Satellite Image Classification

Nearly 10,000 km² of free high-resolution and paired multi-temporal low-resolution satellite imagery of unique locations which ensure stratified representation of all types of land-use across the world: from agriculture to ice caps, from forests to multiple urbanization densities. ​ Those locations are also enriched with typically under-represented locations in ML datasets: sites of humanitarian interest, illegal mining sites, and settlements of persons at risk. ​ Each high-resolution image (Airbus SPOT at up to 1.5 m/pixel) comes with multiple temporally-matched low-resolution images from the freely accessible lower-resolution Sentinel-2 satellites (up to 10 m/pixel, 12 spectral bands). ​ The dataset is accompanied with a paper, datasheet for datasets and an open-source Python package to: rebuild or extend the WorldStrat dataset, train and infer baseline algorithms, and learn with abundant tutorials, all compatible with the popular EO-learn toolbox. The hope is to foster broad-spectrum applications of ML to satellite imagery, and possibly develop the same power of analysis allowed by costly private high-resolution imagery from free public low-resolution Sentinel2 imagery. We illustrate this specific point by training and releasing several highly compute-efficient baselines on the task of Multi-Frame Super-Resolution.

The appearance of the world varies dramatically not only from place to place but also from hour to hour and month to month. Every day billions of images capture this complex relationship, many of which are associated with precise time and location metadata. We propose to use these images to construct a global-scale, dynamic map of visual appearance attributes. Such a map enables fine-grained understanding of the expected appearance at any geographic location and time. Our approach integrates dense overhead imagery with location and time metadata into a general framework capable of mapping a wide variety of visual attributes. A key feature of our approach is that it requires no manual data annotation. We demonstrate how this approach can support various applications, including image-driven mapping, image geolocalization, and metadata verification.

In this competition you will be identifying regions in satellite images that contain certain cloud formations, with label names: Fish, Flower, Gravel, Sugar. For each image in the test set, you must segment the regions of each cloud formation label. Each image has at least one cloud formation, and can possibly contain up to all all four. The images were downloaded from NASA Worldview. Three regions, spanning 21 degrees longitude and 14 degrees latitude, were chosen. The true-color images were taken from two polar-orbiting satellites, TERRA and AQUA, each of which pass a specific region once a day. Due to the small footprint of the imager (MODIS) on board these satellites, an image might be stitched together from two orbits. The remaining area, which has not been covered by two succeeding orbits, is marked black. The labels were created in a crowd-sourcing activity at the Max-Planck-Institite for Meteorology in Hamburg, Germany, and the Laboratoire de météorologie dynamique in Paris, France. A team of 68 scientists identified areas of cloud patterns in each image, and each images was labeled by approximately 3 different scientists. Ground truth was determined by the union of the areas marked by all labelers for that image, after removing any black band area from the areas. The segment for each cloud formation label for an image is encoded into a single row, even if there are several non-contiguous areas of the same formation in an image. If there is no area of a certain cloud type for an image, the corresponding EncodedPixels prediction should be left blank. You can read more about the encoding standard on the Evaluation page. Files train.csv - the run length encoded segmentations for each image-label pair in the train_images train_images.zip - folder of training images test_images.zip - folder of test images; your task is to predict the segmentations masks of each of the 4 cloud types (labels) for each image. IMPORTANT: Your prediction masks should be scaled down to 350 x 525 px. sample_submission.csv - a sample submission file in the correct format

The standard evaluation protocol of Cross-View Time dataset allows for certain cameras to be shared between training and testing sets. This protocol can emulate scenarios in which we need to verify the authenticity of images from a particular set of devices and locations. Considering the ubiquity of surveillance systems (CCTV) nowadays, this is a common scenario, especially for big cities and high visibility events (e.g., protests, musical concerts, terrorist attempts, sports events). In such cases, we can leverage the availability of historical photographs of that device and collect additional images from previous days, months, and years. This would allow the model to better capture the particularities of how time influences the appearance of that specific place, probably leading to a better verification accuracy. However, there might be cases in which data is originated from heterogeneous sources, such as social media. In this sense, it is essential that models are optimized on camera-disjoint sets to avoid learning sensor-specific characteristics that might not generalize accordingly for new imagery during inference. With this in mind, we propose a novel organization for CVT dataset. We split available data into training and testing sets, ensuring that all images from a single camera are assigned to the same set. During this division, we aimed to keep the size of each set roughly similar to the original splits, allowing models to be optimized with similar amounts of data.

The DRIFT dataset includes 25k image patches collected in five European countries sourced from aerial and nanosatellite image archives. Each image patch is associated with three target variables to predict: 1. Canopy height: average height value for pixels containing woody vegetation. 2. Tree count: number of overstory (visible from an overhead perspective) trees in the images. 3. Tree cover fraction: percentage of the image being covered by overstory tree crowns. The DRIFT dataset includes significant shifts between label and visual distributions due to sensor and area differences. Furthermore, vegetation tends to grow to fit the local climate, therefore introducing concept drift in the data: same tree species may appear differently in different subsets. The label distribution also varies among different subsets (countries). The dataset is a good choice for: * image-level regression * domain adaption for regression * remote sensing for forest applications

The HRPlanesv2 dataset contains 2120 VHR Google Earth images. To further improve experiment results, images of airports from many different regions with various uses (civil/military/joint) selected and labeled. A total of 14,335 aircrafts have been labelled. Each image is stored as a ".jpg" file of size 4800 x 2703 pixels and each label is stored as YOLO ".txt" format. Dataset has been split in three parts as 70% train, %20 validation and test. The aircrafts in the images in the train and validation datasets have a percentage of 80 or more in size. Link: https://github.com/dilsadunsal/HRPlanesv2-Data-Set

Read more about the dataset here: https://github.com/ServiceNow/seasonal-contrast

UV6K is a high-resolution remote sensing urban vehicle segmentation dataset. - Images: 6,313 - Vehicle: 245,141 - Resolution: 0.1m - Image Size: 1024x1024