Trees are a fundamental component of livelihoods in drylands and their use goes far beyond the provision of wood products. Dryland trees provide nutritious food and medicinal products, artifact material, cultural aspects and income to the local people, and a higher density of specific trees can improve the physical and economic wellbeing of local communities1. Recent research found that tree cover has generally increased in drylands over the past decades, which is often related to changing management patterns or climatic conditions, such as changes in the rainfall patterns that favor woody over herbaceous species2. However, not all woody species have high nutritional, ecological, cultural or economic value, and the encroachment of particular invasive species is, in some areas, considered as ecosystem degradation3. Yet, climatic changes, such as increased frequency of droughts, and human mismanagement can lead to a reduction in tree density and productivity as well as local extinction of high-value tree species4,5, which can have profound consequences for local livelihoods. Examples include loss of valuable products from the trees, such as fruits, leaves, seeds, and bark- products that local people often use for food and medicinal purposes, and to generate income6.
A prominent example of a highly valuable tree is the baobab, which is an iconic landscape feature of most of the continent’s drylands7,8,9. While the bark, pulp extract and leaves are used in the treatment of fevers, dysentery, and sores, the fruits, seeds and leaves are known for being rich in Vitamin A and C and a source of iron, calcium and zinc8,10,11. The baobab products are traditionally sold in local rural as well as urban markets. In recent years, an international demand for especially baobab fruit powder has increased, and foreign companies have started exporting baobab products, such as juice, cream and powder to global customers9,12,13. The export of baobab fruit powder to Western and Asian markets has seen a remarkable increase, estimated at several hundred tons per year, compared to negligible quantities just two decades ago12. Traditional trees like the baobab are key for local livelihoods and it is therefore important to quantify their abundance at national scales14. This is particularly true in times where plantations of fast growing and easily marketed species, such as Eucalyptus15,16,17, are promoted throughout the tropics, which often do not provide the same level of benefits, such as nutrition.
Our knowledge on the distribution of tree species in Africa is currently either derived from field surveys18,19,20,21, drones22 or hyperspectral data23, or the potential distribution is estimated from niche models21,23,25. Only the latter allows for large-scale assessments, and uses variables defining the growing conditions, such as soil and rainfall, which are used to estimate the possible abundance of certain tree species. This method includes a high level of uncertainty, in particular for species that are closely related to human management, such as the baobab. Consequently, such maps are of limited use for management purposes, as they only provide potential suitable areas for a given species to grow. To effectively monitor tree population dynamics in relation to climate change and management, ensuring adequate regeneration and sustainable management of economic outputs, it is crucial to obtain precise information regarding the abundance of tree species.
Monitoring scattered dryland trees from satellite and aerial images has been challenging until recent times, thus making local field surveys and unmanned aerial vehicles (UAV)-based mapping the most commonly used monitoring tools26,27,28. In the past years, several studies have shown that new methods and sub-meter resolution data can support a monitoring tool of trees at the level of individuals over large areas, thereby improving our knowledge on dryland trees and carbon stocks29. However, such maps do not distinguish between an invasive shrub or a baobab, and merge every woody plant under the variable tree cover. Such tree cover maps are often used to proxy ecosystem services in a given area30, although they are better suited to assess carbon stocks, climate change mitigation and wood products31,32. The creation of maps that include information on specific tree species and their distribution represent a promising avenue for improving assessments of the ecosystem services provided by various trees as well as their contributions to local livelihoods33.
Here we make use of a deep learning model being trained to identify the distinct crown shape of baobab trees in sub-meter resolution satellite imagery to create a database with the location and size of every adult baobab tree (crown size > 10 m2) across the arid and semi-arid Sahel (150–700 mm annual rainfall), from the Atlantic ocean to the red sea, an area spanning 11 countries and 1.5 million km². Using household survey data, we also document the associations between the presence of baobab trees and rural people’s consumption of dark green leafy vegetables. Finally, we discuss the implications of establishing a publicly accessible database, containing such precise geospatial information on all individual baobab trees.
For this study we acquired satellite images covering the entire Sahel at 50 cm resolution during the early dry season32, a period where only Adansonia digitata, Bombax costatum, Ceiba pentandra and Sterculia setigera are large trees which have no or only few leaves. To avoid confusion between these species, we included the shadow of the distinctive branches in the training samples (Fig. 1c-d). We manually labeled 5013 trees used for training the model. The expert labeling was done in Senegal and Mali in regions visited by the authors, and guided by GPS-referenced field photos taken between 2010 and 2015. Only clear examples were trained. The model mapped ~ 2.8 million baobab trees over the entire study area (Fig. 1a). We used independent Skysat data (Fig. 1c-d) to count baobab trees in randomly selected areas, which was compared to our predictions and results in an underestimation bias of 27.1% (Fig. 2a). We subsequently split the mapped baobabs in three size classes (large > 150 m2, medium > 70–150 m2, and small 10–70 m2) and found that large trees can be mapped with a lower uncertainty (17.5%). High densities of baobabs are found in Senegal, Mali, and Sudan. The Kayes region in Mali had the highest density with up to 150 baobabs per km2 (Fig. 1b), which was confirmed by local reports (Fig. S2).
Due to the high density of baobabs (Fig. 1b) and the availability of high quality auxiliary datasets, we selected Senegal to further study the distribution of baobab trees in more detail. We found 665,973 baobabs in Senegal of which 58.8% are small, 26.5% have a medium size, and 14.7% are large trees. Compared with a wall-to-wall assessment of all trees from all species29, baobabs only account for 0.0009% of the number of trees in Senegal.
We then used household survey data from the DHS program (Demographic and Health Survey) from 387 rural household cluster records in Senegal to examine the relationship between rural people’s consumption of dark green leafy vegetables, an important nutritious food group, and the abundance of baobabs. We used the consumption of dark green leafy vegetables (including also other leaves than baobab leaves) as an important sentinel food group, likely to be influenced by the abundance of baobab trees. The variables were recorded as a binary value (consumption/no consumption) in the past 24 hours for children (between 12 and 60 months). We found a significant positive association (p < 0.01) between the abundance of baobabs and the consumption of dark green leafy vegetables (Fig. 2b). No relationship was found if trees of all species were used29, documenting that tree cover alone is not related to the consumption of dark green leafy vegetables in the Sahel.
By using the Google Open Building dataset that includes 8.4 million buildings in Senegal (Fig S3), we found that 78.5% of all rural households have at least one baobab tree within their immediate surroundings (1 km), and only 5.8% of households do not have baobab trees within a distance of 2 km (Fig. 2d). Moreover, most of the mapped baobabs grow in immediate distance to rural households (Fig. 2c) (here rural is defined as a settlement with less than 10,000 people).
We finally used a gradient boosted regression tree analysis to study the variables affecting the spatial distribution of baobabs (Fig. 3). With a total explanatory power of 79.8%, we found that climate factors dominate (62.1%), followed by proximity to rivers and human settlements, which may be interrelated (20.7%), and soil variables (17.2%). Note that climate factors are also interrelated with the presence of human settlements.
We acknowledge that a mapping at the 50 cm scale as presented here cannot currently be repeated on a frequent basis due to image cost and availability, thereby impeding the use of such work as a monitoring tool. However, recent work has shown that cost efficient nanosatellites such as Planet or RapidEye are able to reliably map scattered dryland trees with a crown size > 30 m² as individuals34, which applies to most adult baobab trees (Fig. 4). Species information cannot be readily derived from those images, but a nanosatellite-based tree map can be combined with the benchmark baobab tree species map provided here to track the future status of known baobab trees. Removal, mortality, survival/recovery and even the health status of identified baobab trees can then be tracked in future times using low-cost images. The importance of a regular monitoring tool, not only for tree cover but single tree species, is highlighted by a recent study, which found a high death rate of old baobab trees in different African countries within the past decade35.
Our study demonstrates the feasibility of mapping the precise distribution of individual tree species at a subcontinental scale. This information is crucial in times where human mismanagement and climate change cause the extinction of numerous tree species34,35. Mapping of individual tree species at large scale creates economic opportunities, but inevitably also involves a danger in relation to over-exploitation of local resources. This is in particular problematic for areas where local people strongly depend on the resources provided by trees, which is the case for baobabs in West Africa14,15,19,36,37. Here, a better understanding of where to conserve and invest in long-term sustainable use can support the development of rural communities and promote value chains28,36,37,38. Nevertheless, regardless of the baobab example, the precise mapping of individual tree species is sensitive towards unsustainable economic exploitation, which may then impose a threat to these trees. National and foreign companies interested in selling products derived from fruits and leaves, do not always create opportunities for local livelihoods, but may also lead to disadvantageous situations, such as over-exploitation of resources and by-passing of local markets. In extreme cases, rare tree species possessing valuable timber could be located, extracted, and exported41. This adds a challenge to the concept of open data, and it is important that data, such as the database produced and presented in this study, is responsibly managed by national and local authorities, and used to strengthen and develop rural communities.