Tiny Intelligence at the Edge: A Unified TinyML, Federated Learning and Hybrid Geospatial Architecture for Energy-Aware IoT Disaster Monitoring

Disasters such as floods, earthquakes and wildfires are becoming more frequent, yet the sensingnetworks that could warn of them are built from battery-powered microcontrollers that cannot run conventionalmachine learning. This paper presents a review-driven reference architecture that lets such devices take part indisaster intelligence. We analyse four recent peer-reviewed studies: (i) TDMiL, a framework for federated learningacross heterogeneous microcontrollers using CoAP and CBOR; (ii) a hybrid Random Forest–SVM–GradientBoosting–CNN model that couples artificial intelligence with geospatial analysis for flood, earthquake and wildfiremanagement; (iii) a variational-autoencoder framework that detects abnormal energy behaviour in IoT nodes; and(iv) TinyissimoYOLO, a sub-0.5MB quantised object detector for low-power microcontrollers. From these wederive a four-tier architecture (sensing, edge gateway, fog orchestration, cloud analytics), eight design views, aresource budget for field nodes, and a set of design guidelines. Using only figures reported in the source studies, weshow that an INT8 detector can run in 5.5ms and 196µJ per inference on an accelerator-equipped microcontroller,that straggler-aware aggregation is the safer choice on non-IID data at the price of extra radio traffic, and thatenergy-anomaly monitoring reaches 94.8% accuracy on a public dataset. The paper contains no new experiments; itcloses with a validation roadmap, a threat analysis and open research problems.