Abstract
<jats:p>Mediterranean environments are characterized by highly intermittent hydro-meteorologicalprocesses, including intense rainfall events, flash floods and rapidly evolving atmosphericconditions. These phenomena generate considerable forecasting challenges despite recentadvances in Machine Learning (ML), Deep Learning (DL) and foundation models applied toenvironmental prediction.While most studies focus on comparing forecasting algorithms, the intrinsic relationshipbetween the statistical structure of environmental variables and their forecastability remainspoorly understood. We hypothesize that predictive performance is strongly constrained bythe statistical organization of the observed processes and not solely by model complexity.To investigate this question, we present SAPHIR DATA SERVICE, a cloud-native en-vironmental intelligence platform developed at SPE lab, in the University of Corsica incollaboration with LiSN Lab in the University of Paris Saclay . The platform supportsthe continuous acquisition, storage, visualization and exploitation of heterogeneous envi-ronmental observations originating from meteorological stations, hydrological sensors, IoTmonitoring devices and institutional environmental services.SAPHIR DATA SERVICE relies on a dual-layer data architecture. A real-time databasesupports operational monitoring, environmental surveillance and nowcasting activities, whilea historical database supports long-term analyses, retrospective studies, machine learningtraining and scientific investigations. This architecture enables the simultaneous manage-ment of operational and research-oriented workflows within a unified framework.The backend infrastructure is continuously supervised through Grafana dashboards pro-viding real-time monitoring of acquisition pipelines, database services, sensor status andenvironmental observations. A complementary web frontend provides access to environmen-tal indicators, historical analyses, forecasting products and decision-support services.Beyond data management, the long-term objective of SAPHIR DATA SERVICE is theimplementation of a continuous environmental intelligence pipeline linking observation, in-gestion, learning, prediction and decision support.Within this framework, we investigate whether forecastability can be considered an in-trinsic property of environmental variables and whether it can be explained through theirstatistical signatures. To address this question, we introduce a characterization frameworkcombining temporal autocorrelation, spatial intercorrelation, spatio-temporal structure func-tions, fractional moments, skewness, kurtosis and intermittency.These descriptors are evaluated against forecasting performances obtained from a cata-logue of statistical, machine learning and foundation models, including persistence baselines,ensemble methods and modern deep-learning architectures.The Porto-Vecchio study area provides a real-world Mediterranean testbed for evaluat-ing how environmental statistical signatures relate to achievable forecasting skill within anoperational nowcasting framework. The proposed approach aims to establish a quantitativerelationship between environmental data structure and predictive performance, providingnew perspectives for rare-event forecasting, hydro-meteorological risk management and en-vironmental decision-support systems.</jats:p>