Impact Observatory was founded in 2020 and is headquartered in Washington, D.C. Its map products have been released by institutions or enterprises such as Esri, Microsoft, Amazon and the United Nations, and have been accessed more than one million times.
Impact Observatory, an AI remote sensing company, has completed a $5.9 million seed round of financing led by Esri. The funding will support Impact Observatory’s business activities and accelerate the launch of its AI remote-sensing monitoring products for government and commercial customers. The funds will also be used for research and development, acquiring more remote sensing data sources and improving its global monitoring technology.
Impact Observatory’s DaaS product (Data as a Service) integrates remote sensing data and deep learning AI to achieve near-real-time global mapping with the highest resolution and accuracy in the market. These real-time data provide effective decision support, helping users understand sustainability and environmental risks and predict changes. Currently, Impact Observatory’s core application areas include sustainable agriculture and food security, monitoring carbon emissions, mitigating climate change, and protecting and restoring biodiversity and ecosystems.
Impact Observatory uses deep learning technology to classify global land use and land cover categories using images from Sentinel-2 satellite. It achieves near-real-time, automated monitoring with a spatial resolution of 10 square meters. Impact Observatory claims that this is the first near-real-time LULC fully automated high-resolution world map. The product can provide land cover and land use maps from 2018 to the present, updated daily. Users can customize their areas and time periods of interest and get land use and land cover change maps.
How does Impact Observatory achieve automation and near-real-time monitoring? Microsoft’s cloud computing Azure provides the biggest boost. Impact Observatory uses Azure HPC + AI to expand Esri’s 2020 global land cover map into a time series of annual maps, revealing changes, combining its large data sets, raw images and AI technology to create an accessible visual resource. On average, Impact Observatory processes 21 remote sensing images for each location on Earth - totaling more than 450,000 Sentinel-2 satellite images and 500 TB of satellite data collected each year. All content is compressed into a map of about 60GB, which requires 1000 parallel-running virtual machines. It is now all deployed on Microsoft Azure for rapid implementation. Now, Impact Observatory makes maps faster than satellites collect images.
In terms of data sources, Impact Observatory currently mainly uses images from Sentinel-2 in Europe and Landsat satellites in the United States. Both are the largest open-source satellite data products in the world, and almost all remote sensing image companies use these two data sources for model training. The biggest advantage of open source images is stability, which allows continuous acquisition of time series data for training. The biggest drawback is timeliness: Sentinel-2 provides global images every five days; combined with Landsat data, Impact Observatory can update its map products every two and a half days.
Dr. Brumby, founder of Impact Observatory, said that after this round of financing, Impact Observatory will supplement some commercial data sets and strive to achieve daily updates of monitoring and map products in the future.
The issue of data sources is also worth paying attention to by Chinese remote sensing enterprises. China has always been the largest data acquirer of Sentinel and Landsat. But based on the current global situation, how long will these free and open-source remote sensing data provide for China? This is worth pondering by the industry. Although China has public welfare satellite data sources such as Gaofen, they have always been limited in openness; coupled with the fact that the data quality is not recognized by practitioners; therefore, the development ecology around China’s domestic satellite image data has always been incomplete. And around Sentinel data and Landsat data, a lot of products have been developed, greatly promoting the application ecology of the entire remote sensing information field.