Figure - available from: Nature
This content is subject to copyright. Terms and conditions apply.
Overview of training sites and study area The study area for the wall-to-wall mapping is the westernmost part of the Sahara and Sahel. It represents a typical north–south ecological and climatic gradient, starting in the Sahara Desert in hyper-arid areas (rainfall of 0–150 mm yr−¹) with a sparse vegetation coverage, over arid (rainfall of 150–300 mm yr⁻¹) and semi-arid (rainfall of 300–600 mm yr⁻¹) Sahelian rangelands and croplands, up to sub-humid (rainfall of 600–1,000 mm yr⁻¹) Sudanian lands, where shrublands turn into forests. a, The locations of the manually drawn 89,899 tree crowns used for training the model are shown in red. CHIRPS rainfall⁴³ was used to delineate the rainfall zones. The land use for farmland and urban is from Copernicus Global Land²⁶. In situ data were collected at the field sites around Widou and Dahra in Senegal. Areas of insufficient data quality and beyond rainfall of 1,000 mm yr⁻¹ were masked. b, The region was analysed for sandy (>70% sand content) and non-sandy areas⁴⁴.
Source publication
A large proportion of dryland trees and shrubs (hereafter referred to collectively as trees) grow in isolation, without canopy closure. These non-forest trees have a crucial role in biodiversity, and provide ecosystem services such as carbon storage, food resources and shelter for humans and animals1,2. However, most public interest relating to tre...
Citations
... Considering the irregular distribution of wind turbines within forest areas and the presence of heterogeneous backgrounds on land, identifying onshore wind turbines through conventional offshore wind turbine detection methods presented significant challenges [47]. Thus, satellite images with 2 m or even sub-meter resolutions were commonly used to extract small objects [63]. In this study, we used Jilin-1 satellite images with a resolution ranging from 1.5 to 2.0 m. ...
Wind power plays a pivotal role in the achievement of carbon peaking and carbon neutrality. Extensive evidence has demonstrated that there are adverse impacts of wind power expansion on natural ecosystems, particularly on forests, such as forest degradation and habitat loss. However, incomplete and outdated information regarding onshore wind turbines in China hinders further systematic and in-depth studies. To address this challenge, we compiled a geospatial dataset of wind turbines located in forest areas of China as of 2022 to enhance data coverage from publicly available sources. Utilizing the YOLOv10 framework and high-resolution Jilin-1 optical satellite images, we identified the coordinates of 63,055 wind turbines, with an F1 score of 97.64%. Our analysis indicated that a total of 16,173 wind turbines were situated in forests, primarily within deciduous broadleaved forests (44.17%) and evergreen broadleaved forests (31.82%). Furthermore, our results revealed significant gaps in data completeness and balance in publicly available datasets, with 48.21% of the data missing and coverage varying spatially from 28.96% to 74.36%. The geospatial dataset offers valuable insights into the distribution characteristics of wind turbines in China and could serve as a foundation for future studies.
... Future improvements should focus on expanding LiDAR coverage, incorporating the terrestrial LiDAR data and enhancing segmentation algorithms, which is critical for generating more reliable tree density estimates at large scales. Furthermore, the increasing availability of high-resolution remote sensing imagery, providing detailed canopy information, offers promising opportunities for improved tree identification [16]. Integrating these high-resolution images with LiDAR data in future studies could significantly enhance the precision of tree density estimates. ...
... Remote sensing studies have recently raised awareness about the high woody plant cover and large number of tree individuals present across the Sahelian drylands in northwest Africa (Brandt et al., 2020), despite the dual challenge imposed by low soil fertility and drought stress on plant photosynthesis. Legumes (Fabaceae) are often the dominant plant family in terms of cover and species diversity in Sahelian drylands (Felker, 1981;Sprent & Gehlot, 2010) and are also widely used for livestock feeding. ...
... Modeling studies at a global scale have shown that stomatal conductance should reach peak levels in dry tropical vegetation (savannah trees; Lin et al., 2015 δ 13 C values of Sahelian woody plants suggests that they are capable of achieving high rates of transpiration and carbon assimilation during the short rainy season (Cornwell et al., 2018;Sibret et al., 2021). Sahelian woody vegetation may thus contribute substantially to global primary productivity, despite the severe water and nutrient limitations typical of these drylands (Ahlström et al., 2015;Smith et al., 2019;Brandt et al., 2020). High foliar N content can reflect the production and accumulation of N-based osmolytes, such as proline, that enable plants to endure and sustain more negative internal water potentials during drought (Wink, 2013;Adams et al., 2016), as suggested by the strong negative correlation found between foliar N mass and predawn water potential across species (Table 3). ...
... The possibility of detecting individual trees by performing segmentation on the tree crowns observed allows for a more accurate estimation of tree count and cover [ 21 , 31 -33 ] and can contribute to obtaining more accurate estimations of carbon stocks [ 30 ]. In light of the recent advances in mapping individual tree crowns, particularly in semiarid regions with a low density [ 21 ], Hiernaux et al. [ 34 ] developed allometric equations that can be directly applied using only the derived individual tree crown masks, with the tree crown area as its main input parameter. Although additional allometries have also been developed to include height as a parameter, Tucker et al. [ 19 ] used crown-area-based allometric equations to obtain the carbon stocks of all individual trees in African drylands. ...
... We identified individual tree crowns using the deep learning instance segmentation model developed by Brandt et al. [ 21 ]. This model has been trained and validated in the Sahel, with similar characteristics to our study sites and using VHR imagery at the same pixel resolution, 0.5 m. ...
Quantifying tree resources is essential for effectively implementing climate adaptation strategies and supporting local communities. In the Sahel, where tree presence is scattered, measuring carbon becomes challenging. We present an approach to estimating aboveground carbon (AGC) at the individual tree level using a combination of very high-resolution imagery, field-collected data, and machine learning algorithms. We populated an AGC database from in situ measurements using allometric equations and carbon conversion factors. We extracted satellite spectral information and tree crown area upon segmenting each tree crown. We then trained and validated an artificial neural network to predict AGC from these variables. The validation at the tree level resulted in an R² of 0.66, a root mean square error (RMSE) of 373.85 kg, a relative RMSE of 78.6%, and an overestimation bias of 47 kg. When aggregating results at coarser spatial resolutions, the relative RMSE decreased for all areas, with the median value at the plot level being under 30% in all cases. Within our areas of study, we obtained a total of 3,900 Mg, with an average carbon content per tree of 330 kg. A benchmarking analysis against published carbon maps showed that 9 out of 10 underestimate AGC stocks, in comparison to our results, in the areas of study. An additional comparison against a method using only crown area to determine AGC showed an improved performance, including spectral signature. This study improves crown-based biomass estimations for areas where unmanned aerial vehicle or height data are not available and validates at the individual tree level using solely satellite imagery.
... The inclusion of trees in farming systems is far from a new approach; trees are found on approximately 43% of global agricultural land, 34 although more recent estimates indicate that the prevalence might be much higher. 35 While emerging evidence suggests that on-farm trees can improve diets, 36 on-farm trees do not deliver universally positive or equal nutrition outcomes. The heterogeneous classifications for on-farm tree systems often vary with geographical and cultural contexts and do not often account for the ways in which the trees are used by the household. ...
... This gap hampers our understanding of the specific challenges and socio-ecological impacts in areas with high agricultural spatial concentration, such as Michoacán, Mexico, the world's leading avocado producer (Ramírez-Mejía et al., 2022). Recent technological advances in remote sensing and deep learning have revolutionized our ability to map and analyze land systems, offering new insights into land-use change processes driven by agricultural expansion and intensification globally (Helber et al., 2019;Brandt et al., 2020;Karra et al., 2021). These tools enable the detailed characterization of land system patterns and support the generation of landscape metrics that enhance our understanding of agricultural frontier dynamics (Levers et al., 2018;Pacheco-Romero et al., 2021;Baumann et al., 2022). ...
Agricultural expansion and intensification are major drivers of global biodiversity loss, endangering natural habitats and ecosystem functions, such as pollination. In this study, we analyze the spatiotemporal dynamics of avocado frontier expansion and intensification from 2011 to 2019 and assess their effects on landscape connectivity, focusing on Michoacán, Mexico, the world’s leading avocado exporter. Using high-resolution satellite imagery combined with deep learning based on convolutional neural networks, we delineated avocado orchards and other land use/cover classes, mapped individual avocado tree crowns and irrigation ponds, and identified hotspots of expanding and intensifying avocado production. We used a circuit theory approach to evaluate the effects of avocado expansion and intensification on the connectivity of natural and semi-natural habitats. Our results reveal a rapid increase in avocado orchards, which expanded by 4175 ha—a growth from 27.9% to 37% in area—over the eight-year period. There was also a decline in rainfed agriculture by 3252 ha, and oak-pine forests by 1343 ha. We observed not only the expansion of the avocado frontier into forests but also an intensification of avocado production via increases in high-density plantations, irrigation ponds, and orchards prone to intensive pruning. Moreover, lower-intensity land-use classes, such as rainfed crops, were rapidly converted to avocado orchards. This expansion and intensification have led to increasing isolation of forest fragments. Although we identified routes that could facilitate the movement of species, the dense avocado monocultures continue to threaten the connectivity of natural and semi-natural habitats, causing notable losses of old-growth oak-pine forests and disrupting crucial ecological corridors. Our research underscores the adverse effects of avocado production on land use and landscape connectivity, emphasizing the need for sustainable management practices to ensure the long-term viability of avocado production systems and overall ecosystem functioning.
... The advantages of deep learning models for detailed ecosystem analyses increase with the spatial resolution 29 and might diminish if the pixel resolution is coarser than about 10 m. If the resolution of the image is at least 5 m and ideally sub-metre, trees can be mapped as individual objects 19 . Thus, there is the possibility to map trees at tree level, which enables the mapping and the monitoring of trees outside forests. ...
... They are, thus, generally not detectable with Landsat satellite data 18 . Only a fraction of the Earth's land surface is covered by dense forests 14 , and as of 2024 it remains a challenge to assess tree resources in all non-forest areas globally 19 . ...
... Combining objectbased classification with landscape connectivity modeling has been shown to be a powerful tool [36], but further work needs to be done to create data sets for other key features (e.g., fences) and parametrize them for specific landscapes (e.g., relative impact of wire vs. electric fences). Additionally, these methods can be used to improve vegetation mapping by identifying invasive species [37] and keystone trees [38]. Furthermore, high-resolution satellite imagery is increasingly available for identifying these small landscape features. ...
... While extensive attention has been directed toward forests, defined as areas with a canopy cover exceeding 10 % and spanning more than 0.5 ha (FAO 2004), the significance of solitary trees beyond forested domains cannot be understated. These trees play a pivotal role in land stabilization, combating desertification, safeguarding watersheds, and contributing to wood and economic production (Brandt et al. 2020). Their importance is particularly pronounced in semi-arid and arid territories, which span a third of the world's landmass and sustain approximately one billion inhabitants (Malagnoux et al. 2007;Xia et al. 2021). ...
... For instance, G. Braga et al. (2020) achieved over 90 % precision by employing Mask-RCNN in delineating tree crowns within tropical forests captured by WorldView and Quickbird imagery. Even more noteworthy was the application of U-net in delineating tree crowns across the entirety of West Africa, as demonstrated by Brandt et al. (2020). And, some studies observe that deep learning methods have significant improvements in both accuracy and inference speed (Lassalle et al., 2022;Zhu et al., 2024). ...
... However, the process of generating these training labels is notably time-consuming and requires intensive human labor. As an example, Brandt et al. (2020) trained models using more than 89,000 manually delineated tree crowns. It is therefore crucial to annotate such a massive number of samples to enhance the performance of CNNs for tree crown segmentation over such a vast area. ...
... Skole et al (2021) have articulated a TOF monitoring approach that would expand the current REDD+ framework using high-resolution satellite data and machine learning at the individual tree level. With the increased availability of satellite remote sensing data at resolutions <1 m, and machine learning models that can segment individual trees from the landscape background, TOF measurement, reporting, and verification (MRV) for climate mitigation policy seems possible (Brandt et al 2020, Skole et al 2021b, Mugabowindekwe et al 2023, Reiner et al 2023, Tucker et al 2023. However, Brandt et al (2023) have argued against the need for very highresolution data which can be expensive and difficult to use for change detection, suggesting instead the use of lower resolution observations from platforms such as Sentinel-2 which is readily accessible, low cost and temporally dense. ...
... These results and the approach could be applicable in other regions. The results here are consistent with a growing body of literature regarding individual tree crown mapping using VHR remote sensing (Brandt et al 2020, Reiner et al 2023 and application to carbon stock estimates (Mugabowindekwe et al 2023). However, important challenges remain. ...
To reduce emissions of carbon and other greenhouse gases on a pathway that does not overshoot and keeps global average temperature increase to below the 1.5 °C target stipulated by the Paris Agreement, it shall be necessary to rely on nature-based solutions with atmospheric removals. Without activities that create removals by carbon sequestration it will not be possible to balance residual emissions. Policies that focus solely on reducing deforestation will only lower future emissions. On the other hand, activities that include regeneration or regrowth of tree biomass can be used to create net-zero emissions through carbon sequestration and atmospheric removals now. New methods demonstrated here using high resolution remote sensing and deep machine learning enable analyses of carbon stocks of individual trees outside of forests (TOF). Allometric scaling models based on tree crowns at very high spatial resolution (<0.5 m) can map carbon stocks across large landscapes of millions of trees outside of forests. In addition to carbon removals, these landscapes are also important to livelihoods for millions of rural farmers and most TOF activities have the capacity to bring more countries into climate mitigation while also providing adaptation benefits. Here were present a multi-resolution, multi-sensor method that provides a way to measure carbon at the individual tree level in TOF landscapes in India. The results of this analysis show the effectiveness of mapping trees outside of forest across a range of satellite data resolution from 0.5 m to 10 m and for measuring carbon across large landscapes at the individual tree scale.
