Part II GIS Research Perspectives for Sustainable Development Planning © 2006 by Taylor & Francis Group, LLC 107 7 Advanced Remote Sensing Techniques for Ecosystem Data Collection Alexandr A. Napryushkin and Eugenia V. Vertinskaya CONTENTS 7.1 Introduction 107 7.2 RS-Based Thematic Mapping Methodology 109 7.2.1 General Concept 109 7.2.2 Imagery Interpretation Approach 111 7.3 Thematic Mapping Methodology Implementation 114 7.3.1 The RS Imagery Processing and Interpretation System “LandMapper” 114 7.3.2 Application of “LandMapper” for Anthropogenic Ecosystems Research 116 7.3.2.1 Mapping Hydro Network and Urban Areas of Tomsk City 116 7.3.2.2 Landscape-Ecological Research of Pervomayskoe Oil Field 118 7.4 Conclusion 121 Acknowledgments 122 References 122 7.1 INTRODUCTION The problems of monitoring and ecological control of ecosystems of different natures are becoming more and more urgent. Monitoring of the Earth’s surface has a mul- tidisciplinary character and allows a wide spectrum of issues to be solved. The ecosystem components involved in monitoring are manifold and include, among others, surface waters, soils, vegetation canopy, and anthropogenic landscape components. The latter represent the man-made and man-changed ecosystems and are of primary interest © 2006 by Taylor & Francis Group, LLC 108 GIS for Sustainable Development in the context of monitoring and management problems due to degradation of recent ecological conditions [1]. One of the most important issues solved in the monitoring process is represen- tation of its results as a series of thematic maps indicating the spatial structure of complex ecosystem components [2]. The basic concern of thematic mapping is graphical modeling of ecosystems and providing the information on their conditions for efficient natural resources management. The geoinformation provided by the thematic maps is used for analysis and assessment of natural resource conditions, recording and accounting destructive natural phenomena, studying natural and man- made ecosystems interaction, revealing anthropogenic impact to environment, and assessing its consequences [1,3]. Initial information used for ecosystems thematic mapping is acquired by means of terrestrial and remote monitoring techniques. The former characterize only 1 to 5% of surface and are not efficient to provide sufficient information on large eco- systems. Moreover, when detailed research is conducted, personnel, equipment, and time costs increase dramatically. Remote monitoring techniques provide a number of advantages over the terrestrial techniques, allowing the limitations of the latter to be overcome. In the literature, the concept of remote monitoring or surveying is referred to as remote sensing (RS) [4]. The RS techniques involve detecting and measuring electromagnetic radiation or force fields associated with terrestrial objects located beyond the immediate vicinity of recording instruments, such as radiometers or radar systems mounted on an aircraft or satellite. Remote monitoring, unlike the terrestrial one, allows a large-scale ecosystem to be surveyed with a short repeat cycle. The latter in most cases is a crucial criterion for ecosystem-change research. Generally, RS data represent images much like photos of the sensed surfaces of the objects under surveillance, and in the literature, RS images are often referred to as aerospace imagery [5]. Recently, thematic mapping of ecosystems has been widely implemented through employing geographic information systems (GIS) characterized by advanced capabilities for spatial information storing, manipulating, and processing [6]. Modern GIS provide wide capabilities for both computer-aided thematic mapping and spatial analysis of mapped features and phenomena, allowing derivation of complex quan- titative characteristics indispensable for ecosystem conditions modeling and fore- casting. Commonly, GIS facilities are oriented mainly for vector data handling, while RS-based thematic mapping methodology requires supporting functions of raster image processing. This fact makes urgent the problem of developing efficient and highly integrated software means enabling GIS to implement aerospace imagery processing and facilitate the thematic mapping technologies with use of RS data. In this chapter, the methodology of RS-based thematic mapping is introduced. The implementation of the methodology is based on application of a vector GIS and original image processing and interpretation system “LandMapper” [7], developed at Tomsk Polytechnic University (TPU). The main distinction of the system from its counterparts is adaptive classification procedure (ACP), making the process of image interpretation more flexible and efficient in comparison with existing recog- nition techniques. The chapter considers the basic methodology of image processing and interpretation adopted in the “LandMapper” system and gives the results of its © 2006 by Taylor & Francis Group, LLC Advanced Remote Sensing Techniques for Ecosystem Data Collection 109 application for solving problems of mapping two anthropogenic ecosystems with the use of multispectral imagery acquired from the Russian satellite RESURS-O1. 7.2 RS-BASED THEMATIC MAPPING METHODOLOGY 7.2.1 G ENERAL C ONCEPT Today, thematic mapping technologies making use of RS monitoring data and mod- ern GIS-based tools are of great value, especially when significant interest is taken in research of various aspects of anthropogenic ecosystems. The wide range of anthropogenic issues that can be solved by means of RS-based thematic mapping involve urban areas monitoring [2], land use mapping, anthropogenic load of petro- leum-production territories assessment, snow cover surveying, and flood forecasting. Recently joint use of GIS and thematic maps designed with aerospace imagery proved to be an efficient approach to creating and employing comprehensive models of anthropogenic ecosystems that were indispensable for decision-making. Designing thematic maps with the use of RS imagery consists of a number of steps, including complicated processing of initial imagery, and is, as a rule, a nontrivial task to accomplish. Figure 7.1 illustrates the general scheme of thematic mapping of landscape ecosystems with use of remotely sensed images. According to Figure 7.1, in the methodology of RS-based thematic mapping, the stages of preliminary and thematic processing of imagery may be distinguished. FIGURE 7.1 Thematic mapping with use of remotely sensed imagery. Imagery preliminary processing Receiving ground station (imagery archive) Orbital segment Imagery thematic processing GIS analysis Radiometric and geometric corrections Rectification and georeferencing Interpretation Conversion of raster thematic classes into vector features Spatial analysis and quantitative estimation Forecasting and decision makingGIS modeling Radiochannel Imagery Rectified and georeferenced imagery Thematic maps Ancillary geoinformation Sample data © 2006 by Taylor & Francis Group, LLC 110 GIS for Sustainable Development Initially, imagery acquired from a satellite or aircraft is exposed to multilevel preliminary processing in order to make it usable for comprehensive analysis and facilitate transition from a simple raster image to a complex thematic map model. The preliminary processing involves solving the tasks of geometric and radiometric error correction. The tasks include compensation of radiometric distortion caused by atmospheric effect and instrumentation errors, correction of geometric distortion due to the earth curvature, rotation, and panoramic effect, noise reduction, image registration in a geographical coordinate system (georeferencing) through its recti- fication, and visual properties enhancement by histogram transformation [8]. The thematic and geometric information defining the application domain of the final thematic map is extracted at the stage of imagery thematic processing [5]. In thematic processing, very significant attention is paid to the image interpretation issue. Image interpretation provides revealing thematic knowledge about a studied ecosystem component and its spatial relationships by identifying image features and assigning them appropriate semantic information such as, for instance, landscape cover type. Commonly, two main approaches can be adopted for image interpretation. One is referred to as photointerpretation and involves a human analyst/interpreter extract- ing information by visual inspection of an RS image [5]. In practice, photointerpre- tation is a very laborious and time-consuming process, and its success depends mainly upon the analyst effectively exploiting the spatial and spectral elements present in the image product. Another approach involves the use of a computer to assign each pixel in the image semantic information (land cover type, vegetation, or soil class) based upon pixel attributes. This approach deals with the concept of automated image interpretation–classification. Commonly, the approach appears to be most efficient when applied to multispectral imagery [4] having several bands of data acquired in different not overlapped spectral ranges. In practice, classification is often carried out in so-called supervising mode, requiring the classification procedure to be trained beforehand. Training of the classification procedure relies upon selecting a set of representative elements (pixels) in the image for each informational class (land cover type) and forming training sets to be used further by the procedure as prototypes of extracted classes. Forming training data for supervised classification is one of the important issues in imagery thematic processing. This is carried out by gathering ancillary sample data that helps obtain a prior knowledge of the properties of ecosystem components present in RS imagery. Practically, sample data is acquired from different sources of information about the studied ecosystem — site visit data, topographic maps, air photographs, or even results of initial imagery photointerpretation. The final product of the thematic processing stage is a raster map, each pixel of which is labeled with an appropriate code (label) corresponding to a landscape thematic class. Thus, different groups of equally labeled pixels in a thematic map represent thematically uniform objects recognized in imagery by the classification procedure. Imagery thematic processing is followed by transferring the resultant thematic map into GIS, where it can be integrated with other data acquired from various informational sources, and comprehensive spatial analysis of the data can be con- ducted. Since many GIS software packages basically manipulate vector information, © 2006 by Taylor & Francis Group, LLC Advanced Remote Sensing Techniques for Ecosystem Data Collection 111 the stage of transferring a thematic map into GIS is performed through conversion of the raster map into a set of vector features thematically grouped in layers, each representing a specific class of ecosystem components — water surfaces, vegetation canopy, urban areas. The automated raster–vector conversion is not a straightforward procedure and is implemented by means of applying complex algorithms using “running window” and “tracing contour” principles as well as line generalization techniques [7]. In GIS the extracted vector features are assigned the additional attributive infor- mation. At that stage, the resultant vector thematic map is becoming a valuable informational model of the ecosystem. Such a model can be used efficiently for visualizing, measuring, and analyzing various characteristics of ecosystem compo- nents imaged in initial imagery. In cases when time-series RS imagery has been used for ecosystem thematic mapping, the resultant informational model allows acquiring knowledge for revealing trends of ecosystem change and forecasting its behavior. The RS-based thematic mapping methodology described above is quite common and may be readily adopted in anthropogenic ecosystem research. However, the methodology of RS imagery processing and further thematic analysis can be very specific and can differ considerably in various case studies. In the remainder of this discussion, the imagery thematic processing approach elaborated in the GIS labo- ratory of TPU is considered. 7.2.2 I MAGERY I NTERPRETATION A PPROACH The problem of automated imagery interpretation is still one of the most complicated among those of RS data processing. Among the general problems of automated RS data interpretation, that of efficient image classification techniques synthesis should be addressed. Classification efficiency is commonly defined by the accuracy and computational complexity of the recognition procedures that allow image objects to be categorized and depends on two main factors — conformity of classification decision rule and optimality of feature space. The statistical classification decision rule (CDR) may be represented as function m(X) allowing unambiguous assigning image pixels defined in P-dimensional feature space by respective feature vectors to one of M nonoverlapped classes . Commonly, m(X) returns the index of the class for which X member- ship was proved through finding the largest discriminate function φ i (X) defined for each class [9]. The overall efficiency of a statistical decision rule is determined by a priori knowledge of the imagery classes, classification optimality criterion R(m(X)), and type of discriminate functions adopted. For decision rule synthesis, it is common to employ a Bayesian approach to determining the discriminate functions calculated as a product of the class condi- tional probability density function (PDF) p(X|ω i ) and its a priori probability p(ω i ), with which class ω i membership of X can be guessed before classification [5]. The crucial parameter p(X|ω i ) used in the Bayesian rule may be estimated in different ways, allowing a few CDRs to be derived. The applicability of the derived CDRs Xxj P j == {} ,,1 ω i iM,,= ( ) 1 ω i iM,,= () 1 © 2006 by Taylor & Francis Group, LLC 112 GIS for Sustainable Development may differ, depending on feature vectors X distribution low, as well as the amount and quality of training data used for PDF estimations. The relatively fast parametric Bayesian CDR, making use of the Gaussian (normal) distribution hypothesis, pro- duces good results with only unimodal distributions, whereas nonparametric CDRs, being free of normality constraints, can be efficient with distributions of any form, but at the expense of great computational complexity. In other words, finding a universal CDR effective by accuracy and performance for an arbitrary RS imagery is a big concern. Endeavoring to solve the problem, an idea of adaptive classification approach has been proposed [7]. The approach is based upon employing a few CDRs in the classification procedure and an adaptive decision rule allowing an optimal CDR, in terms of accuracy and performance, to be chosen for classification. In the ACP, synthesis of m(X) rests upon adopting a Bayesian rule that makes use of an empirical risk minimization criterion, R(m(X)), showing the probability of wrong pixel clas- sification. In practice, a common approach for probabilistic description of RS image classes is making an assumption of normal form of PDF p(X|ω i ) for each of M classes and using Gaussian parametrical PDF estimate in the Bayesian decision rule given by: (7.1) in which is sample vector of means, and is sample covariance matrix of class ω i . The approach making use of the parametric estimate (1) is effective when probability distributions are unimodal and/or close to those of normal form that is usually achieved with large training sets. Practically, these constraints may not always be overcome due to lack of prior information and non-normal form of a class features distribution. In such cases, more accurate classification may be obtained with use of a nonparametric approach to multivariate conditional PDF p(X|ω i ) approximation. As a nonparametric estimate, the ACP employs the multivariate analog of Parzen function [10] given by: (7.2) in which n is the number of training samples, P is the number of features, c v is a smoothing parameter; and Φ(u) is a kernel function. It should be noted that the efficiency of the Bayesian approach depends on PDF estimation techniques requiring large training sets to be available. Practically, when the training set size is too small for PDF function to be estimated properly, a simpler decision rule of minimum distance is used by the ACP that does not utilize proba- bilistic description of the RS image classes. pX X X i P ii t i ωπ µ () = () −− () − − − − 2 1 2 2 12 1 ΣΣexp µµ i iM () =,,1 µ i Σ i pX n c xx c iiv i v P vv s v i ω () = − = − ∏ 1 1 Φ = = = ∏ ∑ v P s n i iM 1 1 1,, © 2006 by Taylor & Francis Group, LLC Advanced Remote Sensing Techniques for Ecosystem Data Collection 113 The adaptive decision rule includes a set of discriminate functions corresponding to Bayesian CDR with Gaussian PDF estimate (1), CDR with Parzen PDF estimate (2), and CDR adopting minimum distance principle, respectively. Assuming that φ*(X) is the most effective CDR, the adaptive decision rule m(φ*(X)) can be expressed as follows: (7.3) The adaptive decision rule (3) allows the ACP to choose the most accurate CDR φ*(X) of three functions φ 1 (X), φ 2 (X), φ 3 (X), using minimum empirical risk criterion. Ambiguity between those CDRs having relatively equal values of the parameter (different by any accepted measure of inaccuracy) is resolved through choosing the fastest one. Thus in the classification stage, the ACP reveals the most effective CDR by accuracy and performance for an imagery with arbitrary charac- teristics independently of training set size, and so doing the ACP adapts to the data to be classified, in order to obtain the most accurate results in the shortest time. Unfortunately, the adaptability principle employed in the ACP cannot predefine the overall efficiency of the procedure, since classification success also depends to a large extent upon optimality of the feature space used. Commonly, feature space of an RS imagery is formed by considering the intensity (brightness) values of its pixels in different bands of electromagnetic spectrum (in the case of multispectral imagery) as the components of a multidimensional feature vector. It has been shown that feature space formed by only spectral features allows obtaining accurate clas- sification results for the image areas with relatively uniform intensity distribution [11]; otherwise, the produced classification contains high-frequency noise caused by misclassified pixels. In some works [12] it has been proved that in a RS image the neighbor pixels are spatially correlated, which makes reasonable the idea of using information about pixel context for its classification. So self-descriptiveness of the spectral feature vectors can be improved through extending them with com- plementary components representing the image texture descriptors calculated within the context of the classified pixels. In order to account for image textural information, the ACP utilizes an extended feature space (EFS) when performing classification. The EFS is formed through calculating a textural component of initial image by means of Haralick’s textural analysis approach [12]. The initial image is sequentially scanned by running windows of odd size and textural feature sets are generated. The elements of each textural feature set are computed as the first and second statistical moments of intensity function of initial image pixels falling into current running window of odd size b × b. Since the textural feature sets computed with windows of different size do not contribute equally to discriminating the RS image classes, the ACP performs the feature selection procedure, improving computational efficiency of the EFS classi- fication. The procedure selects the features that are more significant (informative) φφ= () { 1 X , φφ 23 XX () () } , mX X RX i i φφ φ*:* arg min , () () () = () () {} =13 RXφ () () bbb Z×=…,( , , )35 XXXX TX TX TX ZZ TX = {} ×× ×33 55 ,, X TT T bb TX bb bb bb S ××× × =… {} 12 ,,, , (,, ) bZ=…35 © 2006 by Taylor & Francis Group, LLC 114 GIS for Sustainable Development for classification and excludes the rest, using the image classes pairwise separability criterion of Jeffries-Matusita [11]. An original particularity of the ACP is that, once the EFS is built, the further classification of its textural and spectral components is performed separately in an iterative manner. Classification starts from processing textural component of the EFS, in the course of which the different scale textural feature sets are classified sequentially in iterative manner, going from coarser feature sets (calculated in bigger running window) to finer ones. At every iteration, the classification results represent posterior probability maps [5] computed for current textural feature set . The probability maps acquired for feature set are transferred to the next iteration, to be used as prior probabilities for clas- sifying finer scale feature set . The iterations are repeated until the finest feature set is classified. The completion phase of the classification is processing of the spectral feature component of the EFS with use of posterior probability maps calculated at the stage of textural component processing. At each iteration while classifying the image, the ACP employs an adaptive decision rule, finding the best CDR for the data currently processed in order to obtain the most accurate classifi- cation in the fastest way. The principle of the EFS iterative processing adopted in the ACP allows the procedure to overcome the shortcomings of the traditional stacked vector approach for employing textural features for image classification, in which the extended feature vectors are formed by stacking textural and spectral features together [5]. Adopting this approach faces the problem of losing fine spatial details in the resultant thematic map, which makes the approach not very practical, whereas the EFS iterative pro- cessing preserves the finest details in the resultant thematic map. Thus, by employing extended feature space processed in an iterative manner and an adaptive decision rule, the ACP produces better classification results com- pared to traditional image interpretation techniques, as is shown in the following application examples. 7.3 THEMATIC MAPPING METHODOLOGY IMPLEMENTATION 7.3.1 T HE RS I MAGERY P ROCESSING AND I NTERPRETATION S YSTEM “L AND M APPER ” The thematic mapping methodology based on improved imagery interpretation approach has been implemented in the framework of the “LandMapper” system of imagery processing and interpretation developed in the GIS laboratory of TPU. The “LandMapper” system is a software package, which is launched as an additional unit for a vector GIS (MapInfo Professional®, MapInfo Corporation, Troy, New York) providing it with image processing functionality. The general structure of the As can be seen from Figure 7.2, “LandMapper” is based upon vector-raster architecture comprised of two components, Raster (RC) and Vector (VC), respec- tively. The RC provides means for raster data visualization in a GIS environment and implements functions of RS imagery preliminary and thematic processing. The XXXX TX TX TX ZZ TX = {} ×× ×33 55 ,, X bb TX × X bb TX × X bb TX ()( ) − × −22 © 2006 by Taylor & Francis Group, LLC “LandMapper” system is given in Figure 7.2. Advanced Remote Sensing Techniques for Ecosystem Data Collection 115 FIGURE 7.2 General structure of the “LandMapper” system. Subsystem of preliminary processing Subsystem of thematic processing User interface Subsystem of spatial analysis Subsystem of raster data visualizing Subsystem of data exchange Raster component Vector component GIS MapInfo Professional 5.0 Subsystem of raster – vector conversion Subsystem of vector data visualizing and editing © 2006 by Taylor & Francis Group, LLC [...]... null; } } Source: Adapted from Zipf, A and Merdes, M., AGILE Conference Proceedings, Lyon, France, 20 03 © 20 06 by Taylor & Francis Group, LLC GIS for Sustainable Development 1 3 4 6 7 9 10 13 15 16 18 19 21 22 24 25 26 27 30 Spatiotemporal Data Modeling for “4D” Databases 137 Geometry and Feature Therefore, new parents are being introduced into both classes in line 16 These implement the interface TimeDependent... geo-objects of urban areas over historic epochs and act as a basis for the data management components of temporal 3D -GIS (“3D-TGIS” or more colloquial “4DGIS”) to be developed in the future Since the temporal part of this model is a selfconsistent OO-model for temporal structures, it can also be used with 2D geodata The proposed framework is a contribution toward the development of a temporal 3D -GIS. .. only 3D information available for a part of a 3D object or for parts of a facade, this can be expressed by the modeler through the usage of the hierarchical structure of the spatial model by using 3-Cells that only use a 2- Cell (face) The geometry of a 3D object is represented through the geometries (3-cells) of the parts of the 3D object 8.5 AN OBJECT-ORIENTED MODEL FOR TEMPORAL DATA The object-oriented... framework is depicted in Figure 8 .2 The data model introduced so far describes the topology of up to three-dimensional objects The actual geometrical data is integrated by relating multiple versions © 20 06 by Taylor & Francis Group, LLC 126 GIS for Sustainable Development B B B C C G H A A D F A A E D FIGURE 8.3 Graphical representation of the primitives 0-Cell, 1-Cell, 2- Cell and 3-Cell of geometry to the... SPIE, Bellingham, 20 02, 489 12 Haralick, R and Joo, H., A Context Classifier, in IEEE Trans Geoscience Remote Sensing, N24, 1986, 997 © 20 06 by Taylor & Francis Group, LLC 8 Spatiotemporal Data Modeling for “4D” Databases Alexander Zipf CONTENTS 8.1 8 .2 8.3 8.4 8.5 Introduction 123 Spatiotemporal Data Modeling 124 Topological Modeling of Three-Dimensional Geo-Objects 124 Modeling of... has been demanded that a temporal GIS (TGIS) needs to provide functionality for spatiotemporal data storage, data handling, and analysis as well as visualization These functions are usually more complex than in conventional GIS and are still an area of active research 123 © 20 06 by Taylor & Francis Group, LLC 124 GIS for Sustainable Development Within the Deep Map /GIS project a flexible and extensive... should include and perform the analysis of attributes, geometry, and topology equally well For example, there is a lack of research on the possible changes of topological relationships in “4D” space-time in particular within vector-oriented GIS Similarly the sparse availability of functions for inter- and extrapolation within vector-oriented “4D” -GIS is not satisfying Because existing GIS use proprietary... Relational Databases, Research Report 24 6 Department of Informatics, University of Oslo, 1997 9 Skjellaug, B., Temporal Data: Time and Object Databases, Research Report 24 5 April 1997 Department of Informatics, University of Oslo, 1997 10 Skjellaug, B and Berre, A.-J., Multi-dimensional Time Support for Spatial Data Models, Research Report 25 3, May 1997 Department of Informatics, University of Oslo, 1997... Breunig, M., On the way to component-based 3D/4D geoinformation systems Lecture Notes in Earth Sciences, Vol 94, Springer Heidelberg, 20 00 12 Langran, G., Time in Geographic Information Systems, Taylor & Francis, London, 19 92 13 Wachowicz, M., Object-Oriented Design for Temporal GIS, Taylor & Francis, London, 1999 14 Worboys, M., Object-oriented approaches to geo-referenced information, Int J Geogr Inf Syst.,... 335–343 20 Zipf, A and Krüger, S., Ein objektorientierter Framework für temporale 3D-Geodaten AGIT 20 01, Symposium für Angewandte Geographische Informationsverarbeitung, 04–06 Juli 20 01, Salzburg Austria, 20 01 21 Zipf, A and Merdes, M., Is Aspect-Oriented Programming a new paradigm for GIS development? On the relationship of geoobjects, aspects and ontologies, AGILE Conference Proceedings, Lyon, France, 20 03 . components of temporal 3D -GIS (“3D-TGIS” or more colloquial “4D- GIS ) to be developed in the future. Since the temporal part of this model is a self- consistent OO-model for temporal structures,. Three-Dimensional Geo-Objects 124 8.4 Modeling of Thematic Data: The Example of the History of a City 126 8.5 An Object-Oriented Model for Temporal Data 128 8.5.1 Temporal Structure 129 8.5 .2 Temporal. bed map. © 20 06 by Taylor & Francis Group, LLC 120 GIS for Sustainable Development and forest). The class representing forest area was split up into four subclasses: pine sphagnous forest, mixed