WISE: A Lightweight, Weakly-Supervised Model for Onboard Fire Smoke Detection and Localization
Authors: Sha Lu, Yu Sun, Liang Zhao, Jixue Liu, Lin Liu, Jiuyong Li, A. K. Qin, Alejandro Mousist, Stefan Peters
Organizations: Adelaide University, Adelaide, SA 5000, Australia · Swinburne University of Technology, Hawthorn, VIC 3122, Australia · Thales Alenia Space, Cannes, 06150, France
Wildfire smoke detection from satellite imagery is critical for early warning and rapid response. For onboard satellite deployment, detection systems must operate under strict memory and latency constraints while providing spatially informative outputs for downstream decision-making. Existing tile-level classification methods are computationally efficient but lack spatial localization, whereas pixel-level segmentation approaches provide detailed masks yet are typically too computationally demanding for real-time onboard execution. To address this gap, we propose WISE (Weakly-supervised Inference-efficient Smoke Extraction), a deployment-oriented framework for onboard fire smoke detection and localization. WISE leverages only tile-level annotations through a teacher-student distillation strategy, where an offline teacher provides soft spatial supervision to a lightweight WISE-Student optimized for efficient onboard inference. The student jointly predicts tile-level smoke presence and smoke probability maps within a single forward pass, enabling spatially informative detection under strict computational constraints. WISE was evaluated through in-orbit execution aboard the ISS-mounted IMAGIN-e payload. Three model variants achieve average inference times of 0.10 s, 0.14 s, and 0.26 s per tile, indicating near-real-time per-tile inference within onboard resource limits. Ground-based experiments on Landsat 5 and Landsat 8 imagery further indicate effective detection and spatially informative localization. The best-performing variant achieves a mean tile-level F1 score of 0.964 and a mean pixel-level F1 score of 0.750 across 10 runs, while containing only 0.12M parameters and requiring approximately 3 GFLOPs. Together, these results indicate that WISE is a practical candidate for low-latency wildfire smoke monitoring from space under onboard resource constraints.
Figures & tables
Figure 1: Overall WISE pipeline. A teacher model generates Grad-CAM attention maps, which guide training of the lightweight WISE-Student for joint smoke detection and localization during onboard inference.
Figure 2: Multi-layer Grad-CAM activation maps produced by the teacher model trained with tile-level supervision. Grad-CAM maps from Layers 2–4 and their fused result illustrate the emergence of spatial attention from classification-only training.
Figure 3: Architecture of the WISE-Student model. The network jointly produces tile-level smoke prediction and a high-resolution smoke probability map in a single forward pass, with dashed components used only during training.
Model
ResNet-50
WISE-T
WISE-S
WISE-M
WISE-L
Parameters (M)
24.57
0.008
0.030
0.066
0.116
FLOPs (G)
5.566
0.228
0.810
1.748
3.042
Table 1: Comparison of parameter count and FLOPs for the ResNet-50 teacher and WISE model variants.
Metric
WISE-T
WISE-S
WISE-M
WISE-L
ResNet-50
Accuracy
0.840±0.069 (0.910)
0.959±0.014 (0.984)
0.975±0.009 (0.992)
0.976±0.006 (0.986)
0.986
F1 Score
0.767±0.081 (0.865)
0.939±0.021 (0.976)
0.962±0.014 (0.988)
0.964±0.009 (0.980)
0.979
Precision
0.776±0.105 (0.869)
0.939±0.025 (0.974)
0.966±0.014 (0.984)
0.972±0.015 (0.992)
1.000
Recall
0.776±0.115 (0.951)
0.938±0.029 (0.992)
0.959±0.020 (0.992)
0.957±0.017 (0.976)
0.959
FNR
0.224±0.115 (0.049)
0.062±0.029 (0.008)
0.041±0.020 (0.008)
0.043±0.017 (0.024)
0.041
Table 2: Classification performance of WISE variants and the teacher ResNet-50 model. Each cell reports the mean ± standard deviation across 10 runs, and the best result achieved in any run is shown in parentheses. For WISE variants, the best average result for each metric is highlighted in bold.
Model
Params (M)
FLOPs (G)
Accuracy
F1
FNR
CloudScout
1.263
3.533
0.985
0.978
0.023
MobileNetV2
2.882
0.450
0.964
0.944
0.074
PhiFire
26.094
0.913
0.926
0.888
0.120
VIB-SD
0.550
0.651
0.983
0.975
0.031
WISE-M (Ours)
0.066
1.748
0.975
0.962
0.041
Table 3: Tile-level classification performance and efficiency comparison between WISE-M and lightweight baseline models. Classification results are reported as mean values across 10 runs. The best performance among the baseline models is highlighted in bold.
Metric
WISE-T
WISE-S
WISE-M
WISE-L
ResNet-50
Accuracy
0.743±0.10 (0.875)
0.828±0.03 (0.878)
0.851±0.03 (0.894)
0.866±0.04 ( 0.915 )
0.778
F1 Score
0.601±0.07 (0.707)
0.697±0.03 (0.754)
0.730±0.03 (0.778)
0.750±0.04 ( 0.806 )
0.605
IoU
0.461±0.07 (0.593)
0.563±0.03 (0.620)
0.600±0.04 (0.655)
0.627±0.04 ( 0.692 )
0.463
Precision
0.569±0.09 (0.707)
0.676±0.04 (0.739)
0.714±0.04 (0.767)
0.733±0.04 ( 0.789 )
0.514
Recall
0.759±0.08 (0.815)
0.777±0.04 (0.828)
0.795±0.04 (0.838)
0.808±0.03 ( 0.852 )
0.856
Table 4: Pixel-level performance of WISE variants and the ResNet Grad-CAM baseline. Each cell reports the mean ± standard deviation across 10 runs, and the best result achieved in any run is shown in parentheses. The best result for each metric is highlighted in bold.
Metric
VTrUNet
SUNet
FrizCNN
WISE-L (Best Run)
F1
0.826
0.768
0.691
0.806
Precision
0.833
0.739
0.666
0.789
Recall
0.889
0.830
0.770
0.852
Table 5: Pixel-level performance comparison between WISE-L and supervised segmentation models. Best results for each model are reported; the WISE-L value is the best result from a single run in Table 4 , whose 10-run mean F1 is 0.750.
Figure 4: Qualitative results of WISE-L on ten representative examples (A–J). Each example shows the RGB input, the teacher Grad-CAM attention map, and the predicted smoke probability map, with red polygons indicating manually annotated smoke regions.
Model
Grad-CAM Supervision
Tile-Level
Pixel-Level
F1
FNR
F1
WISE-T
With
0.767
0.224
0.601
Without
0.860
0.143
0.406
WISE-S
With
0.939
0.062
0.697
Without
0.910
0.107
0.404
WISE-M
With
0.962
0.041
0.730
Table 6: Impact of Grad-CAM supervision across WISE variants. Results are averaged over 10 runs. Pixel-level evaluation is reported only in terms of F1, as FNR is not commonly used in smoke detection due to the inherent difficulty of exhaustively labeling all smoke pixels.
Model
Batch Size
Overlap
#Tiles
Inference Time (s/tile)
Memory (Peak, MB)
Model Size (MB)
WISE-S
4
50
400
0.10
903
0.14
WISE-M
8
25
187
0.14
767
0.28
WISE-L
4
0
117
0.26
582
0.47
Table 7: Runtime performance of WISE models on the IMAGIN-e payload.
Table A.1: Layer-wise configuration of the WISE model with width multiplier α=0.75 . The model uses a 6-channel 256×256 input and outputs both a tile-level class prediction and a smoke probability map.
We present a deployed system for on-orbit wildfire detection aboard a nine-satellite commercial thermal infrared constellation, operating under demanding joint constraints: sub-megabyte model footprint, sub-150 ms per-batch TensorRT FP16 inference on an NVIDIA Jetson Xavier NX, and an end-to-end alert pipeline targeting under 10 minutes from satellite overpass to fire event communication. The system operates on uncalibrated mid-wave infrared (MWIR) single-band imagery at 200 m ground sampling distance, where fires frequently appear as sub-pixel or single-pixel thermal anomalies under extreme class imbalance -- challenges not addressed by the contextual thermal-thresholding pipelines (MODIS, VIIRS) that currently dominate operational fire monitoring. We present an empirical study of lightweight dense representation learning for this regime using a proprietary nine-satellite MWIR dataset. We compare dense masked autoencoding (DenseMAE) and a hybrid DenseMAE+EMA (exponential moving average) distillation variant, and evaluate representations via linear probing and full-distribution pixel-level average precision (AP) under extreme class imbalance. DenseMAE pretraining enables compact downstream models on the latency-accuracy Pareto frontier: our fastest SSL-pretrained model achieves 0.640 test AP and 0.69 event-level Fire-F1 with 65.34 ms latency per batch and a 0.52 MB engine, without pruning or compression. The best configuration reaches 0.699 AP and 0.744 Fire-F1 below 1 MB, outperforming a supervised baseline (0.650 AP) under comparable constraints.
Matthias Rötzer, Veronika Pörtge, Martin Ickerott +6
Geostationary satellite observations are important for wildfire detection and monitoring. The current study evaluates machine learning models for MTG FCI near-real-time fire detection in 1- and 2-km spatial configurations and compares them with threshold-based algorithms. The models were trained and evaluated using VIIRS fire reference data across diverse ecological regions in Europe, Africa, and the Middle East. The key results are that 1-km models significantly outperform both their 2-km variants and operational threshold products. The constructed 1-km models achieved F1 scores higher by up to 0.36 compared to baseline products. Importantly, the 1-km models detected small fires with higher probability compared to competing models. Finally, the models robustly detected fires up to 260 min earlier than baseline products. To support opensource applications, our trained models are publicly available.
Rapid and accurate wildfire smoke severity assessment from satellite images is essential for emergency response, air quality modeling, and human health risk management. Existing deep learning approaches treat smoke detection as a binary task, producing point estimates without any measure of prediction confidence. We propose a probabilistic framework to categorize a satellite patch into Light, Moderate, and Heavy severity classes and to provide decomposed epistemic and aleatoric uncertainty in a single forward pass. Our architecture uses the backbone of a pre-trained EfficientNet-B3 and a CBAM module with an evidential deep learning head that predicts Dirichlet concentration parameters, directly estimating vacuity (epistemic) and dissonance (aleatoric) without Monte Carlo sampling. Evaluated on 16,298 real satellite patches derived from the Wildfire Detection dataset, our model achieves 93.8% weighted test accuracy (91.1% unweighted) with ECE=0.0274. Selective prediction retaining the most certain 50% of patches achieves 96.7% accuracy. As image quality degrades, uncertainty increases monotonically, and vacuity is a practical scan quality measure. The Moderate class represents transitional smoke conditions that exhibit the highest epistemic uncertainty (mean vacuity = 0.187), confirming the model correctly identifies ambiguous smoke boundary regions. CBAM spatial attention maps localize to structurally distinctive scene regions, and t-SNE demonstrates the clear cluster separation of Light and Heavy smoke.