Spatiotemporal Analytics on Geospatial Imagery
Satellite imagery has historically been a critical source of business intelligence in domains like precision agriculture, defense and intelligence, maritime awareness, insurance, and energy utilities monitoring. However, the recent explosion of remote sensing data has challenged existing business workflows, and analysis of imagery remains expensive and time-intensive.
This past week, Planet’s CEO Will Marshall unveiled the company’s vision for a Queryable Earth to make our imagery more accessible and actionable. We launched Planet Analytics, a new product line that uses machine learning to index physical change and make it searchable, so that customers can get access to the insights that are critical to their businesses. Queryable Earth is an inspiring framework. For me, the exciting work happens “under the hood.” To get to a Queryable Earth will require building new forms of geospatial analytics, to power both Planet native products and those from our partners.
Planet has a unique dataset that is facilitating such advances. We image the entirety of Earth’s landmass every day at 3.7 m resolution and have “tip-and-cue” automatic tasking of 80 cm resolution imagery. Through this trove of imagery, it is possible to create automated analytic capabilities that scale across diverse geographies, identifying and indexing meaningful change over time, what we refer to as spatiotemporal analytics.
Here are three exciting possibilities in spatiotemporal analytics that we are exploring with our imagery:
Understanding Generalizability
Planet captures varied contexts over diverse geographies at multiple resolutions. When detecting and localizing objects on sea and land, or detecting land cover changes, we have the context of the entire planet. For an object class like ships, this could be a collection of ships in ports and open water captured at different sea states and with varied wind conditions. For a land use type, this could be a collection of building footprints in cities and suburbs, by the roadside and by the shoreline. For a land cover class like palm oil trees, this could be plantations in different regions at different stages of growth.
This allows us to create comprehensive datasets for training machine learning models that sample real-world scenarios, moving us towards descriptive analytics. Our imagery source is large and diverse, enabling models that generalize to a wider range of previously unseen scenarios. The result is our ability to provide geospatial insights with high confidence over large regions when deployed in production.
Exploring the Historic Archive
With Planet’s multi-petabyte archive and data pipeline that processes upwards of 10 terabytes of imagery daily, we can annotate static features once and capture varied conditions for them (ie, a building before and after a flood). We can also annotate dynamic features once and capture changes for them (ie, a port with lots of ships moving in and out). In areas with a high probability of clouds, the cadence of the archive is valuable for collecting cloud-free imagery of a particular region.
From here, we can get a dense time-series of observations at every point on the Earth, allowing us to characterize and filter spurious cloud cover and identify meaningful change. For a specific object type, like an oil well pad, or land cover type, like forests, we can reference the archive to determine cyclical change patterns and clearly identify and report on atypical or unexpected deviations.
This timelapse of ground locked imagery leads to a non-IID (independent and identically distributed) dataset, which allows us to develop new analytic techniques at the intersection of statistical learning (ie, Gaussian Processes and Conditional Random Fields) and deep learning with convolutional and residual neural networks.
Making Change Indexable
Identifying where potential changed has occurred (ie, pixels) in remote sensing imagery is straightforward but less useful; identifying forms of meaningful change (like deforestation) is complicated, requires domain expertise, but is of high social and economic utility.
At Planet, we are working to analyze change, extract change features, and store those features in a manner that is easy to search and retrieve. This will make geospatial change features queryable, retrievable, and referenceable. By studying patterns of change, we can create new strategies to collect training datasets that capture signatures of change.
When running predictions to detect change, we start with high recall predictions, collect potential candidates for change, curate them with humans in the loop, and feed the results back to our models for improved performance. This online learning system gives us the ability to continually improve our indexing and henceforth improve our query results over time.
Recent research in machine learning and computer vision using geospatial data has been focused on descriptive analytics — largely because the datasets they rely on are historic, and therefore outdated once published as charts or reports.
By creating a richer set of spatiotemporal analytic capabilities, we can develop a deeper understanding of static and dynamic features in geospatial imagery and identify meaningful change captured in a general context. For instance, you can use a change model trained to detect deforestation in the state of Para in Brazil, and apply it to other regions in the tropical and subtropical moist broadleaf forest biome. Ultimately, this moves us closer to predictive analytics, which is the foundation for a queryable planet.
This is the focus of my team at Planet. Stay tuned as we bring you more on Planet’s work in analytics. We launched Planet Analytics last week. We are actively hiring, so if the scale and scope of this work excites you, check out Planet’s Careers page.
