Statistical background modeling and applications in remote sensing

This thesis addresses some problems in remote sensing and statistical background modeling using the a-contrario theory and artificial neural network models. The work is motivated by two applications in remote sensing where the control of false alarms is an issue due to the large-scale nature of the problems.The identification and measuring of oil depots using satellite images has important commercial and strategic value. Here is addressed the former: the detection of oil storage sites while analyzing vast quantities of satellite images covering a whole country or continent. In this context, to be useful an algorithm needs to achieve a high recall while controlling the number of false detections and with a reduced computational cost. The first method proposed here starts by detecting circular objects, which are then clustered using the a-contrario framework; indeed, oil depots often correspond to a dense group of cylindrical buildings. The approach is completed by an a-contrario patch-matching procedure to recover the missing tanks. Since the method relies heavily on the detection of circular objects, several algorithms for detecting circles in low-resolution satellite images are compared. The a-contrario algorithm is also compared to two neural-network architectures for oil depot segmentation.The second remote sensing application is the detection of hot spots in the daytime using multi-band satellite images that do not have thermal bands. This allows the monitoring of the activity of oil refineries, cement works, and steel mills as well as the activities of volcanoes.The first proposed method is based on an anomaly detection algorithm.A second approach relies on measuring the fitness of the measured radiances in the selected spectral bands to the black body model.Finally, this thesis deals with out-of-distribution (OOD) detection in deep learning methods. A first approach is to supplement the training dataset with extraneous data assigned to an additional out-of-distribution class. Training the model on a segregated dataset helps the model to discriminate out-of-distribution samples, including those the model was never exposed to during training. An unsupervised approach to OOD detection is also presented. For that, a new neural network layer is proposed that enforces a Gaussian embedding for each class. Using this new layer, each target class can be represented by a Gaussian distribution. New samples are then evaluated as belonging to a target class or not by performing a chi-square test for each one. Samples rejected by all tests are considered OOD. This methodology also allows to identify ambiguous samples when validated by more than one class.