Skin cancer is among the most prevalent and deadly types of cancer. Dermatologists mostly use visual cues to diagnose this illness. The classification of multiclass skin cancer is challenging due to the fine-grained variability in the appearance of its several diagnostic categories. Advances in deep neural networks have led to a significant increase in skin lesion classification methods in recent years, with more performante models like vision transformers (ViTs) emerging and improving skin lesion classification performance. Our explainable skin cancer classification with ViT and Grad-CAM (XSC-ViT) is presented in this paper. This uses a specifically designed ViT to classify skin cancer. Our solution integrates L2 regularization in some ViT layers to prevent overfitting and applies the Grad-CAM (Gradient-weighted Class Activation Mapping) method for decision explainability. Our approach demonstrated remarkable performance results for skin cancer classification with an accuracy of 92%, precision of 91.73%, recall of 91.46%, F1 score of 91.51%, and loss of 22.05% when tested on the HAM10000 dataset.

The emergence of developed hyperspectral sensors has improved the spectral and spatial resolutions of the images, leading to a high-level understanding of the remotely sensed scenes. Several methods have been proposed to exploit those resolutions and to extract representative features serving for diverse applications, like classification. For instance, contextual classifiers provide better accuracy using features that represent the spatial context, and not restricting themselves to the raw spectral observations that the image offers. In a previous work, we proposed a method that uses the 3D-Shearlet Transform features along with the spatial and spectral information to better classify a given HSI. We got interesting accuracies that outperformed state-of-the-art methods. In this paper, we propose not to use all the 3D-Shearlet Transform coefficients, but only the most relevant of them. To do so, we resort to use a feature selection method based on the Fisher Discrimination Criterion (FDC). Experimental results showed that the FDC helped select the most representative cubes and outperformed our previously obtained results.

In this study, an effective method used to examine the changes of two optical images captured by Landsat satellite is presented. The proposed method is based on two main parts: A preprocessing step where Shearlet Transform is applied to get a smoother rendering followed by a classification process using Convolutional Neural Network (CNN) to change detection.The proposed method out performed the state-of-the-art methods with an accuracy out of 99,32%.

Local Fisher discriminant analysis (LFDA) is a feature extraction technique that proved efficient to reduce several types of data and succeeded to outperform many state-of-the-art methods. However, due to its supervised nature, LFDA’s efficiency depends on the available labeled samples and declines dramatically when the latter are very few. Hence, we assume that we cannot resort to LFDA to reduce unlabeled data. In this article, we studied to what extent this assumption is true and questioned the possibility of using LFDA to reduce hyperspectral images (HSIs) with no available ground truth. To study the real impact of the labeled information on LFDA’s performance, we replaced the costly expert-made ground truth by different sets of labeled samples that are generated based on the image’s offered spectral and/or spatial information, with no prior knowledge of the captured scene nor of its classes. Our proposed sets proved able to guide LFDA in extracting relevant discriminating features. This proved that LFDA does not depend only on expert-made labeled information and led us to define the unsupervised image-adapted LFDA (uiaLFDA) that can properly reduce an HSI without requiring its ground truth. To do so and to replace the ground truth that LFDA usually requires to reduce an HSI, uiaLFDA defines its own set of labeled samples by simply gridding the image into cells where each cell is considered a class. Our experiments ran on three HSIs proved that uiaLFDA is as efficient as LFDA and, even better, in reducing unlabeled HSIs.

International audience

This paper addresses vegetation monitoring in European agricultural areas using Earth Observation satellites. Due to the small size of typical European fields, two complementary sensors are used, SPOT and NOAA-AVHRR, bringing the spatial and the temporal information respectively. A sub-pixel analysis of NOAA data using one SPOT image is performed to characterize fields with high spatial and temporal resolutions. To be used in an operational context, the method must have realistic data requirements. We define an operational scenario making use of only one SPOT image per site and a one year NOAA sequence, covering a large part of Europe. We first proceed to an unsupervised segmentation of the SPOT image; the NOAA data analysis on test sites provides the temporal evolution of vegetation; then, identification of fields is performed by minimizing a cost function measuring the similarity between the global reflectance observed on NOAA pixels and the reflectance computed from corresponding regions at SPOT resolution.

Describes a fusion process between two different data sources, one providing an accurate spatial information, the other providing time series with a much coarser spatial scale. It is applied in the following remote sensing context: the forecast of cereals production, which is a challenging application of the new generation of Earth observation satellites. These two data types are required since agronomical models must be fed with a daily sampling of cereals reflectances, and since in Europe, fields have a relatively small size. SPOT-XS is wed to provide spatial information at the parcels level, a meso-scale sensor (here, NOAA-AVHRR), which outputs images of large areas every day, provides the temporal information. The combination of these two data sources makes it possible to daily estimate reflectances of main cultivations at the parcels level. The selected approach is as follows: a preliminary learning stage provides the reflectances of each type of cultivation; then operational scenarios are defined to apply the learning information in order to estimate statistics on large areas: using only one SPOT-XS image and meso-scale daily images, a fusion scheme makes it possible to obtain land use identification at high spatial resolution with its temporal behavior.

This paper is concerned with three main problems of image processing occuring on temporal sequences of satellite oceanographic images: approximate localization of the interesting structures on the images; segmentation or determination of the boundary of the structures; and temporal tracking of these boundaries to illustrate their evolution. In this paper, application is done on vortices and results are displayed all over the paper. For this purpose, we propose a new method for optical flow computation and interpretation and a geometric modelling of the structures. Oceanographic images obtained from environmental satellite platforms present a new challenge in computer science. The huge amount of data collected each day and the need for characterizing some specific structures on these images for oceanographic monitoring justify our approach for the detection and tracking of vortices on oceanographic images.