Accurate and consistent monitoring of tree cover is paramount for effective sustainable land management and climate change mitigation strategies. The Glasgow Pact from COP26 emphasizes halting and reversing deforestation by 2030, necessitating high-quality monitoring systems for measurement, reporting, and verification (MRV) of forest area changes. However, existing monitoring systems often neglect trees outside designated forest areas, or their cost prohibits consistent application across diverse regions. The FAO provides forest definitions, but remaining landscapes with trees are grouped into broad categories like "other wooded land" and "other land," obscuring the significant ecological and economic contributions of trees in non-forest settings, particularly in African drylands where trees outside forests are prevalent. Previous studies, using FAO definitions, showed forests cover only 21.4% of Africa, with an additional 14.9% categorized as "other wooded land." The remaining 63.7% is classified as "other land," encompassing various tree complexes outside forests. Accurate quantification of trees in both forested and non-forested landscapes is vital for reducing emissions from deforestation and forest degradation, and increasing sequestration through restoration initiatives. Current approaches, however, are limited by ambiguous definitions, varying MRV techniques across countries, and resource constraints in developing nations. The lack of consistent assessments at the continental scale and across years further complicates matters. While moderate-resolution satellite data (10–30 m) are commonly used, their resolution hinders the identification of individual trees outside forests. Very high-resolution (0.5 m) data and machine learning techniques offer improved accuracy but face limitations in processing, storage, cost, and temporal consistency, making continental or global-scale applications challenging. This study aims to overcome these limitations by leveraging freely available high-resolution PlanetScope nanosatellite imagery to map all forest and non-forest trees across Africa with unprecedented precision, exceeding previous large-scale woody vegetation mapping attempts.

Existing literature highlights the challenges of accurately assessing and monitoring tree cover at large scales, particularly concerning trees outside formally defined forest areas. Studies using FAO definitions show a significant portion of Africa's tree cover falls outside of 'forest' classifications. The limitations of moderate resolution satellite imagery in capturing individual trees and the high cost and processing demands associated with very high resolution data have been widely discussed. The lack of consistent temporal resolution and varying methodologies across different regions further complicates the comparison and interpretation of findings from past studies. This study addresses these issues by employing advanced satellite technology and machine learning techniques to build upon previous research and provide a more complete and accurate representation of African tree cover. The use of a consistent dataset and analysis methodology aims to improve the comparability and reliability of the results compared to past research relying on diverse methodologies, resolutions and time frames.

This study employed high-resolution (3 m) PlanetScope satellite imagery from 2019 to map tree cover across continental Africa. Over 230,000 satellite scenes were mosaicked into 1° x 1° tiles, with image dates selected based on green vegetation phenology to maximize tree leaf cover while minimizing grass interference. Daily PlanetScope imagery enabled the creation of cloud-free composites for a narrow time frame within a single year. A deep learning model, trained with approximately 130,000 manually annotated samples, segmented tree crown cover at 1 m resolution. Trees or tree clusters clearly identifiable as woody plants with shadows were labeled, excluding smaller shrubs. The images were upsampled from 3 m to 1 m to improve model prediction performance. Percent tree cover was derived by aggregating results into 30 m x 30 m grid cells. These cells were grouped into forest (canopy cover >25%) and non-forest (<25%) areas, further categorized by canopy height using a LiDAR-based map. The distribution of trees across different climatic zones was analyzed by summing tree cover based on annual rainfall data. The resulting tree cover map was compared to a state-of-the-art global tree cover map, revealing discrepancies, particularly in low-rainfall areas. The proportion of tree cover in non-forest areas was quantified for each country, highlighting the limitations of moderate-resolution maps in several nations. The very high-resolution tree cover map allowed for detailed analysis of tree cover across different land cover classes using WorldCover maps, revealing the significant contribution of non-forest areas. The number, density, and distribution of trees on croplands were analyzed, considering both individual trees and tree clusters. The study compared tree cover area with FAO statistics, considering the FAO definitions of "forest" and "other wooded land." Model accuracy was validated by comparing the generated tree cover map with canopy height models (CHM) from airborne LiDAR and UAV stereo photogrammetry data from Senegal, DRC, and Mozambique. This validation used different minimum tree height thresholds in the CHM data, demonstrating that the PlanetScope tree cover reliably maps trees above approximately 5-6 m. Finally, a comparison with isolated tree data derived from sub-meter resolution imagery assessed the model's ability to detect individual trees, showing reliable detection for crown sizes above 30 m².

The study produced a high-resolution map of African tree cover in 2019, using PlanetScope nanosatellite imagery and deep learning techniques. Key findings include: 1. A significant portion (29%) of Africa's total tree cover is found outside areas previously classified as forest, predominantly in non-forest landscapes such as croplands and grasslands. This highlights the limitations of existing definitions and mapping approaches that focus solely on formally designated forest areas. 2. The high-resolution map (RMSE = 9.57%, bias = -6.9%) allows for the precise assessment of all tree-based ecosystems at a continental scale, enabling more accurate estimations of carbon stocks, biomass, and land-use impacts. 3. In nine African countries, trees outside forests constitute more than 50% of the total tree cover, emphasizing the need for more comprehensive monitoring practices that go beyond traditional forest classifications. 4. Analysis of tree cover in relation to land cover classes and rainfall patterns showed that the majority of tree cover in dryland regions is located in areas classified as shrublands, grasslands, and deserts. 5. Approximately 433 million individual trees were identified on African croplands, with a mean density of 2.55 trees per hectare. 6. Comparison with FAO statistics revealed discrepancies in the classification of "forest" and "other wooded land," indicating the potential for improved accuracy through high-resolution mapping and refined definitions. 7. Model validation using LiDAR and UAV data showed strong correlation between the PlanetScope-derived tree cover and canopy height models, confirming the accuracy of the methodology. 8. The analysis of isolated tree detection compared to sub-meter resolution imagery demonstrates the reliability of detecting isolated trees with crown sizes larger than 30 m².

The findings of this study challenge conventional views of African tree cover by demonstrating the substantial contribution of trees outside formally defined forest areas. The high-resolution map generated using PlanetScope imagery and deep learning offers a more complete picture of tree cover distribution, crucial for informing land management and climate change mitigation strategies. This comprehensive assessment is especially significant for countries where trees outside forests represent a significant portion of the total tree cover. The discrepancies between the study's results and FAO statistics underscore the need for improved definitions and methodologies for monitoring tree cover. The integration of very high-resolution imagery with advanced machine learning techniques opens new possibilities for accurate, large-scale monitoring and assessment of tree cover changes across diverse ecosystems. The approach holds great promise for more effective monitoring of deforestation, forest degradation, afforestation, and woody encroachment across the African continent.

This study presents a high-resolution map of African tree cover that significantly expands our understanding of tree cover distribution, highlighting the considerable contribution of trees found outside traditionally defined forest areas. This high-resolution data allows for a more comprehensive assessment of carbon stocks, biomass, and land-use impacts. This methodology shows potential for scaling up to global tree cover mapping, advancing MRV initiatives and improving sustainable land management practices. Future research should explore methods for detecting smaller trees and improving the separation of individual crowns within closed canopies. The integration of additional data sources and techniques can improve accuracy further.

While this study provides a significant advancement in high-resolution tree cover mapping, several limitations should be considered. The detection of small, isolated trees with crown sizes below 30 m² remains challenging, hindering the monitoring of newly planted trees and natural regeneration. Only isolated trees can be reliably identified as individuals; further work is needed to effectively separate individual crowns in closed canopies. The current reliance on relatively coarse GEDI LiDAR data for forest type separation introduces uncertainty. Although comparisons with LiDAR data suggest trees and shrubs above 5 m are generally detected, this should not be considered a strict threshold. PlanetScope imagery is not available prior to 2015, limiting the study's ability to assess historical changes. The potential overestimation of tree cover in some forests and shrublands due to the inability to detect gaps between dense tree crowns needs to be addressed. The vast quantity of high-quality manually delineated training data required for global scaling remains a challenge.