Sentinel 2 Image Delivery Platform Forest Industry Informatics - SCION Sam Damesin, Grant Pearse, Ralf Gommers, Andrew Gordon, Jonathan Dash March 2017
Imagine
Currently not possible National natural resource descriptions are inadequate on a national scale: Frequently not spatial Of insufficient resolution Incomplete Out of date Difficult and expensive to acquire This leads to poor decision making, inefficiency, lost productivity and reduced profit.
New technologies provide opportunities New satellites Sentinel 2A (2015) and 2B (2016 2017) European Space Agency Free access Multi-spectral with 13 bands Spatial resolution up to 10 m Covering NZ every 5 days
New technologies provide opportunities New satellites Sentinel 2A (2015) and 2B (2016 2017) European Space Agency Free access Multi-spectral with 13 bands Spatial resolution up to 10 m Covering NZ every 5 days BUT There is no easy mechanism to access, process, and store this imagery.
Project details & Context
This project: proof of concept Objectives: Deliver imagery in an easy-to-use manner for forestry companies and scientists Enable further processing and analysis Compatible with existing tools and workflow Gauge interest for future development Technical information: Build with Python and GDAL (Geospatial Data Abstraction Library) for pre-processing steps Cloud infrastructure: 2-3 servers on Amazon AWS, S3 storage Use GeoServer WMS (Web Map Service) for publishing
Automated data processing pipeline 1. Download Get imagery for NZ Only cloud free 0-40% 2. Pre-process 4 bands Red, Green, Blue and NIR For every collection day Reprojection and multibands mosaic 3. Publish (WMS) Daily mosaic (capture day) With gaps Monthly most recent mosaic No gaps Not ready yet! 4. Analyze Apply cutover detection algorithm as a proof of concept
The longer term vision 1. Data capture Sentinel 2A and 2B 3. Storage and access Cataloguing Access esa 2. Pre-Processing Automated download Radiometric correction Ortho correction Cloud mask Mosaicing 4. Processing Classification Texture Spectral Indices
Potential derived products & users Digital Surface and Terrain model NDVI Natural forests Exotic forests Yields, age, species classification Local and central government Forest owners (large and small) Consultants Wood processors Log buyers Iwi Researchers Change detection
Web Map Service
Server hosting High bandwidth cloud server on Amazon Web Service (EC2) Layers storage on Amazon Web Service (S3) GeoServer Web Mapping Service Open source, standard
Easy access
Example
Example
Land classification & cutover detection
Datasets Sentinel 2 imagery Central North Island 3 dates: Nov. 2016 Dec. 2016 Feb. 2017 4 bands at 10m resolution: Red, Blue, Green Near Infra-Red
Datasets Sentinel 2 imagery Central North Island 3 dates: Nov. 2016 Dec. 2016 Feb. 2017 4 bands at 10m resolution: Red, Blue, Green Near Infra-Red 2 study sites
Step 1: Machine learning training Python code Supervised machine learning Random forest classifier 5 classes: Clouds Grassland/Farmland Harvested area Forested area: young tree Forested area: mature tree
Step 2: Machine learning classification November 2016
Step 2: Machine learning classification November 2016
Step 2: Machine learning classification December 2016
Step 2: Machine learning classification February 2017
Step 3: study sites
Step 3: study sites
Cutover detection application, validation Accuracy validation Manual definition of validation subset (see polygons on the right) Manual classes assignment 98% accuracy for this subset
Cutover detection application Band importance for the classification Band 1 (Red) importance: 31% Band 3 (Blue) importance: 24% Band 4 (NIR) importance: 22% Band 2 (Green) importance: 21% Successfully classify 5 classes on a small dataset Clouds Grassland/Farmland Harvested area Forested area: young tree Forested area: mature tree
Conclusion and next steps
Current status Operational prototype system Automatic pre-processing and storage Daily mosaic of cloud free New Zealand (sensing day) Between 0 and 40% cloud coverage Bands: Red, Green, Blue and NIR at 10m resolution Gaps for cloudy area Monthly up to date mosaic (not ready yet!) No gaps: most recent images used Hosted on Amazon Web Service using GeoServer Published via Web Map Service (standard service) Easily accessible via ArcGIS, QGIS and other GIS software
Future work Internal use for current and future research work, potentially new product development Potentially further develop cutover detection algorithm Seeking feedback from industry on: Level of interest who would access it? What products are of interest? Pathway to support ongoing access?
Acknowledgements Michael Watt, Bryan Graham: approval and funding Jonathan Dash, Grant Pearse: idea and active participation Andrew Gordon, Ralf Gommers: technical inputs
www.scionresearch.com Prosperity from trees Mai i te ngahere oranga Scion is the trading name of the New Zealand Forest Research Institute Limited