Although wildfires play a crucial role in ecological processes in the Mediterranean Basin, they often represent one of the primary disturbances for forests and other ecosystems, entailing landscape and habitat degradation and economic damages. They also determine the consumption of natural carbon reserves and the emission of greenhouse gasses (GHG) correlated with climate change. Accurate information relating to the impact of fire on the forest environment and how its effects are distributed over time and space, both from a qualitative and quantitative point of view, are a key factor for the subsequent implementation of medium and long-term territorial planning, in order to predict and manage irreversible processes of degradation of forests and landscape. In this context, remote sensing provides reliable tools and techniques for monitoring and quantifying the impact of burned areas with reference to satellite platforms. In post-fire monitoring, most studies have been based on optical satellite data acquired using passive multispectral sensors, that are closely correlated with the physiological and biochemical state of the vegetation. For these reasons, vegetation has its unique spectral signature (depending on the species and environmental conditions), and its observation enables its characterization and subsequent monitoring. Anomalies at the spectral level, indeed, conceal anomalies at the plant level. Although their effectiveness for this purpose has been widely proven, optical systems present some limitations, mainly due to their sensitivity to some environmental conditions, such as sunlight and cloud cover, which reduces the frequency of observation at the visible/infrared wavelength bands or the spectral confusion of burned areas with unburned low albedo surfaces (i.e., dark soils, water surfaces, shadow areas), or the premature signal saturation due to the high sensibility to increasing values of leaf area index (LAI). Furthermore, this type of sensors cannot capture many quantitative aspects since these wavelengths do not interact directly with the structure of the objects. Therefore, methods based on data acquired by cloud-independent and structure-dependent sensors at high spatial and temporal resolutions are needed. Among them, Synthetic Aperture Radar (SAR) are active sensors that generates their microwave impulses (2.4-100 cm) and transmits them from its antenna to a target. Calculating the amount of the signal fraction reflected back to the sensor (backscatter) characterizes the target’s spectral radar signature. The penetration capacity of the impulse in the matter is directly proportional to the wavelength. For this reason, the SAR waves can pass through atmospheric particulate or interact with the vegetation cover structure. Therefore, to characterize and quantify the effects of a disturbance on vegetation, radar technology exploits the variations in backscatter caused by the modification of the vegetation cover and soil’s structure and moisture content. However, its processing and interpretation complexity causes this sensor not to be widely used compared to optical ones. Against that background, in this Ph.D. thesis, a complete and open-access workflow aimed at the investigation and mapping of the fire effects on Mediterranean ecosystems in the short term (pre-fire condition assessment; burned area detection; post-fire severity estimation and mapping) and the monitoring of the response of the environment during the first years after the event (post-fire recovery monitoring), was developed. To achieve this, free-available optical (Sentinel-2) and SAR (Sentinel-1) high spatial and temporal data were integrated, assessing the strengths and limitations of each of them and the advantages that are provided by the combination of both information. The first step concerned the construction of an accurate land use/land cover (LULC) map in a heterogeneous Mediterranean forest (located in Portugal) area to have an overview of the qualitative and quantitative state of vegetation present before the fire event. To this end, we applied an exhaustive grid search analysis to set the optimal hyperparameters of a machine learning model (random forest, RF) and the inclusion of different variables 5 (i.e., vegetation indices time-series, optical biophysical variables, and interferometric SAR - InSAR - coherence). This integration allowed reaching an overall accuracy (F-score) of 0.903, observing an improvement of 2.53% when SAR data were integrated into optical information. The second step dealt with the accurate detection of burned areas and delimitation of their perimeters. Two approaches were proposed to fulfil the objective: the first concerned only the use of SAR data (backscatter, dual- and single polarized SAR indices, textures) for an unsupervised detection (using the k-mean algorithm, set using a silhouette score analysis) of burned areas in two different study sites (located in Portugal and Italy respectively), with a reached F-score of 0.803 (Portuguese site) and 0.853 (Italian site); the second approach proposed a multitemporal composite process, by combining both Sentinel-2 and Sentinel-1 images, and a subsequent object-based geographic analysis (GEOBIA) to map burned areas on regional/national scales occurred during the entire fire season (2017) in Portugal, achieving F-score values of 0.914 (when only optical data is used) and 0.956 (combining optical and SAR information). In the third step of the main workflow, the short-term fire effects, in the form of fire severity, and their spatial distribution are estimated. Three approaches were presented, two of which are similar and united by the use of the composite burnt index (CBI) protocol to determine the severity classes in the field and to define the training data of the model, which, in one case (study site located in Portugual), was constituted by the RF algorithm, while in the other case (study site located in Italy) an artificial neural network was built. The RF model reached an F-score of 0.838 when both the datasets were combined (S1 + S2), compared with the values achieved by using SAR (0.513) and optical (0.805). The results obtained using the artificial neural network (F-score > 0.95) gave proof of the great potential in the use of these advanced deep learning models. A third approach involved a spectral mixture analysis (SMA) of optical Sentinel-2 imagery to spectrally characterize and quantify the proportion of the three fraction components indicative of the main physical effects immediately after a forest fire (char, scorched vegetation and green vegetation). For the first years after a fire event, the investigation of temporal and spatial dynamics of the post-fire recovery of different Mediterranean vegetation types characterized the fourth step. Both SAR Sentinel-1 and optical Sentinel-2 time series were analyzed separately according to the fire severity classes (obtained in the previous step), highlighting the complementary and essentiality of both information. Moreover, a burn recovery ratio (BRR), optimized through machine learning regressors for predicting pre-fire conditions, was proposed to estimate and map the spatial distribution of the degree of vegetation recovery. The development of these approaches and managing this amount of data required advanced techniques and solutions of geo-informatics, geo-statistics, geomatic, image processing, and advanced artificial intelligence models. Nevertheless, the whole process was developed and performed, fulfilling the principle of employing freely available data and open-source software and libraries (e.g., ESA SNAP, Scikit-Learn, OTB, Google Earth Engine) mostly executed in Python-script language
Nel bacino del Mediterraneo, sebbene gli incendi svolgano un ruolo cruciale nei processi ecologici, spesso rappresentano uno dei principali disturbi per le foreste e gli altri ecosistemi, comportando degradazione del paesaggio e degli habitat, e danni economici. Gli incendi, inoltre, determinano il consumo di riserve naturali di carbonio e l'emissione di gas serra (GHG) correlati al cambiamento climatico. Informazioni accurate relative all'impatto dell'incendio sull'ambiente forestale, e alla sua distribuzione nel tempo e nello spazio, rappresentano fattori chiave per la successiva attuazione della pianificazione territoriale a medio e lungo termine, finalizzata alla prevenzione e gestione di quei processi irreversibili di degrado degli habitat e del paesaggio. In questo contesto, il telerilevamento satellitare fornisce strumenti e tecniche affidabili per monitorare e quantificare l'impatto delle aree bruciate. Nel monitoraggio post-incendio, la maggior parte degli studi si è basata su dati satellitari ottici acquisiti utilizzando sensori multispettrali passivi, strettamente correlati allo stato fisiologico e biochimico della vegetazione. La vegetazione presenta infatti una firma spettrale univoca (con piccole variazioni a seconda della specie e delle condizioni ambientali), e la sua analisi ne consente la caratterizzazione e il successivo monitoraggio: le anomalie spettrali, infatti, si traducono in anomalie a livello della pianta. Sebbene l’efficacia dei sensori ottici per lo scopo appena descritto sia stata ampiamente dimostrata, essi presentano alcune limitazioni. Queste sono principalmente causate dalla sensibilità ad alcune condizioni ambientali: come luce solare e nuvolosità, che ne riducono la frequenza di osservazione; o dalla confusione di aree bruciate con superfici a bassa albedo (es. suoli scuri, superfici idriche, zone d'ombra); o la sensibilità ai valori crescenti dell'indice di area fogliare (LAI) che comporta una prematura saturazione del segnale. Inoltre, questi tipi di sensori non interagiscono direttamente con la struttura degli oggetti non permettendogli la cattura di molti aspetti qualitativi della copertura vegetale. Pertanto, sono necessari dati acquisiti da sensori che presentino sia indipendenza dalla copertura nuvolosa, sia la capacità di interagire con la struttura. Tra questi, i radar ad apertura sintetica (SAR) sono sensori attivi che generano i loro impulsi a microonde (2,4-100 cm) e li trasmettono dalla loro antenna ad un bersaglio posto sulla superfice terrestre. Il calcolo della quantità della frazione di segnale riflessa verso sensore (backscatter) caratterizza la firma del radar spettrale del bersaglio. La capacità di penetrazione dell'impulso nella materia è direttamente proporzionale alla lunghezza d'onda. Per questo motivo, le onde SAR possono attraversare il particolato atmosferico o interagire con la struttura della copertura vegetale. Pertanto, per caratterizzare e quantificare gli effetti di un disturbo sulla vegetazione, la tecnologia radar sfrutta le variazioni di backscatter causate dalla modifica della copertura vegetale, della struttura del suolo, e del contenuto di umidità degli oggetti osservati. Tuttavia, la sua complessità di elaborazione e interpretazione fa sì che questi sensori non siano ampiamente utilizzati quanto quelli ottici. In tale contesto, la presente tesi di Dottorato propone un workflow completo ed open-source finalizzato all'indagine e mappatura degli effetti del fuoco a breve termine sugli ecosistemi mediterranei (valutazione delle condizioni pre-incendio; rilevamento delle aree bruciate; stima e mappatura della severità post-incendio) e al monitoraggio temporale della risposta della vegetazione durante i primi anni dopo l'evento di incendio (monitoraggio del recupero post-incendio). Per assecondare questi obiettivi, sono stati integrati dati ottici (Sentinel-2) e SAR (Sentinel-1) gratuiti e ad alta risoluzione spaziale e temporale, valutando i punti di forza e i limiti di ciascuno di essi, e i vantaggi forniti dalla combinazione di entrambe le informazioni. Il primo step ha riguardato la realizzazione di un'accurata mappa della copertura e uso del suolo (LULC) in un'area Mediterranea eterogenea (situata in Portogallo) al fine di avere una panoramica dello stato qualitativo e quantitativo della vegetazione presente prima dell'evento di incendio. A tal fine, diverse variabili sono state calcolate ed utilizzate come dati di 7 input (ad esempio, serie temporali degli indici di vegetazione, variabili biofisiche ottiche e SAR interferometrico - InSAR - coerenza). Inoltre, è stata applicata un'analisi exhaustive grid search per impostare gli iperparametri ottimali di un modello di machine learning (foresta casuale, RF). Questa integrazione ha permesso di raggiungere un'accuratezza complessiva (F-score) di 0.903, osservando un miglioramento del 2,53% quando i dati SAR sono stati accoppiate alle informazioni ottiche. Il secondo step ha riguardato il rilevamento accurato delle aree bruciate e la delimitazione dei loro perimetri. Sono stati proposti due approcci per raggiungere questo obiettivo: il primo riguardava solo l'uso di dati SAR (backscatter, indici SAR a doppia e singola polarizzazione, texture) per un rilevamento non supervisionato (utilizzando l'algoritmo k-mean, impostato utilizzando un'analisi silhouette) delle aree bruciate in due diversi siti di studio (situati rispettivamente in Portogallo e in Italia), raggiungendo valori di F-score pari a 0.803 (sito portoghese) e 0.853 (sito italiano); il secondo approccio proponeva un processo di image composite multitemporale, combinando sia le immagini Sentinel-2 che quelle Sentinel-1, e una successiva classificazione ad oggetti (GEOBIA) per mappare le aree bruciate su scala regionale/nazionale da incendi avvenuti durante un’intera stagione degli incendi (2017) in Portogallo, ottenendo valori F-score di 0,914 (quando vengono utilizzati solo dati ottici) e 0,956 (combinando informazioni ottiche e SAR). Nel terzo step, vengono stimati gli effetti dell'incendio a breve termine, sotto forma di severità dell'incendio, e la loro distribuzione spaziale. Sono stati presentati tre approcci di cui due simili e accomunati dall'uso del protocollo Composite Burnt Index (CBI) per determinare le classi di severità sul campo e per definire i dati di addestramento del modello. In un approccio (sito di studio situato in Portogallo), si è utilizzato l’algoritmo RF come modello; mentre in un secondo approccio (sito di studio situato in Italia) è stata implementata una rete neurale artificiale. Il modello RF ha raggiunto un valore di F-score di 0,838 quando entrambi i set di dati sono stati combinati (S1 + S2), rispetto ai valori ottenuti utilizzando solo SAR (0,513) e solo ottico (0,805). I risultati ottenuti utilizzando la rete neurale artificiale (F-score > 0,95) hanno dato prova del grande potenziale nell'uso di questi modelli avanzati di deep learning. Un terzo approccio prevedeva l’applicazione di una spectral mixture analysis (SMA) delle immagini ottiche Sentinel-2 per caratterizzare spettralmente e quantificare la proporzione delle tre componenti frazionali indicativi dei principali effetti fisici riscontrabili immediatamente dopo un incendio boschivo (carbone, vegetazione bruciata e vegetazione verde). Per i primi anni dopo un evento di incendio, l'indagine sulle dinamiche temporali e spaziali del recupero post-incendio di diverse tipologie di vegetazione mediterranea ha caratterizzato il quarto step. Sia le serie temporali SAR Sentinel-1 che quelle ottiche Sentinel-2 sono state analizzate separatamente in base alle classi di gravità dell'incendio (ottenute nello step precedente), evidenziando la complementarietà e l'essenzialità di entrambe le informazioni. Inoltre, è stato proposto un indice burn recovery ratio (BRR) per la stima e la mappatura della distribuzione spaziale del grado di recupero della vegetazione. L’indice è stato ottimizzato nella fase di predizione delle condizioni pre-incendio tramite algoritmo di regressione di machine learning. Lo sviluppo di questi approcci e la gestione di questa quantità di dati hanno richiesto tecniche e soluzioni avanzate di geo-informatica, geostatistica, geomatica, image processing e modelli avanzati di intelligenza artificiale. Tuttavia, l'intero processo è stato sviluppato ed eseguito rispettando il principio dell'utilizzo di dati gratuitamente disponibili e software e librerie open-source (ad es. ESA SNAP, Scikit-Learn, OTB, Google Earth Engine) eseguiti principalmente tramite linguaggio Python
Use of SAR and multispectral satellite images, and remote sensing techniques for monitoring and characterization of consequences of fire on Mediterranean forest vegetation / De Luca, Giandomenico. - (2023 Mar 10).
Use of SAR and multispectral satellite images, and remote sensing techniques for monitoring and characterization of consequences of fire on Mediterranean forest vegetation
De Luca, Giandomenico
2023-03-10
Abstract
Although wildfires play a crucial role in ecological processes in the Mediterranean Basin, they often represent one of the primary disturbances for forests and other ecosystems, entailing landscape and habitat degradation and economic damages. They also determine the consumption of natural carbon reserves and the emission of greenhouse gasses (GHG) correlated with climate change. Accurate information relating to the impact of fire on the forest environment and how its effects are distributed over time and space, both from a qualitative and quantitative point of view, are a key factor for the subsequent implementation of medium and long-term territorial planning, in order to predict and manage irreversible processes of degradation of forests and landscape. In this context, remote sensing provides reliable tools and techniques for monitoring and quantifying the impact of burned areas with reference to satellite platforms. In post-fire monitoring, most studies have been based on optical satellite data acquired using passive multispectral sensors, that are closely correlated with the physiological and biochemical state of the vegetation. For these reasons, vegetation has its unique spectral signature (depending on the species and environmental conditions), and its observation enables its characterization and subsequent monitoring. Anomalies at the spectral level, indeed, conceal anomalies at the plant level. Although their effectiveness for this purpose has been widely proven, optical systems present some limitations, mainly due to their sensitivity to some environmental conditions, such as sunlight and cloud cover, which reduces the frequency of observation at the visible/infrared wavelength bands or the spectral confusion of burned areas with unburned low albedo surfaces (i.e., dark soils, water surfaces, shadow areas), or the premature signal saturation due to the high sensibility to increasing values of leaf area index (LAI). Furthermore, this type of sensors cannot capture many quantitative aspects since these wavelengths do not interact directly with the structure of the objects. Therefore, methods based on data acquired by cloud-independent and structure-dependent sensors at high spatial and temporal resolutions are needed. Among them, Synthetic Aperture Radar (SAR) are active sensors that generates their microwave impulses (2.4-100 cm) and transmits them from its antenna to a target. Calculating the amount of the signal fraction reflected back to the sensor (backscatter) characterizes the target’s spectral radar signature. The penetration capacity of the impulse in the matter is directly proportional to the wavelength. For this reason, the SAR waves can pass through atmospheric particulate or interact with the vegetation cover structure. Therefore, to characterize and quantify the effects of a disturbance on vegetation, radar technology exploits the variations in backscatter caused by the modification of the vegetation cover and soil’s structure and moisture content. However, its processing and interpretation complexity causes this sensor not to be widely used compared to optical ones. Against that background, in this Ph.D. thesis, a complete and open-access workflow aimed at the investigation and mapping of the fire effects on Mediterranean ecosystems in the short term (pre-fire condition assessment; burned area detection; post-fire severity estimation and mapping) and the monitoring of the response of the environment during the first years after the event (post-fire recovery monitoring), was developed. To achieve this, free-available optical (Sentinel-2) and SAR (Sentinel-1) high spatial and temporal data were integrated, assessing the strengths and limitations of each of them and the advantages that are provided by the combination of both information. The first step concerned the construction of an accurate land use/land cover (LULC) map in a heterogeneous Mediterranean forest (located in Portugal) area to have an overview of the qualitative and quantitative state of vegetation present before the fire event. To this end, we applied an exhaustive grid search analysis to set the optimal hyperparameters of a machine learning model (random forest, RF) and the inclusion of different variables 5 (i.e., vegetation indices time-series, optical biophysical variables, and interferometric SAR - InSAR - coherence). This integration allowed reaching an overall accuracy (F-score) of 0.903, observing an improvement of 2.53% when SAR data were integrated into optical information. The second step dealt with the accurate detection of burned areas and delimitation of their perimeters. Two approaches were proposed to fulfil the objective: the first concerned only the use of SAR data (backscatter, dual- and single polarized SAR indices, textures) for an unsupervised detection (using the k-mean algorithm, set using a silhouette score analysis) of burned areas in two different study sites (located in Portugal and Italy respectively), with a reached F-score of 0.803 (Portuguese site) and 0.853 (Italian site); the second approach proposed a multitemporal composite process, by combining both Sentinel-2 and Sentinel-1 images, and a subsequent object-based geographic analysis (GEOBIA) to map burned areas on regional/national scales occurred during the entire fire season (2017) in Portugal, achieving F-score values of 0.914 (when only optical data is used) and 0.956 (combining optical and SAR information). In the third step of the main workflow, the short-term fire effects, in the form of fire severity, and their spatial distribution are estimated. Three approaches were presented, two of which are similar and united by the use of the composite burnt index (CBI) protocol to determine the severity classes in the field and to define the training data of the model, which, in one case (study site located in Portugual), was constituted by the RF algorithm, while in the other case (study site located in Italy) an artificial neural network was built. The RF model reached an F-score of 0.838 when both the datasets were combined (S1 + S2), compared with the values achieved by using SAR (0.513) and optical (0.805). The results obtained using the artificial neural network (F-score > 0.95) gave proof of the great potential in the use of these advanced deep learning models. A third approach involved a spectral mixture analysis (SMA) of optical Sentinel-2 imagery to spectrally characterize and quantify the proportion of the three fraction components indicative of the main physical effects immediately after a forest fire (char, scorched vegetation and green vegetation). For the first years after a fire event, the investigation of temporal and spatial dynamics of the post-fire recovery of different Mediterranean vegetation types characterized the fourth step. Both SAR Sentinel-1 and optical Sentinel-2 time series were analyzed separately according to the fire severity classes (obtained in the previous step), highlighting the complementary and essentiality of both information. Moreover, a burn recovery ratio (BRR), optimized through machine learning regressors for predicting pre-fire conditions, was proposed to estimate and map the spatial distribution of the degree of vegetation recovery. The development of these approaches and managing this amount of data required advanced techniques and solutions of geo-informatics, geo-statistics, geomatic, image processing, and advanced artificial intelligence models. Nevertheless, the whole process was developed and performed, fulfilling the principle of employing freely available data and open-source software and libraries (e.g., ESA SNAP, Scikit-Learn, OTB, Google Earth Engine) mostly executed in Python-script languageFile | Dimensione | Formato | |
---|---|---|---|
De Luca Giandomenico.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Licenza:
DRM non definito
Dimensione
13.41 MB
Formato
Adobe PDF
|
13.41 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.