New Esri Land Cover map shows us how to move fast and share knowledge.
A new map created with machine learning has opened up opportunities for resource managers, land-use planners, scientists and conservationists to update their land cover maps on a weekly basis. Published by Esri, the Land Cover 2020 map uses 10m resolution Sentinel-2 satellite imagery and a land classification model to classify the Earth’s surface into 10 categories including water grass, built area, and crops.
Developed by Impact Observatory, the model was trained with a database of billions of human-labelled image pixels made by the National Geographic Society. As a result, a task that would normally take months took just under a week. This approach to developing land cover maps means they could be created on-demand and more frequently. By knowing what change is happening, and where, in almost near-real-time, the planet’s natural resources can be sustainably managed.
Applications of the map include land-use planning, resource management, conservation planning, and hydrology research. As the technology develops, maps created with machine learning increase the precision of the land cover maps we need to understand the distribution of wild species and their interactions with humans and human activity.
Feedback is welcome.
As with most first releases of new digital technology, there are some criticisms of the Esri Land Cover 2020 map. In a Twitter thread, MD Madhusudan highlighted tea estates in India that are variously labelled as scrub, grass, trees and crops, and an extensive solar farm that’s been classified as “crops” rather than “built area”. Other commentators observed a loss of spatial detail but also acknowledged that given the global scale of this open science product, the final product is still very impressive.
Curious to know how this feedback might be used, I reached out to Steve Brumby and Sam Hyde, co-founders of Impact Observatory. “We are very pleased with the reception and engagement from the user community, which was very encouraging”, said Steve. “A number of regional experts noted interesting places where the map did not agree with other approaches. We greatly appreciate this feedback and are tracking constructive suggestions closely in order to help us identify areas where we could improve our map.”
Sam went on to explain that some of the disagreements may be due to differences in interpretations of category labels. For example, in some cases the ‘built area’ label has been interpreted to mean ‘urban.’ However, as Sam explains, “built area can be urban, but it can also be infrastructure in more rural areas such as a large greenhouse in cropland, or a small rural settlement.”
The process of labelling begins with human annotators using an online platform to draw boundaries around features such as forests, lakes, settlements and farms on photos. The result is a dense collection of labels that enable deep learning algorithms to explore and learn both spatial and spectral characteristics of features and start labelling them.
“How a feature is labelled can be important for practical reasons,” said Steve. “A built area or farming classification may impact local taxation rules. The global community needs to be able to conduct fair comparisons across countries to understand human pressures on the environment, and track how countries are making progress towards their commitments to the big UN conventions on climate, biodiversity, and sustainable development. Our partnerships with Esri and the UN are helping Impact Observatory understand and navigate some of these complexities of global-scale mapping.”
Move fast and make change happen.
Releases like this one from Esri and Impact Observatory represent a new model of science, one of perpetual iteration and improvement. We don’t have time to wait for perfect products to be published. We must publish fast, share openly, and welcome scrutiny. Willingness to learn and improve is a critical characteristic of initiatives that lead to real change.
But, there’s still more to be done.
What other data do we need to make resource management more sustainable and conservation decisions more effective? A more detailed land use map is an imperative addition to planning “toolboxes”. Land cover maps tell us information about physical land types, but they can’t always tell us how that land is being used. What activities are happening in those places? Which farming methods are being used? What are people doing to modify or maintain habitats?
Artificial intelligence, machine learning, and ecological modelling offer opportunities to improve our understanding of what’s happening right now and what the impact of our decisions will be. However, the precision and scale that’s needed to produce global maps that accurately reflect local circumstances will require much more investment to develop the training data sets and models. The better we can predict the impact of planting here, restoring there, or halting development elsewhere, the better our chances are of stopping a catastrophic ecosystem collapse.
Climate change is already altering our environment, and the speed at which these changes happen is increasing. Scientists have warned that decisions made in the next ten years will be the most critical we ever make. There’s no time to wait for the perfect dataset. We have to act today on the information we have and update our strategies as the data and technology improves. The faster we process data, the faster we can get it into the hands of people with influence. And who are the people with influence? That’s all of us.