Corina Vaduva

Latent Dirichlet Allocation for Spatial Analysis of Satellite Images

This paper describes research that seeks to supersede human inductive learning and reasoning in high-level scene understanding and content extraction. Searching for relevant knowledge with a semantic meaning consists mostly in visual human inspection of the data, regardless of the application. The method presented in this paper is an innovation in the field of information retrieval. It aims to discover latent semantic classes containing pairs of objects characterized by a certain spatial positioning. A hierarchical structure is recommended for the image content. This approach is based on a method initially developed for topics discovery in text, applied this time to invariant descriptors of image region or objects configurations. First, invariant spatial signatures are computed for pairs of objects, based on a measure of their interaction, as attributes for describing spatial arrangements inside the scene. Spatial visual words are then defined through a simple classification, extracting new patterns of similar object configurations. Further, the scene is modeled according to these new patterns (spatial visual words) using the latent Dirichlet allocation model into a finite mixture over an underlying set of topics. In the end, some statistics are done to achieve a better understanding of the spatial distributions inside the discovered semantic classes.

A Latent Analysis of Earth Surface Dynamic Evolution Using Change Map Time Series

With a continuous increase in the number of Earth Observation satellites, leading to the development of satellite image time series (SITS), the number of algorithms for land cover analysis and monitoring has greatly expanded. This paper offers a new perspective in dynamic classification for SITS. Four similarity measures (correlation coefficient, Kullback-Leibler divergence, conditional information, and normalized compression distance) based on consecutive image pairs from the data are employed. These measures employ linear dependences, statistical measures, and spatial relationships to compute radiometric, spectral, and texture changes that offer a description for the multitemporal behavior of the SITS. During this process, the original SITS is converted to a change map time series (CMTS), which removes the static information from the data set. The CMTS is analyzed using a latent Dirichlet allocation (LDA) model capable of discovering classes with semantic meaning based on the latent information hidden in the scene. This statistical method was originally used for text classification, thus requiring a word, document, corpus analogy with the elements inside the image. The experimental results were computed using 11 Landsat images over the city of Bucharest and surrounding areas. The LDA model enables us to discover a wide range of scene evolution classes based on the various dynamic behaviors of the land cover. The results are compared with the Corinne Land Cover map. However, this is not a validation method but one that adds static knowledge about the general usage of the analyzed area. In order to help the interpretation of the results, we use several studies on forms of relief, weather forecast, and very high resolution images that can explain the wide range of structures responsible for influencing the dynamic inside the resolution cell.

Semantic Map Generation from Satellite Images for Humanitarian Scenarios Applications

This paper demonstrates how knowledge driven methods and the associated data analysis algorithms are changing the paradigms of user-data interactions, providing an easier and wider access to the Earth Observation data. Some information theory based algorithms are proposed for anomaly and change detection on SPOT images, relative to a widespread humanitarian crisis scenario: floods. The outcomes of these algorithms define an informational representation of the image, revealing the spatial distribution of a particular theme. Using image analysis and interpretation, the multitude of features from a scene are classified into meaningful classes to create sematic maps.

Feature Extraction for Patch-Based Classification of Multispectral Earth Observation Images

Recently, various patch-based approaches have emerged for high and very high resolution multispectral image classification and indexing. This comes as a consequence of the most important particularity of multispectral data: objects are represented using several spectral bands that equally influence the classification process. In this letter, by using a patch-based approach, we are aiming at extracting descriptors that capture both spectral information and structural information. Using both the raw texture data and the high spectral resolution provided by the latest sensors, we propose enhanced image descriptors based on Gabor, spectral histograms, spectral indices, and bag-of-words framework. This approach leads to a scene classification that outperforms the results obtained when employing the initial image features. Experimental results on a WorldView-2 scene and also on a test collection of tiles created using Sentinel 2 data are presented. A detailed assessment of speed and precision was provided in comparison with state-of-the-art techniques. The broad applicability is guaranteed as the performances obtained for the two selected data sets are comparable, facilitating the exploration of previous and newly lunched satellite missions.

Data mining and spatial reasoning for satellite image characterization

High level image understanding and content extraction requires image regions analysis to reveal the spatial interaction between them. This paper aims to engender new attributes for scene description considering the relative position of the objects inside. A visual grammar of the scene is built using an extension for a Knowledge Based Image Information Mining system (KIM). The objects are extracted using statistical models and machine learning through the KIM system, according to the user interest. Further, an affine invariant descriptor of the relative position between two objects is computed. This is the force histogram and it is considered to be a spatial signature which characterizes configurations of regions based on the attraction forces between the composing objects. Thereby, new patterns could be defined using similar object configurations, in order to enhance the effectiveness of the content-based image retrieval inside large databases.

Semantic annotation of high resolution TerraSAR-X images using Information Mining

This paper addresses the problem of High Resolution Synthetic Aperture Radar (SAR) image semantic annotation using a Knowledge Based Information Mining (KIM) System. The authors propose the assessment of the capabilities of KIM to perform an automatic urban classification on TerraSAR-X data. Four test sites have been used in the experiment to prove that the system is generic and data independent. The performance is evaluated by matching the results with GeoEye optical representations of the selected areas. For the evaluation a number of three classes are presented and discussed (water bodies, green urban areas and tall buildings).

Temporal analysis of SAR imagery for permanent and evolving Earth land cover behavior assessment

In the era of constantly increasing Earth Observation (EO) data collections, information extraction and data analysis should be enhanced with a multi-temporal component enabled by the temporal resolution of satellite missions and create handy, yet powerful tools for those applications involving monitoring of land cover. The image time series, as results of the satellite revisiting period, gives you insights not only on a certain area, but also on its representation at different moments of time. In order to limit the issues that might arise due to irregular time sampling of multispectral data, the authors propose a Synthetic Aperture Radar (SAR) image time series for analysis. To this point, the main goal is to mine the satellite image time series (SITS) for understanding the temporal behaviour of an area in terms of evolution and persistency. The paper introduces an analytical approach, combining coherent and no coherent analysis of SAR SITS content. We propose the Latent Dirichlet Allocation model to extract categories of evolution from the SAR SITS and techniques which study statistical and coherent proprieties of the targets to identify the structures with stable electromagnetic characteristics over time, named Persistent Scatterers (PS). The obtained results indicate an evolutionary character hidden inside the persistent class. The results obtained on 30 ERS images encourages further analysis on Sentinel 1 data.

Compound and configurable framework for exploratory earth observation data analysis

The lack of a comprehensive solution for image information mining has often brought confusion and misunderstanding when Earth Observation data based application scenarios were addressed. Considering the variety of dedicated sensors available nowadays, the particularities of the recorded data raises serious issues when explored. Most of the proposed methodologies for data analysis integrate algorithms able to cope with single cases. In order to overcome this limitation, the present paper introduce a compound, configurable framework containing two processing levels, for feature extraction and image classification, that allows different settings depending on the application being handled. The design was proposed such that it facilitates the integration of several methods and algorithm for each level, including a module to serve for validation when reference data is available. The approach is not complete without the interaction with the user, therefore, a human-machine communication strategy was also developed. The validation was performed through a prototype system meeting all the criteria of the defined framework.

A new comprehensive approach for earth observation scene classification using joint image and text analysis

The amount of data available on the internet provides massive additional information for the Earth Observation (EO) imagery. Periodical news, various reports and measurements, pictures or online encyclopedias are just few examples of the existent information. Occasionally, this data offers new perspectives for EO image understanding and interpretation. However, current image analysis do not benefit from the advantage given by external sources. To overcome these drawbacks, the present paper proposes an approach that goes beyond traditional information mining by using a joint image and text analysis. Fast Compression Distance (FCD) is computed to measure the similarities inside a collection of very high resolution images and text files. The main purpose is to discover common patterns within the data, without any a priori assumption, parameter-free, relying on data compression-based techniques. A hierarchical clustering is performed in order to learn about the dependencies between different types of data.

A Framework for Benchmarking of Feature Extraction Methods in Earth Observation image Analysis

In the last few years, thanks to projects like TELEIOS, the linked open data cloud has been rapidly populated with geospatial data some of it describing Earth Observation products (e.g., CORINE Land Cover, Urban Atlas). The abundance of this data can prove very useful to the new missions (e.g., Sentinels) as a means to increase the usability of the millions of images and EO products that are expected to be produced by these missions. In this paper, we explain the relevant opportunities by demonstrating how the process of knowledge discovery from TerraSAR-X images can be improved using linked open data and Sextant, a tool for browsing and exploration of linked geospatial data, as well as the creation of thematic maps.Dimensionality reduction for visualization is widely used in visual data mining where the data is represented by high dimensional features. However, this leads to have an unbalanced and occluded distribution of visual data in display space giving rise to difficulties in browsing images. In this paper, we propose an approach to the visualization of images in a 3D display space in such a way that: (1) images are not occluded and the provided space is used efficiently; (2) similar images are positioned close together. An immersive virtual environment is employed as a 3D display space. Experiments are performed on an optical image dataset represented by color features. A library of dimensionality reduction is employed to reduce the dimensionality to 3D. The results confirm that the proposed technique can be used in immersive visual data mining for exploring and browsing large-scale datasets.In this paper, we evaluate sample selection strategies based on optimum experimental design for SAR image classification. Traditionally, support vector machine active learning is widely used by selecting the samples close to the decision surface. Recently, new methods based on optimum experimental design have been developed. To gain a complete understanding of these selection strategies, a comparative study on three approaches, transductive experimental design, manifold adaptive experimental design and locally linear reconstruction, has been performed for SAR image classification using different features. Among the three approaches,we show that manifold adaptive experimental design performs best and stably in terms of both accuracy and computational complexity.Large volume of detailed features of land covers, provided by High-Resolution Earth Observation (EO) images, has attracted the interests to assess the discovery of these features by Content-Based Image Retrieval systems. In this paper, we perform Latent Dirichlet Allocation (LDA) on the Bag-of-Words (BoW) representation of collections of EO images to discover their high-level features, so-called topics. To assess the discovered topics, the images are represented based on the occurrence of different topics, we name it Bag-of-Topics (BoT). Then, the BoT model is compared to the BoW model of images based on the given human-annotations of the data. In our experiments, we compare the classification accuracy resulted by BoT and BoW representations of two different EO datasets, a Synthetic Aperture Radar (SAR) dataset and a multi-spectral satellite dataset. Moreover, we provide isualizations of feature space for better perceiving the changes in the discovered information by BoT and BoW models. Experimental results demonstrate that the dimensionality of the data can be reduced by BoT representation of images; while it either causes no significant reduction in the classification accuracy or even increase the accuracy by sufficient number of topics.In the context of Earth Observation (EO), image information retrieval systems have gained importance as a way to explore terabytes of archive data. Concurrently, evaluation of these systems becomes a topic. Evaluation has typically been conducted in the form of metrics such as Precision Recall measures, with more recent approaches attempting to include the user in the evaluation process. This paper presents a more user centered evaluation of a CBIR tool in an EO context. The evaluation methodology involved open ended user feedback, which was then inductively categorized, and its distribution and content were analyzed. Results are presented, with conclusions indicating certain aspects of the user experience cannot be obtained from metrics alone, but can be complementary to metrics.This paper presents SAR patch categorization based on feature descriptors within the dual tree complex wavelet transform using non-parametric features, which were estimated for each wavelet based subband, which was additionally transformed using a Fourier transform. Spectral properties of wavelet transform were characterized using thefirst and second moments, Kolmogorov Sinai entropy and coding gainwithin an oriented dual tree complex wavelet transform (2D ODTCWT). A database with 2000 images representing 20 different classes with 100 images per class was used for estimation of classification efficiency. A window size for estimation feature parameters was estimated. A supervised learning stage was implemented with support vector machine using 10 % and 20% of the test images per class. The experimental results showed that the non-parametric features achieved 94.3 % accuracy, when 20 % of database was used for supervised training.This paper presents an application of visual data mining technique to Earth-Observation images for exploring very large image archives. We present a visual data mining workstation solution and create some use cases in order to demonstrate its functionality. This tool allows interactive exploration and analysis of very large, high complexity, and non-visual data sets stored into a database by using human-machine communication. The tool relies on image processing components that transform the image content to primitive feature vectors and a graphical user interface, which allows the exploration of the entire image archive. The use cases are based on Synthetic Aperture Radar images, digital orthophotos and photos in-situ.α-trees provide a hierarchical representation of an image into partitions of regions with increasing heterogeneity. This model, inspired from the single-linkage paradigm, has recently been revisited for grayscale images and has been successfully used in the field of remote sensing. This article shows how this representation can be adapted to more complex data here hyperspectral images, according to different strategies. We know that the measure of distance between two neighbouring pixels is a key element for the quality of the underlying tree, but usual metrics are not satisfying. We show here that a relevant solution to understand hyperspectral data relies on the prior learning of the metric to be used and the exploitation of domain knowledge.The multitude of sensors used to acquire Earth Observation (EO) images have led to the creation of an extremely various collection of data. Along with appropriate methods able to work with great amount of data, informat ion retrieval process requires algorithms to cope with a range of input imagery. Even if the geometry and the manner of creating Synthetic Aperture Radar (SAR) images are totally different than multispectral data, there are attempts of finding a common ground such that optical image indexing algorithms can be applied for SAR data and vice versa. Moreover, new concepts must be defined in order to obtain satisfying results, enabling measurements and comparisons between the extracted features [4]. Regarding this idea, the goal is to develop an application capable to join feature extraction algorithms and classification algorithms . Its success will sustain the integration of a reliable EO data search engine. This paper presents a framework for feature extraction and classification aiming to support EO image annotation. Weber Local Descriptors (WLD), Gabor filter and Support Vector Machine (SVM) are combined in order to define an application to be tested on both SAR and optical data.We introduce a map algebra based on a cochain extension of the Linear Algebraic Representation (LAR), used to efficiently represent and query geometric and physical information through sparse matrix algebra. LAR, based on standard algebraic topology methods, supports all incidence structures, including enumerative (images), decompositive (meshes) and boundary (CAD) representations, is dimension-independent and not restricted to regular complexes. This algebraic representation enjoys a neat mathematical format— being based on chains, the domains of discrete integration, and cochains, the discrete prototype of differential forms, so naturally integrating the geometric shape with the supported physical properties, and provides a mechanism for strongly typed representation of all physical quantities associated with images. It is easy to show that k-cochains form a linear vector space over k-cells, which means that they can used as basic objects in a rich and virtually unlimited calculus of physical properties.In this paper, we present a knowledge-driven content-based information mining system for data fusion in Big Data. The tool combines, at pixel level, the unsupervised clustering results of different number of features, extracted from different image types, with a user given semantic concepts in order to calculate the posterior probability that allows the final search. The system is able to learn different semantic labels based on Bayesian networks and retrieve the related images with only a few user interactions, greatly optimizing the computational costs and over performing existing similar systems in various orders of magnitude.