Development Seed’s PEARL platform employs AI to accelerated land cover mapping
By Sajjad Anwar, Lane Goodman, Martha Morrisey, Nick Ingalls, Vincent Sarago, Vitor George, Sanjay Bhangar, Jeevan Farias, Zhuangfang NaNa Yi, Development Seed
Development Seed and Microsoft AI for Earth are launching PEARL, an AI-accelerated platform for land cover mapping, for experimental use. The platform combines human intelligence and scalable AI for fast, accurate land cover mapping. It leverages capabilities and research under the Microsoft Planetary Computer initiative that reduce the effort to create land cover maps, allowing scientists and researchers to focus on the most pressing environmental and climate research questions.
Two innovations make PEARL possible: Working with the Microsoft team, Development Seed extended Microsoft’s research into cloud infrastructure that makes machine learning model retraining fast enough to be conducted interactively in the browser. They then designed a new user experience that leverages this infrastructure to create intuitive prediction, feedback, and retraining interactions. The result is a tool that allows any user to create and refine a land cover classification model in the browser.
The current experimental platform is available for public use, but its availability is limited to the United States. Its next release will include global coverage using Sentinel-2 imagery.
Rethinking land cover mapping
Accurate, timely and accessible land cover maps are critical for conservation, climate research, and planning. Scientists and analysts currently rely on costly and time-intensive processes to generate bespoke land cover maps. Global land cover datasets exist and are useful for some purposes. However, publicly accessible maps are often out-of-date, low-resolution, or inaccurate, particularly outside of the US and Western Europe.
Artificial intelligence offers an opportunity to accelerate land cover mapping, but the expertise of geospatial analysts is still critical to produce accurate land cover maps. Consequently, Development Seed has invested in human-in-the-loop AI tools that allowed its Data Team to speed up high-quality mapping by as much as 30 times.
Creating a land cover map at speed
PEARL provides imagery and starter models to immediately start mapping an area of interest. The user provides feedback to the model by labelling correctly and incorrectly predicted areas, guided by on screen stats to ensure balanced feedback that will improve the model accuracy. Users can iterate this process for as long as they like. Once completed, they can export the result as a GeoTIFF or an interactive map.
The platform combines machine learning, open data, and open-source software with scalable infrastructure on Microsoft Azure. It does not require any data to be brought to the platform — the user gets access to imagery and models. This enables users to start mapping immediately, instead of engaging in data procurement and complex pre-processing.
A core principle behind the design of PEARL is to provide scientists and researchers easy access to infrastructure that is otherwise expensive, hard to setup and manage. A platform user connects their browser directly to a GPU that runs models and does all the computation. Behind the scenes, a Kubernetes cluster enables scheduling multiple users to perform inference and retraining. Development Seed’s TiTiler provides an imagery service that generates tiles dynamically from mosaics hosted on Azure Blob Storage. The backend infrastructure provides persistent sessions to each user though a REST API and WebSocket connections.
Making machine learning accessible
Building tools to make machine learning more accessible entails developing new but familiar user interactions. The application borrows many of the common patterns of map editing applications, giving users an improved retraining sample selection experience. Map creators can select individual points or draw freehand shapes to define image areas for retraining.
Retraining sometimes may not be enough for a high-quality map, which is why the team introduced map refinements to make final adjustments. Users can treat classes and previous checkpoints as a brush, filling in freehand shapes. In this way, the final map can sample from any of the trained results, and users can remove unwanted pixel noise or minor aberrations.
Key metrics are surfaced throughout the map creation process. The model metadata cards provide the user with context to help select a starter model. This involved the creation of a new metadata schema for models that captures information like class distribution of the training dataset, labels, label source, imagery source, geographic location of the training dataset, and per class performance over the hold out test dataset.
AI that improves over time
This initial version of PEARL contains two FCN segmentation starter models trained with nine and four land cover classes based on labelled data from the Chesapeake Conservancy’s dataset. Both of these starter models have a global F1 score of just under 90%. When users provide feedback, the model is retrained by updating the parameters of the last layer of the model using the point labels provided by users. Users can improve the performance of the model for a local area and can even define new LULC classes.
As people use PEARL they produce new model checkpoints that work better for their area of interest. These checkpoints can then be applied to larger areas of interest. Future versions of PEARL will include more starter models that cover more regions and classes. The data team also intends to allow users to save and share their checkpoints and to contribute high performing models to the starter model library.