Methane is a major anthropogenic greenhouse gas and a priority target for near-term climate mitigation. Satellite observations provide increasingly detailed evidence, but products acquired at different spatial and temporal scales do not describe the same physical quantity or inference level. Sentinel-5P provides repeated regional rasters of column-averaged methane (XCH4), whereas Carbon Mapper instruments provide targeted plume-event observations, source attribution, and reported emission-rate estimates. The objective of this thesis is to determine how these complementary products can support cross-scale methane hotspot screening and patch-level source classification without conflating regional concentration context with plume- or facility-level detection. The study develops a reproducible geospatial framework over the conterminous United States using Python, Google Earth Engine, and the Carbon Mapper API. It queries and maps Carbon Mapper plume-event records, compares selected Tanager events with temporally adjacent Sentinel-5P rasters, and constructs 36 monthly Sentinel-5P products for January 2023–December 2025. Valid-observation and fixed-threshold exceedance counts are combined in bivariate classes and sampled at the locations and months of 5,261 Carbon Mapper plume-event records. Local-background calculations provide complementary anomaly and standardised-enhancement products at several spatial scales. The final stage is a CNN patch classifier whose predictor channels are derived from monthly Sentinel-5P rasters and whose labels are constructed from month-specific Carbon Mapper source evidence. It estimates source-associated patch probability; it does not detect individual plumes or identify facilities. Of the 5,261 Carbon Mapper plume-event records analysed, 4,869 could be assigned a monthly Sentinel-5P bivariate class. Among those classified records, 96.5% of the summed reported emission-rate estimates came from cells containing at least one valid retrieval above 1900 ppb during the event month. For the 2,202 patches in the independent September–December 2025 test period, the classifier obtained an area under the precision–recall curve of 0.648, compared with a prevalence baseline of 0.173, and an area under the receiver-operating-characteristic curve of 0.814. At the validation-selected threshold of 0.38, precision was 0.643, recall 0.526, and F1 = 0.579. These values describe patch classification under the constructed labels; they are not facility-level detection probabilities. The contribution is a reproducible geospatial and machine-learning workflow that preserves coordinate reference systems, units, NoData semantics, temporal support, label provenance, and measurement meaning.
Il metano è un importante gas serra di origine antropica e un obiettivo prioritario per la mitigazione climatica nel breve periodo. Le osservazioni satellitari forniscono informazioni sempre più dettagliate, ma prodotti acquisiti a scale spaziali e temporali diverse non descrivono la stessa grandezza fisica né lo stesso livello di inferenza. Sentinel-5P fornisce campi regionali ripetuti della concentrazione colonnare media di metano (XCH4), mentre gli strumenti Carbon Mapper forniscono osservazioni mirate di eventi di plume, attribuzione della sorgente e stime del tasso di emissione riportate. L'obiettivo della tesi è determinare come questi prodotti complementari possano supportare lo screening multiscala degli hotspot di metano e la classificazione delle sorgenti a livello di patch, senza confondere il contesto regionale di concentrazione con il rilevamento di plume o impianti. La tesi sviluppa un quadro geospaziale riproducibile per gli Stati Uniti contigui usando Python, Google Earth Engine e l'API di Carbon Mapper. Il flusso interroga e rappresenta i record di eventi di plume Carbon Mapper, confronta eventi Tanager selezionati con campi Sentinel-5P temporalmente adiacenti e costruisce 36 prodotti mensili Sentinel-5P per gennaio 2023–dicembre 2025. I conteggi delle osservazioni valide e dei superamenti di una soglia fissa sono combinati in classi bivariate e campionati nelle posizioni e nei mesi di 5.261 record Carbon Mapper. Il calcolo del fondo locale produce inoltre raster di anomalia e di arricchimento standardizzato a più scale. La fase finale è un classificatore CNN di patch: i canali predittori derivano dai raster mensili Sentinel-5P e le etichette sono costruite da evidenze mensili di sorgenti Carbon Mapper. Il modello stima la probabilità che una patch sia associata a una sorgente; non rileva plume individuali né identifica impianti. Tra i record con supporto mensile classificato, il 96,5% della somma dei tassi di emissione riportati ricade in celle con almeno un superamento di soglia. Sul test indipendente di 2.202 patch, il classificatore raggiunge AUPRC 0,648 rispetto al valore di riferimento di prevalenza pari a 0,173, AUROC 0,814, precisione 0,643, richiamo 0,526 e F1 = 0,579. Questi valori descrivono la classificazione delle patch secondo le etichette costruite, non una probabilità di rilevamento a scala di impianto; analogamente, una classificazione negativa non dimostra l'assenza di una sorgente. Il contributo principale è un flusso riproducibile che integra dati raster e vettoriali, campionamento raster–vettore, analisi tra sensori e apprendimento automatico, mantenendo espliciti sistemi di riferimento, unità, supporto temporale, valori NoData e provenienza delle etichette.
Cross-scale methane hotspot screening and patch-level source-associated classification with Sentinel-5P and carbon mapper
ZAHED KACHAEE, SADRA
2025/2026
Abstract
Methane is a major anthropogenic greenhouse gas and a priority target for near-term climate mitigation. Satellite observations provide increasingly detailed evidence, but products acquired at different spatial and temporal scales do not describe the same physical quantity or inference level. Sentinel-5P provides repeated regional rasters of column-averaged methane (XCH4), whereas Carbon Mapper instruments provide targeted plume-event observations, source attribution, and reported emission-rate estimates. The objective of this thesis is to determine how these complementary products can support cross-scale methane hotspot screening and patch-level source classification without conflating regional concentration context with plume- or facility-level detection. The study develops a reproducible geospatial framework over the conterminous United States using Python, Google Earth Engine, and the Carbon Mapper API. It queries and maps Carbon Mapper plume-event records, compares selected Tanager events with temporally adjacent Sentinel-5P rasters, and constructs 36 monthly Sentinel-5P products for January 2023–December 2025. Valid-observation and fixed-threshold exceedance counts are combined in bivariate classes and sampled at the locations and months of 5,261 Carbon Mapper plume-event records. Local-background calculations provide complementary anomaly and standardised-enhancement products at several spatial scales. The final stage is a CNN patch classifier whose predictor channels are derived from monthly Sentinel-5P rasters and whose labels are constructed from month-specific Carbon Mapper source evidence. It estimates source-associated patch probability; it does not detect individual plumes or identify facilities. Of the 5,261 Carbon Mapper plume-event records analysed, 4,869 could be assigned a monthly Sentinel-5P bivariate class. Among those classified records, 96.5% of the summed reported emission-rate estimates came from cells containing at least one valid retrieval above 1900 ppb during the event month. For the 2,202 patches in the independent September–December 2025 test period, the classifier obtained an area under the precision–recall curve of 0.648, compared with a prevalence baseline of 0.173, and an area under the receiver-operating-characteristic curve of 0.814. At the validation-selected threshold of 0.38, precision was 0.643, recall 0.526, and F1 = 0.579. These values describe patch classification under the constructed labels; they are not facility-level detection probabilities. The contribution is a reproducible geospatial and machine-learning workflow that preserves coordinate reference systems, units, NoData semantics, temporal support, label provenance, and measurement meaning.| File | Dimensione | Formato | |
|---|---|---|---|
|
Cross-Scale Methane Hotspot Screening and Patch-Level Source-Associated Classification with Sentinel-5P and Carbon Mapper.pdf
accessibile in internet per tutti
Dimensione
17.81 MB
Formato
Adobe PDF
|
17.81 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/261404