Project Outcomes
Automated deforestation detection algorithms have the potential to be an extremely powerful tool used to better understand current ecological trends, locate areas in need of restoration, and to prevent and mitigate illegal logging, which accounts for about 15-30% of the lumber industry world-wide. One of the main reasons illegal logging is so prevalent is because it is very difficult to detect and enforce. Having algorithms that can scour vast areas of forest and detect change as it happens would help deter illegal logging and protect forests.
While deforestation detection is a problem that has been solved in the past with deep learning, those models typically only utilize imagery with RGB (red green blue) signals. However, satellites record light of many wavelengths which can be useful for determining plant health and diversity. The project goal was to develop a model that could use this multispectral data to predict which segments of the forest need protection.
The first approach was to train the model using an off-the-shelf dataset, consisting of ~750 image pairs and a corresponding mask of the true deforested areas which acted as a ground truth reference. A U-Net architecture was used to take in the images and produce a novel mask, a prediction of the deforested area, which is then compared to the ground truth reference. The model is trained by repeating this process over all the examples.
Initial results using this method were promising but not perfect. It was clear from the results that more data was necessary. The next step then became acquiring enough data to train the model sufficiently.
A new dataset was developed using satellite imagery from Sentinel 2, a satellite launched in 2015 by the European Space Agency. It consisted of ~54,000 image pairs, a significant improvement on the previous dataset.
After training the model, the results are promising but further refinement is needed. When locating large areas of deforestation, the model can perform accurately. However, when only a small amount of deforestation is present, it overgeneralizes and predicts no deforestation at all. Further development could include using a “course to fine” approach – first ruling out areas that are true negatives, then zeroing in on what remains to find the true positives.
Ground Truth Deforestation Compared to Predicted Deforestation from AI Model (Theo McArn / Organized Matter)
No Comments.