University of Copenhagen, Denmark, researchers mapped every tree across Africa. (Image Credit: lesjbohlen/pixabay)

University of Copenhagen, Denmark, researcher Florian Reiner and his team used high-res satellite imagery and machine-learning models to map every tree in Africa. Their method may help improve worldwide deforestation monitoring.

The team used 3 m satellite imagery from the PlanetScope constellation of nanosatellites. Each image has 5-band scenes with “atmospherically corrected surface reflectance values for the Blue, Red, Green, and Near-infrared-red bands.” In addition, they developed an algorithm that downloaded and mosaiced over 230,000 satellite images onto 1 x 1” grid tiles. They also applied a histogram matching algorithm via Landsat reference images to decrease sharp edges between scenes.

To create optimal tree cover visibility, the team chose a date range in which the deciduous and evergreen trees had full foliage. The MMODIS/Terra phenology product determined the local mean days for senescence, mid-greendown, and dormancy thresholds. It also features an indicator to determine whether deciduous or evergreen vegetation dominates most of the tile. That data came from the Copernicus Dynamic Land Cover map.

Reiner and his colleagues developed a deep learning framework in Python for segmenting the tree crown in the PlanetScope images. They also used a UNet model, trained with a batch size of eight and a 512 x 512 patch, featuring batch normalization and self-attention. The team wrote in the paper, “To enhance the training data, image augmentation was performed with multiple transformations including flipping, cropping, affine transformations, and linear contrast enhancement.”

This UNet model was trained using over 130,000 manually labeled training samples, which include merged canopy clusters and individual trees. It provided a tree with a label if it identified as a wooden plant with its shadow in Google Earth. Labeling was achieved in two steps. The first one used random areas with varying ecosystems, like savannas, croplands, shrublands, woodlands, and forests across Africa. Afterward, the researchers trained a model and predicted tree cover. The researchers wrote, “We conducted a visual inspection and thereafter a second round of labelling in areas where the model did not perform well.”

They obtained samples from 21 countries and covered different forest types, rainfall, and ecosystems through western, eastern, and southern Africa. Modifying the framework by incorporating a dynamic resampling feature into the deep learning pipeline made it possible to upscale the imagery from 3m to 1m during training. This improved the prediction quality by “preserving the high fidelity of the manual training data. The training labels were polygons with many vertices at sub-pixel resolution, and this detail would be lost when rasterizing them to 3 m. By rasterizing the labels to 1 m, and training with upsampled 1 m data, the model can instead use contextual clues such as shadows to produce sub-pixel predictions at 1 m.”

The researchers suggest using this technique to monitor worldwide tree coverage. It can improve humanity’s ability to detect and map tree cover change at a high temporal frequency, extending deforestation monitoring systems and allowing it to track the removal of individual trees and thickets.

Have a story tip? Message me at: http://twitter.com/Cabe_Atwell