This document primarily lists resources for performing deep learning (DL) on satellite imagery. To a lesser extent Machine learning (ML, e.g. random forests, stochastic gradient descent) are also discussed, as are classical image processing techniques.
- https://github.com/chrieke/awesome-satellite-imagery-datasets
- A modern geospatial workflow
- geospatial-machine-learning
- Datasets
- Online computing resources
- Interesting dl projects
- Production
- Image formats and catalogues
- State of the art
- Online platforms for Geo analysis
- Techniques
- Useful references
- As part of the EU Copernicus program, multiple Sentinel satellites are capturing imagery -> see wikipedia.
- Sentinel-hub provides access to a range of Sentinel data and may be the best overall source of imagery + data.
- Descartes Labs Platform hosts sentinel data
- Playground to explore data
- raw data - requester pays
- Paid access via sentinel-hub and python-api.
- GBDX also has Sentinel imagery.
- Example loading sentinel data in a notebook
- Long running US program -> see Wikipedia
- Imagery on GCP, see the GCP bucket here, with imagery analysed in this notebook on Pangeo
- https://en.wikipedia.org/wiki/WorldView-3
- 0.3m PAN, 1.24 MS, 3.7m SWIR
- Some publicly available data here used in the 3D modelling notebook here.
Kaggle hosts several large satellite image datasets (> 1 GB). A list if general image datasets is here. A list of land-use datasets is here. The kaggle blog is an interesting read.
Each sample image is 28x28 pixels and consists of 4 bands - red, green, blue and near infrared. The training and test labels are one-hot encoded 1x6 vectors. Each image patch is size normalized to 28x28 pixels. Data in .mat Matlab format. JPEG?
- Sat4 500,000 image patches covering four broad land cover classes - barren land, trees, grassland and a class that consists of all land cover classes other than the above three Example notebook
- Sat6 405,000 image patches each of size 28x28 and covering 6 landcover classes - barren land, trees, grassland, roads, buildings and water bodies.
- https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/data
- 3-5 meter resolution GeoTIFF images
- 12 classes including - cloudy, primary + waterway etc
- 1st place winner interview - used 11 custom CNN
- https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection
- Rating - medium, many good examples (see the Discussion as well as kernels), but as this competition was run a couple of years ago many examples use python 2
- 3 (i.e. RGB) and 16-band (400nm - SWIR) images
- 45 satellite images covering 1km x 1km in both 3-band and 16-band formats.
- 10 Labelled classes include - Buildings, Road, Trees, Crops, Waterway, Vehicles
- Interview with 1st place winner who used segmentation networks - 40+ models, each tweaked for particular target (e.g. roads, trees)
- Deepsense 4th place solution
- https://www.kaggle.com/c/airbus-ship-detection/overview
- Rating - medium, most solutions using deep-learning, many kernels, good example kernel.
- I believe there was a problem with this dataset, which led to many complaints that the competition was ruined.
- https://www.kaggle.com/c/draper-satellite-image-chronology/data
- Rating - hard. Not many useful kernels.
- Images are grouped into sets of five, each of which have the same setId. Each image in a set was taken on a different day (but not necessarily at the same time each day). The images for each set cover approximately the same area but are not exactly aligned.
- Kaggle interviews for entrants who used XGBOOST and a hybrid human/ML approach
- Satellite + loan data -> https://www.kaggle.com/reubencpereira/spatial-data-repo
There are a variety of datasets suitable for land classification problems.
- http://weegee.vision.ucmerced.edu/datasets/landuse.html
- This is a 21 class land use image dataset meant for research purposes.
- There are 100 RGB TIFF images for each class
- Each image measures 256x256 pixels with a pixel resolution of 1 foot
- Landsat -> free viewer at remotepixel and libra
- Optical, radar, segmented etc. https://aws.amazon.com/earth/
- SpaceNet and article here. Also example semantic segmentation using Raster Vision
- Several people have uploaded datasets to Quilt
- https://developers.google.com/earth-engine/
- Various imagery and climate datasets, including Landsat & Sentinel imagery
- Python API but all compute happens on Googles servers
- UK met-odffice -> https://www.metoffice.gov.uk/datapoint
- NASA (make request and emailed when ready) -> https://search.earthdata.nasa.gov
- NOAA (requires BigQuery) -> https://www.kaggle.com/noaa/goes16/home
- Time series weather data for several US cities -> https://www.kaggle.com/selfishgene/historical-hourly-weather-data
Generally a GPU is required for DL. Googles colab is free but limited compute time (12 hours) and somewhat non persistent,
- Free to use
- GPU Kernels (may run for 1 hour which limits usefulness?)
- Tensorflow, pytorch & fast.ai available
- Advantage that many datasets are already available
- Read
### Clouderizer
- https://clouderizer.com/
- Clouderizer is a cloud computing management service, it takes care of installing the required packages to a cloud computing instance (like Amazon AWS or Google Colab). Clouderizer is free for 200 hours per month (Robbie plan) and does not require a credit card to sign up.
- Run projects locally, on cloud or both.
- SSH terminal, Jupyter Notebooks and Tensorboard are securely accessible from Clouderizer Web Console.
- GPU available
- https://aws.amazon.com/ec2/?ft=n
- GPU available (link?)
- Focus on CNTK?
- https://azure.microsoft.com/en-us/free/?b=16.45
- https://docs.microsoft.com/en-us/azure/machine-learning/preview/scenario-aerial-image-classification
- ML engine - sklearn, tensorflow, keras
- Collaboratory (notebooks with GPU as a backend for free for 12 hours at a time),
- Tensorflow available
- pytorch can be installed, useful articles
- https://www.floydhub.com/
- Cloud GPUs
- Jupyter Notebooks
- Tensorboard
- Version Control for DL
- Deploy Models as REST APIs
- Public Datasets
- https://www.paperspace.com/
- 1-Click Jupyter Notebooks
- GPU on demand
- Python API
- https://www.crestle.com/
- Cloud GPU & persistent file store
- Fast.ai lessons pre-installed
- https://www.azavea.com/projects/raster-vision/
- An open source Python framework for building computer vision models on aerial, satellite, and other large imagery sets.
- Accessible through the Raster Foundry
- Example use cases on open data
- https://github.com/mapbox/robosat
- Generic ecosystem for feature extraction from aerial and satellite imagery.
- A fork of robotsat
- https://github.com/datapink/robosat.pink
- https://github.com/trailbehind/DeepOSM
- Train a deep learning net with OpenStreetMap features and satellite imagery.
- https://github.com/nshaud/DeepNetsForEO
- Uses SegNET for working on remote sensing images using deep learning.
- https://github.com/developmentseed/skynet-data
- Data pipeline for machine learning with OpenStreetMap
- Basic https://blog.keras.io/building-a-simple-keras-deep-learning-rest-api.html with code here
- Advanced https://www.pyimagesearch.com/2018/01/29/scalable-keras-deep-learning-rest-api/
- https://github.com/galiboo/olympus
- https://www.tensorflow.org/serving/
- Official version is python 2 but python 3 build here
- Another approach is to use Docker
TensorFlow Serving makes it easy to deploy new algorithms and experiments, while keeping the same server architecture and APIs. Multiple models, or indeed multiple versions of the same model, can be served simultaneously. TensorFlow Serving comes with a scheduler that groups individual inference requests into batches for joint execution on a GPU
- Allows exposing model via rest API
- https://modeldepot.io
- ML models hosted
- We certainly want to consider cloud optimised GeoTiffs https://www.cogeo.org/
- https://terria.io/ for pretty catalogues
- Remote pixel
- Sentinel-hub eo-browser
- Large datasets may come in HDF5 format, can view with -> https://www.hdfgroup.org/downloads/hdfview/
- Climate data is often in netcdf format, which can be opened using xarray
- The xarray docs list a number of ways that data can be stored and loaded.
- Specification describing the layout of a catalogue comprising of static files. The aim is that the catalogue is crawlable so it can be indexed by a search engine and make imagery discoverable, without requiring yet another API interface.
- An initiative of https://www.radiant.earth/ in particular https://github.com/cholmes
- Spec at https://github.com/radiantearth/stac-spec
- Browser at https://github.com/radiantearth/stac-browser
- Talk at https://docs.google.com/presentation/d/1O6W0lMeXyUtPLl-k30WPJIyH1ecqrcWk29Np3bi6rl0/edit#slide=id.p
- Example catalogue at https://landsat-stac.s3.amazonaws.com/catalog.json
- Chat https://gitter.im/SpatioTemporal-Asset-Catalog/Lobby
- Several useful repos on https://github.com/sat-utils
What are companies doing?
- Overall trend to using AWS S3 backend for image storage. There are a variety of tools for exploring and having teams collaborate on data on S3, e.g. T4.
- Bucking the trend, Descartes & Airbus are using a google backend -> checkout gcsts for google cloud storage sile-system
- Just speculating, but a serverless pipeline appears to be where companies are headed for routine compute tasks, whilst providing a Jupyter notebook approach for custom analysis.
- Traditional data formats aren't designed for processing, so new standards are developing such as cloud optimised geotiffs and zarr
- This article discusses some of the available platforms -> TLDR Pangeo rocks
- Pangeo - open source resources for parallel processing using Dask and Xarray http://pangeo.io/index.html
- Descartes Labs -> access to EO imagery from a variety of providers via python API -> not clear which imagery is available (Airbus + others?) or pricing
- DigitalGlobe have a cloud hosted Jupyter notebook platform called GBDX. Cloud hosting means they can guarantee the infrastructure supports their algorithms, and they appear to be close/closer to deploying DL. Tutorial notebooks here. Only Sentinel-2 and Landsat data on free tier.
- Planet have a Jupyter notebook platform which can be deployed locally and requires an API key (14 days free). They have a python wrapper (2.7..) to their rest API. No price after 14 day trial.
This section explores the different techniques (DL, ML & classical) people are applying to common problems in satellite imagery analysis. Classification problems are the most simply addressed via DL, object detection is harder, and cloud detection harder still (niche interest).
- Very common problem, assign land classification to a pixel based on pixel value, can be addressed via simple sklearn cluster algorithm or deep learning.
- Land use is related to classification, but we are trying to detect a scene, e.g. housing, forestry. I have tried CNN -> See my notebooks
- Land Use Classification using Convolutional Neural Network in Keras
- Sea-Land segmentation using DL
- Pixel level segmentation on Azure
- Deep Learning-Based Classification of Hyperspectral Data
- A U-net based on Tensorflow for objection detection (or segmentation) of satellite images - DSTL dataset but python 2.7
- Monitor water levels, coast lines, size of urban areas, wildfire damage. Note, clouds change often too..!
- Using PCA (python 2, requires updating) -> https://appliedmachinelearning.blog/2017/11/25/unsupervised-changed-detection-in-multi-temporal-satellite-images-using-pca-k-means-python-code/
- Using CNN -> https://github.com/vbhavank/Unstructured-change-detection-using-CNN
- Siamese neural network to detect changes in aerial images
- https://www.spaceknow.com/
- LANDSAT Time Series Analysis for Multi-temporal Land Cover Classification using Random Forest
- Change Detection in 3D: Generating Digital Elevation Models from Dove Imagery
- Change Detection in Hyperspectral Images Using Recurrent 3D Fully Convolutional Networks
- PySAR - InSAR (Interferometric Synthetic Aperture Radar) timeseries analysis in python
- Wikipedia article on registration -> register for change detection or image stitching
- Traditional approach -> define control points, employ RANSAC algorithm
- Phase correlation used to estimate the translation between two images with sub-pixel accuracy, useful for allows accurate registration of low resolution imagery onto high resolution imagery, or register a sub-image on a full image -> Unlike many spatial-domain algorithms, the phase correlation method is resilient to noise, occlusions, and other defects. Applied to Landsat images here.
- A typical task is detecting boats on the ocean, which should be simpler than land based challenges owing to blank background in images, but is still challenging and no convincing robust solutions available.
- Intro articles here and here.
- DigitalGlobe article - they use a combination classical techniques (masks, erodes) to reduce the search space (identifying water via NDWI which requires SWIR) then apply a binary DL classifier on candidate regions of interest. They deploy the final algo as a task on their GBDX platform. They propose that in the future an R-CNN may be suitable for the whole process.
- Planet use non DL felzenszwalb algorithm to detect ships
- Segmentation of buildings on kaggle
- Identifying Buildings in Satellite Images with Machine Learning and Quilt -> NDVI & edge detection via gaussian blur as features, fed to TPOT for training with labels from OpenStreetMap, modelled as a two class problem, “Buildings” and “Nature”.
- Deep learning for satellite imagery via image segmentation
- Building Extraction with YOLT2 and SpaceNet Data
- A subset of the object detection problem, but surprisingly challenging
- From this article on sentinelhub there are three popular classical algorithms that detects thresholds in multiple bands in order to identify clouds. In the same article they propose using semantic segmentation combined with a CNN for a cloud classifier (excellent review paper here), but state that this requires too much compute resources.
- This article compares a number of ML algorithms, random forests, stochastic gradient descent, support vector machines, Bayesian method.
- DL..
- https://medium.com/the-downlinq/super-resolution-on-satellite-imagery-using-deep-learning-part-1-ec5c5cd3cd2
- https://modeldepot.io/joe/vdsr-for-super-resolution
- Image fusion of low res multispectral with high res pan band. Several algorithms described in the ArcGIS docs, with the simplest being taking the mean of the pan and RGB pixel value.
- Does not require DL, classical algos suffice, see this notebook and this kaggle kernel
- https://github.com/mapbox/rio-pansharpen
- Wikipedia DEM article and phase correlation article
- Intro to depth from stereo
- Map terrain from stereo images to produce a digital elevation model (DEM) -> high resolution & paired images required, typically 0.3 m, e.g. Worldview or GeoEye.
- Process of creating a DEM here and here.
- ArcGIS can generate DEMs from stereo images
- https://github.com/MISS3D/s2p -> produces elevation models from images taken by high resolution optical satellites -> demo code on https://gfacciol.github.io/IS18/
- Automatic 3D Reconstruction from Multi-Date Satellite Images
- Semi-global matching with neural networks
- Predict the fate of glaciers
- monodepth - Unsupervised single image depth prediction with CNNs
- Stereo Matching by Training a Convolutional Neural Network to Compare Image Patches
- Terrain and hydrological analysis based on LiDAR-derived digital elevation models (DEM) - Python package
- Phase correlation in scikit-image
- Simple band math
ndvi = np.true_divide((ir - r), (ir + r))but challenging due to the size of the imagery. - Example notebook local
- Landsat data in cloud optimised format analysed for NVDI with medium article here.
- QGIS- Create, edit, visualise, analyse and publish geospatial information. Python scripting and plugins.
- Orfeo toolbox - remote sensing toolbox with python API (just a wrapper to the C code). Do activites such as pansharpening, ortho-rectification, image registration, image segmentation & classification. Not much documentation.
- https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0
- https://github.com/taspinar/sidl/blob/master/notebooks/2_Detecting_road_and_roadtypes_in_sattelite_images.ipynb
- Geonotebooks with Docker container
- Sentinel NetCDF data
- Open Data Cube - serve up cubes of data https://www.opendatacube.org/
- Process Satellite data using AWS Lambda functions
- OpenDroneMap - generate maps, point clouds, 3D models and DEMs from drone, balloon or kite images.