Technologies

5. Edge Computing and Distributed Learning

Cloud4Business designs edge computing architectures that run artificial intelligence models directly on the peripheral nodes of the infrastructure — environmental IoT sensors, agro-meteorological stations, water treatment plants, low-power embedded devices. This architectural choice enables real-time adaptive management decisions, with latency below 200 milliseconds, even in rural, mountain or post-emergency contexts lacking continuous connectivity. It is a non-negotiable requirement for operational environmental monitoring, where a network delay can translate into a failure to respond to a critical hydrological, ecological or climatic event.

On this basis we implement federated learning protocols for the continuous improvement of AI models from data distributed across multiple sites, without the raw data — ecological, financial, territorial — ever leaving the local infrastructure of the entity that produces it. Training takes place in a decentralised manner on the individual nodes; only the updated model parameters, anonymous and aggregable, are shared with the central orchestrator. This scheme allows multiple entities — public administrations, water utility operators, river basin authorities, farms — to contribute to the training of a shared model while retaining full legal and technical control over their sensitive data, in full consistency with the GDPR and with the principles of digital sovereignty to which we adhere.

Edge AI< 200 ms latencyFederated learningGDPRDigital sovereignty

6. AI Framework and Development Pipeline

For the development, training and deployment of predictive models we use PyTorch and TensorFlow as our reference frameworks, on heterogeneous hardware infrastructure: on-premise GPU servers for the intensive training of deep models and low-power edge nodes for field inference. The pipelines integrate, within a single environment, Copernicus Sentinel-1 satellite imagery (synthetic aperture radar for analysis of soil moisture, flooding, biomass) and Sentinel-2 (multispectral for vegetation indicators, water quality, habitat classification), time series from IoT sensor networks over LoRaWAN and MQTT protocols, and geo-referenced data from environmental registers, regional GIS systems and INSPIRE infrastructures.

The learning architectures are selected on the basis of the nature of the data: convolutional neural networks (CNN) for the analysis of remotely sensed imagery and land cover classification, LSTM recurrent networks for the modelling of environmental and hydrological time series, random forest for regression and classification on multi-source tabular data. All pipelines are optimised for multi-scale ecological data, where spatial and temporal variability is the rule and not the exception. Applications in production and under development include the monitoring of surface water quality, the early detection of invasive species, the mapping of ecosystem functionality, and the assessment of the performance of Nature-Based Solutions according to the BACI (Before-After-Control-Impact) protocol, which makes it possible to causally attribute the observed effects to the renaturalisation intervention, distinguishing them from the background variability of the context.

The pipeline software is released as open-source: see the full record on Zenodo (DOI: 10.5281/zenodo.19278318) and on the Open Science page.

PyTorchTensorFlowSentinel-1/2LoRaWANMQTTCNN/LSTMRandom forestBACI