Can machine learning ensembles detect nighttime anomalies from Black Marble imagery?

Why thermal-only detection misses weaker anomalies

Nighttime lights from NASA's Black Marble product suite capture thermal and light emission signals from anomalous events such as fires, volcanic eruptions, and gas flaring. Existing detection approaches rely primarily on thermal bands, which limits sensitivity to weaker signals. The VIIRS Day/Night Band spans the visible to near-infrared region, has a large dynamic range, and is sensitive to weaker instances with less distinct thermal signals, but in prior work it was used only as a confirmatory feature rather than as a primary detection channel.

The paper's stated aim is therefore to jointly model Black Marble M-band and DNB signals and derive a generalized, spatio-temporally robust ensemble of anomaly detectors. The intended payoff is threefold: higher detection confidence, detection of weaker signals, and lower false detections.

A two-step workflow turns detections into training labels

High-quality labels for nighttime anomalies are scarce, so the framework avoids training on existing thermal-threshold datasets, which the authors argue would bias detectors toward those thresholds and restrict them to a subset of bands. Instead, step 1 models the non-anomalous background from a small volume of unlabeled data and flags deviations. Autoencoders minimize pixel-level multi-spectral reconstruction error, while variational autoencoders add a KL-divergence term so the latent space follows a standard normal distribution. Anomalies are detected from total reconstruction error on unseen scenes from the same region.

Clouds complicate this step: the magnitude of total reconstruction error is positive for anomalies and negative for clouds, so high negative per-band deviations are removed as possible cloud cover while high positive deviations are retained as candidate anomalies. The authors chose this over cloud masking because the VIIRS cloud mask is prone to errors at night and currently mislabels nighttime thermal anomalies as clouds. Detections over water are also masked because of higher nighttime brightness temperature.

Step 2 uses those detections as labels. Set-1 detections train regional detectors that are applied to a small unseen set of observations from the same region (set 2); the models are then refined or retrained on both sets to create the final generalized regional detector. Set-1 and set-2 detections of the final regional detectors are jointly trained across VNP46 tiles into a generic model applied to unseen regions and seasons. The authors state that this jointly trained model, referred to as the joint model, is the most important to assess for operational suitability.

Five detectors and what separates them

The ensemble contains five detectors. Two come from step 2 contrastive learning with Siamese networks that share weights across identical subnetworks: one trained with contrastive loss and one with triplet loss, producing embeddings that map an 8-D sample input into a 2-D feature space before fully connected layers classify it as anomaly or background. A fully connected neural network trained on the filtered set-1 and set-2 samples with a weighted loss serves as a machine learning baseline. Two simple outlier detectors use the most responsive bands: a top-10 percentile detection on M-11, and a top-2 percentile detection on M-11 plus DNB.

Inputs differ across detectors, and this matters for interpretation. The Siamese networks and the fully connected network require M-10 to M-16, DNB, and the World Settlement Footprint 2019 layer. The M-11 outlier detector needs only M-11. The M-11 plus DNB detector needs two inputs. Ancillary VIIRS land-water mask and World Settlement Footprint layers filter out background features such as clouds, water, and urban nightlights. Because anomalies are rare, all filtered detection instances are used for the anomaly class while background instances are randomly sampled, and losses are weighted inversely to class frequency.

Joint thermal and light detectors lead on overlap

Step-1 detectors were evaluated for pixel-level co-location with the VIIRS NightFire product. Overlap percentages show the combined autoencoder at 92.66% for tile h05v05, 85.75% for h08v06, and 76.83% for h20v10; the combined variational autoencoder is nearly identical at 92.87%, 85.69%, and 76.83%. The DNB-only detector is lower at 77.21%, 83.46%, and 67.58%. The joint models are highest: joint autoencoder at 93.84%, 96.15%, and 84.83%, and joint variational autoencoder at 93.84%, 96.21%, and 84.83%.

For the generalized model, the paper reports overlap and probability of extra detections across all five detectors for each tile. The first three tiles (h05v05, h08v06, h20v10) show region-based training performance, while h07v05, h05v04, and h02v07 test generalization to unseen regions for gas flares, fire detection, and volcanic eruptions respectively. The authors report that machine learning methods outperform the single-band approaches, and that performance does not suffer when models are extended across years and seasons or to unseen regions, indicating generalizability and robustness to spatio-temporal variation.

Among detectors, the contrastive-loss embedding model has the highest overlap score across all regions and event classes, and the authors note it also has the highest confidence because it uses all bands. The single-band outlier detector follows, with lower confidence because it requires only one band. The fully connected baseline ranks next, with higher confidence than the single-band detector because it uses all bands. The triplet-loss model performs comparably to the fully connected network for fires and volcanoes but poorly on gas flares and controlled burning, both single-pixel events covering very small spatial areas. The authors explain this by the spectral distribution: fires and volcanoes have very high signal in M-10 to M-16, whereas gas flare signals are predominantly constrained to M-10 and M-11, so the triplet model misclassifies gas flares because it lacks strong M-12 to M-16 signals and appears to place strong emphasis on those bands.

Confidence tiers separate strong and weak signals

The framework does not output one binary anomaly map. When DNB detections agree with thermal bands, they form the joint high-confidence set. When no co-occurring thermal band signal is observed, the result is defined as the moderate-confidence set, filtered with ancillary bands to retain likely anomalies. A high-confidence detection set is also derived from detections found by at least two of the five approaches.

The DNB-only layer matters because it lowers the detection threshold, enabling detection of weaker anomalies without a distinct thermal emission signal without increasing false positives. The authors report that including it increases the fraction of anomalies observed for flaring in tile h08v06, and that information gain from DNB varies by region: high for flaring in h08v06 relative to thermal bands, less informative for flares over h22v05. Examining a regional Landsat composite, they found high spatial correlation between these DNB-only detections and visible signatures of gas flaring infrastructure, suggesting that after urban settlement signals are masked out, these are possible flaring activities missed by thermal bands. They caution that when electric lighting co-occurs with such detections, VIIRS-based detections may not be sufficient to separate contamination from true anomaly.

Where the detectors transfer and where they do not

The study uses daily top-of-atmosphere Black Marble observations (VNP46A1) at 500 m spatial resolution, comprising six moderate-band brightness temperature bands (M-10 to M-16) and DNB radiance, plus ancillary cloud and land-water masks and the World Settlement Footprint 2019 layer. Study regions correspond to VIIRS tiles susceptible to fires, persistent gas flaring sites, and retrospective volcanic eruption sites. The study duration covers cloud-free observations from peak fire season (forest fire h05v05, 38 nights; controlled burning h20v10, 92 nights) and brightness temperature extremes for gas flaring (h08v06, 32 summer nights and 42 winter nights). Generalization was explored in additional regions: fires in the Western United States (h05v04, 53 nights), flaring in the Permian basin (h07v05, 77 nights), and Kilauea volcano in Hawaii (h02v07, 231 nights). Anomaly detection and iterative training were performed over 2018, with generalization tests in 2019 to 2021.

The authors state that the approach extends over land and ocean and is class-agnostic across fires, volcanoes, and gas flares. The clearest boundary is the triplet model's poor performance on gas flares and controlled burning, which the authors attribute to those events being single-pixel with very small spatial area and to the model's emphasis on M-12 to M-16 signals. They suggest that splitting anomalous classes by peak signal bands may improve triplet model performance.

No ground truth, so agreement stands in

The primary limitation the authors identify is the lack of ground truth data for nighttime anomaly monitoring, which limits model tuning and evaluation. Their response is to create multiple models and estimate detection confidence from model agreement. They are explicit that this is not ground truth but an indicator of detector reliability.

Several extensions are proposed rather than demonstrated. Anomaly detection is treated as binary classification, but multiclass classification could be explored based on anomaly type indicated through per-band deviation. Other architectures could be assessed, including segmentation, convolutional layers to incorporate spatial information, and few-shot labeling with contrastive learning embeddings. Expert labeling of detections might improve models through a semi-supervised approach, and retraining on missed detections could improve performance as the iterative process repeats. One explored approach failed: nearest-neighbor classification of the embeddings, due to computational cost.

The authors also frame the labeling method itself as a contribution, arguing that machine learning-generated labels can alleviate the labeled data scarcity that plagues remote sensing, and that small data volumes can serve as training data when features are carefully extracted. Operationalizing the detectors on daily VIIRS imagery would enable longer-term assessment of model performance and a value-added data product for applications such as natural hazard monitoring and energy extraction.

Which detector to trust for which event

The paper's most decision-relevant result is not a single accuracy figure but a ranking that depends on the event class. The contrastive-loss embedding model leads on overlap across all regions and event classes and carries the highest confidence because it uses all bands. The single-band M-11 outlier detector is competitive on overlap but has lower confidence by construction. The fully connected baseline sits between them. The triplet-loss model is comparable to the baseline for fires and volcanoes but fails on gas flares and controlled burning, the single-pixel events whose signal is confined mainly to M-10 and M-11.

That split is the practical threshold this paper offers: for events with strong M-10 to M-16 signal, the triplet model is usable, but for small-area, spectrally narrow events such as gas flares, the all-band contrastive model or the joint thermal-plus-light detections are the better choice. The confidence tiers reinforce the same logic. High-confidence detections require agreement between thermal and light signals or between at least two detectors, while DNB-only detections form a moderate-confidence set that extends sensitivity to weaker anomalies at the cost of relying on a single signal type. Readers weighing these detectors for operational use should treat the choice as conditional on event class and on whether the priority is maximum overlap or a defensible confidence level.

Sources used in this answer