Guy Thoonen

Contextual Subpixel Mapping of Hyperspectral Images Making Use of a High Resolution Color Image

This paper describes a hyperspectral image classification method to obtain classification maps at a finer resolution than the images original resolution. We assume that a complementary color image of high spatial resolution is available. The proposed methodology consists of a soft classification procedure to obtain landcover fractions, followed by a subpixel mapping of these fractions. While the main contribution of this article is in fact the complete multisource framework for obtaining a subpixel map, the major novelty of this subpixel mapping approach is the inclusion of contextual information, obtained from the color image. Experiments, conducted on two hyperspectral images and one real multi source data set, show excellent results, when compared to classification of the hyperspectral data only. The advantage of the contextual approach, compared to conventional subpixel mapping approaches, is clearly demonstrated.

Classification of heathland vegetation in a hierarchical contextual framework

Heathlands in Western Europe have shown dramatic declines over the last century and therefore have been given a high conservation priority in the Habitats Directive of the European Union (EU). Accurate surveying and monitoring of heathland habitats is essential for appropriate conservation management, but the large heterogeneity of vegetation types within habitats as well as the occurrence of similar vegetation across habitat types hinders a straightforward, automated mapping based on aerial images. In such a case, a context-dependent classification algorithm is expected to be superior to traditional classification techniques. This article presents a novel approach to map the conservation status of heathland vegetation by using a hierarchical classification scheme that describes the structural dependencies in the field between the basic vegetation and the land-cover types that habitats are composed of. These dependency relationships are included as contextual information in the classification process, using a tree-structured Markov random field (TS-MRF) technique with a tree that reflects the hierarchy of the classification scheme. Results of this approach for a heathland area in Belgium were compared with results from more conventional classification approaches. Validation of the results showed that the structure of the scheme contained important spatial relationships, which were further reinforced by using the contextual classification strategy, especially for the most detailed level of the classification scheme. Accuracy increased and the classification results were more suitable for visual interpretation.

Automatic threshold selection for morphological attribute profiles

In this article, an automatized procedure for selecting informative values of the thresholds, essential for the construction of morphological attribute profiles, is proposed. To this end, connected component analysis is performed on a preliminary supervised or unsupervised classification result that does not involve contextual information. Subsequently, after extracting the relevant attributes from each of the connected components, the threshold values are found by grouping the attribute vectors using a clustering algorithm. In our experiments, we demonstrate the effect of image scaling on the selected thresholds. In addition, we show the advantage of using our automatic threshold selection approach with respect to manual selection, by both monitoring redundancy and performing a classification experiment.

Accuracy assessment of contextual classification results for vegetation mapping

Abstract A new procedure for quantitatively assessing the geometric accuracy of thematic maps, obtained from classifying hyperspectral remote sensing data, is presented. More specifically, the methodology is aimed at the comparison between results from any of the currently popular contextual classification strategies. The proposed procedure characterises the shapes of all objects in a classified image by defining an appropriate reference and a new quality measure. The results from the proposed procedure are represented in an intuitive way, by means of an error matrix, analogous to the confusion matrix used in traditional thematic accuracy representation. A suitable application for the methodology is vegetation mapping, where lots of closely related and spatially connected land cover types are to be distinguished. Consequently, the procedure is tested on a heathland vegetation mapping problem, related to Natura 2000 habitat monitoring. Object-based mapping and Markov Random Field classification results are compared, showing that the selected Markov Random Fields approach is more suitable for the fine-scale problem at hand, which is confirmed by the proposed procedure.

Automatic forensic analysis of automotive paints using optical microscopy

The timely identification of vehicles involved in an accident, such as a hit-and-run situation, bears great importance in forensics. To this end, procedures have been defined for analyzing car paint samples that combine techniques such as visual analysis and Fourier transform infrared spectroscopy. This work proposes a new methodology in order to automate the visual analysis using image retrieval. Specifically, color and texture information is extracted from a microscopic image of a recovered paint sample, and this information is then compared with the same features for a database of paint types, resulting in a shortlist of candidate paints. In order to demonstrate the operation of the methodology, a test database has been set up and two retrieval experiments have been performed. The first experiment quantifies the performance of the procedure for retrieving exact matches, while the second experiment emulates the real-life situation of paint samples that experience changes in color and texture over time.

Habitat reporting of a heathland site: Classification probabilities as additional information, a case study

Abstract Heathlands are man-made habitats and their decline during the last century can be contributed to shifts in both agricultural and management practices as well as to hydrological and atmospheric changes. As a result, many heathland sites, including the Kalmthoutse Heide in Belgium, were included in the European Natura 2000 program, a network of protected areas across the European Union. To assure an accurate mapping of the Kalmthoutse Heide and other Natura 2000 sites in Belgium a classification framework for habitat status reporting with remote sensing data and in particular high resolution hyperspectral imagery was started. In this study we propose a simple and fast context based method for mapping heathland heterogeneity using the intermediate, otherwise redundant, classification probabilities as generated by a hard classification algorithm. Our study proved to be successful in using intermediate classification probabilities as a valuable source of ecological information. The delineated areas have been shown to be statistically sound and robust compared to a neutral model. The technique is not limited to a particular hard classification technique and can easily be adopted into current vegetation monitoring efforts. The resulting maps provided accessible maps which can support management of the protected site and enhance the accuracy of EU reportage as required by the habitat directive.

Spatial hyperspectral image classification by prior segmentation

In this paper, we propose a technique to incorporate spatial features in the classification of hyperspectral data by means of a prior segmentation of the dataset. The key idea of the technique is that each pixel is not classified individually, but that the regions obtained from the prior segmentation are classified as a whole. The proposed technique is validated on a hyperspectral dataset of a heathland area in Belgium. Experimental results show that we can achieve larger and spatially smoothed regions, while the overall classification success rate is comparable to the pure spectral classification results.

Domain adaptation with Hidden Markov Random Fields

In this paper, we propose a method to match multitemporal sequences of hyperspectral images using Hidden Markov Random Fields. Based on the matching of the data manifold, the algorithm matches the reflectance spectra of the classes, thus allowing the reuse of labeled examples acquired on one image to classify the other. This allows valorization of spectra collected in situ to other acquisitions than the one they were acquired for, without user supervision, prior knowledge of the class reflectance in the new domain or global information about atmospheric conditions.

Two-stage fusion of thermal hyperspectral and visible RGB image by PCA and guided filter

Nowadays, advanced technology in remote sensing allows us to get multi-sensor and multi-resolution data from the same region. Fusion of these data sources for classification remains challenging problems. In this paper, we propose a novel algorithm for hyperspectral (HS) image pansharpening with two-stage guided filtering in PCA (principal component analysis) domain. In the first stage, we first downsample the highresolution RGB image to the same spatial resolution of original low-resolution HS image, and use guided filter to transfer the image details (e.g. edge) of the downsampled RGB image to the original HS image in the PCA domain. In the second stage, we perform upsampling on the resulting HS image from the first stage by using original high-resolution RGB image and guided filter in PCA domain. This yields a clear improvement over an older approach with one stage guided filtering in PCA domain. Experimental results on fusion of a low spatial-resolution Thermal Infrared HS image and a high spatial-resolution visible RGB image from the 2014 IEEE GRSS Data Fusion Contest, are very encouraging.

Subpixel mapping of hyperspectral data using high resolution color images

This article introduces a method that uses information from a high spatial resolution color image to perform subpixel mapping of a lower spatial resolution hyperspectral image. The method uses a modified majority voting approach, that makes a distinction between pure and mixed pixels, in order to combine a classification map of the hyperspectral data with a segmentation map of the color image. Experiments, conducted on two hyperspectral images, show excellent results, when compared to conventional classification of the hyperspectral data only. The developed classification scheme results in much higher accuracies and leads to large visual improvements, including well-defined class transitions and fairly homogeneous objects.