1. Introduction
In the semi-arid Sudano-Sahelian ecological zone of West Africa, trees maintained by farmers on their farmed plots are an important element of the local livelihood [
1]. Farmers use trees for fuelwood for their own use as well as for sale to supplement farm incomes. Wood fuel in Kano has traditionally been derived from trees grown and maintained by farmers in the farmed parklands surrounding the city. The ‘parkland’ landscape is defined by the large variety of trees grown and maintained on farmland, which are used for a very wide variety of purposes, including fuel wood, timber for building materials, food, fodder, fibre, and medicines [
1,
2]. Additionally, the large areal extent of farmed parkland landscapes in the Sudano-Sahelian ecological zone makes them an important component of the global climate system, as they store and sequester large amounts of carbon in the woody biomass and soils [
3,
4]. Higher demand for fuelwood due to high population growth combined with predictions of higher temperatures and decreased rainfall pose a serious challenge to tree stocks. Therefore, spatial, and quantitative assessments of tree species are especially urgent since climate change and intensified land use in recent decades have put increasing pressure on tree cover.
In addition to pixel-based approaches to classifying tree species, object-based image analysis is effective for classifying objects at multiple scales. This means that tree crowns of different sizes can be delineated separately, from individual tree crowns to large clusters of tree crowns. Numerous studies have found high accuracy and low error for the classification of tree species using GEOBIA compared to pixel-based approaches [
5,
6]. Over the last few years, there has been enormous development in remote sensing with the launch of high-spatial-resolution commercial satellites. The use of traditional statistical analysis of single pixels is not appropriate for high resolution satellite images, as the pixel under consideration and its neighbouring pixels may differ spectrally but belong to the same land cover class [
7]. This high spectral variability within the same land cover class creates a “salt-and-pepper” effect during classification. As human beings normally recognise patterns in a landscape by their spatial relationship to neighbourhood objects, it is useful to use spatial and contextual information for the characterisation of land use classes, along with spectral information [
8]. Spatial relationships between adjacent pixels in the form of texture provide important information for identification of individual objects, which are building blocks of the original features of interest [
9]. In this way, homogeneous objects based on spatially connected groups of pixels with similar spectral characteristics can be identified. Image segmentation is the process by which homogeneous image objects are created by aggregating groups of pixels with regard to spectral and spatial characteristics. The term ‘homogeneous’ implies that within-object variance is low compared to that between objects, and those identified objects also contain additional information about geometry (size and shape), contextual, and textural aspects besides spectral information [
10]. These homogeneous objects reflect real-world objects of interest [
7].
Many studies have used high-spatial-resolution satellite images for tree crown delineation [
5,
11,
12]. Bunting and Lucas [
5] extracted and classified different tree crown species in Australian mixed forests using the Compact Airborne Spectrographic Imager (CASI) hyperspectral data through GEOBIA. Rasmussen et al. [
12] used QuickBird imagery for extracting tree crowns in Northern Senegal, and Karlson et al. [
11] used WorldView-2 data for tree cover extraction in Burkina Faso using GEOBIA. In an agroforestry landscape, there is a variety of deciduous trees with varying crown sizes and ages; therefore, GEOBIA is well suited for such tree crown cover mapping. Remote sensing has been successfully used for tree species mapping using airborne hyperspectral systems [
13,
14], but the high cost and small footprint of these airborne systems restrict their usage for large areas. Therefore, there has been a growing interest in the use of very high resolution space-based satellite remote sensing images for the identification of tree species [
6,
13,
15,
16,
17] because they provide timely, repetitive, and large area coverage from local to global scales. Karlson et al. [
15] investigated the capability of multi-seasonal WorldView-2 imagery to map five dominant tree species at the object level in central Burkina Faso using the Random Forest (RF) classifier.
There have been many studies using satellite data with different machine learning methods for tree species classification from satellite data, including Support Vector Machine (SVM) [
6,
18], K-nearest neighbours (KNN) [
19,
20,
21], Random Forest (RF) [
6,
22,
23], Logistic Regression (LR) [
20,
24], Extra Gradient Boosting (XGBoost) [
25,
26,
27], Multi-Layered Perceptron (MLP) [
28,
29,
30], Light Gradient Boosting (LightGBM) [
25], Gaussian Naïve Bayes (GNB) [
31], Gradient Boosting (GB) [
32]. However, most studies are based on very few species, and the accuracy levels achieved are generally not above 80%. For example, Karlson et al. [
15] tested five (only four native) tree species with one machine-learning classifier, Random Forest. Producer accuracy in the dry season was below 80%, except for the distinctive species
M. indica, which has dark green shiny leaves, and a non-native Eucalyptus species, which has a very distinctive compact crown and blue-green leaves. Lelong et al. [
16] examined two machine learning algorithms, SVM and RF, but achieved a relatively low kappa index of 0.71 for identifying four tree species in Senegal. Most previous studies have also combined different sensing systems, such as optical satellite images combined with Lidar or airborne images, or combined multiple dates. To the best of our knowledge, no study has compared different machine-learning methods for tree species classification. Moreover, our study uses only a single sensor.
The objective of the study is to test and evaluate a cost-effective method for detailed tree species classification in the agroforestry landscape of West Africa, using Kano, Nigeria, as a case study. WorldView-2 imagery is used, as a single image covers a large area at a high level of detail, and airborne or UAV imagery is not generally available in a developing country environment. An evaluation of nine different machine learning methods is performed to suggest the best-performing and most cost-effective method for detailed classification of tree species over large areas. This will enable effective rural afforestation programmes, which need an accurate inventory of existing tree stocks, and contribute to a sustainable rural economy where farm trees have multiple and diverse uses.
4. Discussion
Previous research has shown that remote sensing based tree species mapping in tropical dryland ecosystems is possible with the help of machine learning methods. The accuracy obtained by this study exceeds that of most other studies, uses only a single date remote sensing image (WorldView-2), and examines a larger number of different species than previous studies. The study identified SVM and MLP as the most accurate machine learning methods, with an OA of 82% and K = 0.79 that were substantially higher than the other methods tested. A study in Senegal by Lelong et al. [
16] examined two machine learning methods, SVM and RF, and found SVM to have higher accuracy. Additionally, similar to our study, they found SVM to be superior to RF for a small number and unequal distribution of samples. The accuracies achieved by our study compare well with other similar studies, such as Lelong et al. [
16], whose results were based on only four species and had a highest kappa index of 0.71, and Karlson et al. [
11], whose results were limited to only four indigenous tree species and used a single machine learning method, Random Forest. They achieved accuracies of OA = 78% and K = 0.74 for dry season imagery. While our results are significantly better than these, Karlson et al. [
11] did obtain accuracies comparable to ours (with an OA of 83% and K = 0.76) when multi-seasonal imagery was used, but for only four different tree species.
The current study demonstrates that accurate species identification can be achieved with machine learning methods for a range of species in agroforestry landscapes. The tree species studied here are among the most important species in the West African agroforestry landscape.
Parkia biglobosa, which is used for soup stock as well as fibre, and
Faidherbia albida, used for dry-season fodder, have been shown in a recent study by Usman et al. [
72] to be fast declining.
The sample data for some species were limited (
Table 2) in this study. However, the SMOTE sampling was used to increase the frequency of individual tree species with low frequencies (
Supplementary Table S1), thereby avoiding overfitting of machine learning models. For example, four tree species—Mangifera indica, Vitellaria paradoxa, Faidherbia albida, and Tamarindus indica—have tree counts less than 30 trees. After applying SMOTE, the tree counts for those species substantially increased (
Supplementary Table S1), which avoided wrong predictions and overfitting of machine learning models. In previous studies, e.g., Lelong et al. [
16] had less than 30 field samples for three out of six species sampled, and Karlson et al. [
11] had less than 10 field samples for three out of five native species sampled. In our study, out of nine species that we sampled, the average number of field samples was 24, although the method of compensation (SMOTE) we used increased the overall number from 210 to 325, as mentioned in
Supplementary Figure S1. Nonethless, that study was restricted in scope as it used field measurement to obtain data over limited areas, compared with the over 100 km
2 covered by the single WV2 image used in this study.
The other species studied here, including
Azadirachta indica,
Piliostigma reticulatum,
Anogessus leiocarpus, and
Diospiros mespiliformis, were shown by Usman et al.’s [
36] study to be actively regenerating. These four important fuelwood species were identified by MPL with 87/72/80/81% accuracy by MLP, and 86/72/86/81% accuracy, respectively, by the SVM classifier. The huge dependence on wood as fuel in Nigeria, where the latest available figures (National Bureau of Statistics, 2011) suggest that 95% of the energy used for cooking is from wood, may explain the increased abundance of these species. The
Vitellaria paradoxum, or shea butter, tree provides emollients and fats for a wide range of modern food, medicinal, and cosmetic products. This species was identified with 90% accuracy by the two best classifiers, MLP and SVM. Such trends need to be documented accurately over large areas in order to understand and manage possible threats to the local economy, as well as identify opportunities for growth.
WorldView-2 and WorldView-3, with their unique spectral band configurations including red edge, near infrared, and shortwave infrared bands at very high spatial resolution, have proved to be capable of mapping tree species in the West African agroforestry landscape [
15,
16]. As other recent very high resolution sensors such as WV3 and Pleiades have similar spectral bands, spatial resolution, and swath width to the WV2 images used in this study, little advantage is expected from using them. This study has demonstrated that a single image, along with a robust machine-learning tool such as MLP or SVM, can provide highly accurate tree species inventories over large areas.
5. Conclusions
The comprehensive examination of methods for tree species classification presented here can assist state and rural authorities in undertaking rapid and cost-effective rural surveys of the agroforestry landscape in Nigeria. This will permit a better understanding of the pressures currently facing Nigeria’s dryland ecosystems. Violent outbreaks in recent years among migrant pastoralists stemming from land shortages are related to trends in tree species, as declines in Faidherbia albida, traditionally used as dry-season fodder, are removed from farmland to counter predation by cattle. Rural households are susceptible to climatic fluctuations and trends as well as the current massive growth in the rural population. Land fragmentation due to traditional inheritance customs requires more farmland trees to supply additional households with wood fuel, as farmers indicate that they rarely buy wood. The disappearance of non-fuelwood species in recent decades is of concern due to their importance in the local household economy.
This study demonstrates two machine-based learning models that provide over 80% accuracy and can be applied to a single date of WorldView-2 imagery, to identify the most common farmland tree species in West African farmed parkland. Rural afforestation programmes would benefit from accurate inventories of current stocks of tree species and their regional variations.
Due to the longevity of trees, it is unlikely that such tree species inventories as described by this study would need to be repeated on a regular basis. More important would be to extend the survey to wider areas and repeat it perhaps once per decade. However, because a farmed parkland landscape with the same tree species exists throughout the semi-arid zone of West Africa and, to a lesser extent, southern Africa, the findings should have broader applicability.