Satellite Images Computer Vision

Table of Contents

Expertise

We are part of a large consortium of universities and companies with advantage expertise in processing satellite images with computer vision and artificial intelligence. Our expertise spans many areas:

Available Dataset Knowhow

We know existing datasets, their strengths and deficiencies. As an example consider a newly announced dataset (worldstrat) that allow the development of machine learning algorithms of various kind, as superresolution. This allows us to bring together our expertise in advanced deep learning algorithms and apply it to satellite images. From the original paper (open review link), (PDF Downdload)

The largest and most varied such publicly available dataset, at Airbus SPOT 6/7 satellites’ high resolution of up to 1.5 m/pixel, empowered by European Space Agency’s Phi-Lab as part of the ESA-funded QueryPlanet project, we curate 10,000 sq km of unique locations to ensure stratified representation of all types of land-use across the world: from agriculture to ice caps, from forests to multiple urbanization densities. We also enrich those with locations typically under-represented in ML datasets: sites of humanitarian interest, illegal mining sites, and settlements of persons at risk. We temporally-match each high-resolution image with multiple low-resolution images from the freely accessible lower-resolution Sentinel-2 satellites at 10 m/pixel.

This is just one of multiple dataset that we use in our research with our research partners. In Figure 1 you can see a visual summary of the construction of the dataset.
Figure 1: Structure of the construction of the worldstrat database. Image taken from [https://openreview.net/pdf?id=DEigo9L8xZA](https://openreview.net/pdf?id=DEigo9L8xZA).

As an example here some recent useful and interesting datasets that can be used for machine learning and generally any computer vision task (list partially adapted from github link). Images and demos are taken from the dataset websites to give an example of the possibilities.

Recent Satellite Image Datasets

Object Detection Datasets

  • Airbus Aircraft Detection (Airbus, Mar 2021)
    Aircraft bounding boxes, 103 images of worlwide airports (Pleiades, 0.5m res., 2560px).

Figure 1: An example of counting vehicles. Image taken from https://gdo152.llnl.gov/cowc/mundhenk_et_al_eccv_2016.pdf.

Semantic Segmentation Datasets

3. Semantic Segmentation

  • FloodNet (University of Maryland, Jun 2021)
    2343 image chips (drone imagery), 10 landcover categories (background, water, building flooded, building non-flooded, road-flooded, …). Paper: Rahnemoonfar et al., 2021

  • LoveDA (Wuhan University, Oct 2021)
    5987 image chips (Google Earth), 7 landcover categories, 166768 labels, 3 cities in China. Paper: Wang et al., 2021

  • Sentinel-2 Cloud Mask Catalogue (Francis, A., et al., Nov 2020) 513 cropped subscenes (1022x1022 pixels) taken randomly from entire 2018 Sentinel-2 archive. All bands resampled to 20m, stored as numpy arrays. Includes clear, cloud and cloud-shadow classes. Also comes with binary classification tags for each subscene, describing what surface types, cloud types, etc. are present.

  • Open Cities AI Challenge (GFDRR, Mar 2020) .
    790k building footprints from Openstreetmap (2 label quality categories), aerial imagery (0.03-0.2m resolution, RGB, 11k 1024x1024 chips, COG format), 10 cities in Africa.

  • DroneDeploy Segmentation Dataset (DroneDeploy, Dec 2019)
    Drone imagery (0.1m res., RGB), labels (7 land cover catageories: building, clutter, vegetation, water, ground, car) & elevation data, baseline model implementation.

  • SkyScapes: Urban infrastructure & lane markings (DLR, Nov 2019)
    Highly accurate street lane markings (12 categories e.g. dash line, long line, zebra zone) & urban infrastructure (19 categories e.g. buildings, roads, vegetation). Aerial imagery (0.13 m res.) for 5.7 km2 of Munich, Germany. Paper: Azimi et al. 2019

  • SpaceNet 5: Automated Road Network Extraction & Route Travel Time Estimation (CosmiQ Works, Maxar, Intel, AWS, Sep 2019)
    2300 image chips, street geometries with location, shape and estimated travel time, 3/8band Worldview-3 imagery (0.3m res.), 4 global cities, 1 holdout city for leaderboard evaluation, APLS metric, baseline model

  • SEN12MS (TUM, Jun 2019)
    180,748 corresponding image triplets containing Sentinel-1 (VV&VH), Sentinel-2 (all bands, cloud-free), and MODIS-derived land cover maps (IGBP, LCCS, 17 classes, 500m res.). All data upsampled to 10m res., georeferenced, covering all continents and meterological seasons, Paper: Schmitt et al. 2018

  • Slovenia Land Cover Classification (Sinergise, Feb 2019)
    10 land cover classes, temporal stack of hyperspectral Sentinel-2 imagery (R,G,B,NIR,SWIR1,SWIR2; 10 m res.) for year 2017 with cloud masks, Official Slovenian land use land cover layer as ground truth.

Research Partnership for Fast Technology Customisation

We partner with researchers and companies with a highly advanced expertise in satellite image analysis. In the figure below you can see an example of an application developed by Graniot that has uses satellite images to do precision agriculture and can extract information from images and compare them between different moments of time. We work with Graniot to adapt this technology to different use cases. We can customise it to your use cases, by starting from a sound and working base.
Figure 2: The interface of the web app that Graniot has developed for precision agriculture. Image courtesy of Graniot.

Their technology allow a comparison of the images and the extracted information at different moments in time (depending on the availablility of the images of course). In Figure 3 you can see for example comparison of ground humidity of a piece of land.
Figure 2: Comparison of ground humidity at two different points in time. Image courtesy of Graniot.