This work addresses the problem of incremental change detection in Sentinel-1 C-band SAR time series with the concrete use-case of monitoring the construction of “The Line” (NEOM, Saudi Arabia). The goals are (i) to detect, per pixel, statistically significant step-like changes in backscatter and (ii) to estimate the change date with an associated confidence measure while operating incrementally (i.e., updating a compact state as new images arrive). Two pipelines are developed and implemented. The first pipeline (Cumulative Sums with Bootstrapping — BCSUM) computes residuals against a temporal baseline, accumulates evidence via cumulative sums (span statistic), and selects candidate pixels by thresholding excursion magnitudes. For candidates, a permutation/bootstrapping calibration is performed to produce an empirical confidence rank and a normalized effect-size; a combined score is used to accept detections and extract change times. The second pipeline (Double-Hedged CUSUM) maintains streaming statistics with Welford’s algorithm, normalizes residuals online, and runs two one-sided CUSUM accumulators (positive and negative). Alarms are emitted when accumulators exceed control limits; alarm timing yields change dates and reset logic permits detection of multiple events. Both approaches are engineered to address operational constraints, including an O(1) state, masked-array handling for missing/no-data samples, and candidate-restricted bootstrapping to limit computational requirements. Experiments on synthetic data and three real use-cases (The Line, Lake Poopó, Upsala Glacier) demonstrate that the methods are effective and complementary: BCSUM provides statistically validated, low-false-alarm detections with robust effect-size scoring, while Double-Hedged CUSUM yields low-latency, memory-efficient alarms suitable for continuous monitoring. Practical trade-offs between sensitivity, false-alarm control, and computational cost are discussed, and further directions are outlined for operational deployment.
Questo lavoro si occupa del rilevamento incrementale dei cambiamenti nelle serie temporali SAR in banda C di Sentinel-1, utilizzando come caso d’uso il monitoraggio della costruzione di “The Line” (NEOM, Arabia Saudita). Gli obiettivi principali sono due: (i) identificare, per ogni singolo pixel, cambiamenti significativi nella retrodiffusione e (ii) stimare la data di questi cambiamenti, associando una misura di confidenza, il tutto operando in modalità incrementale. Questo significa aggiornare uno stato man mano che arrivano nuove immagini. Sono state sviluppate e implementate due pipeline. La prima pipeline (Cumulative Sums with Bootstrapping, BCSUM), calcola i residui rispetto a una baseline temporale, accumula evidenze tramite somme cumulative e seleziona i pixel candidati applicando una soglia alle escursioni di magnitudine. Per i candidati, viene eseguita una calibrazione tramite bootstrapping per generare un punteggio di confidenza che viene poi utilizzato per accettare le rilevazioni e estrarre le date dei cambiamenti. La seconda pipeline, il Double-Hedged CUSUM, mantiene statistiche in streaming, normalizza i residui online e gestisce due accumulatori CUSUM unidirezionali (positivo e negativo). Vengono emessi allarmi quando gli accumulatori superano determinati limiti; le date degli allarmi forniscono le tempistiche dei cambiamenti e la logica di reset consente di rilevare eventi multipli. Entrambe le soluzioni sono progettate tenendo conto di vincoli operativi: uno stato O(1), gestione di array mascherati per campioni mancanti o senza dati, e bootstrapping limitato ai pixel candidati per contenere il carico computazionale. Esperimenti condotti su dati sintetici e tre casi reali (The Line, Lago Poopó, Ghiacciaio Upsala) dimostrano che i metodi sono efficaci e complementari: BCSUM fornisce rilevazioni statisticamente valide con pochi falsi allarmi e un robusto punteggio dell’effetto, mentre il Double-Hedged CUSUM offre allarmi a bassa latenza e bassa memoria, adatti al monitoraggio continuo. Vengono discussi i trade-off pratici tra sensibilità, controllo dei falsi allarmi e costo computazionale, e sono indicate direzioni future.
Incremental change detection in Sentinel-1 SAR image streams : monitoring the construction of The Line in Neom, Saudi Arabia
GEMMA, FRANCESCO
2024/2025
Abstract
This work addresses the problem of incremental change detection in Sentinel-1 C-band SAR time series with the concrete use-case of monitoring the construction of “The Line” (NEOM, Saudi Arabia). The goals are (i) to detect, per pixel, statistically significant step-like changes in backscatter and (ii) to estimate the change date with an associated confidence measure while operating incrementally (i.e., updating a compact state as new images arrive). Two pipelines are developed and implemented. The first pipeline (Cumulative Sums with Bootstrapping — BCSUM) computes residuals against a temporal baseline, accumulates evidence via cumulative sums (span statistic), and selects candidate pixels by thresholding excursion magnitudes. For candidates, a permutation/bootstrapping calibration is performed to produce an empirical confidence rank and a normalized effect-size; a combined score is used to accept detections and extract change times. The second pipeline (Double-Hedged CUSUM) maintains streaming statistics with Welford’s algorithm, normalizes residuals online, and runs two one-sided CUSUM accumulators (positive and negative). Alarms are emitted when accumulators exceed control limits; alarm timing yields change dates and reset logic permits detection of multiple events. Both approaches are engineered to address operational constraints, including an O(1) state, masked-array handling for missing/no-data samples, and candidate-restricted bootstrapping to limit computational requirements. Experiments on synthetic data and three real use-cases (The Line, Lake Poopó, Upsala Glacier) demonstrate that the methods are effective and complementary: BCSUM provides statistically validated, low-false-alarm detections with robust effect-size scoring, while Double-Hedged CUSUM yields low-latency, memory-efficient alarms suitable for continuous monitoring. Practical trade-offs between sensitivity, false-alarm control, and computational cost are discussed, and further directions are outlined for operational deployment.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Gemma_Executive Summary_02.pdf
accessibile in internet per tutti a partire dal 24/10/2026
Descrizione: Executive Summary
Dimensione
1.85 MB
Formato
Adobe PDF
|
1.85 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Gemma_Tesi_01.pdf
accessibile in internet per tutti a partire dal 24/10/2026
Descrizione: Testo della Tesi
Dimensione
9.1 MB
Formato
Adobe PDF
|
9.1 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/10589/245137