AI and satellites in forest surveillance: classification and source fusion
Forest surveillance through satellite observation has incorporated, in recent years, automatic inference architectures that enable the classification of thermal hotspots with greater precision than approaches based exclusively on fixed radiometric thresholds. Optical and thermal infrared sensors aboard low-orbit platforms provide multispectral data which, processed by supervised classification models trained on geotagged historical series, allow active fires, controlled agricultural burns, industrial reflections and instrumental anomalies to be distinguished from one another. This discrimination capability is operationally relevant because it reduces the verification burden on emergency coordination centres. One of the persistent technical problems in forest early-warning systems is the false positive rate, which can significantly degrade operational efficiency when automated alerts do not incorporate sufficient spatial or temporal context. Deep learning architectures applied to surface reflectance data, combined with auxiliary information layers such as digital terrain models, vegetation moisture indices or reanalysis meteorological data, have demonstrated in validation environments a statistically significant reduction in these false alarms compared to conventional threshold-based detection methods. Source fusion constitutes another technical axis of interest in this domain. The integration of data from high-revisit-rate low-orbit constellations with higher spatial resolution imagery from medium-orbit or geostationary platforms makes it possible to construct surveillance products with complementary temporal and geometric coverage. This fusion requires a rigorous pre-processing chain that includes atmospheric correction, geometric registration and radiometric normalisation across sensors with differing spectral characteristics, conditions that largely determine the quality of the final classification product. From a regulatory and operational standpoint, the use of these systems in the European context is conditioned by the Earth observation data management frameworks established within the Copernicus programme, which defines access protocols, latency and product quality standards for emergency management services. The integration of automatic inference models into certified operational workflows additionally requires traceability of the classification process to allow system decisions to be audited by competent authorities, which raises specific requirements regarding the explainability and documentation of models deployed in mission-critical environments.
NASSAT - Network Satellite Systems