Classification of deforestation drivers
Project type: prototyping
Region: West-Kalimantan, Indonesia
Duration: 9 months (October 2021 - June 2022)
Collaboration partner: World Resources Institute
Rainforest deforestation is a major environmental challenge, causing loss of biodiversity, erosion and the release of stored carbon into the atmosphere contributing to climate change. Counteracting rainforest deforestation begins with monitoring individual events and finding the underlying primary drivers.
The University of Maryland devised the Global Analysis and Discovery (GLAD) alerts for detecting changes in rainforests, providing both time and location information. These alerts are openly available at Global Forest Watch, but lack the classification of what initiated the change. The goal of this project is to classify the GLAD alerts into primary deforestation drivers using satellite imagery.
The project is a collaboration between S[&]T Norway and the World Resources Institute (WRI).
The GLAD alerts provided by University of Maryland show the location of the forest disturbance, but don’t identify the driver. A clear example is shown in this before/after image, where GLAD alerts (red dots) are shown after a deforestation event. The reflectance of the deforestation for different drivers is very similar, which makes the classification a hard problem. Fortunately, the different drivers have specific patterns, and we can rely on the surrounding context to be able to perform a classification.
Example of GLAD alerts showing forest disturbance events.
Deforestation driver classification is a very challenging task for many reasons. To mention few:
Clouds,haze and cloud shadows.
Quickly changing temporal patterns.
Scarce training data.
Ambiguous patterns (e.g. dirt roads and small rivers look very similar).
The basic concept is simple: semantic segmentation of the land cover into different classes, such as forest, large scale agriculture, small scale agriculture, mining, roads, selective logging, etc.
However, many advanced tricks need to be used to achieve the quality requested by the customer. Some of the major concepts we apply are:
Pre-trained backbone networks.
Semi-supervised training with pseudo-labeling and consistency losses.
Domain specific augmentations (e.g., artificial haze and clouds). See an example of our cloud augmentation below. The left side shows the base image, the right side depicts various generated clouds and haze.
Example of cloud augmentation performed during the project.
Results are continuously shared and evaluated with our partner (WRI). Their experts share their domain knowledge about the field data and S[&]T contributes with its machine learning expertise. The continuous evaluation lets us efficiently iterate on model development and data collection to target what is most important for the project goals.
We publish some of our select results on the ESA Living Planet Symposium 2022 in the form of a poster. As an example the figure below shows detected deforestation related to large-scale agriculture. The project demonstrating feasibility in West-Kalimantan ends in 2022. Our next goal is to provide full coverage for Malaysia and Indonesia.