1. Introduction
Peatlands are characterized by the accumulation of partially decayed organic matter formed from plant debris under waterlogged conditions [
1]. They provide a wide range of ecosystem services including carbon sequestration/storage, climate change mitigation, improvement of water quality and runoff regulation, and the provision of a landscape with cultural, recreational and livelihood values. Globally, peatlands hold an estimated 650 billion tonnes of carbon on 3% of the Earth’s land surface, the equivalent to more than half of the carbon in the atmosphere or the carbon stored by Earth’s vegetation [
2]. For their multiple benefits, the need for peatland conservation is widely recognized (i.e., the United Nations Framework Convention on Climate Change, Ramsar Convention on wetlands, the Convention on Biological Diversity, United Nations Convention to Combat Desertification) but has been hampered by short term economic priorities and national development policies [
3,
4,
5]. Large areas of peatlands have already been degraded (estimated 20–25%), and remaining areas are quickly disappearing as a result of logging and plantation development, conversion to residential and industrial zones, climate change impacts and accidental burning [
2,
6,
7,
8,
9]. Not only do these land use changes reduce biodiversity, they also turn peatlands into net emission sources of greenhouse gases (GHGs) at a faster rate since draining of peatlands release greenhouse gases such as carbon dioxide from the carbon stored within peat soils.
For sustainable management of remaining peatlands, a better understanding of fundamental variables including their spatial distribution and extent is required. However, considerable uncertainties about these variables remain, particularly in the tropics. Generally, peatland research has focused strongly on boreal and temperate peatlands, with tropical peatlands receiving much less attention. The past two decades have seen increased interest in tropical peatland research [
10,
11,
12], but most studies have focused on Southeast Asia where an estimated 56% of tropical peatlands exist [
13]. By comparison, African peatlands are wholly understudied. Indeed, there remains a basic uncertainty about the existence–extent and distribution–of peatlands in Africa. For instance, a very large peatland, approximately 14,550,000 ha in area and storing an estimated 30.6 billion tonnes of carbon, was discovered only recently in the Congo Basin [
14]. In Ghana, the Greater Amanzule landscape has been reported as a tropical biodiversity hotspot undergoing rapid development from agricultural plantation and urbanization [
15,
16]. Also, with an influx of oil and gas activities on the Greater Amanzule landscape from 2011, a complex array of pressures for peatland conversion are expected to intensify over at least the next decade [
16]. However, to date, the spatial distribution and extent of peatland on the landscape is not fully known. Generally, there is limited research on practical and cost-effective remote sensing application (e.g., analysis on the sensitivity of tropical peatland to image features) for tropical peatland mapping. Developing detailed, comparable and robust tropical peatland maps in Africa is therefore an urgent priority–to inform policymakers and conservation practitioners and to provide critical information for landscape planning and for enhancing authorities’ capacity in monitoring, reporting, and verification (MRV).
Mapping allows both the location and quantification of the extent of pristine peatlands, as well as identification of areas at risk due to their proximity to degraded zones [
17]. In addition, mapping functioning peatland areas can provide data to inform spatially explicit and realistic restoration and protection goals. However, landscape mapping of tropical peatlands remains a challenge, especially at regional, national and global scales, and this has resulted in their consistent under-representation, or complete omission from, many global vegetation maps [
18,
19].
Methodologically, optical data from high spatial resolution sensors such as Sentinel-2 (10 m multispectral imagery) and Landsat (30 m multispectral imagery) have been the primary and most successful tool for mapping peatlands [
20,
21,
22,
23,
24,
25,
26] mainly due to their spectral detail. The enhanced spectral capabilities of optical sensors help to derive numerous band ratios and indices, such as spectral vegetation indices (VIs) for monitoring vegetation species [
27,
28,
29]. VIs have several advantages over stand-alone spectral bands, including decreased effect of soil background on canopy reflectance, enhanced variability of spectral reflectance of target vegetation, reduced effect of atmospheric conditions and canopy geometry, and shading [
30,
31]. The application of optical sensing is however constrained by the frequent cloud coverage in the tropics. For monitoring high cloud coverage areas, some developments have been reported in using radar products [
32,
33,
34]. Radar can penetrate cloud cover and is also sensitive to variable soil moisture conditions which makes it suitable for wetland mapping [
19]. It offers detailed information on the often difficult to detect characteristics of vegetation such as moisture, roughness and shape [
35]. Additionally, data from the Shuttle Radar Topographic Mission (SRTM) have been used to successfully identify hydrological landscape units in cloud-persistent areas (e.g., [
36,
37,
38]). SRTM data provides an estimate of elevation and is useful for identifying large-scale topographical boundaries within tropical landscapes.
Recent classification approaches favour the integration of data from multiple sensors for improved landscape characterization (e.g., [
36,
39,
40,
41]). Because of their complementarity, optical, radar and topographical data fusion presents an increased opportunity to map peatlands at fine scales in equatorial zones affected by cloud cover, although the choice of appropriate features from these datasets remains a challenge. Past efforts to partially map the Greater Amanzule peatland (predominantly mangrove) have relied on a single source of data [
42,
43], plus participatory GIS and ground referencing methods [
15,
16]. In the present study, optical, radar and elevation remote sensing data and their various combinations are used to classify the entire Greater Amanzule landscape. We determine the optimal classification approach for the Greater Amanzule tropical peatland based on the integration of image features derived from Sentinel-2, Sentinel-1 and SRTM data–thus to advance geospatial methodologies for mapping tropical peatland. Specifically, the study (i) defines the extent and distribution of peatland on the Greater Amanzule landscape; (ii) proposes a framework for extracting appropriate Sentinel-2, Sentinel-1 and SRTM image features for tropical peatland mapping; (iii) assesses how classifications of different combinations of Sentinel-2, Sentinel-1 and SRTM image features compare; and (iv) determines the sensitivity of land cover types to multi-source data features.
3. Results
3.1. Selection of Optimal Feature Variables
After executing the RFE, the optimal number of features was obtained, as shown in
Figure 4. This represents subsets of feature variables that combine to produce the best classification accuracy. The results show that all features considered for the classifications of S2, S1, and S1+ were relevant, thus, the highest accuracy for these datasets was obtained at the point where all features were used for the classification. For the other three datasets—S2+, S2+S1+, S2+S1+DEM—cross-validation scores increased significantly in the early stages and peaked when the number of features was 20, 25 and 29, respectively. Beyond the maxima for these three datasets, cross validation scores fluctuated with increasing feature numbers, signifying the presence of irrelevant or redundant features which were not increasing classification accuracy. The analysis revealed that the S2+, S2+S1+ and S2+S1+DEM datasets contained one, eight and seven redundant features respectively which needed to be removed. The relevant features of the S2+ dataset included nine original bands, 10 vegetation features and one texture feature. For the S2+S1+ dataset, relevant features included eight original bands, 10 vegetation features, six texture features and one temporal feature. For S2+S1+DEM, relevant features included 10 original bands, 10 vegetation features, six texture features, two temporal features and one elevation feature. Our analysis showed that the cross-validation score increased with increased number of image features. Accuracy scores followed a similar pattern where S1 < S1+ < S2 < S2+ < S2+S1+ < S2+S1+DEM. This shows that the presence of more relevant feature variables can result in improved classification accuracy.
Table 4 shows the optimal features retained for classification.
3.2. Land Cover Classification Accuracy
Sections of the six land cover classifications (S2, S2+, S1, S1+, S2+S1+, S2+S1+DEM) are presented in
Figure 5. Maps were visually similar for S2 and S2+, S1 and S1+, as well as S2+S1+ and S2+S1+DEM, the differences of which were only revealed in quantitative accuracy assessment.
In general, land cover classifications were relatively accurate, with all overall accuracies approaching or exceeding 90%, except for S1 and S1+, as expected (
Table 5). Of the different datasets, S2+S1+DEM produced the highest overall accuracy (94%), followed by S2+S1+ (92%) and S2+ (91%). Full error matrices of all classifications are available as
Supplementary Materials (
Tables S1–S6).
Despite the visual similarities of the classified images, a McNemar test of significance showed that accuracies of all datasets were significantly different from each other (
Table 5). This finding reaffirms the contention that accurate land cover mapping requires the use of relevant features, and here feature optimization holds considerable value for the mapping community.
The UA and PA results are presented in
Table 6. The S2+S1+DEM stands out with generally better results. The worst were found for S1 and S1+. To investigate the superiority of the S2+S1+DEM classification against the other datasets, we compared class UAs and PAs (
Table 6). A total of 101 table cells (84%) showed improvement in either PA or UA with S2+S1+DEM compared to the other datasets, whereas 8 cells (7%) showed another dataset was better than S2+S1+DEM. This demonstrates the robust nature of the S2+S1+DEM for discriminating different land cover classes–it performed better when compared to the other datasets. For land cover classes in which the UA or PA did not improve by at least 10% over the other datasets, accuracies were already generally high (>70%).
F-scores from all datasets are presented in
Figure 6. Apart from S1 and S1+, all datasets achieved high accuracies in differentiating the peatland classes (mangrove swamp, mixed swamp, palm swamp and bog plain), with an F-score between 0.91 and 0.99. S2+S1+DEM had the best F-score for all the land cover classes, between 0.80 and 0.99, which indicated that this approach has strong potential for land cover classification, particularly for tropical peatland mapping. The results also reaffirm the ability of the RF machine learning algorithm to map complex landscapes accurately [
32,
65,
66,
67,
68].
3.3. Feature Importance
A summary of the most important features for each dataset is presented in
Figure 7. The results varied considerably, depending on the feature types used in training the RF classifier.
When the classifier was trained with the full dataset (S2+S1+DEM), elevation, was the most important predictor variable, thereby highlighting the important role of topography in peatland delineation. Although the S1 and S1+ datasets were relatively ineffective for classification on their own, the radar derived features were consistently important predictor variables in multi-source datasets, VH, VV and VH standard deviation in particular. Red Edge 2 was the most important variable for the S2 dataset, but was deemed irrelevant in S2+, S2+S1+, and S2+S1+DEM, showing that optimizing feature sets by removing irrelevant features is an important step to avoid assumptions that can lead to reduced classification accuracy. From this point on, we focus on S2+S1+DEM, owing to its general superiority over the other datasets.
The sensitivity of the 12 land cover classes to the 36 features of the S2+S1+DEM (described in
Table 3 in
Section 2.4) is demonstrated in
Table 7. The original Sentinel-1 bands (VV and VH) were consistently important predictors for most of the vegetation classes (both peatland and non-peatland vegetation classes) on the landscape. Elevation feature derived from the SRTM data was also very important for discriminating vegetation types, indicating that the spatial distribution of vegetation types is greatly influenced by topographic information. One image feature which also stood out as being very important for differentiating unvegetated from vegetated areas is Sentinel-2’s SWIR 2 band. Results on how the other datasets discriminated the various land covers are presented as
Supplementary Materials (
Tables S7–S11).
3.4. Classification of the Greater Amanzule Tropical Peatland Using the S2+S1+DEM Dataset
The extent and distribution of the Greater Amanzule landscape are presented in
Table 8 and
Figure 8, respectively. Water aside, peatland classes constituted a significant proportion of the landscape (23% of total area without water)—dominated by mixed swamp, palm swamp, bog plains, and mangroves, respectively. The result clearly demonstrates a largely vegetated landscape, with patches of built-up land making up only 5273 ha (˂1%) of the total landscape. While this may suggest a largely undeveloped landscape, it is important to note that about 50,713 ha (8.7% total area) of the vegetated areas are plantations of coconut, rubber and oil palm—this represents land use conversion from natural forest and/or peatland. In the absence of clearly defined boundaries and management strategies, the plantation development in Greater Amanzule should be an environmental concern since similar land conversion has proved to be a major threat to tropical peatlands in South East Asia [
74,
75,
76].
Patches of the landscape showed bare surfaces, mostly comprised of land cleared for development. Field observations showed road construction works and the development of oil and gas industries. Although limited in number, roads have already been constructed across some of the peatland of the Greater Amanzule (e.g., between Ellonyi and Kengen, and between Alabokazo and Sanzule, on the Ankobra River [
15]). No studies have yet been conducted to ascertain the specific impact of roads on the peatland vegetation, hydrology or soil properties, although observation of palm swamp and mangrove die back at Sanzule and Kamgbunli respectively, during the Alabokazu-Sanzule road construction [
15], suggests that roads could be having a negative impact on the peatland. Roads can divert or impede water, act as a barrier to groundwater and channel flow, and eventually lead to the degradation of peatland vegetation or carbon cycling. The siting of large oil and gas facilities in the peatland [
15] may have a negative impact, if not controlled. This has the potential to drive up population, increasing demand for land and eventually peatland, and construction work may have a negative impact on the hydrology of the area.
We estimate that the total peatland area in 2019 is 60,187.04 ha. Within the landscape, mangrove occurred in patches and predominated around the Ankobra River, Bakanta, and Miemia. A large block of mixed swamp could be found along the border of Ghana and Cote D’Ivoire, on the River Tano, extending onto the Aby lagoon. Another large block of mixed swamp was found around the stilt village, Nzulezo. Palm swamp and bog plain occurred in patches, with the highest concentration around Nzulezu. These are areas currently without any formal management regime in place and will thus require effort from stakeholders, particularly local communities and government agencies, to ensure their conservation.
In terms of the distribution of other land cover classes, coconut was highly concentrated in the western part of the landscape while rubber was highly concentrated in the east. The concentration of oil palm was high at the Cote D’Ivoire boundary of the landscape in the west, and sparse vegetation surrounded built-up areas.
4. Discussion
In this study, effort was focused on developing a robust framework for mapping the Greater Amanzule tropical peatland of Ghana, using multi-source satellite imagery and a RF algorithm within the GEE environment. The proposed framework provides a systematic technique for extracting appropriate feature variables for tropical peatland classification. This has been developed by integrating original spectral bands, plus derived vegetation indices, texture and temporal features, as well as ancillary elevation data features into a single composite dataset. Multi-sensor satellite imagery has complementary characteristics which enabled improved detection of peatland and non-peatland classes. These land cover types are difficult to map using single source datasets due to structural complexity and high heterogeneity. Our results are consistent with earlier studies that combined optical and radar dataset for land cover mapping, showing that the combination produces higher overall accuracy over individual sensor dataset (e.g., [
36,
49,
59]). The introduction of the elevation data improved the accuracy of the optical-radar data combination significantly (
Table 4), thus confirming that classification enhancement may occur when a primary dataset such as SRTM is integrated with other datasets for peatland classification [
77]. The lowest overall accuracy was observed with the Sentinel-1 only products (S1, S1+); this is also consistent with previous studies (e.g., [
39,
59,
78]). Despite the cloud penetration advantage of Sentinel-1 data, it failed on its own to distinguish various land cover classes. This suggests that the best way to maximize the utility of such data for various land cover classes discrimination is to combine it with optical datasets as demonstrated in this study. Da Silva
et al. [
79] also suggested advanced techniques such as SAR polarimetry to optimize SAR for the discrimination of diverse land cover classes.
The overall accuracy of the S1 dataset (70%) reported in our studies was relatively high when compared to other land cover classifications that utilized SAR and RF classifier (e.g., [
59]). Our findings are however consistent with similar studies in wetland areas [
39,
80,
81]. Accuracy of the S1 dataset improved significantly, by 8%, when additional Sentinel-1 features were extracted and combined with the original bands for classification. Likewise, the overall accuracy of the S2 dataset improved significantly when combined with additional Sentinel-2 extracted features. The latter observation is contrary to observations made by Tavares et al. [
59] who noted decreased accuracy when other optical features were combined with the original bands. The decreased accuracy observed in their case might be due to the presence of irrelevant or redundant features. It is therefore important to optimize datasets by eliminating such redundant features for an improved accuracy in the case of integrating more features.
The relevance of the multi-sensor features for the delineation of land cover components of the tropical landscape using RF classifier is illustrated in
Table 7 and
Figure 7. Elevation ranked as the most important feature when the landscape was classified with the full dataset (S2+S1+DEM), illustrating the importance of topographical and landform position in peatland occurrence and identification. Peatlands develop under long-term water saturation of the soil and are found in areas where large amounts of water are available or flowing (e.g., rivers, depression). Elevation models are known to be useful for identifying hydrological landscape units [
14,
36,
59]. This was further demonstrated when the individual classes were considered; elevation proved the most important feature for delineating water. We concur with Lidzhegu et al. [
38] that topographic information derived from the SRTM can better offer different geomorphologic characteristics which reflect the habitats of different vegetation types and can help in their identification. Despite the Sentinel-1 classifications (S1, S1+) proving relatively inaccurate overall, Sentinel-1 features were consistently rated highly in combined datasets. The addition of Sentinel-2, especially the NDWI feature, clearly leads to a more accurate distinction of peatland and non-peatland classes—this is consistent with observations made by Slagter et al. [
80] who also reported the importance of NDWI for wetland delineation. In our study, texture features computed from the Sentinel-1 image were of relatively low importance and were often removed as redundant features. This could be because of the presence of other features such as the original Sentinel-1 bands and the standard deviations of the VV, VH and NDVI bands which played similar roles thus rendering the Sentinel-1 GLCM texture features less useful.
The sensitivity of land cover types to classification features was demonstrated in
Table 7. For example, when discriminating tropical peatland classes—mangrove, palm and mixed swamp—Sentinel-1’s VV, VH and the standard deviation of the VH bands acquired higher importance scores than the optical and elevation features considered. This may be because microwaves from Sentinel-1 penetrate forests and interact with different parts of trees to produce substantial volume scattering. As the importance of VH and VV was high, it is likely that volume scattering (especially for VH in medium- and high-vegetated peatlands), double-bounce scattering (especially for VV in low- and medium-vegetated peatlands) and specular reflection (especially for VV in non- and low-vegetated peatlands) contributed to accurate classification of peatland classes [
80]. When distinguishing between rubber and oil palm, elevation was among the most important features. Elevation was also the most useful feature for separating natural forest from peatland forest (e.g., mixed swamp, palm swamp). This demonstrates that the right combination of multi-sensor features is important for the discrimination of diverse land cover types as they maximize the complementarity of the optical spectral sensitivity and the radar structural/geometric characteristics.
Our analysis estimates Ghana’s Greater Amanzule peatland at 60,187 ha, comprising mangrove, mixed swamp, palm swamp and bog plain. This is a relatively large tropical peatland with no formal/legal protection [
15,
16]. To date, community and NGO efforts to manage sections of the peatland have tended to focus on mangrove forest. The results (
Table 8) however show that mangrove occupies the smallest area of the peatland classes on the landscape. This underlines the need to broaden the scope of management foci to include the other predominant peat classes. This is important because peatlands function as hydrological landscape units; the hydrological connectedness means damage to one part can have wide-reaching consequences on the whole system. Conservation efforts that concentrate on only certain parts of a peatland unit may therefore allow peatland degradation due to activities outside the conserved zones, thus reducing the effectiveness of that conserved area in achieving its conservation goals. An option to manage the unit would be to extend the boundaries of existing conserved areas to the hydrological boundaries of the peatlands that they encompass. Our analysis suggests that plantation development may be a major threat to the Greater Amanzule peatland. This development bears close resemblance to experiences with peatland landscapes in Southeast Asia, where plantation development, predominantly rubber and oil palm, has been reported as the major threat to peatlands [
12,
74,
75,
76,
82,
83]. This points to the need to manage plantation development on the Greater Amazule landscape to ensure its sustainability. Available reports though suggest that plantation developers have not been actively engaged in the current efforts by NGOs and communities in their informal management of the landscape [
15,
16]. At present, the Greater Amanzule peatland is relatively intact, hence the need to fully investigate the threat and conservation priorities of the landscape to aid management decisions. Research is also needed to evaluate the carbon stock of the Greater Amanzule peatland to complement the work done by Ajonina et al. [
46] and Asante and Jengre [
45] who investigated carbon stock in sections of the peatland. This will aid understanding of the potential of the landscape to attract climate change mitigation funding to the benefit of fringe communities. More broadly assessing the current and future potential of the Greater Amanzule peatland to supply ecosystem services will help underline its importance and motivate public protection and conservation of its unique ecosystem functions and services [
84]. The results from this study can be used as a baseline for onward analysis of land cover change to understand the impact of plantation development and to simulate future land uses. For instance, the distribution of coconut plantations in the study area is reported to have been affected by a coconut disease that killed most coconuts in the eastern part of the landscape. Time series analysis may again help to quantify the impact of the disease, and subsequently help to prepare for similar situations in the future [
16]. Finally, even though our proposed framework is implemented on a relatively large area (e.g., when compared to [
74,
83]), we still recommend its application on an even larger area to better understand how the model will perform on more complex and heterogenous landscapes.