RescueNet Model

• RescueNet Model is a family of integrated frameworks that enable real-time disaster response, remote sensing damage assessment, and multi-objective resource optimization.
• It employs techniques such as heuristic multi-agent reinforcement learning, decision-theoretic dispatch, and attention-based segmentation for efficient emergency management.
• Evaluations demonstrate reduced response times, cost-effective resource allocations, and high segmentation accuracy through deep neural architectures and meta-heuristic optimization.

RescueNet Model is a terminology used to refer to a family of models, datasets, and algorithmic frameworks dedicated to emergency response, remote sensing-based disaster damage assessment, multi-agent rescue resource optimization, and clinical decision support in rescue operations. Its usage encompasses heuristic reinforcement learning agents, decision-theoretic dispatch systems, end-to-end neural architectures for damage segmentation and classification, multi-objective resource allocation, semantic feature extraction for UAV data transmission, and robust health management software. The diversity of applications and technical contributions is reflected in the distinct problem domains covered in the literature.

1. Heuristic Multi-Agent Coordination for Disaster Response

Early work in the RescueNet lineage involves heuristic multi-agent reinforcement learning (MARL) for the dynamic coordination of volunteers in disaster scenarios (Nguyen et al., 2018). Here, agents represent volunteers acting within a grid-based discretization of a real-world disaster zone, abstracted as a Markov game with state, action, transition, and reward elements:

$G = \langle N, S, A, P, R, \gamma \rangle$

where $N$ is the number of agents, $S$ the state space representing spatial distributions, $A$ the allowed moves, $P$ the transition kernel, $R$ the per-agent reward, and $\gamma$ the discount factor. A Q-learning update rule augmented with a heuristic prioritization function

$$Q(s, a) \leftarrow Q(s, a) + \alpha [r + \gamma \max_{a'} Q(s', a') - Q(s, a)]$$

guides agents toward minimizing Manhattan distances between volunteers and victims:

$h = |r_a - r_v| + |c_a - c_v|$

This reduces the search space and accelerates convergence. Social network data—most notably Twitter posts—are mined and classified (using SVMs) to identify victim and volunteer coordinates, which are then mapped onto the grid for agent scheduling. The rescue task is formalized as a binary allocation matrix $X$, optimizing the assignment cost:

$\min C = \sum_{t} \sum_{i,j} d_{ij} x_{ij}$

Performance metrics include average episode time, reward rate, and rescuing cost, with the ResQ algorithm exhibiting superior efficiency (reward rate of 37.9, rescuing cost of 0.026). This tightly-coupled data-driven MARL approach provides real-time coordination primitives for large-scale disaster rescue.

2. Online Decision-Theoretic Pipeline for Emergency Dispatch

RescueNet methodologies further extend into the domain of online emergency responder dispatch, underpinning their agent decision frameworks with a Semi-Markov Decision Process (SMDP) paradigm (Mukhopadhyay et al., 2019). States encompass incident queues, responder positions, and environmental vectors (including traffic conditions):

$S = \{ I^t, R^t, E^t \}$

Actions $A$ consist of dispatch and idling choices, with transitions $p_{ij}(a)$ and response delay costs $\rho(i, a)$. To update incident predictions in real time, an online log-linear survival analysis mechanism is used:

$$\log(\tau_i) = \beta_1 w_{i1} + \dots + \beta_m w_{im} + y$$

where $\beta$ parameters are adapted via stochastic gradient ascent:

$$\beta^{(p+1)} = \beta^{(p)} + \alpha \nabla L(\beta^{(p)}, D')$$

Responders’ route-time predictions are enhanced via LSTM-based traffic speed forecasting and A* Search with ALT. Instead of full policy computation, localized Monte Carlo Tree Search (MCTS) simulates incident chains with candidate actions filtered by cost thresholds within a stochastic horizon. Empirical results show response times and system latencies drop significantly (0.034–0.384 sec per decision), demonstrating high scalability and real-time tractability.

3. Remote Sensing Damage Segmentation and Assessment

RescueNet appears prominently as an end-to-end neural architecture for building segmentation and damage level assessment from satellite (xBD dataset) and UAV imagery (Gupta et al., 2020, Rahnemoonfar et al., 2022). The core technical structure is as follows:

Architecture:

• ResNet50 backbone with atrous convolutions and an Atrous Spatial Pyramid Pooling (ASPP) module for multi-scale encoding
• Dual segmentation heads (simple upsampling or encoder–decoder with skip connections)
• Change detection head operating on the difference of post- and pre-disaster features

Loss Function:

The localization aware loss consists of binary cross-entropy for segmentation and a selective, foreground-only categorical cross-entropy for damage classification:

$$L(y_i, \hat{y}_i) = \begin{cases} -\log(\hat{y}_{il}) + \sum_{k \in C} [-\log(\hat{y}_{ik})], & y_{il} = 1 \ -\log(1 - \hat{y}_{il}), & y_{il} = 0 \end{cases}$$

Optionally supplemented by Dice loss for boundary accuracy.

Dataset Innovations:

RescueNet (UAV version) offers multi-class, pixel-level segmentation masks (water, buildings with 4 damage levels, blocked/clear roads, vehicles, trees), facilitating transfer learning and multi-task scene understanding. Annotation via V7 Darwin ensures rigorous quality control.

Model Benchmarks and Metrics:

Evaluations employ F1 score for localization and harmonic means of damage class F1s (xview2 challenge metric). Attention U-Net and PSPNet architectures exhibit top per-class IoU scores (frequently >98%). Notably, transformer-based Segmenter models yield high performance utilizing patch-level embeddings.

4. Semantic Feature Extraction for Efficient UAV Data Transmission

RescueNet has been leveraged as a dataset for benchmarking semantic extraction and efficient data transmission in UAV-assisted disaster damage assessment (Hewawiththi et al., 14 Dec 2024). The approach uses a two-stage process:

• Stage 1: PSPNet with ResNet-50 backbone produces semantic segmentation masks $M = ResNet(I)$.
• Stage 2: Lightweight FCN generates a binary mask $B$, trained end-to-end with Gumbel-Softmax activation for differentiable selection of critical regions.
• Final representation: $y = M \odot B$ is transmitted, reducing size by up to 92% (from 177.78 kB to 13.728 kB), with only slight increases in classification error (30.00% to 41.33%). High Jaccard index (~0.993) ensures semantic fidelity.

Loss is a weighted combination:

$$\text{Loss} = w_s \cdot L_\text{sparsity} + w_c \cdot L_\text{categorical}$$

This methodology prioritizes efficient, actionable data transmission under bandwidth constraints for rapid disaster scene assessment.

5. Multi-Objective Resource Allocation in IoT-Enhanced Rescue Logistics

Recent advances introduce highly optimized meta-heuristic frameworks for emergency resource allocation in IoT environments (Xu et al., 15 Mar 2024). The Multi-Objective Shuffled Gray-Wolf Frog Leaping Model (MSGW-FLM) hybridizes Grey Wolf Optimization and Shuffled Frog Leaping Algorithm, augmented with NSGA-II multi-objective ranking. Novel mechanisms are:

• Population memeplex splitting and intra-memeplex evolution
• Quality Rank (QR) assignment using Pareto frontiers and crowding distance
• Crossover and Levy flight operators for diversity and local optima escape

Operating in a multi-cycle, rolling planning window with real-time IoT data streams, MSGW-FLM adapts allocations to pedestrian distributions and supply points. Comparative analyses demonstrate consistently superior hypervolume and diversity metrics over established benchmarks like NSGA-II, IBEA, MOEA/D across 28 multi-objective problems. Pseudocode executes parameter initialization, memetic evolution, archive update, and termination described in Algorithm 3 of the referenced paper.

6. Multi-label Damage Classification and Attention Mechanisms

Damage assessment in hurricane contexts further leverages RescueNet as a dataset for multi-label classification (Liu et al., 3 Jul 2025). A ResNet-based feature extractor ($X = f(I)$) is combined with a class-specific residual attention (CSRA) module:

$F_i = \text{GAP}(X) + A_i \odot X$

where $A_i$ is the spatial attention tensor for class $i$, and $\odot$ denotes elementwise multiplication. This mechanism supports simultaneous, independent classification of multiple overlapping damage types.

Performance is quantified via mean average precision (mAP, 90.23% for ResNet152+CSRA), overall precision, recall, and F1. Tested on 4,494 aerial images with broad damage annotation, the attention-enhanced framework consistently outperforms baselines (VGG19, EfficientNet, ViT-B16+CSRA).

7. AI-Driven Rescue Health Management

In medical emergency scenarios, RescueNet materializes as an AI-powered clinical decision support system for health status assessment (Ahammed et al., 7 Aug 2024). The architecture integrates:

• Pre-trained ML (Extreme Gradient Boosting, SVM, KNN, Random Forest, ANN, Logistic Regression, Naïve Bayes)
• Feature selection via recursive feature elimination with cross-validation (RFECV)
• Hyperparameter optimization (Grid/Random Search, Hyperband, Bayesian Optimization)
• Extensive historical rescue data (273,000 events, 452 attributes per case)

Probabilities for six major complications (e.g., cardiovascular, respiratory, neurological) are provided:

$P(\text{complication}) = f(x; \theta)$

where $x$ is a health feature vector. Python-based modular software is optimized for usability, pressure environments, and robust processing. Real use cases confirm accurate discrimination and probability estimation, critical for time-sensitive emergency treatment.

8. Lightweight Semantic Segmentation Decoders for Remote Sensing

Innovations in decoder architecture for real-time semantic segmentation have also been validated on RescueNet (Chen et al., 15 Apr 2025). LightFormer employs:

• Cross-scale Feature Fusion Module (CFFM) with soft adaptive weighting:

$$O = \frac{e^\alpha X + e^\beta Y}{e^\alpha + e^\beta}$$

• Lightweight Channel Refinement Module (LCRM): channel-split into self-attention and convolution branches; reduces parameters and FLOPs by up to 75%
• Spatial Information Selection Module (SISM): parallel paths with large/small receptive fields, channelwise pooling and modulation:

$O = X + \alpha X_s + \beta X'$

In RescueNet UAV inference (EfficientNet-B3 backbone, 1024×1024 window, stride 128), LightFormer achieves mean Intersection over Union (mIoU) improvements up to 1.2–1.4% over hybrid CNN–Transformer decoders, while using only 4.6–15.9% of their FLOPs and parameters. Its ability to capture both global context and fine object details facilitates accurate disaster assessment under strict computational budgets.

---

RescueNet Model constitutes a collection of advanced technical solutions for disaster response, encompassing agent coordination, decision-theoretic pipelines, neural architectures for segmentation and multi-label assessment, IoT-driven rescue logistics, semantic data extraction, and intelligent health management. Its flexible, modular designs and dataset innovations have enabled rapid progress in real-time, scalable, and high-precision disaster scene understanding and resource allocation, contributing significantly to contemporary humanitarian aid and emergency response research.