Claude Cariou

Hyperspectral image segmentation using a cooperative nonparametric approach

In this paper a new unsupervised nonparametric cooperative and adaptive hyperspectral image segmentation approach is presented. The hyperspectral images are partitioned band by band in parallel and intermediate classification results are evaluated and fused, to get the final segmentation result. Two unsupervised nonparametric segmentation methods are used in parallel cooperation, namely the Fuzzy C-means (FCM) method, and the Linde-Buzo-Gray (LBG) algorithm, to segment each band of the image. The originality of the approach relies firstly on its local adaptation to the type of regions in an image (textured, non-textured), and secondly on the introduction of several levels of evaluation and validation of intermediate segmentation results before obtaining the final partitioning of the image. For the management of similar or conflicting results issued from the two classification methods, we gradually introduced various assessment steps that exploit the information of each spectral band and its adjacent bands, and finally the information of all the spectral bands. In our approach, the detected textured and non-textured regions are treated separately from feature extraction step, up to the final classification results. This approach was first evaluated on a large number of monocomponent images constructed from the Brodatz album. Then it was evaluated on two real applications using a respectively multispectral image for Cedar trees detection in the region of Baabdat (Lebanon) and a hyperspectral image for identification of invasive and non invasive vegetation in the region of Cieza (Spain). A correct classification rate (CCR) for the first application is over 97% and for the second application the average correct classification rate (ACCR) is over 99%.

Application of unsupervised nearest-neighbor density-based approaches to sequential dimensionality reduction and clustering of hyperspectral images

In this communication, we address the problem of unsupervised dimensionality reduction (DR) for hyperspectral images (HSIs), using nearest-neighbor density-based (NN-DB) approaches. Dimensionality reduction is an important tool in the HSI processing chain, aimed at reducing the high redundancy among the HSI spectral bands, while preserving the maximum amount of relevant information for further processing. Basically, the idea is to formalize DR as the process of partitioning the spectral bands into coherent band sets. Two DR schemes can be set up directly, one based on band selection, and the other one based on band averaging. Another scheme is proposed here, based on compact band averaging. Experiments are conducted with hyperspectral images composed of an AISA Eagle HSI issued from our acquisition platform, and the AVIRIS Salinas HSI. We evaluate the efficiency of the reduced HSIs for final classification results under the three schemes, and compare them to the classification results without reduction. We show that despite a high dimensionality reduction (< 8% of the bands left), the clustering results provided by NN-DB methods remain comparable to the ones obtained without DR, especially for GWENN in the band averaging case. We also compare the classification results obtained after applying other unsupervised or semi-supervised DR schemes, based either on band selection or band averaging, and show the superiority of the proposed DR scheme.

Nearest neighbor-density-based clustering methods for large hyperspectral images

We address the problem of hyperspectral image (HSI) pixel partitioning using nearest neighbor - density-based (NN-DB) clustering methods. NN-DB methods are able to cluster objects without specifying the number of clusters to be found. Within the NN-DB approach, we focus on deterministic methods, e.g. ModeSeek, knnClust, and GWENN (standing for Graph WatershEd using Nearest Neighbors). These methods only require the availability of a k-nearest neighbor (kNN) graph based on a given distance metric. Recently, a new DB clustering method, called Density Peak Clustering (DPC), has received much attention, and kNN versions of it have quickly followed and showed their efficiency. However, NN-DB methods still suffer from the difficulty of obtaining the kNN graph due to the quadratic complexity with respect to the number of pixels. This is why GWENN was embedded into a multiresolution (MR) scheme to bypass the computation of the full kNN graph over the image pixels. In this communication, we propose to extent the MR-GWENN scheme on three aspects. Firstly, similarly to knnClust, the original labeling rule of GWENN is modified to account for local density values, in addition to the labels of previously processed objects. Secondly, we set up a modified NN search procedure within the MR scheme, in order to stabilize of the number of clusters found from the coarsest to the finest spatial resolution. Finally, we show that these extensions can be easily adapted to the three other NN-DB methods (ModeSeek, knnClust, knnDPC) for pixel clustering in large HSIs. Experiments are conducted to compare the four NN-DB methods for pixel clustering in HSIs. We show that NN-DB methods can outperform a classical clustering method such as fuzzy c-means (FCM), in terms of classification accuracy, relevance of found clusters, and clustering speed. Finally, we demonstrate the feasibility and evaluate the performances of NN-DB methods on a very large image acquired by our AISA Eagle hyperspectral imaging sensor.

Dense registration of CHRIS-Proba and Ikonos images using multi-dimensional mutual information maximization

We investigate the potential of multidimensional mutual information for the registration of multi-spectral remote sensing images. We devise a gradient flow algorithm which iteratively maximizes the multidimensional mutual information with respect to a differentiable displacement map, accounting for partial derivatives of the multivariate joint distribution and the multivariate marginal of the float image with respect to each variable of the mutual information derivative. The resulting terms are shown to weight the band specific gradients of the warp image, and we propose in addition to compute them with a method based on the k-nearest neighbours. We apply our method to the registration of Ikonos and CHRIS-Proba images over the region of Baabdat, Lebanon, for purposes of cedar pines detection. A comparison between (crossed) single band and multi-band registration results obtained shows that using the multidimensional mutual information brings a significant gain in positional accuracy and is suitable for multispectral remote sensing image registration.

Affinity propagation for large size hyperspectral image classification

The affinity propagation (AP)1 is now among the most used methods of unsupervised classification. However, it has two major disadvantages. On the one hand, the algorithm implicitly controls the number of classes from a preference parameter, usually initialized as the median value of the similarity matrix, which often gives over-clustering. On the other hand, when partitioning large size hyperspectral images, its computational complexity is quadratic and seriously hampers its application. To solve these two problems, we propose a method which consists of reducing the number of individuals to be classified before the application of the AP, and to concisely estimate the number of classes. For the reduction of the number of pixels, a pre-classification step that automatically aggregates highly similar pixels is introduced. The hyperspectral image is divided into blocks, and then the reduction step is applied independently within each block. This step requires less memory storage since the calculation of the full similarity matrix is not required. The AP is then applied on the new set of pixels which are then set up from the representatives of each previously formed cluster and non-aggregated individuals. To estimate the number of classes, we introduced a dichotomic method to assess classification results using a criterion based on inter-class variance. The application of this method on various test images has shown that AP results are stable and independent to the choice of the block size. The proposed approach was successfully used to partition large size real datasets (multispectral and hyperspectral images).