On-Orbit Real-Time Wildfire Detection Under On-Board Constraints

Technical Summary: On-Orbit Real-Time Wildfire Detection Under On-Board Constraints

Problem Statement
The paper addresses the critical challenge of detecting wildfires from space under extreme operational constraints. While satellite-based monitoring is essential for global fire management, existing architectures face fundamental trade-offs: geostationary platforms offer high temporal revisit but coarse spatial resolution (2–3 km), missing small fires; Low Earth Orbit (LEO) sensors offer finer resolution but limited daily overpasses, often missing peak fire activity windows. OroraTech's OTC-P1 mission attempts to bridge this gap with a nine-satellite thermal infrared constellation, but it imposes a unique set of joint constraints that current state-of-the-art models cannot satisfy:

1. Strict Latency: An end-to-end alert pipeline must operate in under 10 minutes from satellite overpass to communication, necessitating on-board inference rather than downlink-and-process architectures.
2. Resource Constraints: Inference must run on an NVIDIA Jetson Xavier NX (10 W mode) with a model footprint under 1 MB and inference latency under 150 ms per batch.
3. Data Characteristics: The system operates on uncalibrated, single-band Mid-Wave Infrared (MWIR) imagery at 200 m ground sampling distance. Fires often appear as sub-pixel or single-pixel thermal anomalies, creating extreme class imbalance (fire pixels constitute ~0.018% of valid pixels).
4. Noise: The pipeline uses raw digital numbers (DN) without radiometric calibration to minimize latency, increasing susceptibility to sensor noise and artifacts.

Current operational pipelines (e.g., MODIS, VIIRS) rely on heuristic, contextual thermal-thresholding methods that struggle with these specific constraints, particularly the sub-pixel nature of early fires and the lack of multi-spectral calibration.

Methodology
The authors propose a deployment-oriented deep learning pipeline centered on DenseMAE (Dense Masked Autoencoder), a lightweight self-supervised learning approach tailored for dense segmentation on uncalibrated MWIR imagery.

• Architecture: DenseMAE is a fully convolutional masked autoencoder. Unlike standard Vision Transformers (ViT) that tokenize images, DenseMAE uses a staged convolutional encoder with a single early downsampling step (to $H/2 \times W/2$) followed by dilated convolutional blocks to increase the receptive field efficiently. It produces dense per-pixel embeddings rather than global image tokens.
• Pretraining Objective: The model is pretrained using a masked reconstruction objective. Random blocks of the input are masked, and the encoder-decoder pair learns to reconstruct the missing thermal patterns from the surrounding context. The loss is computed only on masked, valid pixels, explicitly excluding "no-data" regions.
• Hybrid Ablation: The authors investigated a hybrid variant combining masked reconstruction with an Exponential Moving Average (EMA) teacher-student consistency loss. However, empirical results showed this provided only marginal and inconsistent gains, leading the authors to favor the reconstruction-only approach.
• Downstream Transfer: After pretraining, the decoder is discarded. The encoder is transferred to the wildfire segmentation task using lightweight heads. Two transfer strategies were evaluated:
  1. Frozen Encoder: Training only a lightweight head (linear probe or small context head) on top of fixed embeddings.
  2. Full Fine-tuning: Jointly optimizing the encoder and head end-to-end.
• Refinement: To address the sub-pixel nature of fires, a high-resolution (HR) refinement head is employed. This module re-injects high-resolution features from the encoder stem into the final prediction stage to improve the localization of single-pixel anomalies.
• Training Strategy: To handle extreme class imbalance, training batches are biased toward positive patches. The system uses binary cross-entropy loss, AdamW optimizer, and mixed precision.

Key Contributions

1. Deployment Pipeline: Introduction of a deep-learning pipeline for on-orbit wildfire detection that operates on uncalibrated, single-band MWIR imagery under strict latency (<150 ms) and footprint (<1 MB) constraints.
2. DenseMAE: Proposal of a lightweight, staged convolutional masked autoencoder that learns dense MWIR representations, enabling sub-megabyte deployment models without pruning or compression.
3. Empirical Study: A systematic evaluation of self-supervised pretraining (MAE vs. hybrid EMA) and transfer learning (frozen vs. full fine-tuning) under extreme class imbalance, using pixel-level Average Precision (AP) on the full test distribution.
4. Pareto Frontier Achievement: Demonstration that dense self-supervised representations enable compact models that sit on the latency–accuracy Pareto frontier, outperforming supervised baselines under comparable constraints.

Results
The system was evaluated on a dataset of 839 annotated scenes from OroraTech's OTC-P1 mission, split into 70/15/15% train/val/test.

• Performance vs. Baseline: The best-performing DenseMAE configuration (embedding dimension 64, full fine-tuning, DW-Res head + HR refinement) achieved a 0.699 Test AP and 0.744 Fire-F1. This outperforms the strong supervised baseline (U-Net++ with HR refinement), which achieved 0.650 AP under similar constraints.
• Latency and Footprint:
  • The fastest configuration (embedding dimension 32, TRT-Head) achieved 0.640 AP with 65.34 ms latency (TensorRT FP16, batch size 8) and a model engine size of 0.52 MB.
  • The best accuracy configuration remained under 1 MB (0.91 MB) with 141 ms latency.
• Ablation Findings:
  • Increasing embedding dimensionality from 32 to 64 consistently improved AP.
  • Full fine-tuning consistently outperformed frozen-encoder transfer (0.689 vs. 0.612 AP for emb64), indicating that SSL initialization is most beneficial when the downstream model can adapt end-to-end.
  • Hybrid EMA distillation did not provide consistent full-distribution gains over reconstruction-only pretraining.
• External Validation: The model was compared against the VIIRS 375 m active-fire product (a standard operational reference) on coincident observations. The study found strong cross-sensor agreement, with the model detecting 103 joint fire clusters. The model identified 33 clusters missed by VIIRS (often weak thermal signals near the detection limit) and 3 clusters detected only by the model, suggesting the ability to find small or short-lived fires below VIIRS thresholds.

Significance and Claims
The paper claims that lightweight self-supervised representations enable accurate and efficient on-orbit fire detection where traditional methods fail due to hardware and data constraints. By achieving sub-megabyte model sizes and sub-100 ms inference latencies on edge hardware, the system facilitates an end-to-end alert pipeline capable of detecting ignitions within minutes of satellite overpass. This represents a significant step toward reducing the gap between fire onset and operational response. The authors modestly note that while the system improves robustness to common sensor artifacts, confusion with sunglint remains a challenge, and future work may require multi-channel inputs or physics-informed features to address these edge cases. The study validates that dense SSL pretraining is a viable and superior strategy for extreme class imbalance and resource-constrained remote sensing tasks.