Search-and-Rescue robotic operations require trustworthy AI systems where severe overconfidence (D-ECE > 0.9) could compromise life-critical decisions. We present a multi-domain calibration framework demonstrating improvements across synthetic, simulated, and real SAR domains. Our approach uses heuristic-based ground truth generation, enabling calibration assessment without expensive manual annotation. Crucially, we reveal that explainability method choice directly impacts calibration quality-LayerCAM achieves optimal performance with superior sparsity (0.044 ± 0.029) and calibration (D-ECE: 0.136) by creating focused attention maps that enable reliable confidence assessment. Diferent CAM methods produce distinct attention regions, which afect how calibration is computed and validated, making joint optimization essential for safety-critical deployment. The framework provides foundations for EU AI Act Article 13 transparency requirements while acknowledging the need for expanded validation before operational use.
Model overconfidence occurs when neural networks assign high probability scores to incorrect
predictions [
Nixon et al. established D-ECE as the standard detection calibration metric [
SAR applications face unique explainability challenges where operators must quickly understand
AI recommendations under time pressure [
This paper introduces four contributions addressing these human-centered trustworthiness challenges: 1. Heuristic-Based Ground Truth Generation: A filename -based heuristic procedure to supply calibration labels at scale, removing the need for manual annotation in this study. 2. Multi-Domain Calibration Framework: Novel approach with validation across synthetic (D_LLM), simulated (D_SIM), and real (D_REAL) domains. 3. Cross-Domain Calibration Analysis: Empirical evidence that calibration improvements are achievable across diferent data collection methodologies. 4. Transparency-Enhanced Implementation: Technical framework addressing EU AI Act Article 13 transparency requirements.
Our evaluation demonstrates calibration improvements across domains while acknowledging expanded validation needs before safety-critical deployment. Critically, explainability and calibration are interdependent—diferent CAM methods create distinct attention regions that directly influence calibration computation, making joint optimization essential.
Throughout this manuscript, domain refers to data-source modality (D_LLM, D_SIM, D_REAL) unless referring to broader application contexts.
Table 1 summarizes our contributions and their assessed novelty levels relative to existing work.
CAM techniques provide visual explanations by highlighting image regions that contribute most to CNN predictions, essential for understanding AI decisions in safety-critical SAR operations.
For Grad-CAM [
LayerCAM [
1 ∑,︁ , , = ReLU ︃( )︃
, =1 =1 which tends to yield finer localization—useful in SAR scenes with small, critical targets.
EigenCAM applies Principal Component Analysis (PCA) to activation maps but often produces fragmented attention patterns that can be dificult for human operators to interpret in time-critical SAR scenarios.
Model calibration measures the alignment between predicted confidence and actual accuracy—a critical safety requirement in life-critical operations.
Expected Calibration Error (ECE) [
ECE =
∑︁ || ⃒⃒ acc() − conf( )⃒⃒ .
Detection Expected Calibration Error (D-ECE) [
D−ECE =
∑︁ || ⃒⃒ prec() − conf( )⃒⃒ .
Interpretation Thresholds: D-ECE < 0.15 indicates well-calibrated systems suitable for operational
deployment, while D-ECE > 0.9 represents dangerous overconfidence requiring immediate attention
before safety-critical use [
We introduce a heuristic-based approach to calibration ground truth generation, addressing SAR data scarcity through filename-based pattern recognition for systematic calibration labels.
SAR Data Scarcity: Collecting diverse SAR training datasets is problematic due to high costs, safety risks, environmental variability, and annotation challenges in adverse conditions, hindering model generalization.
Synthetic Video Solution: Our framework leverages Sora, difusion-based model producing up to
60-second realistic videos with complex multi-entity scenarios and cinematic quality suitable for SAR
(1)
(2)
(3)
(4)
training. It also employs DeepSeek Janus-Pro (7B parameter transformer), for high-quality
frame-byframe generation, ofering flexible integration into custom pipelines. Finally, Gemini Pro + Veo allows
us to generate 8-second clips with synchronized audio, enabling simulation of radio communications,
victim calls, and environmental sounds [
Mission-Critical Scenarios: We generate diverse SAR scenarios including rugged terrain traversal, victim detection (thermal views, burial states), sensor degradation (smoke, dust, weather), and varying environmental conditions (day/night, indoor/outdoor).
We implement domain-specific calibration ground truth generation using systematic filename pattern analysis as detailed in Algorithm 1. ◁ Special images assumed positive ◁ Non-standard files assumed positive ◁ General model
Rationale for Deterministic Labeling: Traditional calibration approaches require extensive manually-labeled validation sets, which are prohibitively expensive and time-consuming for SAR applications. Our heuristic labeling provides a systematic alternative for calibration assessment when manual annotation is infeasible.
1. Core Purpose: Generate balanced positive/negative samples for computing calibration metrics without expensive manual labeling. These labels serve as proxies for actual detection outcomes, enabling systematic calibration assessment across thousands of images that would otherwise require expert annotation. 2. Scientific Methodology: The filename patterns in our datasets encode temporal and spatial information that correlates with real SAR search patterns. 3. Calibration Application: These heuristic labels enable: (a) D-ECE computation by providing accuracy baselines for confidence comparison, (b) temperature scaling parameter optimization through gradient-based methods, and (c) cross-domain consistency analysis across D_LLM, D_SIM, and D_REAL environments.
Limitation Acknowledgment: This heuristic approach represents a practical compromise between annotation cost and calibration assessment needs. Future work should validate these findings with larger expert-annotated datasets and explore LLM-based label generation methods.
Validation: Maritime SAR studies show 218% improvement in mean Average Precision when
synthetic data augments real datasets [
Our framework addresses SAR multi-domain operations with human operator understanding.
Domain Architecture and Data Collection:
Our multi-domain approach reflects the realistic deployment pathway for SAR AI systems, progressing from synthetic training data through simulation validation to real-world application. • D_LLM Domain (Synthetic): Synthetic frame sequences (≈ 163 samples per CAM method, ≈ 1,141 total). These are generated using state-of-the-art AI systems: Sora for 60-second realistic video sequences, DeepSeek Janus-Pro (7B parameter transformer) for detailed scene understanding, and Gemini Pro+Veo for 8-second clips. Generated scenarios include collapsed buildings with realistic debris patterns, challenging terrain navigation, and atmospheric efects (smoke, dust, varying weather conditions). • D_SIM Domain (Simulated): Simulated environments (≈ 35 samples per CAM method, ≈ 245 total) from physics-based simulation environments using Unity3D game engine and NVIDIA Isaac Sim platform. These platforms provide realistic rubble dynamics, accurate thermal signature simulation, and particle efects for dust and debris. • D_REAL Domain (Real-world): Real-world SAR operations (≈ 73 samples per CAM method, ≈ 511 total)
Uneven sample distribution reflects real-world SAR data scarcity, requiring synthetic augmentation for safety-critical deployment.
Human-Centered Calibration Process: For each domain, we implement temperature scaling with human oversight:
^ = ( / )
where is the domain-specific temperature parameter optimized on synthetic ground truth. We
evaluate calibration quality using D-ECE [
A critical finding of our research is that explainability and calibration are not independent concerns—the choice of CAM method fundamentally afects how well-calibrated the resulting confidence estimates become.
Mathematical Foundation: Diferent CAM methods alter calibration computation through their spatial attention distribution patterns. For a given CAM method producing attention map , the calibration-weighted confidence becomes: ^ = ∑︀, (, ) · (, )
∑︀, (, ) ∑︀, I[, > ] = ∑︀, I[, > 0] (5) (6) (7) where (, ) represents the pixel-wise prediction confidence. This means that the spatial distribution of attention directly influences the final confidence estimate used for calibration assessment.
Sparsity-Calibration Relationship: We hypothesize the explanation sparsity correlates with calibration quality: Lower sparsity indicates concentrated attention, enabling reliable confidence assessment and better human interpretation.
Comparative Method Analysis: • LayerCAM: Typically produces more focused attention; see Table 3. • GradCAM: Often yields more difuse attention; see Table 3.
• EigenCAM: Can produce fragmented attention patterns; see Table 3.
Practical Impact: Diferences in explanation focus can materially afect how closely confidence aligns with accuracy, underscoring that explanation choice and calibration should be considered jointly in safety-critical SAR scenarios.
Our empirical evaluation (Table 2) provides initial evidence that calibration improvements can be
achieved across diferent data collection methodologies. Moreover, these consistent improvements
across diferent data collection methodologies suggest that systematic calibration approaches may
be transferable. Our results achieve the established benchmark threshold of D-ECE < 0.15 across all
domains [
LayerCAM emerges as the optimal method for SAR robotics applications, achieving superior performance in both explanation focus (sparsity: 0.044 ± 0.029) and calibration quality (D-ECE: 0.136), outperforming gradient-based methods like GradCAM (sparsity: 0.324 ± 0.121) and eigenspace approaches. This finding suggests that methods with architectural advantages in spatial attention (LayerCAM’s layer-specific focus) may achieve better calibration across diferent data collection approaches, though expanded validation is needed.
The quantitative results presented in Tables 2 and 3 are corroborated by visual evidence demonstrating that our calibration framework preserves spatial attention quality while dramatically improving confidence reliability. Figures 1 and 2 illustrate how the 84%+ calibration improvement (D-ECE reduction from 0.927 to 0.136) maintains operational efectiveness for human-AI collaboration in SAR scenarios.
Figure 1 demonstrates a critical finding. Calibration enhancement transforms dangerous overconfidence without degrading the spatial attention patterns essential for human operator decision-making. The uncalibrated attention map exhibits severe overconfidence (D-ECE: 0.95) that could lead to false security in life-critical situations, while the calibrated version achieves appropriate confidence levels (DECE: 0.15) while preserving identical target localization accuracy. This validates our model-dominance hypothesis—architectural calibration solutions can address overconfidence while maintaining the spatial intelligence that makes these systems operationally valuable.
Having demonstrated calibration improvements and visual preservation of spatial attention quality, we now present transparency framework capabilities against regulatory requirements. Our approach provides foundations for regulatory compliance through systematic transparency measures (Table 4).
The visual validation evidence (Figures 1 and 2) supports transparency requirements by demonstrating interpretable calibration processes, while the quantitative metrics provide measurable trustworthiness characteristics aligned with EU AI Act Article 13 and NIST AI-RMF frameworks.
Our multi-domain monitoring capabilities (D_LLM, D_REAL, D_SIM) with quantitative calibration metrics (85% D-ECE improvement) demonstrate substantial progress beyond foundational concepts toward operational monitoring systems. However, operational deployment requires comprehensive regulatory assessment, expanded validation datasets, and formal compliance certification.
Our multi-domain calibration framework demonstrates systematic improvements across synthetic, simulated, and real SAR domains, with LayerCAM achieving optimal performance (sparsity: 0.044±0.029, D-ECE: 0.136) by transforming dangerous overconfidence (D-ECE > 0.9) into well-calibrated predictions (D-ECE < 0.15). The heuristic-based approach provides O(1) complexity ground truth generation while addressing EU AI Act transparency requirements.
Deployment and Limitations: The framework supports progressive deployment requiring systematic field validation and regulatory assessment. Current limitations include reliance on heuristic labeling, which we only qualitatively verified against expert annotations—expanded validation is needed for deployment. Uneven sample distribution (D_LLM: 1,141; D_REAL: 511; D_SIM: 245) limits statistical power, necessitating domain-specific validation. The framework shows broader applicability for medical imaging, autonomous systems, and industrial inspection.
Impact and Future Work: The 85.3% calibration improvement across domains ofers clear practitioner guidance: choose LayerCAM for focused, well-calibrated explanations. Priority research areas include dynamic calibration, multi-modal integration, and comprehensive safety assessment. Our framework represents a foundational step requiring continued validation before operational deployment. As AI increasingly supports life-critical decisions, appropriate confidence calibration becomes both a technical and ethical imperative.
This work was supported by the project "Explainable Trustworthy AI for Data Intensive Applications" (EXTRA - BRAIN), Grant no. 101135809 - HORIZON-CL4-2023-HUMAN-01-CNECT.
During the preparation of this work, the author(s) used OpenAI GPT-4 and Claude Sonnet 4 for grammar and spelling check; formatting assistance (LaTeX error correction).