thesis.

Project Description

Anomaly detection is an essential task in the analysis of remote sensing data, focusing on finding spatial irregularities that differ from normal background patterns. As high-dimensional satellite data becomes more abundant, deep learning models (especially generative models) have proven highly effective at understanding complex data distributions to spot these unusual structures without supervision. Specifically, diffusion models have gained traction as robust solutions for reconstructing images and finding anomalies.

In remote sensing, identifying anomalies offers a data-centric alternative to more traditional physics- or threshold-based techniques. This is particularly valuable when noise, cloud cover, or missing data make direct modeling challenging. By learning from “normal” distributions, generative AI can recognize statistically notable shifts in the data, allowing researchers to catch rare or unforeseen occurrences over diverse areas. However, even though this type of detection successfully pinpoints anomalies, it still fails at explaining the actual nature or identity of the anomaly.

To overcome this limitation, we introduce a multi-channel anomaly detection system (see Figure 1) built on diffusion models and completed with a classification component. This proposed framework is specifically engineered to simultaneously locate and categorize unusual patterns within multispectral Sentinel-3 data.

Figure 1: Workflow of the anomaly detection framework applied to SLSTR data. (a) is the complete framework, while (b) describes the partial diffusion process and how anomaly maps are generated.

Find below an example of coldspots anomalies identified by our framework and one of our segmentation algorithm.

Figure 2: Visualization of anomalies detected by our algorithm used for coldspot segmentation (a) against clouds detected by the SLSTR flag (b), the thermal infrared F1 band (c), and the short-wave infrared S4 band (d).