Landscape restoration and greening in Africa

As a reaction to ongoing environmental change, many local land restoration projects have emerged that aim to prevent or reverse land degradation, combat climate change through carbon sequestration or improve the local climate. However, the contribution of these projects to the greening of Africa at larger scales is still unknown due to the absence of a (public) complete database of land restoration projects, the lack of monitoring and the low survival rate of planted vegetation. Here, we use climate independent greening time series to detect local greening hotspots in Africa. We find that 2.1% of Africa, an area of roughly 400 000 km2, experiences local greening, especially in semi-arid environments. We show that various forms of sustainable land management (SLM) lead to significant local greening and demonstrate that some forms, e.g. active revegetation, are more effective than others, e.g. natural regeneration. This study, therefore, provides a first continental-scale insight in the greening potential of land restoration, which is needed for a thorough understanding of the effectiveness of SLM.


Introduction
Despite the increasing efforts to halt environmental change [1], land degradation continues to affect an estimated 1.3-3.2 billion people worldwide [2,3], of which the majority lives in developing countries [4]. African drylands such as the Sahel are particularly vulnerable to the effects of land degradation and climate change due water scarcity and population growth [5][6][7]. Recognizing these pressures we put on key ecosystems, the United Nations declared 2021-2030 to be the Decade of Ecosystem Restoration to support efforts preventing global ecosystem degradation and increase awareness of the importance of restoration [8]. In addition, several organisations have taken the ambitious plan to restore millions of hectares of land in the coming decades by planting billions of trees in African drylands as well as other parts of the world (e.g. the Bonn Challenge, the African Forest Landscape Restoration initiative (AFR100) or the African Great Green Wall initiative [6]). Furthermore, land restoration and tree planting projects have also become a widely recognized approach to combat climate change through carbon sequestration [9][10][11] or by changing the biophysical properties of the land surface [12][13][14][15]. Some researchers even argue that land restoration may be one of the most effective methods for climate change mitigation [16].
As a result, the number of land restoration projects in Africa has rapidly increased over the years [1]. Monitoring these projects is often done with remote sensing products such as the Normalized Difference Vegetation Index (NDVI) [17] or the Enhanced Vegetation Index (EVI), as field measurements are often time-intensive, costly and, in case of remote or large areas, practically impossible. Using these vegetation indices in combination with, for example, a before/after control/impact type of designs [18,19] has been used to estimate the changes in greenness and vegetation productivity of projects with a known location. However, the lack of a complete and publicly accessible database of land restoration projects in Africa, the considerable number of organizations that work on this [1] and the reported low survival rates of planted vegetation [20][21][22][23], make these methods less suitable to evaluate the greening of land restoration projects and their climate change mitigation and adaptation potential on a continental scale.
On top of that, the African continent is not only affected by greening due to small-scale processes such as land restoration, but also by largescale and long-term greening and browning trends. In the 1970s and 1980s Africa has experienced severe large-scale droughts, which are now attributed to the El Niño-Southern Oscillation and changes in sea surface temperature [24,25]. Contrarily, observations over the last decades have shown an overall increasing trend in vegetation cover across Africa [14,26], which is likely caused by an increase in global CO 2 concentration [26] and an increase in precipitation due to changes in sea surface temperature [27][28][29][30].
Due to the co-existence of small-scale greening caused by land restoration, and this large-scale 'background' greening, simply monitoring changes in vegetation indices does not tell us the effectiveness of land restoration projects or its contribution to the greening of Africa, but rather shows the combined effect of land management and natural climate variability. To compensate for this effect, previous studies have used vegetation-rainfall relations [31], which can partly account for background trends because it provides an indicator of vegetation or ecosystem functioning. Yet, vegetation-rainfall relationships are complex and often differ over biomes, making it less suitable to study vegetation productivity over large scales [32,33]. In addition, vegetationrainfall relations do not consider background trends other than rainfall variability, while literature suggests that CO 2 fertilization causes roughly 70% of the observed greening [26].
Alternatively, we can use spatial-context to separate the small-scale or 'local' greening from large-scale background trends. In this approach, it is assumed that background trends due to natural climate variability act on a much larger scale than a land restoration project [34]. For example, if the greening at a land restoration project is the result of natural climate variability rather than the project itself, surrounding areas will likely show a similar amount of greening as the project area. Contrastingly, if the greening is caused by the project, it is expected that the project area shows a larger amount of greening than surrounding areas. For this reason, the background trends can be removed from vegetation index time series by comparing observed greening trends with surrounding areas.
Here, we apply this spatial-context method to NDVI and EVI time series in Google Earth Engine [35] to (1) create a map of local greening hotspots for Africa, (2) compare the spatial distribution of local greening to background greening, and (3) compare local greening to a publicly available database of sustainable land management (SLM) projects to determine the effectiveness of land restoration. Although spatial-context methods have been used before to detect deforestation [34,36], land degradation [37,38] or burned areas [39] in forested as well as grassland areas, this is, to our knowledge, the first time that such a spatial-context method is used to detect local greening hotspots on a continental scale in Africa.

Input data and study area
We used four different input variables (table 1). The main input data consists of NDVI [40] and EVI [41] time series data (extended data figure 1), which we used as an indicator for vegetation productivity. The main analysis is performed using NDVI and EVI data from the Moderate-Resolution Imaging Spectroradiometer (MODIS) [40] because it is a good compromise between spatial and temporal resolution and the readily availability of quality controlled vegetation index composites. We used Landsat-7 data on a small sample area to explore the effects of a higher spatial resolution on our results (extended data figure 10). Furthermore, we used land cover (45) and aridity index (AI) data to provide some insight into potential causes of local greening. The AI is a measure of dryness and can be defined as a 30-year average fraction between precipitation and potential evapotranspiration [42], which we used to divide the study area into hyper-arid, semi-arid, dry subhumid and humid regions [43]. To evaluate the effect of land restoration practices on the amount of local greening, we used 434 SLM projects from the World Overview of Conservation Approaches and Technologies (WOCAT) database [44] within the study area. This database contains often-used land restoration techniques, such as tree planting or assisted natural regeneration, but also techniques like sustainable agriculture and water harvesting. Here, the whole database of SLM projects is used to study the process of land restoration, meaning we used a broader definition of the term land restoration than other studies (e.g. UNCCD [7] or IUCN [45]). If a single SLM project contained multiple locations, we considered it as multiple projects, resulting in 628 project locations. We categorized these projects into 11 categories and 55 subcategories based on their description, where the categories of 'revegetation' and 'natural regeneration' are highlighted in the results. We refer to the combination of all categories simply as 'sustainable land management' (SLM) (table 2). The study area consists of areas on the African continent with a median NDVI higher than 0.15 or EVI higher than 0.11, which results in similar case study boundaries for the NDVI and EVI. This way, areas with a too low vegetation cover are not considered in the calculations to prevent noisy results. The masked areas mainly consist of the Sahara Desert and constitute 36% of the African continent (extended data figure 1).

Category Description Project categories included
Sustainable land management 'The use of land resources, including soils, water, animals and plants, for the production of goods to meet changing human needs, while simultaneously ensuring the long-term productive potential of these resources and the maintenance of their environmental functions' [44]. WOCAT mainly focusses on preventing and reducing land degradation All WOCAT SLM projects, e.g. runoff harvesting, alternative cooking methods, riverbank restoration, planting trees, area closure, fire management, cover crops, agroforestry and conservation agriculture.

Revegetation
Active planting of vegetation species to accelerate vegetation regrowth.
Planting (fruit) trees, shrubs and grasses, implementing vegetation strips, and projects described as restoration. Natural regeneration A passive method of regreening, where vegetation cover is increased through natural regrowth by using, for example, area closure, grazing management or management of invasive species.
Assisted natural regeneration, combating invaders, farmer managed natural regeneration, area closure, grazing management and bush thinning.

Spatial-context approach
To separate background trends from the NDVI and EVI time series, we used a spatial-context approach. For each pixel, we determined the vegetation index time series over the 2001-2021 period, after which we calculated neighbourhood averaged time series over a square-shaped neighbourhood around the pixel. A square centre with a radius of 1 km, corresponding to a square of 2 × 2 km, was not included in the mean to reduce the influence of the original time series on the mean neighbourhood time series. Next, we subtracted the neighbourhood time series form the centre pixel time series to create 'spatially corrected' time series. Creating this spatially corrected time series for the NDVI and EVI allows us to evaluate the changes in greenness compared to surrounding areas, thus separating greening trends resulting from small-scale processes and land management from those caused by natural climate variability.
In this study, we applied the spatial-context method to three neighbourhood radiuses of 25 km, 10 km and 5 km, corresponding to squares of 50 × 50 km, 20 × 20 km and 10 × 10 km, respectively. We highlight the results of the 25 km radius, which captures the effect of large land restoration efforts without crossing multiple AI classes. The results for 10 km and 5 km radiuses are included in the supplementary data.

Definition of local greening
We applied the Breaks For Additive Seasonal and Trend (BFAST) algorithm [49][50][51] for Google Earth Engine [52] to the spatially corrected NDVI and EVI time series. BFAST decomposes the time series into a trend, seasonal and remainder component, by fitting a linear/harmonic model to the time series. Unlike other decomposition algorithms, BFAST can detect significant changes, called breakpoints, in the components, resulting in a piecewise linear harmonic model. We applied BFAST using a seasonal harmonic model order of 3, a minimum spacing between two breakpoints of 0.15 (fraction of total time series length, i.e. 3 year) and a maximum of one breakpoint. Here, we use BFAST instead of linear regression because we expect land restoration to show a sudden change in greenness compared to surrounding areas rather than a gradual change, which BFAST can capture in the form of a breakpoint. We therefore expect the breakpoint to represent the moment a project, or another process, will start to affect vegetation cover.
We defined local NDVI/EVI greening as pixels where: (1) the BFAST algorithm detects a breakpoint in the trend of the spatially corrected NDVI/EVI time series, (2) the computed BFAST trend after the breakpoint is positive, significantly different from zero (p = 0.05) and larger than before the breakpoint, and (3) the original (centre pixel) NDVI/EVI time series shows a positive linear trend after the breakpoint ( figure 1(A)). We included this last condition, as pixels could theoretically show a greening trend compared to its surroundings, even though the pixel itself is browning. Next, we defined local greening hotspots in Africa as areas that simultaneously experience local NDVI greening and local EVI greening. By combining the NDVI and EVI, we aim to reduce noise and therefore improve the accuracy of this spatial context method.

Calculation of background trends
To calculate the background trends, we applied linear least squares regression to the original NDVI and EVI data between 2001 and 2021 within the study area ( figure 1(B)). We then assigned background greening to pixels that show a significant positive trend for both the NDVI and EVI (p = 0.05). Similarly, browning trends show a significant negative trend. Areas that have a positive NDVI trend and a negative EVI trend or vice versa, and areas without a significant trend are considered not to have a background trend.

Local greening of SLM
Next, we used the WOCAT project database to evaluate the regreening effects of SLM projects. Therefore, we used local greening instead of the often-used background greening, to reduce the effects of large-scale processes such as natural climate variability, which makes it more likely that observed greening is due to the changes in land management. Because WOCAT only contains point coordinates of projects instead of boundaries, we cannot directly calculate the amount of local greening inside the project. Instead, we computed the percentage of local greening pixels in a circle around the project's geo-tag multiple times, using a radius of 5000, 4000, 3000, 2000, 1000 and 500 m. Next, we computed for each project the percentage of greening pixels over the locations in the study area within the same country, AI class and land cover class as the project (extended data figure 2). We then determined whether SLM projects cause a significant increase in local greening using the two-sided ttest for independent samples, assuming unequal variances and a significance level of 5%.

Spatial distribution of local greening in Africa
Applying the spatial-context method ( figure 1(A)) to the African continent with a neighbourhood radius of 25 km, 2.1% of the study area (roughly 400 000 km 2 ) shows local greening over the last two decades ( figure 1(C)). Most of these areas have a breakpoint towards the end of the study period, with peaks around 2015 and 2018, suggesting that many areas start to show an increase in greenness compared to its surroundings around these years ( figure 1(E)), although it should be noted that the used settings of the BFAST algorithm do not allow for the detection of breakpoints after 2018. We can also observe that the local greening is not evenly distributed across the continent, as a large part of the local greening can be found in the Sahel, Kenya, Tanzania, and regions in southern Africa (extended data figure 2). Similar spatial distributions can be found for a neighbourhood radius of 10 km and 5 km, although the total area classified as local greening as well as the number of adjacent local greening pixels is smaller (extended data figure 8). For a 10 km and 5 km radius, respectively 1.9% and 1.8% of the study area shows local greening, compared to 2.1% for a 25 km radius.
We also compared the local greening hotspots to NDVI and EVI background trends ( figure 1(B)) to provide more insight into the contributions of local greening to the greening of the African continent. Overall, a larger area shows background greening (32.4% of the study area) than local greening (2.1% of the study area), especially in more humid areas such as the Congo Basin in central Africa (figure 1(D); extended data figure 7). These areas do usually not show a breakpoint when BFAST is applied to the spatially corrected NDVI time series, suggesting that a large part of the background trends, especially in humid areas, is not caused by small scale or abrupt processes, but by longer large-scale processes. In other, dryer areas such as in Botswana and Namibia, we observe strong background greening combined with a large amount of local greening pixels, suggesting a combination of large-scale and smallscale processes. Although most local greening in the study area is located in areas that also show a background greening trend (45.1%) or no background trend (41.6%), also a considerable amount of local greening is present in areas that show a long-term browning trend (13.3%) ( figure 1(F)). The combination of local greening and background browning suggests a long term linear browning trend, with a sudden increase in greenness, compared to surrounding areas, at the end of the time-series. A similar spatialcontext method can of course be used to detect local browning, which occurs at 1.9% of the study area (extended data figure 9).

Drivers and properties of local greening hotspots
We compared local greening to an aridity and land cover classification and find that 39.1% of the local greening can be found in semi-arid regions, while these regions only cover 26% of the study area (figure 2(H); extended data table 1). Here, 3.0% of the area is found to be greening compared to its neighbourhood. Humid areas, on the other hand, account for only 32.0% of the local greening, while covering more than half of the study area. This pattern is also visible in the distribution of the local greening over the land cover types, as most local greening occurs in shrublands, grasslands and savannas, but also areas classified as cropland sometimes show local greening (figure 2(H); extended data table 1).
By visually interpreting high resolution satellite imagery, we also aim to provide some more insight into potential drivers of local greening. Land restoration practices such as in Kenya, show greening between 2014 (figure 2(A) and 2019 ( figure 2(B)), while surrounding areas are untreated. Similar results can be found in South-Africa, where local greening pixels are located roughly inside protected areas [53] (figure 2(F)). However, also reduction in open-water surface area, such as in Lake Chad, can cause unwanted classification of local greening (figures 2(C) and (D)). This is caused by the low NDVI value of water and, consequently, a local increase in NDVI when the water retreats. In addition, we observed a reasonably large amount of local greening in agricultural areas such as in the north of Nigeria ( figure 3(C)), which is also visible in the histogram ( figure 3(H)) and may result from intentional greening efforts.  figure 2). For readability, we did not include the land use and aridity classes that contained less than 5% of the greening pixels. The complete histogram is attached in extended data table 1.

SLM as driver of local greening
As most of the local greening can be found in semi-arid environments, most SLM projects in these regions also show generally a high percentage of local greening ( figure 3(A)). Furthermore, even though projects in more hyper-arid or humid regions show less local greening, SLM projects show an overall significant increase in local greening compared to the areas with similar aridity and land cover for all distances around the project we included in this study ( figure 3(B)). Revegetation projects, including the planting of (fruit) trees, shrubs and grasses, have a significant higher percentage of local greening in an area of 500 and 1000 m around the project's geo-tag, while the other radiuses show an increased, yet not significant, local greening. Especially planting trees and fruit trees seems to have a large effect (extended data figure  11). Natural regeneration projects, such as farmer managed natural regeneration, assisted natural regeneration or area closure, appears to have a lower effect on the amount of local greening than revegetation projects for all area sizes around the project. Yet, there is a significant increase in local greening around 5000, 4000 and 3000 m around the project. Other categories of SLM, such as water harvesting, erosion prevention or agriculture management have a more mixed effect on local greening (extended data figure 11).

Discussion and conclusions
In this study, we used a spatial-context approach in Google Earth Engine, to separate small scale greening caused by land restoration and SLM, from background trends due to natural climate variability. With this method, we showed that that 2.1% of the African continent experienced local greening over the 2001-2021 period, especially in semi-arid environments. In more humid regions, we saw less local greening, even though these regions showed significant background greening. This matches with our expectations from the spatial-context method, because changes in land management may result in a larger vegetation cover increase in sparsely vegetated semi-arid regions than in regions that are already densely vegetated. In humid areas it is less likely that small scale processes result in such a large increase in vegetation cover that it would be detected as a breakpoint in the NDVI and EVI time series.
In addition, our results also suggests that even though SLM as a whole has a positive effect on the amount of local greening, revegetation and tree planting appears to be more effective than natural regeneration over smaller areas. Natural regeneration does, however, show a significant increase in local greening pixels around a larger area than active revegetation. This coincides with other studies, that found a faster recovery of highly degraded land due to active restoration practices such as tree planting compared to more passive methods [55,56]. On the other hand, natural regeneration is, although slower, often much cheaper [57] and can therefore result in the restoration of much larger areas, with a more natural species composition [58] if enough time is available. Therefore, natural regeneration projects may prove to be more effective once a longer study period is available. In addition, we want to emphasize that greening is often not the only goal of SLM and land restoration projects [1], as people also implement land restoration for biodiversity conservation, income generation, legislations or cultural reasons [59]. A lack of local greening found for some projects in this study does therefore not necessarily mean that the SLM project is not at all effective or failed, as greening may not have been the main goal of the project.
Yet, several uncertainties that should be kept in mind when interpreting the spatial-context approach. As we defined local greening as a sudden increase in greenness compared to surrounding areas, it does not only show changes in land restoration, but every type of small-scale greening, including changes in water level or agriculture. This calls for further development of the spatial-context method such that these different causes can be split, allowing for the evaluation of land restoration only. In addition, we evaluated the SLM projects based on a circle around a point coordinate. An exact size and boundary of each project would result in a more accurate evaluation. Furthermore, the moderate spatial resolution (250 m) of MODIS limits the detection of small SLM projects and spatial heterogeneity within these projects. Because 7% of the projects within the WOCAT database reported a size smaller than the spatial resolution of MODIS, there lies great potential in highresolution satellite imagery or microwave and LiDAR observations, which can detect vegetation optical depth, forest structure or even individual trees and shrubs outside forested areas [60][61][62]. Unfortunately, higher-resolution imagery usually comes with a lower temporal resolution or shorter data range, which limited their use in this study. In addition, applying the spatial-context method to higher-resolution Landsat-7 data showed similar patterns, but smaller areas of local greening, probably caused by the lower temporal resolution of Landsat-7. This suggests that this data will not result in a more sensitive detection of local greening. Similarly, the spatial-context method is unable to detect greening over areas larger than the used neighbourhood (approximately 2500 km 2 ), due to uniform greening of the neighbourhood. Fortunately, only 4% of the SLM projects reported to have a size larger than 2500 km 2 , although it should be noted that 38% of the projects did not report any project size and oftentimes it is not clear whether this is the actual size of the project or of the region in which it is implemented. Finally, studying seasonal changes of greenness due to land restoration more in depth may provide useful information for policy makers when implementing land restoration projects.
These days, tree planting and land restoration projects are seen as an effective method for climate change mitigation, and are therefore widely included in climate change policy proposals, resulting in ambitions tree planting projects across the world [16,63]. However, if not implemented correctly, land restoration can also have severe negative consequences on the environment [64], such as a decrease in biodiversity through monocultures or non-native vegetation [65,66], changes in water availability through increased evapotranspiration [67][68][69], or the destruction of native ecosystems through displacement of land use [70]. The increasing interests in land restoration thus asks for more research on the cost and benefits of land restoration [71,72]. However, due the co-existence of small-scale greening and background trends, in combination with the lack of an available (complete) dataset of land restoration projects, it is not known to what extend land restoration actually causes greening. In this study, we provided insight into the hotspots of local greening, as well as an objective meta-evaluation of different types of SLM. We show that implementing SLM projects in semi-arid areas can indeed result in local greening. If the goal of projects is to increase vegetation cover, active revegetation may provide good results on smaller scales, while natural regeneration has the potential to regreen larger areas. Policymakers should therefore carefully match the project approach to its goals, keeping all positive as well as negative consequences of vegetation changes in mind. Yet, even though this research provided a useful monitoring and evaluation tool for land restoration projects, we want to stress the importance of documenting and monitoring the projects its implementation, which would improve the accuracy of the evaluation. We therefore argue that more research should focus on the creation of a complete and open-access database of land restoration projects, before large-scale implementation does more harm than good.

Data availability statements
The data that support the findings of this study are openly available.
The data that support the findings of this study are openly available at the following URL/DOI: https://doi.org/10.6084/m9.figshare.21155335.