Assessing Trees in Global Drylands: Continental scale mapping of tree cover at individual tree level

Global drylands cover 41% of the world and support more than 2 billion people. Due to low annual rainfall, drylands are characterised by sparse vegetation, with trees and shrubs scattered across the landscape. These Trees Outside Forest play a key ecological role and provide vital ecosystem services to local communities. However, their extent, distribution and dynamics are not well understood at a global scale, as their scattered nature impedes their detection in current medium-resolution tree cover maps. There is an urgent need for an improved assessment of non-forest trees, with wide applications for conservation ecology, sustainable land management, agroforestry, and tree-based climate change mitigation.

This thesis aims to develop new methods to assess dryland tree cover at single tree level and continental scale. The thesis is based on three first-authored articles. Common to all papers is an underlying approach applying state-of-the-art deep learning techniques to a new data source: 3 m resolution PlanetScope imagery. PlanetScope is a constellation of approximately 130 nanosatellites that provide daily 4-band optical imagery of the global land area, including a near-infrared band suitable for detecting vegetation.

In the first paper, a convolutional neural network model is used to segment tree crown cover in PlanetScope imagery. Trained with manually delineated tree crowns, the model learns to detect individual scattered trees across a variety of dryland ecosystems. An image preparation framework is developed for the large-scale generation of PlanetScope mosaics, as well as an end-to-end prediction and processing pipeline. Applied at continental scale, these produce a map of all African tree cover at 1 m resolution, including both forest and non-forest trees. The results demonstrate a large underestimation of non-forest trees in existing global tree cover maps and show that more than a quarter of African tree cover is found outside areas previously classified as forest by current land cover maps.

The second paper introduces a framework to track individual cropland trees over multiple years, enabling the detection of tree-level changes. The challenge of high uncertainty due to variable image quality is addressed with a confidence-based change classification system. A case study of Tanzanian cropland trees tracked over five years demonstrates the ability to detect individual tree changes at national scale. In the third paper, the tree tracking framework is applied to monitor large farmland trees in India from 2010 to 2022, integrating additional imagery from the older RapidEye constellation. This reveals an 11% reduction in large cropland trees between 2010/2011 and 2018, contrasting with a previous national narrative of increasing tree cover and expansion of agroforestry. The losses of large cropland trees are corroborated by qualitative interviews with local farmers, who explain that trees were removed due to changes in land management and a transition to flooded rice cultivation.

In conclusion, this thesis develops new methods for the mapping and monitoring of scattered trees and advances the state of knowledge on non-forest trees. The frameworks developed in the core papers also contribute to several further studies mapping trees and other objects and lay the groundwork for global studies of dryland tree cover and dynamics at single tree level.