A key limitation of supervised learning is the ability to handle data from unknown distributions. Often, such methods fail when presented with samples from a source not represented in the training data. This work proposes an effective way of controlling the behavior of a neural network in the presence of out-of-distribution examples. For this, the training dataset is supplemented with extraneous data assigned to an additional out-of-distribution class. The extraneous data may come from a different dataset or be even noise. By applying a Gaussian mixture model on the latent representation, and by taking advantage of the ability of these models to generalize well, the method described thereafter performs well. Training the model on a segregated dataset helps the model to distinguish out-of-distribution data, including the ones the model were never confronted to during training.

Change detection on satellite image series is a much more complex problem than detecting changes in color or illumination. First, it is subject to common problems inherent to image series: images are generally not well aligned, changes in atmospheric conditions impact the colors and overall contrast of images, and clouds can partially or totally obstruct the areas of interest. Secondly, noise in satellite images also causes changes in images that can disrupt change detection algorithms, especially on low contrast areas. In this paper, we study the noise characteristics of PlanetScope images and their impact on change detection. We then present a general algorithm, published online, that performs change detection robust to noise.

ESA's Sentinel- 2 satellites have been in orbit for five years, acquiring huge amounts of data all over the world. They are a formidable tool for mass detection as they are freely available. Given their importance in the energetic transition and their spread over countries or continents, wind turbines are natural candidates for such studies. We propose an automatic wind turbine detector for low resolution satellite images based on an a contrario approach, exploiting the geometry of wind turbines' shadows and hubs. Our experiments show promising detection rates, improving the state of the art in the proposed conditions.

Unsupervised learning based on Contrastive Learning (CL) has attracted a lot of interest recently. This is due to excellent results on a variety of subsequent tasks (especially classification) on benchmark datasets (ImageNet, CIFAR-10, etc.) without the need of large quantities of labeled samples. This work explores the application of some of the most relevant CL techniques on a large unlabeled dataset of aerial images of building rooftops. The task that we want to solve is roof type classification using a much smaller labeled dataset. The main problem with this task is the strong dataset bias and class imbalance. This is caused by the abundance of certain types of roofs and the rarity of other types. Quantitative results show that this issue heavily affects the quality of learned representations, depending on the chosen CL technique.

The energy sector is a key industry in the global economy and monitoring oil storage provides valuable insights into the economic state of a country. Our aim is to detect oil tank farms as accurately as possible using Sentinel-2 images. An a contrario clustering method is used to group by density the result of a circle detection step. Then, a patch-match procedure is used to complete the tank detection. Although most existing methods are designed to work on high-resolution images, the proposed method is designed for low-resolution images; we also propose an adaptation to high-resolution images.

Neural network methods are nowadays used for a wide variety of tasks, in particular in computer vision. Among those tasks, semantic segmentation aims at labeling every pixel in an image, giving a good understanding of the scene. In this paper, we propose to use two neural network architectures designed for semantic segmentation to detect oil tank depots in Sentinel-2 images. We compare the methods to an unsupervised algorithm designed to solve the same problem.

There has been this last decade a major improvement in earth observation imagery, both in terms of resolution and revisit rate. This technological leap comes with an increasing demand for detecting objects in satellite images. In this context, vehicle detection generates a great deal of interest in the scientific and industrial community. Indeed, detecting vehicles can have a range of applications from traffic monitoring to parking lots occupancy rate estimation. This interest is demonstrated in the wide range of public vehicle datasets that already exist. However, these datasets rely on the use of aerial images, acquired from planes, that are different in many aspects from satellite images. Machine learning algorithms trained on these datasets are therefore likely to underperform on satellite images. Therefore, we introduce a large-scale dataset for VEHicle detection on SATellite images (VehSat). To this end, we collected 4544 crops of satellite images, from 4 different satellites and on 8 different areas. In total, 36851 vehicles were annotated in various contexts and resolutions. To build a baseline for vehicle detection in satellite images, we also evaluate state-of-the-art object detection algorithms on VehSat. Results show that, although performing better than human annotators on crowdsourcing platforms, machine learning algorithms still have a lot of room for improvement due to the challenging nature of the dataset.

This paper presents a method for estimating the occupancy ratio of parkings lots from satellite images. The algorithm takes as input a series of PlanetScope images along with a mask indicating where the parking is positioned and returns for each image an occupancy ratio. The method is generic, doesn't require any calibration and can easily be extended. For validating our results, we have created a PlanetScope image series dataset that associates each image to a ground occupancy ratio. This ground truth is estimated thanks to a camera that permanently films and records the parking. We observe a strong correlation between the estimated occupancy rate and the ground truth occupancy rate. Qualitative analysis shows that our method correctly generalizes to other parkings.

Monitoring oil stocks provides valuable insights on the balance between production and demand of petroleum products. The identification of oil depots is important for estimating storage capacities and measuring oil stocks. To achieve this purpose, we present an oil tank detector. As oil tanks are generally circular, we use as a first step a circle detection algorithm. However, this approach tends to generate a considerable amount of false detections, as circular shapes are not necessarily tanks. To reduce these false detections, we take advantage of the fact that oil tanks are generally densely grouped together, and filter out isolated detections using a clustering algorithm. The clustering method uses the a contrario framework which gives a way to control the number of false detections per image. The method is illustrated on Sentinel-2 images.

These last years, earth observation imagery has significantly improved. Public satellites such as WorldView-3 can now produce images with a Ground Sample Distance of 31cm, reaching a resolution equivalent to aerial images. Perhaps more importantly, the revisit frequency has also been greatly enhanced: providers such as Planet can now acquire images of a given ground location on a daily basis. These major improvements are fueled by an increasing demand for frequent objects detection. An application generating a particular interest is vehicle detection. Indeed, vehicle detection can give to public and private actors valuable data such as traffic monitoring and parking occupancy rate estimations. Several datasets, such as DOTA or VehSat, already exist, allowing researchers to train machine learning algorithms to detect vehicles. However, these datasets focus on relatively high definition and expensive aerial and satellite images. In this paper, we will present a method for detecting vehicles on medium resolution satellite images, with a GSD comprised between 1 and 5 meters. This approach can notably be used on Planet images, allowing to monitor traffic of an area on a daily basis.