Abstract
We present an adaptive system for the automatic assessment of both physical and anthropic fire impact factors on periurban forestries. The aim is to provide an integrated methodology exploiting a complex data structure built upon a multi resolution grid gathering historical land exploitation and meteorological data, records of human habits together with suitably segmented and interpreted high resolution X-SAR images, and several other information sources. The contribution of the model and its novelty rely mainly on the definition of a learning schema lifting different factors and aspects of fire causes, including physical, social and behavioural ones, to the design of a fire susceptibility map, of a specific urban forestry. The outcome is an integrated geospatial database providing an infrastructure that merges cartography, heterogeneous data and complex analysis, in so establishing a digital environment where users and tools are interactively connected in an efficient and flexible way.
Introduction
Urban and periurban forests, such as London’s Hyde Park or Rome’s Castel Fusano, are essential for maintaining air quality and providing recreational spaces in metropolitan areas. However, these environments face significant fire risks that require sophisticated management strategies. Effective fire prevention must balance the preservation of biodiversity with the need to clear fuel loads, as neglect can lead to catastrophic fires threatening urban infrastructure.
The complexity of fire risk assessment stems from the interplay of physical factors (vegetation type, moisture content), meteorological conditions, and anthropic (human) behaviors. Traditional fire maps often lack the resolution necessary for local forest management. This research addresses this gap by proposing a high-resolution, integrated geospatial database that combines "hard" satellite data with "soft" social and historical information to estimate fire susceptibility accurately.
Methodology
The proposed system utilizes a multi-stage pipeline to convert raw radar data into actionable fire susceptibility maps.
SAR Acquisition and Coregistration
The system employs X-band Synthetic Aperture Radar (SAR) data from the Cosmo-SkyMed constellation. SAR technology is advantageous because it can penetrate clouds and smoke and operates independently of solar illumination. To facilitate interpretation, rectified SAR images are coregistered with optical imagery from Google Earth. This process creates a multi-dimensional data structure where radar backscattering is aligned with standard RGB channels.
The alignment is achieved by maximizing a cross-correlation function between the radar and optical datasets, ensuring precise spatial mapping of the forest features.
Segmentation and Moisture Estimation
A discrete level set segmentation method is used to partition the images into thematic zones, such as specific tree canopies. This algorithm identifies subregions by minimizing a cost function that balances image fit with regularization. Furthermore, the system estimates relative humidity and moisture content by correlating the average gray-level values of these SAR segments with ground-truth meteorological data, specifically dew point and temperature.
Feature Extraction and Classification
To overcome the "speckle" noise inherent in SAR imagery, the system extracts Speeded Up Robust Features (SURF) to describe the texture and spatial properties of the pixels. These descriptors are then processed by a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel to classify the forest into categories such as "oak forest" or "burned area."
Data and Study Areas
The study focuses on two primary locations:
- Hyde Park, London: Used to demonstrate the coregistration of SAR data with optical imagery.
- Castel Fusano, Italy: A forest near Rome used for training and validation. Data included high-resolution (0.6m to 1m) Cosmo-SkyMed X-SAR images, ancillary thematic maps (soil, vegetation, historical fire records), and meteorological data from the Aviation Digital Data Service.
Results and Discussion
The system demonstrated high performance in identifying forest types and previously burned regions in the Castel Fusano area. The SVM achieved a training accuracy of 99.10% across 30,000 instances. In testing on independent X-SAR images, the model identified oak tree areas with 90.03% accuracy and non-oak areas (roads, burned patches) with 89.40% accuracy.
Key insights include the finding that texture-based descriptors (SURF) are significantly more effective than simple intensity values for species identification in noisy SAR data. Additionally, the unique sensitivity of X-band radar to moisture allows for the estimation of fuel moisture, a critical driver of fire ignition. The inclusion of anthropic factors—treating human habits as layers in the geospatial grid—represents a significant advancement in modeling the social reality of urban fire risks.
Conclusion
While the system provides a robust framework for fire susceptibility mapping, challenges remain regarding speckle noise and the integration of more complex behavioral data. Future work will focus on generalizing the SURF-based descriptors to diverse forest compositions and climates, and further refining the modeling of human-induced fire risks.
Summary updated