Mapping the Abundance of Multipurpose Agroforestry Faidherbia albida Trees in Senegal

: Multi-purpose Faidherbia albida trees represent a vital component of agroforestry parklands in West Africa as they provide resources (fodder for livestock, fruits and ﬁrewood) and support water lifting and nutrient recycling for cropping. Faidherbia albida trees are characterized by their inverse phenology, growing leaf ﬂowers and pods during the dry season, thereby providing fodder and shedding leaves during the wet season, which minimizes competition with pastures and crops for resources. Multi-spectral and multi-temporal satellite systems and novel computational methods open new doors for classifying single trees and identifying species. This study used a Multi-Layer Perception feedforward artiﬁcial neural network to classify pixels covered by Faidherbia albida canopies from Sentinel-2 time series in Senegal, West Africa. To better discriminate the Faidherbia albida signal from the background, monthly images from vegetation indices were used to form relevant variables for the model. We found that NDI54/NDVI from the period covering onset of leaf senescence (February) until end of senescence (leaf-off in June) to be the most important, resulting in a high precision and recall rate of 0.91 and 0.85. We compared our result with a potential Faidherbia albida occurrence map derived by empirical modelling of the species ecology, which deviates notably from the actual species occurrence mapped by this study. We have shown that even small differences in dry season leaf phenology can be used to distinguish tree species. The Faidherbia albida distribution maps, as provided here, will be key in managing farmlands in drylands, helping to optimize economic and ecological services from both tree and crop products.


Introduction
Faidherbia albida (Del.) A. Chev. (synonym Acacia albida Del.) trees belong to the mimosoideae sub family of the Fabaceae family and are a key component of West African agroforestry parklands, where crops are grown under a discontinuous cover of scattered trees. Faidherbia albida trees provide resources such as fodder for livestock, fruits, firewood, wood products for construction and traditional medicine [1][2][3][4][5][6]. The deep-rooted trees have access to deep water layers and mineral resources. In addition, Faidherbia albida trees contribute to symbiotic nitrogen fixation, and thus serve multiple ecosystem services, such as water lifting, nutrients recycling and carbon sequestration [7]. What makes Faidherbia albida trees distinct is their unique leaf phenology, which is characterized by leafing out during the dry season and shedding leaves in the rainy season. Due to this "inverse 3 of 16 from the leaf-out (May/June) and senescence (October) phases provide a higher accuracy as compared to the use of single image [44][45][46].
So far, most attempts to conduct tree species mapping from remote sensing have focused on forest environments, and mapping tree species of scattered woody populations in drylands on a large-scale remains a challenge. The scattered distribution of agroforestry parkland trees in the Sahel makes the mapping different from trees in forest stands, as the signal from a pixel can be influenced by the soil or herbaceous background, especially when the pixel is at the edge of the canopy. Despite the improvement in spatial, temporal and spectral resolution of sensor systems, remote sensing has not yet been widely applied for the mapping of Faidherbia albida on a country scale. For instance, Triboulet et al. [47] pioneered this area by using pixel value thresholding, but due to the lack of field data the final result was not validated. Lelong [48] utilized WorldView-3 imagery to classify 7 tree species including Faidherbia albida in Senegal. However, these studies were limited to a local scale and a snapshot in time. It is still a challenging endeavour to achieve species-level detection across large geographic areas due to restrictions related to data costs and the workload associated with the data analysis.
Previous studies have demonstrated that large savanna trees can be identified with the 10-m resolution of Sentinel-2 [49]. Classification algorithms are rooted in computer science and statistical learning [29], for example, machine learning algorithms like Random Forest and neural networks [50][51][52]. Considering trade-offs between data cost and spatiotemporal resolution requirements, this study investigates the possibility of mapping the distribution of Faidherbia albida over Senegal based on publicly available Sentinel-2 time series data. The constellation of Sentinel-2A and 2B multi-spectral satellite systems were launched in 2015 and 2017 as a part of European Commission's Copernicus Programme. This system, with a 5-day revisiting time under the same viewing conditions [53], provides an unprecedented opportunity for mapping Faidherbia albida trees. The high temporal resolution may capture the inverse phenology of Faidherbia albida, contrasting with other woody and herbaceous vegetation types in the semi-arid Sahel. Our specific objectives are (1) to select dates and bands from Sentinel-2 images suited to map Faidherbia albida trees, (2) to use a Multi-Layer Perception (MLP) artificial neural network fed with a number of features to generate a 10-m resolution Faidherbia albida distribution map over Senegal and (3) to compare this distribution map with a potential occurrence map derived from an SDM model.

Study Area and Sample Data
Most of Senegal extends in the Sahelian and Sudanian regions with annual rainfall ranging between 100 to 1000 mm ( Figure 1). The western part of Senegal comprises a vast area of subsistence rain-fed cultivation, called the peanut basin. Livelihoods are based on traditional farming systems rotating millet grown for on-farm consumption and groundnuts as cash crops. The peanut basin is the most important crop production area in the country and is characterized by intensively cultivated parklands [54] comprising randomly scattered trees and woody shrubs within cropped fields. The two major soil types are arenosols and calcisols, according to the FAO classification [55]. They are characterized by a high sand content and low organic matter. Among tree species forming the parklands, Faidherbia albida is widely present across the peanut basin. Generally, the natural habitat of Faidherbia albida species occurs throughout most areas in West, East and Southern Africa where annual rainfall ranges between 400 and 900 mm [56]. As a leguminous nitrogen-fixing tree species, Faidherbia albida adds nitrogen and organic matter through leaf fall, and hence improves soil fertility and boosts crop productivity [57]. Consequently, Faidherbia albida occurs widely in croplands as a result of human management promoting the presence of this species on deep light sands or sandy clays [58], which is also the case for the peanut basin in Senegal [11]. Faidherbia albida rarely occurs in pristine savannas but is mainly disseminated by livestock in croplands. vegetation, bare soil and shrubs. In total, 2310 samples were used as reference points for a binary classification of two classes encompassing 516 points representing Faidherbia albida, and 1794 point for the other class. The Sentinel-2 pixels that contain the reference points were sampled and used for building the model. Since the sampled pixels are often a mixture of the tree canopy and the immediate surroundings of the tree, the Faidherbia albida mapping is thus conducted at a canopy level with a tree canopy minimum mapping unit of one pixel. Field samples were collected in two different semi-arid regions of Senegal (Niakhar and Nioro; Figure 1) in 2018 and 2019 using an optimized sampling strategy based on a landscape heterogeneity classification [59,60]. At each sampling site, an exhaustive inventory of all trees was conducted in a 1-hectare plot. The coordinates of each individual tree were recorded with a GPS. The collection of tree species contains over 70 common species, including 6612 sample points in the Nioro area where Faidherbia albida accounted for 3.07% (203) of the samples. A total of 9258 trees were collected in Niakhar among which Faidherbia albida accounted for 41.82% (3872). Since the spatial resolution of Sentinel-2 is 10 m, we only selected Faidherbia albida samples with a canopy diameter greater or equal than the Sentinel-2 pixel size as the reference data used for the building the MLP model. The individual tree canopy diameter information was derived from a previous study [61]. We further selected and added samples of other tree species, cropland, water, urban, herbaceous vegetation, bare soil and shrubs. In total, 2310 samples were used as reference points for a binary classification of two classes encompassing 516 points representing Faidherbia albida, and 1794 point for the other class. The Sentinel-2 pixels that contain the Remote Sens. 2022, 14, 662 5 of 16 reference points were sampled and used for building the model. Since the sampled pixels are often a mixture of the tree canopy and the immediate surroundings of the tree, the Faidherbia albida mapping is thus conducted at a canopy level with a tree canopy minimum mapping unit of one pixel.

Sentinel-2 Data Pre-Processing
Sentinel-2 top of atmosphere data are available in the Google Earth Engine (GEE) cloud computing platform. All available Sentinel-2 images from January 2017 to December 2019 with cloud cover less than 50% were fused to monthly composites. Cloud masking was done according to the flags of the Sentinel-2 QA band. The spectral noise caused by scenespecific illumination conditions and residual cloud effects is expected to be reduced by using 3-year average values from the period 2017-2019. Vegetation indices and bands (next section) in specific months and their dynamic changes were selected as feature collection for the following image classification based on Recursive Feature Elimination. The detailed Sentinel-2 data processing is described in the following sections.

Vegetation Indices
The Sentinel-2A and B instruments sample 13 spectral bands: visible and near-infrared at 10-m, red edge and SWIR at 20-m and atmospheric bands at 60-m spatial resolution. The bands with a resolution >20 m have been discarded from the current study due to the mismatch between pixel resolution and tree size, and the six bands of 20-m resolution were resampled to 10-m by the nearest neighbour, leaving 10 bands for the monthly composites. To better isolate the signal inherent to photosynthetic activity from other signals within spatially mixed pixels, 10 vegetation indices, which are widely applied to Sentinel-2 imagery [62], were selected (Table S1). As some of those vegetation indices were derived from the same spectral bands, considerable redundancy in information exists. To determine the similarity of indices, the pair-wise correlation coefficient was calculated (Table S2). Based on the high co-variability, those indices were categorized into 3 classes. The first class is composed of NDVI, NDVI74, PSSR a , EVI, MCARI, S2REP and EVI2. The second class is composed of NDI54 and GNDVI and the third class includes only MTCI. To select only one of each group for feature collection, the correlation coefficient between vegetation index values of Faidherbia albida and other species (Table S3) was taken into consideration, with a lower coefficient value indicating a higher discrepancy between Faidherbia albida and other tree species. Hence only NDVI and NDI54 were retained, and the MTCI was omitted because its monthly values showed a high standard deviation during the growing season ( Figure S1H). The initial feature bands used to construct the model are the monthly values of each band and 2 vegetation index bands, yielding in total 12 × (2 + 10) = 144 bands.
The foliage production of Faidherbia albida trees is high in the dry season and low in the rainy season, and their foliage production reaches a peak in December ( Figure 2). As biochemical and structure properties of woody plants can influence the spectral signature, the inverse phenology of Faidherbia albida potentially makes the tree distinguishable from other deciduous or evergreen tree species. The green herbaceous layer and understory impact the Faidherbia albida canopy pixels in Sentinel-2 images, so this study discriminates Faidherbia albida trees not only based on monthly band/vegetation index values but also uses seasonal dynamics. This was done by calculating all possible monthly image combinations (by subtraction of images) producing 792 dynamic images, leading to a total of 936 feature bands. Remote Sens. 2022, 14, x FOR PEER REVIEW 6 of 17

Feature Selection
A recursive feature elimination was then applied to reduce redundant features. The Random Forest classifier ranks the importance of features through their impact on the model accuracy. Recursive feature elimination can be used to rank the feature by looping through Random Forest classifiers and removing the least important feature until the accuracy remains stable. Here we generated collections of features and a 4-fold cross-validation was utilized to test the performance of the various feature collections on the MLP model. Cross-validation is a universal method for assessing the accuracy of a classifier based on a limited data sample. The dataset was generated from the field-observed data in combination with the Sentinel-2 imagery, and the whole dataset was split into 5 individual sets. A test set was held out for final evaluation. Then, three of the remaining datasets were used as training datasets, initially used to fit the model. The last dataset is used as a validation dataset, which provides an unbiased evaluation of a model fit on the training dataset while tuning the parameters. The process is repeated 4 times. The feature collection including the fewest number of features achieving the same performance as using all features was selected for the final model.
The performance metrics of recall and precision were used to evaluate models using different feature collections [63]. Precision is defined as the fraction of relevant instances among the retrieved instances, i.e., the number of true positives divided by the total number of elements labelled as belonging to the positive class. Recall is defined as the fraction of relevant instances that were retrieved, i.e., the number of true positives divided by the total number of elements that actually belong to the positive class. The true positives refer to the elements correctly labelled as belonging to the positive class, here, Faidherbia albida trees. The elements labelled as belonging to the positive class is the sum of true positives and false positives, the latter which are instances incorrectly labelled as belonging to the class. The elements that actually belong to the positive class are the sum of true positives and false negatives, which are the elements that were not labelled as belonging to the positive class but should have been. Recall and precision scores are often inversely related. Taking the harmonic average of precision and recall leads to the F-measure (a measure that combines precision and recall): Precision and recall weigh equally in F-measure and a higher F-measure represents a superior model.

Feature Selection
A recursive feature elimination was then applied to reduce redundant features. The Random Forest classifier ranks the importance of features through their impact on the model accuracy. Recursive feature elimination can be used to rank the feature by looping through Random Forest classifiers and removing the least important feature until the accuracy remains stable. Here we generated collections of features and a 4-fold crossvalidation was utilized to test the performance of the various feature collections on the MLP model. Cross-validation is a universal method for assessing the accuracy of a classifier based on a limited data sample. The dataset was generated from the field-observed data in combination with the Sentinel-2 imagery, and the whole dataset was split into 5 individual sets. A test set was held out for final evaluation. Then, three of the remaining datasets were used as training datasets, initially used to fit the model. The last dataset is used as a validation dataset, which provides an unbiased evaluation of a model fit on the training dataset while tuning the parameters. The process is repeated 4 times. The feature collection including the fewest number of features achieving the same performance as using all features was selected for the final model.
The performance metrics of recall and precision were used to evaluate models using different feature collections [63]. Precision is defined as the fraction of relevant instances among the retrieved instances, i.e., the number of true positives divided by the total number of elements labelled as belonging to the positive class. Recall is defined as the fraction of relevant instances that were retrieved, i.e., the number of true positives divided by the total number of elements that actually belong to the positive class. The true positives refer to the elements correctly labelled as belonging to the positive class, here, Faidherbia albida trees. The elements labelled as belonging to the positive class is the sum of true positives and false positives, the latter which are instances incorrectly labelled as belonging to the class. The elements that actually belong to the positive class are the sum of true positives and false negatives, which are the elements that were not labelled as belonging to the positive class but should have been. Recall and precision scores are often inversely related. Taking the harmonic average of precision and recall leads to the F-measure (a measure that combines precision and recall): Precision and recall weigh equally in F-measure and a higher F-measure represents a superior model.

Mapping Faidherbia albida Trees Using Artificial Neural Networks
The Multi-Layer Perception (MLP) model was adopted for the binary classification of Faidherbia albida trees. MLP is an artificial neural network with multiple densely connected layers between the input and output layers. The densely connected layers in MLP models combine all the features of the previous layer, which is suitable for complex classification tasks. We used the artificial neural network instead of other non-parametric classifiers, such as Random Forests or Support Vector Machines, which have been applied in previous tree species classifications, because of the irregular spectral signatures of the dataset and the uneven composition of species in the data structure. Considering the image resolution and the size of tree canopy, any given Faidherbia albida pixel will usually be a mix of reflectance from Faidherbia albida canopies and background vegetation or bare soil. Therefore, even the same tree species would show considerable intra-species spectral variability among training and testing data. Neural networks as a general-purpose technique for classification hold the potential to map Faidherbia albida in such a challenging case. The MLP model is created with three layers, and the learning rate, batch size and epochs are fine-tuned through grid search. The input of the MLP is the feature collection selected by cross-validation. The output of the MLP is the probability that a pixel includes a Faidherbia albida canopy, with values ranging from 0 to 1. Here, pixel values > 0.5 were mapped as Faidherbia albida canopies.

Post Processing
To comply with the nature of parklands and savanna ecosystems where Faidherbia albida trees are predominantly scattered without forming a closed canopy, all clusters of predicted Faidherbia albida pixels covering more than 1000 m 2 were masked. This threshold was selected based on Brandt et al. [61], where very high spatial resolution image analysis revealed that <0.0001 of all single tree canopies surpass this size in the study area.
We combined the Faidherbia albida canopy map derived from MLP with the woody canopy map produced by Zhang et al. [49], which is also based on time-series Sentinel-2 data. A map representing the relative cover of Faidherbia albida related to all woody plant cover was generated by calculating the fraction of the Faidherbia albida cover (per hectare) and the woody canopy cover (per hectare) map. To explore the agreement between our remote sensing based Faidherbia albida distribution map and the potential occurrence based on ecological modelling, we obtained a Faidherbia albida potential occurrence map that was provided by Kindt [64] at a 5-km resolution. The occurrence map was generated from 9 bioclimatic layers, 2 topographic layers and 3 soil maps by the R package BiodiversityR. Details about the data sources and resolution of the layers can be found in Table S4. This map consists of a unitless Faidherbia albida habitat suitability index varying from 0 (lowest) to 910 (highest). The absence-presence threshold for suitability of Faidherbia albida was provided by the SDM model as 534, meaning that suitability values above 534 correspond to the likely presence of the species with increasing values associated with higher likelihood of species presence.

Intra-Annual Difference in Vegetation Indices
Monthly values of NDVI and NDI54 for pixels covering Faidherbia albida and other tree species were analysed for Nioro and Niakhar. It is noticeable that even pixels covering the same group of tree species show varying index values in the two study regions. This is likely because the majority of the sampled pixels in Sentinel-2 images are mixed pixels, and the index value is the result of the reflectance from the tree canopy and a complex set of other covers underneath such as crops and herbaceous vegetation or bare soil, especially in the wet season when the Faidherbia albida trees are defoliated. However, the values of pixels with Faidherbia albida show the same seasonality in both regions different to other tree species (Figure 3). Specifically, pixels with Faidherbia albida have a lower NDVI value in the rainy season and a higher NDVI value in the dry season as compared to the pixels representing other tree species (Figure 3A), which means that the relative difference of Faidherbia albida NDVI values between the dry and wet seasons is smaller as compared to other tree species. While all tree species reach their NDVI peak in October, the NDVI value of pixels with Faidherbia albida decrease much slower after the peak month as compared to other tree species. As for NDI54 ( Figure 3B), Faidherbia albida pixels have the highest values in November while the pixels of other tree species show a peak in October. Consequently, the mapping of Faidherbia albida ideally should not only depend on the absolute index values of individual months, but also take the dynamics between different months into consideration. in the rainy season and a higher NDVI value in the dry season as compared to the pixels representing other tree species ( Figure 3A), which means that the relative difference of Faidherbia albida NDVI values between the dry and wet seasons is smaller as compared to other tree species. While all tree species reach their NDVI peak in October, the NDVI value of pixels with Faidherbia albida decrease much slower after the peak month as compared to other tree species. As for NDI54 ( Figure 3B), Faidherbia albida pixels have the highest values in November while the pixels of other tree species show a peak in October. Consequently, the mapping of Faidherbia albida ideally should not only depend on the absolute index values of individual months, but also take the dynamics between different months into consideration.

Mapping Faidherbia albida Using the MLP Model
By conducting a recursive feature elimination on the original 936 features (Section 2.2) we tested subsets of feature combinations as a function of the model mean accuracy using a 4-fold cross-validation approach (Figure 4). The evaluation of the Faidherbia albida mapping was performed by mean accuracy of all iterations using a test set. Using only the seasonal dynamic feature of NDI54 between November and August, the classifier achieves an average F-measure of 60%. When increasing the features to 40, the total variability in the feature layers reaches the same level of accuracy (precision, recall and the Fmeasure) as when using all features, indicating that 40 features are sufficient. Using 40 features, the model achieves an overall F-measure of 89%, with a fairly well-balanced rate of 91% precision and 85% recall, showing that the pixels of Faidherbia albida canopies are captured well (Table S5). The final selected features for the MLP model (Table 1) show that the seasonal dynamic features account for more than 50% of all selected features, and the typical span of the two months forming the difference feature is around 5 months, usually representing one month in the start of the dry season and the other month at the end of the dry season. The acquired times of the selected data by recursive feature elimination are February (start of Faidherbia albida senescence), July (leaf-off) and October (leafout). In terms of feature type (vegetation index or single band reflectance), the NDVI and NDI54 dominate the selected features.

Mapping Faidherbia albida Using the MLP Model
By conducting a recursive feature elimination on the original 936 features (Section 2.2) we tested subsets of feature combinations as a function of the model mean accuracy using a 4-fold cross-validation approach (Figure 4). The evaluation of the Faidherbia albida mapping was performed by mean accuracy of all iterations using a test set. Using only the seasonal dynamic feature of NDI54 between November and August, the classifier achieves an average F-measure of 60%. When increasing the features to 40, the total variability in the feature layers reaches the same level of accuracy (precision, recall and the F-measure) as when using all features, indicating that 40 features are sufficient. Using 40 features, the model achieves an overall F-measure of 89%, with a fairly well-balanced rate of 91% precision and 85% recall, showing that the pixels of Faidherbia albida canopies are captured well (Table S5). The final selected features for the MLP model (Table 1) show that the seasonal dynamic features account for more than 50% of all selected features, and the typical span of the two months forming the difference feature is around 5 months, usually representing one month in the start of the dry season and the other month at the end of the dry season. The acquired times of the selected data by recursive feature elimination are February (start of Faidherbia albida senescence), July (leaf-off) and October (leaf-out). In terms of feature type (vegetation index or single band reflectance), the NDVI and NDI54 dominate the selected features.

Comparing Faidherbia albida Canopy Cover Maps with Potential Occurrence Maps
A Faidherbia albida canopy map at 10-m resolution covering Senegal was generated based on 3 years of Sentinel-2 data. To illustrate the mapping at the national scale, the perpixel mapping was aggregated to 1 hectare showing the percentage of Faidherbia albida cover ( Figure 5A). Faidherbia albida trees are mainly distributed in the western cropland areas ( Figure 5A) where annual precipitation ranges from 300 to 600 mm. Faidherbia albida trees are to a less extent also found in the croplands of the northern part of the country, where rainfall is less than 300 mm/year. For most of the cropland areas, the percentage of Faidherbia albida cover is in the range between 2-15%. Our remote sensing-based distribution agrees to some extent with the potential occurrence map from ecological modelling ( Figure 5B) in which the potential occurrence is shown by SDM suitability values. While similar patterns are shown in western Senegal, the southeast, the Cap Vert peninsula and the limited cropland areas along the north-western coast do not match with our map.

Comparing Faidherbia albida Canopy Cover Maps with Potential Occurrence Maps
A Faidherbia albida canopy map at 10-m resolution covering Senegal was generated based on 3 years of Sentinel-2 data. To illustrate the mapping at the national scale, the perpixel mapping was aggregated to 1 hectare showing the percentage of Faidherbia albida cover ( Figure 5A). Faidherbia albida trees are mainly distributed in the western cropland areas ( Figure 5A) where annual precipitation ranges from 300 to 600 mm. Faidherbia albida trees are to a less extent also found in the croplands of the northern part of the country, where rainfall is less than 300 mm/year. For most of the cropland areas, the percentage of Faidherbia albida cover is in the range between 2−15%. Our remote sensing-based distribution agrees to some extent with the potential occurrence map from ecological modelling ( Figure 5B) in which the potential occurrence is shown by SDM suitability values. While similar patterns are shown in western Senegal, the southeast, the Cap Vert peninsula and the limited cropland areas along the north-western coast do not match with our map. The remote sensing-based map of Faidherbia albida cover was directly compared with the SDM suitability score within different land use/cover types ( Figure 6). The SDM map at a 5-km resolution was resampled to 1 ha using the nearest neighbour. Faidherbia albida trees mainly occur in areas with suitability values above 560 and are almost absent below this threshold except in croplands where they occur even with SDM values of 300. Faidherbia albida trees were barely found in the north/north-eastern shrub and open forest areas of Senegal ( Figure 5), even though the SDM suitability was around 600. When suitability values exceed 680, we observe that the high suitability score relates to low Faidherbia albida cover (<0.25%). This discrepancy is linked to the Senegal land cover, which is related to the near coastal areas around the Dakar peninsula (Figures 1 and 5). The remote sensing-based map of Faidherbia albida cover was directly compared with the SDM suitability score within different land use/cover types ( Figure 6). The SDM map at a 5-km resolution was resampled to 1 ha using the nearest neighbour. Faidherbia albida trees mainly occur in areas with suitability values above 560 and are almost absent below this threshold except in croplands where they occur even with SDM values of 300. Faidherbia albida trees were barely found in the north/north-eastern shrub and open forest areas of Senegal ( Figure 5), even though the SDM suitability was around 600. When suitability values exceed 680, we observe that the high suitability score relates to low Faidherbia albida cover (<0.25%). This discrepancy is linked to the Senegal land cover, which is related to the near coastal areas around the Dakar peninsula (Figures 1 and 5).
By combing the woody cover map from Zhang et al. [49] and our results, we calculated the relative contribution of Faidherbia albida trees to the overall woody plant cover per hectare (Figure 7). The map of this relative contribution shows the degree of dominance of Faidherbia albida among woody plant cover. The higher the contribution of Faidherbia albida to the woody plant cover, the lower is the woody plant diversity. Furthermore, in areas where woody cover is nonzero, the contribution of Faidherbia albida to cover as a function of all woody species cover is illustrated in Figure 7B. The relative contribution of By combing the woody cover map from Zhang et al. [49] and our results, we calculated the relative contribution of Faidherbia albida trees to the overall woody plant cover per hectare (Figure 7). The map of this relative contribution shows the degree of dominance of Faidherbia albida among woody plant cover. The higher the contribution of Faidherbia albida to the woody plant cover, the lower is the woody plant diversity. Furthermore, in areas where woody cover is nonzero, the contribution of Faidherbia albida to cover as a function of all woody species cover is illustrated in Figure 7 B. The relative contribution of Faidherbia albida in croplands is on average 7.23%, with several areas being dominated by Faidherbia albida.

Mapping Faidherbia albida
This study presents a method to estimate Faidherbia albida cover across Senegal from Sentinel-2 spatiotemporal signatures at 10-m resolution. Previously, only Lelong et al. [48] used very high resolution Worldview-3 imagery to map 7 tree species including Faidherbia albida. Our approach was based on the assumption that the classification could take ad-

Mapping Faidherbia albida
This study presents a method to estimate Faidherbia albida cover across Senegal from Sentinel-2 spatiotemporal signatures at 10-m resolution. Previously, only Lelong et al. [48] used very high resolution Worldview-3 imagery to map 7 tree species including Faidherbia albida. Our approach was based on the assumption that the classification could take advantage of the distinct and inverse phenological behaviour of Faidherbia albida trees as compared to other tree species. However, we found that this phenological behaviour was not seen in satellite time series from Sentinel-2. A possible explanation is that the herbaceous layer growing below the canopy of Faidherbia albida trees during the non-leaf season dominates the signal. It impedes the capacity of satellite time series to capture the inverse phenology during the leaf-off period. As for the leaf-on period, the difference is also unexpectedly small and varies with location and time. Since most remote sensing applications in mapping forest types or tree species rely on the phenological differences in satellite time series between the target and the background [45], the mapping of Faidherbia albida trees is more challenging than initially assumed.
Given this situation, we maximized the input capacity by generating a total of 936 features, which include monthly and seasonal features of both multiple vegetative indices and bands. Instead of directly using sparse temporal features and single vegetation indices [43,65], we made use of recursive feature elimination to reduce data redundancy. The accuracy stabilized when 40 features were included, mainly from the senescence month of Faidherbia albida and including both spectral bands and vegetation indices [44,45]. Features related to seasonal dynamics were most important, which aligns with the results of [43,66]. NDVI, NDI54 and the second SWIR band of Sentinel-2 are often picked as high-ranking features, indicating the importance of the red, red edge and SWIR bands for tree species mapping [29]. Additionally, the background reflectance causes different values for different regions with Faidherbia albida trees ( Figure 3A). This effect can be mitigated by using images of seasonal changes in addition to monthly composites (Table 1).
We did not use a convolutional neural network (CNN) because CNNs rely on features which are extracted by sliding windows within an image, and the widow size N is usually larger than the tree canopy pixel (usually one pixel). Thus, the importance from a given tree canopy pixel might be diluted on the extracted window's feature. Contrastingly, the input for the MLP is the labelled feature collection, which is derived from single pixels in an image, which is considered as more suitable when using Sentinel-2 imagery.
In the past decades, mapping forest cover and type have achieved satisfactory outcomes globally and operationally using remote sensing techniques [37,39]. Mapping tree species at large scale has however only progressed slowly, mainly because the required remote sensing datasets are expensive to acquire, such as LiDAR and other very highresolution data [35,40]. Furthermore, the Faidherbia albida agroforestry multipurpose trees studied here grow in drylands where trees grow as individuals, and such trees outside of forest have been underestimated [61]. As such, our study represents a first attempt to conduct a national scale mapping of scattered non-forest tree species. It would be interesting to further map the distribution of Faidherbia albida trees using higher spatial resolution images like PlanetScope, which are now provided as a pre-processed product in an analysis-ready data format of monthly composites of 4.7-m spatial resolution free of charge from the Norwegian International Climate & Forests Initiative (NICFI). However, PlanetScope data has a lower spectral resolution (only four classical bands of R, G, B and NIR) as compared to Sentinel-2. The degraded spectral information may thus decrease the mapping capacity. With a monthly temporal resolution and spatial resolution of 4.7 m, recurrent CNN techniques could boost the mapping by interpreting temporal and textural information of individual tree canopies simultaneously [67].

Comparison with Ecological Niche Modelling Products
Our results show that Sentinel-2 image time series are able to map the actual distribution of Faidherbia albida trees, which is a significant advance as compared to SDM-based potential species occurrence maps [64]. Our results indicate that Faidherbia albida trees are mainly found in croplands where they are the dominating tree species. The relation in Figure 6 showed that when the land suitability exceeds the threshold defined by the SDM model, there is indeed often an observed occurrence of Faidherbia albida trees. The average Faidherbia albida cover is always higher in croplands than for other land use/cover types because farmers are well aware of the benefits of Faidherbia albida trees being an integral part of agroforestry systems [68][69][70]. In the sparsely populated south-eastern part of Senegal where land cover is mainly open woodland and shrublands, we did not predict to find any Faidherbia albida, while the SDM model predicts a high suitability, likely because human management was not carefully considered in the applied model.

Conclusions
A 10-m resolution Faidherbia albida canopy map was generated for Senegal based on a time series of Sentinel-2 images from the period of 2017 to 2019, used to form 12 monthly composites covering the phenological phases of the Faidherbia albida tree. Ground observations of tree species from two different regions in Senegal were used to generate a robust model mapping the occurrence of Faidherbia albida trees. Compared to an existing map showing the potential occurrence of Faidherbia albida trees from a suitability modelling approach (species distribution modelling; SDM), our study took advantage of the unique phenology of Faidherbia albida trees to separate this tree species from other trees. We showed that the actual distribution does not always agree with the potential distribution. We also showed that Faidherbia albida dominates tree cover in some cropland areas, supporting the well-known benefits of the tree in agroforestry systems of West Africa (termed parklands). Faidherbia albida is managed and utilised by farmers in different ways as this multi-purpose tree does not only provide fuel wood and fodder like other species, but also affects crop traits for improved productivity and yield by shedding leaves in the crop growing season. The derived 10-m map of Faidherbia albida occurrence could be used as a base-map in support of an empirically based understanding of the role of Faidherbia albida in parkland ecosystems in relation to crop yields. Ultimately, this could serve as a vantage point for spatially explicit analyses of the extent to which the use/management of this tree resource is optimally implemented. Overall, 1418 pixels were sampled and averaged over the study period, including 57 Faidherbia albida pixels and 112 pixels with other tree species in the Nioro region, and 459 Faidherbia albida pixels and 790 pixels with other species in the Niakhar region; Table S1: Vegetation indices used in the study. Their formulation and the related spectral bands (band numbering is referring to the Sentinel-2 sensor system); Table S2: Correlation coefficient of monthly vegetation index of Faidherbia albida; Table S3: Correlation coefficient between vegetation index values of Faidherbia albida and other species; Table S4: Climatic, geomorphologic, and pedologic layers used for the Species Distribution Modelling [71]; Table S5: Results from the 4-fold cross-validation using a test set of 40 features.