Periodic Reporting for period 1 - TEMA (Trusted Extremely Precise Mapping and Prediction for Emergency Management)
Okres sprawozdawczy: 2022-12-01 do 2024-05-31
TEMA will greatly improve Natural Disaster Management (NDM, e.g. for wildfires, floods) by automating precise semantic 3D mapping and disaster evolution prediction to achieve NDM goals in near-real-time. It will analyze and fuse many heterogeneous extreme data sources: smart drone and in-situ sensors, remote sensing data, topographical data, meteorological data/predictions and geosocial media data (text, image and videos). TEMA will focus on the extreme nature of the data, due to their varying resolution and quality, very large volume and update rate, different spatiotemporal resolutions and acquisition frequencies, real-time needs and multilingualism. It will develop an integrated, ground-breaking NDM platform, focusing on real-time semantic extraction from multiple heterogeneous data modalities and sources, on-the-fly construction of a meaningful semantically annotated 3D disaster area map, prediction of disaster evolution and improved communication between service providers and end-users, through automated process triggering and response recommendations. Semantic analysis computations will be distributed across the edge-to-cloud continuum, in a federated manner, to minimize latency. Extreme data analytics will be performed in a trustworthy and transparent way, by greatly advancing state-of-the-art AI and XAI approaches. The constantly updated 3D map and the disaster evolution predictions will form the basis for an advanced, interactive, Extended Reality (XR) interface, where the current situation will be visualized and different response strategies will be dynamically evaluated through simulation by NDM personnel. The innovative, scalable and efficient TEMA platform will provide precise NDM support, based on extreme data analytics. It will be validated on two critical disaster use-cases (wildfires and floods), in four EU countries, and will form the basis for the TEMA NDM-Analytics-as-a Service (NDM-AaaS) model.
Summary of TEMA Project Activities (M1-M18)
WP1: Project Management
Established management structure, communication protocols, and Consortium Agreement.
Organized and documented project meetings.
Managed financial distribution, budget monitoring, and reporting.
Coordinated technical planning, deliverables, and templates.
WP2: User Requirements and Specifications
Identified end-user requirements and completed technical specifications.
Ensured legal and ethical compliance for new data collection.
WP3: Analytics and AI Development
Developed explainable AI methods and decentralized DNN frameworks for robust fire detection.
Enhanced real-time fire and flood detection using optimized models and new datasets.
Improved sentiment analysis and topic identification in social media.
WP4: Prediction and Information Fusion
Advanced 3D smoke and flood modeling for better evacuation strategies.
Developed information fusion and decision support mechanisms to aid disaster response.
WP5: Visualization and Augmented Reality
Created precise Digital Twins and initiated geovisual analytics system.
Planned AR-based interactive visualization prototypes for real-time disaster management.
WP6: Hardware Integration
Focused on integrating TEMA hardware components and validating experimental setups.
WP7: Dissemination and Collaboration
Established dissemination strategies, designed the TEMA website, and engaged in communication activities.
Assessed market potential and developed IP frameworks.
Formed AI.BIG Cluster with other Horizon Europe projects for AI and Big Data in disaster management.
WP1: Project Management
Established management structure, communication protocols, and Consortium Agreement.
Organized and documented project meetings.
Managed financial distribution, budget monitoring, and reporting.
Coordinated technical planning, deliverables, and templates.
WP2: User Requirements and Specifications
Identified end-user requirements and completed technical specifications.
Ensured legal and ethical compliance for new data collection.
WP3: Analytics and AI Development
Developed explainable AI methods and decentralized DNN frameworks for robust fire detection.
Enhanced real-time fire and flood detection using optimized models and new datasets.
Improved sentiment analysis and topic identification in social media.
WP4: Prediction and Information Fusion
Advanced 3D smoke and flood modeling for better evacuation strategies.
Developed information fusion and decision support mechanisms to aid disaster response.
WP5: Visualization and Augmented Reality
Created precise Digital Twins and initiated geovisual analytics system.
Planned AR-based interactive visualization prototypes for real-time disaster management.
WP6: Hardware Integration
Focused on integrating TEMA hardware components and validating experimental setups.
WP7: Dissemination and Collaboration
Established dissemination strategies, designed the TEMA website, and engaged in communication activities.
Assessed market potential and developed IP frameworks.
Formed AI.BIG Cluster with other Horizon Europe projects for AI and Big Data in disaster management.
Summary of TEMA Project KPI Achievements (M1-M18)
Semantic/Instance Segmentation Accuracy
Burnt Region: AUTH's CNN-I2I model improved mIoU by 8.56% over SotA.
Flood Segmentation: AUTH's ST++ method increased mIoU by 3.5%; DLR-DFD's methods improved IoU by 23% (burnt regions) and 10% (floods).
Object Detection Accuracy
Fire Detection: AUTH's RT-DETR model increased mAP by 6.2% over SotA.
Data Scarcity Solution: ATOS created a synthetic dataset to enhance training.
Image Classification Accuracy
Forest Fire: AUTH's EfficientNet B1 increased CCR by 14%; additional methods (OvO, IAMSP, QoI) showed significant accuracy gains (13.21% and 16.69% over SotA).
Sentiment Analysis Accuracy in Social Media
Novel Method: AUTH's strategy improved accuracy by 9.3% over SotA.
Aspect-Based Emotion Analysis (ABEA): PLUS developing advanced sentiment analysis, quantitative evaluation pending.
Topic Identification Accuracy in Social Media/News
Joint Topic-Sentiment Approach: PLUS improved topic quality by 11% and coherence by 14% over BERTopic.
Visual Analysis Speed
Real-time Performance: AUTH's models for object detection and segmentation (RT-DETR, YOLOv6, CNN-I2I) achieved high FPS (33-145 FPS). DLR-DFD's methods significantly increased speed of data analysis.
Social Media Analysis Speed
Real-time Analysis: AUTH's BERTbase model can analyze 50-100 posts/second. PLUS's pipelines handle multimodal information extraction, but real-time analysis is currently not feasible.
Semantic/Instance Segmentation Accuracy
Burnt Region: AUTH's CNN-I2I model improved mIoU by 8.56% over SotA.
Flood Segmentation: AUTH's ST++ method increased mIoU by 3.5%; DLR-DFD's methods improved IoU by 23% (burnt regions) and 10% (floods).
Object Detection Accuracy
Fire Detection: AUTH's RT-DETR model increased mAP by 6.2% over SotA.
Data Scarcity Solution: ATOS created a synthetic dataset to enhance training.
Image Classification Accuracy
Forest Fire: AUTH's EfficientNet B1 increased CCR by 14%; additional methods (OvO, IAMSP, QoI) showed significant accuracy gains (13.21% and 16.69% over SotA).
Sentiment Analysis Accuracy in Social Media
Novel Method: AUTH's strategy improved accuracy by 9.3% over SotA.
Aspect-Based Emotion Analysis (ABEA): PLUS developing advanced sentiment analysis, quantitative evaluation pending.
Topic Identification Accuracy in Social Media/News
Joint Topic-Sentiment Approach: PLUS improved topic quality by 11% and coherence by 14% over BERTopic.
Visual Analysis Speed
Real-time Performance: AUTH's models for object detection and segmentation (RT-DETR, YOLOv6, CNN-I2I) achieved high FPS (33-145 FPS). DLR-DFD's methods significantly increased speed of data analysis.
Social Media Analysis Speed
Real-time Analysis: AUTH's BERTbase model can analyze 50-100 posts/second. PLUS's pipelines handle multimodal information extraction, but real-time analysis is currently not feasible.