Multitemporal Earth Observation Image Analysis (eBook)

Remote Sensing Image Sequences

Nesrine Chehata, Cl ment Mallet (Herausgeber)

eBook Download: EPUB

2024 | 1. Auflage
272 Seiten
Wiley (Verlag)
978-1-394-30664-0 (ISBN)

Infos

Earth observation has witnessed a unique paradigm change in the last decade with a diverse and ever-growing number of data sources. Among them, time series of remote sensing images has proven to be invaluable for numerous environmental and climate studies.

Multitemporal Earth Observation Image Analysis provides illustrations of recent methodological advances in data processing and information extraction from imagery, with an emphasis on the temporal dimension uncovered either by recent satellite constellations (in particular the Sentinels from the European Copernicus programme) or archival aerial images available in national archives.

The book shows how complementary data sources can be efficiently used, how spatial and temporal information can be leveraged for biophysical parameter estimation, classification of land surfaces and object tracking, as well as how standard machine learning and state-of-the-art deep learning solutions can solve complex problems with real-world applications.

Clément Mallet is a senior scientist in remote sensing for land-cover mapping issues at the LaSTIG Laboratory (Gustave Eiffel University, IGN, French Mapping Agency), France. He is also Editor-in-Chief of the ISPRS Journal of Photogrammetry and Remote Sensing.

Nesrine Chehata is a senior lecturer in geomatics and AI for Earth observation at ENSEGID-Bordeaux INP, France. She is also President and Co-founder of AGEOS (African Association for Geospatial Development), President of FNAACC (National Forum of Climate Change Adaptation Actors in Tunisia) and an IEEE GRSS senior member.

Textauszug

1
Broader Application of the Time-SIFT Method: Proof-of-Concept of 3-D-Monitoring Study Cases with Various Spatiotemporal Scales

Denis FEURER1, Sean BEMIS2, Guillaume COULOUMA1, Hatem MABROUK3, Sylvain MASSUEL4, Romina Vanessa BARBOSA5, Yoann THOMAS5, Jérôme AMMANN6 and Fabrice VINATIER1

1LISAH, INRAE, IRD, Institut Agro, University of Montpellier, France

2Department of Geosciences, Virginia Tech, Blacksburg, USA

3Institut National Agronomique de Tunisie (INAT), Tunis, Tunisia

4G-EAU, AgroParisTech, Cirad, INRAE, IRD, Institut Agro, University of Montpellier, France

5CNRS UMR 6539 LEMAR, UBO, IRD, Ifremer, IUEM Plouzané, France

6CNRS UMR 6538 LGO, IUEM Plouzané, Université de Bretagne Occidentale, France

1.1. Introduction

Photography, aerial photography and satellite earth observation, along with the rapidly expanding accessibility of uncrewed aerial vehicle (UAV)-based remote sensing and smartphones, are a rich source of potential image time series that allow the monitoring of a wide range of phenomena across a broad scope of scientific domains. Some recent advances in remote sensing hence renewed the tools for the monitoring of 3-D change. This is particularly notable for the monitoring of sub-metric 3-D changes such as geomorphic studies, vegetation or erosion monitoring (e.g. Marcus and Fonstad 2008; Brodu and Lague 2012; Eltner et al. 2016). For a long time, studies focusing on these scales relied on topographic ground surveys, which are very cumbersome and costly. Terrestrial laser scanning provides denser topographic data than can be acquired from typical ground survey methods, such as a total station, but still has the drawback of being a ground-based technique with the corresponding limitations (e.g. occlusions, implementations, difficult coverage of large areas). Dense airborne laser scanning allows the collection of such topographic data, but the relative novelty of this technique results in a lack of historical depth. Meanwhile, the last decade saw the rapid expansion of methods from the field of computer vision adopted into the earth sciences community (Westoby et al. 2012; Fonstad et al. 2013; Bemis et al. 2014). Structure from Motion (SfM) photogrammetry is now widely used to provide detailed 3-D descriptions of various natural and anthropogenic objects of interest or systems (see, for example, reviews from Carrivick et al. 2016; Eltner et al. 2016; Smith et al. 2016). Moreover, SfM has allowed new insights to be extracted from archival aerial imagery (e.g. Verhoeven 2011; Salach 2017; Sevara et al. 2017) and hence access to an unprecedented historical depth for the monitoring of 3-D change. Still, in order to compare data acquired at different epochs, we must ensure that multi-temporal data share a common geometric reference. Co-registration of data of different epochs is indeed – by nature – the prerequisite of diachronic studies. In some experimental designs, permanent reference points can be established to provide common registration points for a multi-temporal study. However, there are a wide range of examinations for which it is not possible to establish reference points prior to imagery collection. For example, study areas with difficult access or environmental sensitivity may preclude the establishment of semi-permanent reference features. Furthermore, a wealth of temporal data already exists in archival imagery where it would be impossible for new studies to establish ground control references prior to imagery collection.

Archival aerial imagery is the most ancient source of multi-temporal Earth observation data, with coherent archives that go back to the early years of the 20th century (Cowley and Stichelbaut 2012). Methods for processing time series of the Earth’s surface were first focused on these kinds of data. These methods can differ according to the available data, but even for the most favorable cases – when all the camera calibration and ancillary data is available – we may still need to perform additional co-registration, as described by Fischer et al. (2011). Two main strategies exist for co-registration of multi-temporal 3-D datasets. One approach relies on a posteriori co-registration of the final 3-D models, whereas the other approach utilizes a suite of ground control points (GCPs) that can be found in all images. The first category of studies usually requires the user to have external fine topographic data available, such as a lidar-derived topographic dataset. This is the approach used by Bakker and Lane (2017), who noticed that the propagation of linear errors could not be limited during the SfM processing and that these errors resulted in spatial differences between the digital elevation model (DEM) of the different epochs. In such cases, authors perform co-registration with an a priori knowledge of stable zones. The “stable zones” approach has also been favorably used for UAV-based studies (e.g. Haas et al. 2016). The second category of studies seeks to have or create a set of GCPs that may be unambiguously found in all images of the multi-temporal dataset. This strategy seems to be the most used in the photogrammetric community, with remarkable recent works that proposed to automatically determine stable linear features (Nagarajan and Schenk 2016), or more recently, automatically derived multi-temporal GCPs (Giordano et al. 2018).

Despite the remarkable methods proposed by these previous works, a non-negligible amount of non-image information is needed, either in the form of external data or in the form of photogrammetric expertise, which significantly reduces the potential for a broader use of existing multi-temporal stereoscopic data. In this context, Feurer and Vinatier (2018a) and Feurer and Vinatier (2018b) proposed a new method, which allows a user to establish a unique geometric reference shared by all images of a multi-temporal stereoscopic image dataset. This method takes advantage of the invariance properties of feature detection algorithms such as the Scale Invariant Feature Transform (SIFT; Lowe 2004) to automatically link together images of different epochs and is hence called Time-SIFT. The idea of the Time-SIFT method has its roots in the breakthrough that SfM algorithms proposed in photogrammetry, with the provision of dense sets of matched feature points. Indeed, as noticed by several authors (Westoby et al. 2012; Fonstad et al. 2013), the fact that SfM processing of multi-view stereoscopic imagery relies first and foremost on image information only opens up new possibilities in photogrammetric processing. Among these new possibilities, archival aerial imagery was a compelling candidate because the “classical” photogrammetric workflows require a lot of ancillary data (calibration certificates) in addition to the images themselves, and images with a specific geometric processing if scanned from analogue photography. The study by Feurer and Vinatier (2018a, 2018b) showed that the Time-SIFT method allowed for the provision of stackable multi-temporal DEM from almost image information only. Using imagery that contains temporal change to establish a multi-temporal geometric reference seems counterintuitive as we may expect the change captured between epochs to contaminate the registration or reconstruction. Therefore, the Time-SIFT method requires further testing using test cases with a range of spatiotemporal scales. The initial paper (Feurer and Vinatier 2018b) focused on one particular spatiotemporal scale and phenomena, and the use of archival aerial imagery to examine prior 3-D changes linked to human activity. The broader suitability of the method for other spatiotemporal scales and phenomena depends on the ability of the invariance properties of the SIFT-like algorithm to provide consistent results across varied implementations. This possible invariance is expected to result in a trade-off between the spatiotemporal scale of the image data set and the spatiotemporal scale of the studied processes.

In this context, our study aims to check the potential of the Time-SIFT method on varied test cases with different spatiotemporal scales, from the millimeter/minute to the kilometer/decade. The expected success criteria are the ability of the Time-SIFT method to build a complete (with all images and all epochs) and consistent (all data representing the same 3-D scene at the global scale) multi-temporal 3-D block, as well as provide informative 3-D data allowing for the description and characterization of temporal change associated with the targeted phenomena.

Considering the relatively low effort required by the Time-SIFT method (which at a minimum requires only image data), a significant gain should be expected relatively to other existing methods, either in terms of working and/or processing time, when using GCPs or when external registration data are feasible, or in terms of achievable results when these data are not available, which implies that classical methods would not be applicable.

The chapter is divided into two sections. The first section provides a highlight and a detailed description of the Time-SIFT method. The second section is dedicated to the description of five different test cases, presented in five different subsections. The chapter ends with a summary of the findings of these different test cases and concluding remarks.

1.2. The Time-SIFT...

Produkt-Details

• Erscheint lt. Verlag: 17.7.2024
• Sprache: englisch
• Themenwelt: Naturwissenschaften ► Geowissenschaften ► Geologie
• Themenwelt: Technik
• ISBN-10: 1-394-30664-4 / 1394306644
• ISBN-13: 978-1-394-30664-0 / 9781394306640