GEMSTONE CASE STUDY: Global Economic Monitoring using Satellite Data and AI/ML Technology
Intro
Over the last year, SpaceKnow worked on the GEMSTONE project, supported through the ESA InCubed program. GEMSTONE stands for Global Economy Monitoring System delivering Transparency and Online Expertise and aims at developing cutting-edge artificial intelligence (AI) and machine learning (ML) based economic activity indicators in the form of indices, derived from satellite data.
Illustrative image
What is behind it?
Supervised and unsupervised algorithmic workers provide detections over the monitored areas of interest and perform data fusion and aggregation into economic indices. Specifically, the SpaceKnow Research Team has developed eight new cutting-edge algorithms for detecting the following raw materials and man-made structures: Coal, Iron, Lithium, Wood, Open-Pit Mines, Containers, Oil Tanks, and Roads.
A unique combination of spectral unmixing and deep learning neural networks (DNNs) ensures the highest quality output, in which both precision and recall metrics are superior compared to using these methods separately. Algorithmic workers of some materials use a fusion of more DNNs as a deployed model, for instance in the case of coal, the first model was a binary DNN utilizing all coal classes’ annotations merged into one general coal class; the second model was a multiclass DNN differentiating three classes (coal in storages, coal in mines and coal with an artificial surface). Both of the fusion models used all twelve Sentinel-2 bands.
Coal Algorithmic Worker prediction map together with underlying imagery visualization with coal in storages (green), coal in mines (magenta), and coal with an artificial surface (white).
Consequently, these algorithmic solutions are deployed over a vast number of precisely selected locations, and the outputs of the analyses are later aggregated into specific indices.
The outputs in the form of indices are presented in a user-friendly dashboard, with no additional processing needed on the customer’s side, as shown in the example below. Alternatively, the indices can be also delivered to the end-users using an API (Application Programming Interface).
Through the GEMSTONE project, customers will have instant access to both location-specific knowledge as well as country-wise or industry type aggregated information that allows querying activity at a ‘topic’ level (coal facilities, oil wells, etc). Clients can also provide locations of interest (ex. the points of a specific company or product supply chain) and SpaceKnow data can be accumulated at those locations.
Nagoya use case
Nagoya is one of Japan’s largest ports with many supplies stored outdoors, making it an ideal testbed to showcase the power of SpaceKnow’s algorithms.
In the figure below, Nagoya city and related port analysis is presented. The timeline at the bottom shows the evolution of the area covered by each of the segmented elements. In this big picture, the outputs of the SpaceKnow segmentation algorithms are observed: orange for urban, green for non-urban, purple for roads, and blue for water.
Nagoya Analysis: the big picture
Zooming in to Nagoya port, the more detailed view below shows oil tank detections. Single oil tanks are highlighted in red, whereas oil tanks blocks are highlighted in blue. Thanks to the sub-polygons drawn around the oil tanks we have precise and deep insight into smaller areas of Nagoya port and even separate time series showing the development of oil tank counts.
A detailed view on the Oil Tanks detections
In the following figure, subpolygons with detected containers and woodpiles are shown.
Containers and Wood detections
Zayed City: Roads detection
The road algorithm, unlike other GEMSTONE algorithms, uses PlanetScope medium resolution optical imagery. For the purposes of the GEMSTONE Pilot dashboard, a newly constructed Zayed City district in Abu Dhabi was selected. Zayed City is a huge construction project with an area of approximately 40 square kilometers. The newly developed roads algorithm was deployed over this area, on imagery ranging from 2017 to 2022, which was when the construction activity peaked. During this period, the roads algorithm detected a substantial expansion of the road network in the area, growing from 0.233 square kilometers (0.61% of total area) in May 2017 when the construction started and only main roads were observed, to 2.51 square kilometers at the end of the monitoring period in January 2022.
The following figure shows the proprietary SpaceKnow Analytics platform, where the polygon above Zayed city was created. The time series below shows a clear trend of expanding the road network. There are, however, some temporal drops observable in the time series, caused by dust coming from the nearby desert, which at some point may partially cover the roads and thus makes the detections harder for the algorithm. Also, some images may have been taken in non-optimal atmospheric conditions, again making detections harder and causing seemingly temporarily disappearing roads. The overall trend is however very clear.
Figure: Screenshot from SpaceKnow Analytics platform with the road analysis
The following figures show a few images from the monitoring period with the roads detected and highlighted in pink color. The expansion of the roads network is clearly visible even on the optical data.
July 2017, 0.359 sq km of roads detected
August 2019, 1.22 sq km of roads detected
October 2021, 2.51 sq km of roads detected
The roads algorithm and indices derived from it have a great potential in monitoring urbanization and road networks on large scale, without a human in the loop, especially in areas when ground truth is not easily available (developing countries).
Interested in bringing these insights into your organization? Reach out to us at info@spaceknow.com
