Convolutional Neural Networks May Allow Mapping Of Every Tree On Earth

Sign up for daily news updates from CleanTechnica on email. Or follow us on Google News!

---

What can you do with a massive database of high-resolution satellite images covering more than 1.3 million square kilometers of the western Sahara and Sahel regions of West Africa? Well, a group of researchers has used artificial intelligence (AI) neural networks to map the location and size of more than 1.8 billion individual tree canopies.

The results?

It may soon be possible to map the location and size of every tree worldwide. This step forward in observational capabilities is important, as it can alter the way we consider, monitor, plan, and manage global terrestrial ecosystems.

Writing in Nature, Martin Brandt and team analyzed more than 11,000 images at a spatial resolution of 0.5 meters. Their goal was to identify individual trees and shrubs with canopy diameters of 2 meters or more. Never before have trees been mapped at this level of detail across such a large area (although it should be noted that this method required an input of approximately 90,000 manually digitized training points, which may be untenable for some replication studies).

The team completed this huge task using AI, exploiting a computational approach that involves what are called fully convolutional neural networks. Training data consisted of satellite images in which the visible outlines of tree and shrub canopies were manually traced. Through these samples, the computer learned how to identify individual tree canopies with high precision in other images.

Their product is a wall-to-wall mapping of all large trees across the whole of southern Mauritania, Senegal, and southwestern Mali.

Chip in a few dollars a month to help support independent cleantech coverage that helps to accelerate the cleantech revolution!

What are Convolutional Neural Networks?

The spatial resolution of most satellite data is relatively coarse. The pixels from each individual image generally conform to areas on the ground that are larger than 100 square meters and often larger than one square kilometer. A Convolutional Neural Network, on the other hand, is a deep learning algorithm which can absorb an input image, assign importance — i.e. learnable weights and biases — to various aspects/objects in the image, and be able to differentiate one from the other.

These neural networks have layers that essentially perform convolutions, in which convolutional filters move over the spatially structured input to produce outputs known as feature maps. The innovation of convolutional neural networks is the ability to automatically learn a large number of filters in parallel to a training dataset under the constraints of a specific predictive modeling problem, such as image classification. The result is highly specific features that can be detected anywhere on input images.

In the case of the Brandt study, this deep-learning method was designed to recognize tree canopies on the basis of their characteristic shapes and colors within a larger image. A review of the study in Nature says this result is a “striking demonstration of this transformation in terrestrial remote sensing.”

Because convolutional neural networks rely on the availability of training data, the satellite images of visible outlines of tree and shrub canopies were manually traced. To improve canopy separation, the Brandt team used a weighting scheme in training their convolutional neural network but still resorted to a “canopy clump” class to describe aggregated canopy areas of more than 200 meters.

In this way, Brandt and colleagues were able to offer detailed information on the location and size of every individual canopy. The high level of detail includes for the regions with annual rainfall greater than 600 millimeters. It exposes local spatial variability in trees that is generally associated with contrasting soil types, water availability, land use, and land-use history.

What are Terrestrial Ecosystems & Why are They Important?

The functionality of ecosystems is important for energy dissipation, ecosystem service provisioning, resilience to global change, and adaptive capacity. Terrestrial ecosystems are the 3rd largest global carbon pool only after the ocean and the geological carbon pool. Carbon sequestration by terrestrial ecosystems refers to the conversion of atmospheric CO2 into the carbonaceous components by the plants or geological processes that can be stored as carbohydrates, soil organic matter, and carbonate minerals.

Once the CO2 has been transferred into these materials, it is effectively locked until decomposition occurs. These processes can be affected by many factors, such as species characteristics and age of vegetation, climatic conditions, land use, and soil type.

Defined for the most part by their woody plants, terrestrial ecosystems change in nuanced ways among grasslands, shrublands, savannahs, woodlands, and forests. Variations and gradations in tree and shrub density move from low-density, low-stature woody plants to those with taller trees and overlapping canopies. Accurate information on the woody-vegetation structure of ecosystems is, therefore, fundamental to our understanding of global-scale ecology, biogeography, and the biogeochemical cycles of carbon, water, and other nutrients.

Overview of  Tree Canopy Mapping with Help from Neural Networks

The framework for the detection of tree crowns in satellite imagery of very high spatial resolution (panchromatic and pan-sharpened NDVI 31 images at 0.5 m) used a supervised deep-learning neural networks approach to detect single woody plants. As savanna trees tend to be scattered, their crowns stand out as objects with a high NDVI value — in contrast to their surroundings, which have low NDVI values in the dry season.

Visually, tree crowns were the easiest to identify in the satellite images, and 89,899 individual trees along a north–south gradient were manually delineated and annotated. The resulting spatial database includes each detected tree, its crown size, mean annual rainfall, land use, and soil.

The Brandt team detected:

• over 1.8 billion individual trees/13.4 trees per hectare
• a median crown size of 12m along a rainfall gradient from 0 to 1,000 mm per year
• a canopy cover increase from 0.1%/0.7  trees per hectare in hyper-arid areas
• 1.6%/9.9 trees per hectare in arid areas
• 5.6%/30.1 trees per hectare in semi-arid zones
• 13.3%/47 trees per hectare in sub-humid areas

Final Thoughts

The Brandt assessment suggests a global way to monitor trees outside of forests and to explore their role in mitigating degradation, climate change, and poverty.

As more innovation in remote sensing becomes available, detail about vegetation structure will be derived from a variety of sources, such as light detection and ranging (LIDAR), radar, and high-resolution visible and near-infrared sensors. Advances in satellite-derived high-resolution data on tree canopy size and density should be able to contribute to the inventory and management of forests and woodland, deforestation monitoring, and assessment of the carbon sequestered in biomass, timber, fuel wood, and tree crops.