One of the objectives of the PERIVALLON project is to develop AI-based geospatial intelligence solutions applied to earth observation imagery captured by satellites, in order to detect activities potentially related to environmental crimes, particularly those related to illicit waste management. In this framework, one of the main requirements for a regional authority like ARPA Lombardia, which is one of the end users among the project partners, is to efficiently analyse large areas to support surveillance activities against illicit or non-compliant waste management practices.
Access to High-Resolution Satellite Data
The PERIVALLON project was granted access to very high-resolution satellite data through the Copernicus Contributing Missions (CCM) Service by the European Space Agency (ESA). This grant provides access to both archived and newly acquired data.
To test and demonstrate PERIVALLON’s solutions, a specific request was recently submitted to the Copernicus Contributing Missions (CCM) Service for a new acquisition of very high-resolution satellite data over a designated Area of Interest (AOI) in Lombardy. The selected AOI is located in central Lombardy, mainly in the Province of Bergamo, for a total area of 510 Km2, covering (totally or partially) 78 municipalities with different characteristics. Notably, part of this AOI has previously been analyzed using traditional photo interpretation methods based on past satellite data, enabling multitemporal assessments and comparisons.
A few days after the request to the CCM system, the AOI was successfully acquired by WorldView-3 satellite (acquisition date: 13/08/2024, spatial resolution 30 cm).
AI_SAT_LAND Model Results
The high-resolution satellite data received by the CCM system was pre-processed by ARPA. After that, the waste detection model (AI_SAT_LAND), developed by POLIMI in PERIVALLON, performed the inference on the entire imagery dataset.
ARPA’s evaluations on data retrieval and processing workflow and on AI_SAT_LAND model outputs, confirmed the very good quality of the imagery acquired from CCM (WV-3, 30 cm resolution) and the timeliness of AI_SAT_LAND model in delivering results. These factors are crucial for real-world applications, where rapid processing of large datasets is essential.
Within the entire AOI, the AI_SAT_LAND model, applied to the 13/08/2024 WV-3 data, found more than 4700 tiles (squared areas where the model detected waste areas, classified with different priority scores).
Along with the square tiles, another useful output provided by the AI_SAT_LAND model is the CAMS (class activation maps) polygon file, that represent the areas on which the algorithm focuses in the classification process.
In “never seen” areas, the outputs provided are very effective in guiding attention from a broad scale to the most critical situations at a local level.
The outputs provided are also highly effective in monitoring the status of previously identified “potential critical sites”, and then, if it is the case, in supporting the evaluations prior to direct control or inspections, also supported by very updated imagery before accessing the specific site.
Thanks to PERIVALLON, it is now possible to automatically analyze a vast territory in a few hours, a task that would take weeks to complete manually.
Benefits and Future Directions
The methods developed seem effective for strengthening the “surveillance approach” over wide areas. Regarding the performance of the model, no issues are posed by false-negatives (i.e., the potentially interesting areas are in general well-classified by the algorithm), while some potential false-positives can arise (example: some misclassification over rooftop and in densely industrialized areas) leading to scored tiles high cardinality (situation however preferable because precautionary).
From an operational perspective, in the context of the typical Lombardy territory, the number of the “interesting tiles” (e.g. score > 0.7) appears rather high, because the AI-SAT_LAND algorithm tends to classify correctly all the areas with presence of waste or other materials.
To facilitate integration into the operational monitoring and control workflows of end users (Environmental Agencies, Law Enforcement Agencies, etc.), a key focus is on the prioritization of the detections, and the development of a robust risk assessment, based also on external information and different datasets: this evolution is covered in PERIVALLON under the developments of the PERIVALLON Platform.
Future developments include conducting a more formal evaluation of the outcomes and benefits of the PERIVALLON solutions for end users. Additionally, expanding the analysis to include additional satellite datasets in the Lombardy region (future inputs for PUC1-Italy 2nd and 3rd iterations) and other European territories with varying characteristics will further enhance the project’s capabilities.
These advancements will contribute to the ongoing development and refinement of the PERIVALLON platform, ensuring its continued effectiveness in supporting environmental monitoring and enforcement efforts.
Written by Dario Bellingeri and Dario Lombardi
ARPA Lombardia, UO CREO – Centro Regionale di Earth Observation