Semantic Segmentation of Methane Plumes

By Michelle, Kelvin, and Yasmeen

Introduction and Background

01

Methane Leaks

• Extend the annotation of the AVIRIS hyperspectral image dataset
• By using HyperSTARCOP model, we can reduce false detection rate

• Increased attention on methane leak detection
• Need for automated method due to recent satellite missions (PRISMA, EnMAP, EMIT)

HyperSTARCOP Benefits

• Can help refine methane plume quantification

​

• Can generalize data from other hyperspectral sensors
• Fully autonomous methane detection on satellites

Methane Signature and Enhancement Methods

​

• Illustration of the presence of the methane signal
• The hyperspectral sensors show the wavelengths
• The multispectral instruments show the whole band ranges
• Water vapour absorption is typically excluded from the data analysis

Methane Signature and Enhancement Methods

​

• Current detection methods rely heavily on manual visual inspection

• Methane is detectable in two specific near-infrared range (1600-1850 nm and 2100-2500 nm)
• Hyperspectral imagers cover methane ranges with fine resolution

Machine Learning for Methane Detection

• HyperSTARCOP uses real world methane plumes
• Trains and compares models for multispectral and hyperspectral data

​

• Some methods lack automation and struggle with performance in new locations.

Data and Methodology

02

The Datasets

AVIRIS dataset (aerial data from Permian basin, US)

• Used for mag1c model output and simulating bands (spectral res: 5-10nm, spatial res: 3 - 10 m)
• Sampled chips (512 x 512) from images, same amount of positive (plumes) as negative (no plume) classes.
• Half of all negative samples are high confounders based on mag1c output.
• Train: Sampled 128 x 128 px with overlap from those chips
• Test: data from different days, kept as 512 x 512

​

The Datasets

EMIT dataset (satellite sensor)

• Testing generalisation
• Sensors with different disturbance
• Different regions around the world
• Used L1 processing of EMIT to be similar to on-board sensors

Feature Extractions (multispectral)

Varon ratio (S,B)

• S is signal band and B is background band
• c is found in a Least Square fit OR

​

Sanchez ratio (S)

• Use Multiple linear regression to find the background info (S_MLR) from a combination of other bands

​

​

Multispectral input

​

Varon input : V(B7,B5), V(B8,B5), and V(B7,B6)

Sanchez input : S(B7,B7_mlr) ,S(B8,B8_mlr) using B1-B2 and B4-B6 as background bands for MLR, and SWIR band B1

(Varon+Sanchez) : V(B7,B5), V(B8,B5), S(B7,B7_mlr)

Feature Extractions (hyper-spectral)

​

Mag1c:

• Subtract average background reflectance from each pixel.
• Compares each pixel's spectrum to the expected methane absorption spectrum (black line)

​

​

Improvements in Mag1c:

• Sparsity Regularization: Reduces noise by enforcing sparsity in detection.
• Albedo Correction: Adjusts for surface reflectance, improving plume visibility.
• Filters: Minimize high noise sensitivity in urban areas and remove low probability detections

​

​

Models

• Pretrained on ImageNet dataset for MultiSTARCOP and Trained whole model for HyperSTARCOP
• Same architecture for each model, just different input channels
• 6.6 million parameters

MobileNet v2

• Bottleneck layers that lowers computation without losing much performance
• Increase depth (# of channels) and perform depthwise convolution

U-net Decoder

• Takes the encoded data and recursively upsample it.
• Skip Connection to preserve information lost in encoding
• Conv layers to refine and extract features

Experimental setup

• Adam with lr = 0.001, ReduceLROnPlateau by 0.5 and patience = 4 for 15 epochs, and used WeightedRandomSampler to oversample minor class.
• Loss functions
• MultiSTARCOP uses weighted binary cross-entropy loss; plume weighted by 15.
• HyperSTARCOP multiplies the non-weighted binary cross-entropy loss computed over the whole tile with the mag1c product.

​

Evaluation Setup

• Turn output into binary map (> 0.5 = plume) (for F1, precision etc) .
• Prediction output contains plume if more than 10 active pixels.
• Plume is captured if it has 1 pixel overlap with the annotation.

​

• Able to computed similar input products as HyperSTARCOP inputs for EMIT.
• Don’t have to rescale EMIT resolution; expect model to be scale agnostic.

Results

03

1. MultiSTARCOP model: trained on the simulated WorldView-3 data
2. HyperSTARCOP model: trained with the hyperspectral data from the AVIRIS sensor
3. evaluation of the HyperSTARCOP model for zero-shot detection of methane leaks in the hyperspectral data from the EMIT sensor

MultiSTARCOP model is generally able to only detect the area of the plume with the higher gas concentration

HyperSTARCOP model is capable of detecting the entire plume, including the areas of lower concentration in the plume tail

Multispectral Model Performance:

• The MultiSTARCOP model outperforms the baseline in all metrics.
• The "Varon+Sanchez ratios" variant achieves the best AUPRC and F1 scores, while the "Varon ratios" variant has the lowest false positive rate.
• Performance degrades significantly for smaller plume events, highlighting the challenge of detecting these.

Hyperspectral Model Performance

• outperforms the baseline
• both variants improve F1 scores for both strong and weak plumes while reducing false positives

From what I observed, the models sometimes has difficulty detecting the weak plumes.

• The model is semantic segmentation (classifying each pixel) which makes detecting weak methane plumes difficult.
• Also, the pixel-based predictions get converted into a tile classification, so weaker signals might not get detected, leading to reduced sensitivity on weaker plumes.

More complex classification methods could potentially improve the detection of these weak plumes.

My thoughts on what can be improved

For this project, we could consider using more complex classification methods like:

• ensemble models
• recurrent neural networks (RNNs)
• attention-based models / attention-based neural networks

These methods could integrate both pixel-level and spatial context, improving sensitivity without amplifying noise.

Conclusion

04

The model addresses the known limitation of matched filter methods, which produces many false positive detections.

​

The model they made reduced the false positive rate per tile by over 41.83% in contrast to the baseline, at the cost of small drop in the number of captured plumes.

• Explore general, sensor-agnostic systems, that would be able to detect signals of arbitrary gas signatures from hyperspectral data.

​

• Since the model would be deployed on-board a satellite, we would need to evaluate the speed of the trained models in a constrained environment (for real-life applications, in case of a natural disaster).

Thanks!

CREDITS: This presentation template was created by Slidesgo, and includes icons by Flaticon, and infographics & images by Freepik