Modeling the Spatial Distribution of Acacia decurrens Plantation Forests Using PlanetScope Images and Environmental Variables in the Northwestern Highlands of Ethiopia

Plantation forests are forests established by planting or deliberate seeding to achieve principally economic goals [1] such as timber, energy, fiber, and non-woody forest products [2]. They are also established for soil and water conservation and carbon sequestration in the process of afforestation and reforestation [3]. According to the FAO [2] report, plantation forests cover about 131 million ha worldwide and account for 3% of the global forest. Between 1990 and 2020, their area increased by 55.8 million ha, with the biggest jump (21.2 million ha) recorded between 2000 and 2010.

The impact of plantation forests on the environment will depend on what land use they replace [4]. If they are established on frequently cultivated land for a long time or degraded lands, they may provide substantial opportunities for biodiversity conservation [5] and deliver vigorous ecosystem service [6], but plantations converted from the natural forest have adverse impacts on biodiversity [7]. Plantation forests are essentially significant in fragmented landscapes, where they may account for a large amount of remaining forest habitat [8] and can serve as corridors between habitats [4].

Small-scale plantation forests initiated by farmers on degraded lands have become important in Ethiopia, particularly since the mid-1990s [9]. An estimated area of 754,900 ha of the country is covered by small-scale plantation forests, and of this, 84.7% (639,400 ha) is found in the Amhara Region [9]. Critical levels of land degradation and reduced productivity forced farmers to start planting trees, often with predominantly exotic and fast-growing species at the expense of crop production [10,11,12].

In the Northwestern Highlands of Ethiopia, specifically in the Fagita Lekoma District, the study site and one of the districts in the Amhara Region, growing Acacia decurrens plantations on small-size farmlands have been rapidly increasing and are widely planted because of the economic and environmental benefits. This species is preferable due to its advantages of a fast-growing rate and adaptability to degraded and acidic soil conditions [13]. Additional reasons that motivate farmers to plant this species are its use as fuel and construction wood, animal fodder, charcoal production and availability of market, and soil fertility maintenance [14]. It also creates job opportunities for landless community parts and supports local livelihoods and rural developments, especially when managed by smallholders [13,15]. Land-use change from cultivated land and grassland into Acacia decurrens plantation has been common in the district in the past three decades and resulted in an increase in the forest cover of the district by more than 250% between 1887 and 2015 [16], and around 400% between 2006 and 2017 [17]. A range of ecosystem services can be obtained from plantation forests established on degraded lands that require restoration [18]. New plantation forests generated from former agricultural land can improve ecosystem service [19].

The rapid expansion of plantation forests significantly affects climate, hydrology, biodiversity, and the terrestrial carbon cycle. The expansion of plantation forests can amend the understory climate condition and soil properties [20], and water quality [21]. These changes occur as a result of changes in temperature, rainfall, land use type, and storm frequency and magnitude [4]. In addition, plantation forests can contribute strongly to regulating the environment, biodiversity, and socioeconomic functions, especially carbon sequestration [22]. To monitor such dynamics, remote sensing is an essential and effective source of data [23].

Remote sensing technologies drive developments in forest resource assessments and monitoring at various scales. They enable the provision of airborne and spaceborne data with a higher spatial resolution, frequent coverage, and expanded spectral coverage [24]. The remote sensing-based assessment of forest study is repetitive, affordable, competent, and non-destructive for monitoring [25]. Recently, in complementarity with field data, it has shown great adaptability in environmental studies such as droughts [26], floods [27], the spread of invasive species [28], disturbance [29], and other human-induced pressures [30]. The contribution of satellite data is becoming impressive to monitor the spatial distribution and temporal dynamics of plantation forests [31]. Optical data are spectrally sensitive to different species and can distinguish phenological characteristics unique to a particular plantation [32]. Different plantation forests can have distinct implications for the local ecosystem service [33].

The geographic distribution of species is dynamic at accelerating rates because of anthropogenic pressures, the introduction of non-native species, and climate change [34]. To understand the distribution of introduced or expanding species, researchers often map the suitability status of the habitat or the potential occurrence probability of species using different techniques such as expert opinion [35], mathematical models [36], or machine learning algorithms [37]. These methods help researchers and decision-makers to identify priority areas for environmental conservation [38], examine landscape planning approaches on the management and restoration of protected areas [39], assess species distribution under changing anthropogenic or environmental conditions [40], investigate the impact of environmental changes on the biodiversity [41], and model the invasion status of invasive species [28,42].

The species distribution model (SDM) is a popular technique in ecology and conservation biology to assess the impact of land use and climate changes on biodiversity distribution [43], predict species diversity and composition patterns over space and time [44], and provide spatially explicit and compressive maps that are specifically important to understand the distribution level and extent of a given species. It combines observations of species occurrence or abundance with environmental variables [45]; its performance depends on the collected data during field surveys and exists as simple presence/absence records, which are crucial to train and validate the model [46]. Moreover, the accuracy of SDM varies among algorithms [47], and integrating multiple algorithms is more reliable to get robust estimations of species distribution [48].

The use of SDM algorithms in combination with remotely sensed datasets is effective for mapping plant species across different management levels at local and regional scales [49]. In Sub-Saharan Africa, agricultural landscapes are highly fragmented and this is one of the challenges complicating their mapping [50]. Fragmented parcels of land with verities of coverage highlighting the need of methods based on high resolution satellite imageries [51].

PlanetScope (PS) satellite constellation can achieve daily coverage with a spatial resolution of 3–5 m, visible to near-infrared and atmospherically corrected imagery [52], which has been successfully applied in many fields, for example, rubber plantation mapping [53], forest canopy height estimation [54], biomass estimation [55], leaf area index production [56], cropland mapping [51], and crop yield prediction [57]. It provides effective spatial data for the extraction of plantation forest and agricultural information in the tropical and subtropical regions [58], and offers a good opportunity to overcome challenges in mapping smallholder agricultural fields [51].

Understanding the interplay between the spatial distribution of a species and its environmental determinants is a fundamental concept in ecology and conservation [59]. Establishing plantation forests on agricultural or degraded land presents significant prospects for biodiversity conservation [60]. Consequently, the impact of plantation forests is contingent upon their spatial extent of landscape coverage [61] and the specific land use they replace [4]. Despite prior studies in the study area focusing primarily on land use/land cover changes across all classes [11,16,17,62,63], there is a noticeable gap in research specifically addressing the species-level identification of Acacia decurrens plantations through the utilization of high-resolution satellite imagery and environmental variables combined with machine learning algorithms. Examining such a spatial pattern is crucial in any study aiming to ensure the provision of goods and services [61]. Therefore, the objectives of this research were (i) to model the spatial distribution of Acacia decurrens plantation forests using high-resolution satellite imagery, (ii) to evaluate the performance of SDM algorithms for Acacia decurrens distribution modeling, and (iii) to identify the relative importance of predictor variables for Acacia decurrens distribution modeling. Modeling and understanding the spatial spread of such species are high priorities for resource managers to assess the environmental implications of the sustained use of plantation forests and to scale up for the other degraded areas of the country based on scientific findings and with great attention. This is because in Ethiopia, it is believed that small-scale tree plantations can contribute to addressing issues related to sustainable agricultural land use, mitigating the negative impacts of deforestation, and meeting the needs for the livelihood and energy of the growing population [64]. Having comprehensive information about its distribution enables effective control and management.

The assessment of multicollinearity was a crucial phase aimed at identifying and addressing strong correlations among multiple predictor variables used in the identification process. The implementation of the VIF method, recognized for discerning collinearity among predictor variables [78], resulted in the exclusion of almost three-fourths of the total variables due to collinearity issues, potentially decreasing the efficiency of prediction and increasing the uncertainty of the SDM [96]. When VIF exceeds 10, it serves as an indicator of collinearity issues within the model [97]. Furthermore, collinearity represents a significant concern that can potentially result in the incorrect identification of relevant predictor variables [98]. The high correlation among many of the input predictor variables can be attributed to the fact that observations were made within a relatively local scale and due to the similarity in spectral vegetation indices. It is worth noting that all the reduced variables were vegetation indices and raw bands, implying a relatively higher similarity between the generated vegetation indices. This result aligns with the findings of [99], possibly due to the limited spectral resolution of the PS image with only four bands.

The use of a combination of SDM algorithms, such as GLM, MARS, BRT, RF, SVM, and CART, in a complementary manner, along with the incorporation of accuracy estimators based on presence/absence data, enables a more effective representation of the spatial distribution of species at a local scale. The overall accuracy of the algorithms is relatively good, with the values exceeding 0.8 for AUC and 0.6 for TSS (Table 2). Predictive accuracy pertains to the ability of the algorithms to gauge the disparity between observed and predicted values [100]. RF exhibits an AUC above 0.9 and a TSS above 0.8, signifying near-perfect agreement. BRT and SVM demonstrate AUC values above 0.8 and TSS values above 0.7, indicating substantial agreement. The result aligns with the findings of Maxwell et al. [101], who found that machine learning outperformed regression algorithms for species identification. RF attains higher values than other algorithms for both evaluation metrics. This aligns with previous studies that have shown RF’s superior performance in species identification [28,102,103,104,105], remote sensing image classification [106,107], and data mining [108] compared to other algorithms. This is because the RF algorithm generates predictions by creating thousands of trees and aggregating their results through averaging [86]. This approach allows the algorithm to prevent overfitting, thereby enhancing predictive performance and reducing variance [109]. Thus, RF proves to be a robust technique for modeling species distribution prediction, as supported by previous studies [110,111,112,113]. In other studies, it has been noted that generative methods such as RF tend to yield improved results with small sample sizes, possibly due to their faster convergence toward their higher asymptotic error when compared to discriminative methods [114].

The performance of all the applied algorithms was effective, as indicated by the mentioned measures, enabling their inclusion in the ensemble modeling. The spatial extent of Acacia decurrens plantation in the study area was 22.44% during 2022. Worku et al. [63] reported that 17.97% of the district was covered by the plantations during 2017. The proportion of plantations was described as 33.9% in one of the watersheds in the district during 2017 [17]. Another study by Wondie and Mekuria [13] stated that 25.6% of the district was covered by forestlands. The difference in the extent between this study and previous studies might be attributed to differences in the satellite data properties, methodology used, and study time.

The ensemble model unveiled the variable importance and response curves in the prediction of Acacia decurrens distribution. The results indicate that CVI, ARVI, and GI are ranked as the first, second, and third variables, respectively, with values of 39.3%, 16%, and 7.1% based on the AUC metric, signifying the relatively strong influence of vegetation indices. These indices played a predominant role in determining the distribution and proliferation of Acacia decurrens. Vegetation indices played a crucial role in identifying tree species, demonstrating the highest relative importance, and can effectively serve as a classification variable to differentiate evergreen trees [115]. The study by Anna [116] also reported similar variable importance, with vegetation indices being the best predictors of tropical evergreen species. This result is consistent with the description of vegetation index variables exerting a more substantial influence than the bioclimatic variables, significantly contributing to defining the distribution range and landscape patterns in the Chelodina longicollis model, with a total contribution of 50.75 compared to 36.94 for the 11 bioclimatic variables [117]. Moreover, Engler et al. [99] also indicated that variables derived from remote sensing are significantly crucial for mapping the spatial distribution of both broadleaved and coniferous tree species at high-resolution data. Elevation, distance to roads, and mSAVI occupy the fourth, fifth, and sixth ranks, respectively, making significant contributions to the species’ distribution. Elevation significantly influences the distribution of plant species [118,119], and in particular, the decision of where and whether to establish plantations depends on environmental factors such as elevation [120]. Further, the findings of Altamirano and Lara [121] also indicate that plantation forests tend to be located in areas characterized by moderate elevation levels and a short distance from roads. Rainfall holds the seventh rank in terms of its relative importance in determining the distribution of this species. The lower ranked predictor variables, specifically, temperature and slope, also play a role in the distribution of Acacia decurrens. However, in relative terms, they are not as influential as other predictor variables in the model, as indicated by the AUC metrics. This may be attributed to the size of the study area in relation to the resolution of the utilized data.

The response curves in Figure 6 show that the probability of Acacia decurrens occurrence generally increases with higher values of vegetation indices. Among the environmental variables, the probability of occurrence increases with higher elevation, keeping other variables constant at their mean value. The probability of occurrence is rare at low-elevation areas of the district because these areas are more suitable for agricultural practices, especially small-scale irrigation activities. This is clearly seen in Figure 7, where the distribution of Acacia decurrens is rare in the Northeastern and Western parts, where small-scale irrigation activities are being carried out along the Guder and Tinbil rivers, respectively. Conversely, the probability of occurrence decreases with an increase in distance from roads and slope (Figure 7). The community cultivates Acacia decurrens to generate income by selling the standing trees or producing charcoal, and road access plays a vital role in facilitating this. Access to roads reduces input costs and in certain situations, leads to higher prices for plantation products [122].