Edinburgh Research Explorer Rapid Mangrove Forest Loss and Nipa Palm (Nypa fruticans) Expansion in the Niger Delta, 2007-2017

: Mangrove forests in the Niger Delta are very valuable, providing ecosystem services, such‘as carbon storage, ﬁsh nurseries, coastal protection, and aesthetic values. However, they are under threat from urbanization, logging, oil pollution, and the proliferation of the invasive Nipa Palm ( Nypa fruticans ). However, there are no reliable data on the current extent of mangrove forest in the Niger Delta, its rate of loss, or the rate of colonization by the invasive Nipa Palm. Here, we estimate the area of Nipa Palm and mangrove forests in the Niger Delta in 2007 and 2017, using 567 ground control points, Advanced Land Observatory Satellite Phased Array L-band SAR (ALOS PALSAR), Landsat and the Shuttle Radar Topography Mission Digital Elevation Model 2000 (SRTM DEM). We performed the classiﬁcation using Maximum Likelihood (ML) and Support Vector Machine (SVM) methods. The classiﬁcation results showed SVM (overall accuracy 93%) performed better than ML (77%). Producers (PA) and User’s accuracy (UA) for the best SVM classiﬁcation were above 80% for most classes; however, these were considerably lower for Nipa Palm (PA—32%, UA—30%). We estimated a 2017 mangrove area of 801,774 ± 34,787 ha ( ± 95% Conﬁdence Interval) ha and Nipa Palm extent of 11,447 ± 7343 ha. Our maps show a greater landward extent than other reported products. The results indicate a 12% (7–17%) decrease in mangrove area and 694 (0–1304)% increase in Nipa Palm. Mapping e ﬀ orts should continue for policy targeting and monitoring. The mangroves of the Niger Delta are clearly in grave danger from both rapid clearance and encroachment by the invasive Nipa Palm. This is of great concern given the dense carbon stocks and the value of these mangroves to local communities for generating ﬁsh stocks and protection from extreme events.


Introduction
Mangrove ecosystems are intertidal regions at the land-sea, fresh-salt water interface; hence, they have characteristics of both zones. Mangrove systems also have peculiar properties, such as anaerobic conditions, salinity fluctuation, and tidal influence [1]. Mangrove ecosystems in the world are of great ecological, economic, and social importance [2][3][4] because they are important carbon stores, supportive of the lives of coastal human populations who depend on fisheries. Mangrove forests also serve as a direct means of coastal protection from storm surges, tidal waves, and, over longer time scales, provide resilience to climate-change driven sea level rise. The ecosystem services provided by mangrove forests are under threat from natural and anthropogenic factors. The superimposition of mangroves with the high population density of coastal communities has resulted in rapid and A further identified source of the degradation of Nigeria's mangroves is the alien invasive mangrove palm, Nypa fruticans, a mangrove palm commonly known as the Nipa Palm, native to the Indian Ocean coast. Non-native species are known to proliferate in damaged ecosystems, and can lead to a change in community structure and biodiversity by altering the function of the ecosystem [22]. Nipa Palm was introduced in the Calabar estuary of Nigeria, to the east of the Niger Delta, in 1906 to serve as a plantation for beautification and beach erosion control [23]. However, Nipa Palm spreads rapidly through its water-borne seeds, and has proliferated rapidly along Nigeria's coast. This spread has been assessed by poor management of plantations, as well as actions that degrade mangroves, letting the Nipa Palm in, including unsustainable logging activities, dredging, and oil pollution [24,25]. Nipa Palm spread is further compounded by the lack of use of this palm as a resource, as it is in Asia and Oceania where its fruit is eaten and used to produce juice, and palm fronds used for roof construction [26]. Therefore, anecdotal studies suggest Nipa Palm is gradually replacing mangrove forests in the Niger Delta, especially in areas of high exploitation [27]. Various attempts have been made to use Nipa Palm's potential products, which include its sap (which is sugar rich), its tannins, its potential for bioenergy products, and its palm leaves as a building material [24,28,29]. However, the location of the Nipa Palm along the tidal channels places it at an unfavorable location for transportation of the resource generated. The United Nations published a report in 2007 on possible control measures of Nipa Palm infestation in Nigeria [30]. There has also been a proposal for the utilization of Nipa Palm as a renewable source of energy [24]. Cultural management of Nipa Palm spread is undertaken by the removal of Nipa Palm seedlings in some parts of the Niger Delta, but there are no reports of a large-scale management plan of Nipa Palm invasion in the Niger Delta. Monitoring the extent of native species and the spread of invasive species can provide the key to controlling alien species and managing the conservation of the native species [31]. Therefore, monitoring vegetation over space and time is essential for successful management of Nipa Palm, as well as for other threats to mangroves forests in the Niger Delta.
Satellite remote sensing offers an appropriate method for mapping land cover (LC) and LC change. Remote sensing data splits broadly into two types: passive or active [32]. Passive sensors use reflected solar radiation to monitor the observed region, while active sensors emit signals that interact with surface or sub-surface features [33]. Passive sensors such as optical data are sensitive to the greenness of an ecosystem, while active sensors have the penetrative power to retrieve forest structure including the woody biomass. Passive sensors can be excellent at monitoring productivity and the degree of canopy cover, given their sensitivity to the contrast between green vegetation and non-vegetated surfaces. However, optical data are often masked by cloud cover, which is always a problem in coastal tropical regions [5]. In contrast, active sensors such as Synthetic Aperture Radar (SAR) record the surface structures [34][35][36]. Medium and long-wavelength (C-band or longer) Synthetic Aperture Radar (SAR) data do not suffer from cloud cover, and obtain complementary information to optical data about forested systems. In particular, this is true of longer-wavelength SARs (L-band or longer), as their data penetrate the forest canopy and given information on forest structure. However, these SAR data have been more rarely used for classification of mangrove systems because of poorer data availability, more complex processing, lower resolution and more noisy data, and higher costs, than optical data (which are often free). However, the release by Japanese Aerospace Exploration Agency (JAXA) of the Kyoto and Carbon mosaics of Advanced Land Observatory Satellite Phased Array L-band Synthetic Aperture Radar (ALOS PALSAR) data has made long-wavelength SAR data far easier to use [37].
Although these two sensor types can be used separately, more recent studies have fused the two to achieve higher accuracy in analysis due to their combined individual strengths [11,35,38,39]. The Landsat and Sentinel-2 and sensors have been acquiring optical data at moderate (10-30 m) resolution since the 1970s and 2015, respectively, and are openly available. Although most scenes captured by these satellites over mangrove regions are cloudy, optical data can still be used through the stitching together of cloud-free portions of many images to make cloud-free composites, and in combination with SAR data.
There are various factors involved in choosing the classification algorithm to use for optimal LC discrimination from remote sensing data. LC classification can be done using unsupervised or supervised methods. Unsupervised classification predicts different classes based on statistics from the spectral characteristic of the satellite products; while supervised methods predict LC types using ground control points as training data [40]. A simple and common supervised classification method is the Maximum Likelihood (ML) classifier; a parametric method that assumes a normal distribution of the multispectral data, and that is highly computationally efficient [41,42]. The Support Vector Machine (SVM) classifier, contrary to ML, is non-parametric, making no assumption about the data distribution. SVM identifies the optimum boundary between training classes utilizing the edge of the class distribution [43,44]. This optimum boundary can also be referred to as the Optimum Separation Hyperplane (OSH), while the training samples at the class distribution edge can be referred to as the support vectors [40,45]. The accuracy of SVM has been shown to be better than ML, Artificial Neural Networks (ANN), and decision trees, especially when training datasets are comparatively small [42,46,47]. However, the use of SVM and ML in detecting fringe invasive species is largely unstudied. Mapped results are validated for accuracy by comparing with other mapped products, using accuracy assessments or field verification [48].
The development of a spatiotemporal map of mangrove cover for Nigeria could be valuable as it would form a baseline for mangrove conservation and restoration. Mangrove extent in the Niger Delta has been estimated by James et al. (2007) and Fatoyinbo and Simard (2013), both using an unsupervised classification ISODATA method to estimate mangrove area from Landsat Enhanced Thematic Mapper Plus (ETM+) [49,50]. These studies show the ability of remote sensing options in estimating mangrove area without the need for training data. However, the use of ground truth data and fused radar/optical data has not been explored in the estimation and change detection of coastal vegetation in the Niger Delta.
Here therefore, we estimate mangrove forest and Nipa Palm cover in the Niger Delta between 2007 and 2017. We compare the accuracy of ML and SVM in estimating coastal vegetation, in order to provide a steer to other studies. Specifically, we aimed to: 1. Compare the two different types of land cover classification in estimating mangrove area; 2. Estimate the extent of mangrove and Nipa Palm and; 3. Detect and report the change of mangrove and Nipa Palm area between 2007 and 2017.
From our knowledge of the literature [42,44,45,51], we hypothesized that SVM will have a better accuracy than ML in predicting LC classes using fused data. We also predicted from our knowledge of the region that mangrove area will have reduced in the delta due to deforestation, and Nipa Palm extent increased, over the study period.

Study Area
The Niger Delta study region ( Figure 1) spans from the Benin River estuary in the west to the Calabar river estuary in the east. Economic activity is primarily farming within these regions while fuelwood and fisheries account for minor sources of income [19]. Increased urbanization is occurring including road construction, port establishment and settlement building [27]. Mangroves species zonation is influenced by saline conditions [52][53][54]. However, Nipa Palm interrupts this zonation along mangrove fringe and inland sections where exploitation for timber or fuelwood has taken place [23].

Field Data and Sampling Strategy
We carried out three field surveys for eight months across the Niger Delta: September 2016-January 2017 and June-August 2017. We adopted a multiclass method consisting of 6 broad vegetation classes, chosen from our knowledge of the region during fieldwork and the literature ( Figure 2, Table 1). These classes were chosen in order to focus on mangrove and Nipa Palm. The other classes gave a general overview of other land cover types in the region.
During the fieldwork, we collected 567 ground control points (GCPs) across the East-West Highway (which connects the entire coastal state); during surveys on a boat, and during sample collection ( Figure 1). We used an opportunistic sample scheme to collect GCPs due to security issues ranging from kidnapping, oil bunkering and militancy [55]. We recorded the type of land cover class, GPS coordinates using a Garmin eTrex 20x device (≤10 m error); and at the end of each day, we used Google Earth Pro application to delineate the areal extent of each of land cover class. Areas of interest during GCP selection were islands off the Calabar estuary and along the Imo river estuary where a clear transition of Nipa Palm, to mangrove forests, then to agricultural lands or terra forma forests was observed. Another essential region were creeklets in Rivers State where settlements were surrounded by a transition of mangrove forests and rainforests. We selected points with distinct spectral characteristics in order to improve accuracy and validation of the classification output. Despite not having GCPs in 2007, we used Google Earth Pro application to locate the GCPs collected during the field campaigns and assess if they were a different or the same class in 2007 using the timeline tool. We removed GCPs where their locations were obscured by cloud in the image data. We used 417 stable GCPs in the supervised classification in 2007.
The GCPs over the entire region were uneven because of the proportional cover of the different classes over the Niger Delta region. Hence, we collected more training pixels for classes with more coverage to classify the land classes in relation to their proportion. We divided the GCPs between training and testing in a ratio of 70% for training the classification algorithm and 30% for testing map accuracy.    The SRTM acquired data at both X and C-SAR bands designed for single-pass operation interferometry [56]. This data filled gaps and voids with elevation data primarily from the Terra Advanced Space borne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model Version 2.0 (GDEM 2) and secondarily from the U.S. Geological Survey (USGS) Global Multi-resolution Terrain Elevation Data (GMTED) 2010 [57]. We also used the SRTM DEM in order to aid the LC classification of the Niger Delta because we know that such sea-influenced vegetation will change with elevation above sea level [58]. Mangrove and Nipa Palm vegetation occur at the intertidal zones of the coast. Hence, this will significantly aid in differentiating coastal vegetation from other vegetation types. We downloaded the SRTM DEM 1 arc sec global version 3 data for twelve tiles covering the Niger Delta using Earth Explorer.

Synthetic Aperture Radar (SAR)
The Japanese Aerospace Exploration Agency (JAXA) has successfully launched EO missions in order to monitor disaster, cultivated land, increase data archives and tropical rainforests [59]. These missions include the Advanced Land Observing Satellite 1 and 2 (ALOS and ALOS-2). ALOS-2 was a sequel to ALOS "DIACHI", launched in May 2014 [60]. ALOS-1 and 2 both used the Phased Array type L-band Synthetic Aperture Radar (PALSAR). The L-band SAR scenes we used were acquired in Fine Beam Dual (FBD) polarization mode with off-nadir angle averaging 34.3 • for ALOS PALSAR and multi-looked pixel size of 0.8 arc sec (~25 m).

Landsat Data
Landsat ETM+ 7 Collection 1 Tier 1 Digital Number values from the study region was downloaded and pre-processed to remove cloud cover using the Google Earth Engine (GEE) [62].

Image Processing
All spatial analyses were undertaken within ENVI version 5.1 [63], QGIS [64], Google Earth Engine (GEE) [62], Google Earth Pro (GEP) [65] and ArcGIS [66]. Three datasets were used for LC classification: SRTM DEM, ALOS PALSAR and Landsat 7 ETM+ ( Figure 3). We imported ground control points (GCPs) into GEP to generate polygons based on information from field surveys. QGIS and ArcGIS were used to convert the polygons to shape file format, re-project and dissolve to the land cover classes. GEE was used to pre-process and download Landsat data. ENVI was used to pre-process digital elevation model and radar data, carry out classification and post classification. Details of these processes are given in the sections below.

Synthetic Aperture Radar (SAR)
We mosaicked individual data tiles to form a single image file of the Nigerian coastline in both bands and geo-referenced with the projection Universal Transverse Mercator (UTM) Zone 31 North and World Geodetic System (WGS) 1984 datum. The image file of both bands was registered using 41 GCPs in Landsat bands of Hansen et al., (2013) and using a first order polynomial approach (RMS error = 0.27) which was sufficient for data analysis [67].
Calibration of ALOS PALSAR backscattering on the image through two stages. We first converted the digital number (DN) value on the original image to decibel (dB) (Equation (1)) based on the coefficients and equations from Shimada et al., (2009) [68]. We used single year tiles (2007 and 2017) for the LC classification.
Conversion of DN to σ 0 (dB) = 10 × (log 10 (DN 2 )) − 83, Speckle associated with Synthetic Aperture Radar (SAR) data reduces the efficiency of class characterization by affecting the radiometric and textural qualities. However, filtering processes can reduce the impact of speckle. These filtering techniques can reduce or eliminate the information contained in the image, in particular resulting in a smoothing out of the (real) hard boundary between two land cover types [69,70]. Adaptive filters have been developed to attempt to reduce this problem: and thus, we used the Enhanced Lee Filter to reduce the speckle. Adaptive filtering makes a choice about how to average a pixel based on its neighborhood. The Enhanced Lee Filter determines the grey level for each pixel by computing the weighted sum of the center pixel value, the mean value, and the variance calculated in a square kernel surrounding the pixel. This filter is used primarily to suppress speckle by smoothening image data without removing edges or sharp features in the images while minimizing the loss of radiometric and textural information [71].
We carried out the filtering process in this study used a 7 × 7 window on both HH and HV bands to minimize the loss of information in the image, which gives the best results for classification based on published results [71,72]. After the conversion of digital value to the backscatter coefficient (σ 0 ) and speckle reduction performed, the ratio of HV to HH was calculated after transformation into the power domain (P) (Equation (2)). This ratio has been shown to contain further information useful for classification, by giving information on the comparative strength of different scattering mechanisms.

Landsat Data
A cloud-free composite was created using the ee.Algorithms.Landsat.simpleComposite() method, which allows the creation of a composite from and image collection (stacked images). This algorithm selects the median value for each band of each pixel. We ensured median composites were cloud free over the period by building up the collection using pixels with less than 5% cloud cover over the time-series. In order to assess two different years, we created a composite of Top of the Atmosphere (TOA) data from 2005 and 2007; and from 2015 and 2017. We also calculated the cloud score of the scenes in the Landsat collection used in the creation of a cloud free composite. The GEE code can be assessed here: https://github.com/ebukanwobi/Landsat-Scene-Code-Niger-Delta through a GIT repository.

Texture Measure
We performed first order occurrence statistics on the DEM, and all ALOS PALSAR and Landsat 7 ETM+ bands. We applied a 3 × 3 and 7 × 7 window size for all image texture analysis for comparison. The window size has the advantage of capturing the heterogeneity of pixel values over edges of different land cover classes, especially mangrove-Nipa Palm interface. However, texture measures with higher window sizes can reduce the quality of the image for analysis. We selected texture measures based on their established ability to characterize vegetation structure. We calculated three first order texture measures (data range, mean and variance) using the ALOS PALSAR and Landsat 7 bands. We chose these measures because they had the highest ROI separability between mangrove and Nipa palm.

Layer Stacking
We stacked all EO data for both years (2007 and 2017) using a 30 m scale ( Table 2).

Supervised Classification
We implemented ML and SVM to classify the layer stacked data for 2007 and 2017 into six classes. We masked any 'no data' pixels within bands for the ML to avoid errors during statistics computation. We also tested ML classification for various band combinations for comparison including: The kernel function in SVM involves a choice of how to represent the data in higher dimensional space [73]. This dimensional space could be radial basis function, sigmoid and polynomial. The penalty parameter (C) is the degree of how much error given in the classification. A higher C will result in minimal error [74]. Gamma is a function of how the distance between training data affect the similarity of the training points. This only applies to non-linear kernel and we used the default value (inverse of the number of bands used in the classification-31 bands). We tested two types of kernel type for SVM classification based on results of Yang, (2011) who tested various parameters of SVM in a LC classification [75]. In order to carry out a change detection analysis, we classified the same set of data for both years, 2007 and 2017.

Accuracy Assessment
We assessed the performance of the classification results by carrying out post classification confusion matrices and overall accuracies using testing pixel (Table 1) independent of the training pixels. We used identical training and testing pixels for the different types of classifiers to minimize bias separately for both years. We placed particular interests in the following confusion matrix variables: Overall accuracy (a measure of how accurate the classified classes), Producer's accuracy (PA; the probability of how accurate each of the classes were classified), and User's accuracy (UA; the probability that a certain class prediction belongs to that class) [76].
We used prior knowledge of the study site from both fieldwork and communication with the locals, as well as Google earth timeline images. Areas of focus for the visual search included roads; urban areas embedded within larger forest, and forest-urban-mangrove boundary. We also reported the individual confusion matrices of the different land cover classes.
We used the PA and UA to calculate 95% confidence intervals (CI) for areas of each land cover classes. This method of 95% CI for areas was based on Olofsson et al., (2014) using stratified random sampling based on single year classification [77].

Change Detection
We performed change detection analysis of the resultant LC types using the change detection statistics tool in ENVI. This tool analyses the change from a base initial image for each class. It does this by evaluating the number and percentage of pixels change in classes between the initial and final images. The time intervals investigated in this study were 2007 (initial) to 2017 (final).

Accuracy Assessment
Overall accuracies from the SVM classifications for 2017 (93%) was greater than the ML method (77%). Using ML method, classification based on Landsat bands (66.8%) performed better than SAR bands (44.4%) ( Table 3). Accuracy improved using a combination of Landsat and SAR bands (67.4%). Inclusion of texture measures also improved LC classification with 3 × 3 window size (77%) performing slightly better than the 7 × 7 (76.9%). Using a combination of SRTM DEM, SAR, Landsat and texture measures using 3 × 3, the overall accuracy (92-93%) did not differ amongst the kernel types tested in the SVM method (Table 3). Surface water; urban regions and agricultural lands were easily detected with PA above 90% for all SVM classifiers (Tables 3, A1 and A2). The mangrove class had the lowest producer's accuracy (65%) in the ML, but fortunately had much better accuracies (>90%) when using the SVM methods (Table 3). We also discovered that while the highest producer's accuracy (100%) on the Nipa Palm invasive species was in the ML method, it also had a very low user's accuracy (5%) ( Table 3), with further investigation showing this was due to overestimation ( Figure 4). The accuracy of Nipa Palm classification was low using all other methods, with even the best SVM method (RBF) achieving a User's and Producer's Accuracy of 32% and 30% respectively. However, as these were relatively balanced, we hope that the area estimate of Nipa Palm from the classification remains reasonably accurate, and certainly these low accuracies will feed into the confidence interval estimates so useful ranges of area are still available from the analysis (See Table 6).
We decided that the best classification results were from the SVM method under the Radial Basis Function kernel type, which had the highest classification and user's accuracy for mangrove (89%, 93%) and Nipa Palm (32%, 30%). There was high confusion of Nipa Palm with both surface water and mangrove forests because of the class transition between the land cover classes (Table 4).
For 2007, we only used the best performing classification algorithm from 2017 (SVM with a radial basis function kernel), which resulted in a similarly high overall accuracy of 93.4% (Table 3). We discovered that both SVM polynomial and RBF kernel classifiers in 2007 had similar classification results (Table 3). However, we chose the RBF because it had a higher producer's accuracy for mangrove (Table 3). Accuracies per class were similar to the 2017 classification. Surface water, urban regions, mangrove forests and agricultural lands were easily detected with producer's accuracy above 80% while Nipa Palm had a PA of 42%. This was because, Nipa Palm was mostly confused with mangrove forests (25% of classified pixels) and surface water (26%) ( Table 5).
We also visually compared the different classification results (Figure 4). Across three regions where visual evidence was presence based on fieldwork: Calabar estuary with heavy Nipa Palm invasion (Figure 4a), Oproama community with no Nipa Palm invasion (Figure 4b) and Imo River Estuary with intact stands but medium invasion (Figure 4c). ML overestimated Nipa Palm vegetation, predicting its presence even in areas where they are non-existent. SVM classifiers performed better in separating Nipa Palm from mangrove and other LC classes, with its distribution matching the author's expectations from areas they have visited during the fieldwork, and based on the literature.

Classification Results
As stated above, we chose an optimal mapping method for 2007 and 2017 (SVM with RBF), and this represented the best LC maps for the Niger Delta ( Figure 5 (Table 7). We also reported mangrove area based on the coastal division of Nigeria by Hughes and Hughes, (1992) [78] ( Table 8).
We observed that mangrove forests extended about 60 km in the western Niger Delta basin (Delta state- Figure 5a). We also calculated a mangrove extent of 40 km in the central Niger Delta basin (Bayelsa State- Figure 5b); 60 km inland around the eastern Niger Delta basin (Rivers state- Figure 5c); 20 km in Imo River (Akwa Ibom state- Figure 5d) and about 3 km along the Cross River estuary (Figure 5e). We observed from the classification maps that agricultural lands were around settlements and rain forests close to river water shed (Figure 5f). Along the coast, we observed Nipa Palm fringes followed by a longer strip of mangrove forests before transitioning to tropical forests. We observed most of the Nipa Palm fringes along Imo river estuaries especially in Kono Creek and Ete Creek. These Nipa Palm fringes were also associated with urban settlements such as Opobo, Rivers State and Ikot Abasi, Akwa Ibom State ( Figure 6).

Change Detection of Mangrove and Nipa Area
Comparing the two final maps, there was a 12% (range 10-14%) loss in forest over the entire delta and an increase of 11% in agricultural land (range 10-13%) and 50% in urban regions (range 45-55%) ( Table 6). There was a 12% decrease in mangrove area (range 7-17%) and a likely large increase in Nipa Palm area (600%), but with large uncertainty (range 29-1304%) between 2007 and 2017 (Table 4). We observed that~50% of Nipa Palm area in 2017 invaded mangrove forests (Figure 7) and~6% of mangroves were converted to farmlands or urban regions. The largest relative decrease in mangrove area was in Delta state (18%) and the lowest decrease in Rivers state (6%) ( Table 7). The highest Nipa Palm increase was also observed in Rivers state. Based on the coastal division by Hughes and Hughes, (1992) [78]; we observed an 8% loss of mangrove area at the Cross River estuary and 14% loss in the Niger Delta basin (Table 8).
We also analyzed the loss of mangrove and Nipa Palm spread over some local regions notably Benin River, Imo River and Calabar estuary ( Figure 6). There was a 15% reduction in mangrove area and over fivefold increase of Nipa Palm area in the Benin River estuary. Imo river estuary had a 40% decrease in mangrove area with over 50% of Nipa Palm colonization from these regions. We recorded a reduction of 9% in mangrove area in the third region along the Calabar estuary.

Comparison with Global Mangrove Datasets
We compared our regional map to global maps (  [10,11,79,80]. Data for comparison were acquired using the UN Ocean Data Viewer [81]. It is important to note that these maps were from different years ( Table 9) and represents mangrove forests area in Nigeria. We compared two regions along the Niger Delta, Oproama Creek (Figure 8a) and Benin River estuary (Figure 8b) in our study and the maps generated by the GMW showing similarity in mangrove forests in the Niger Delta. This similarity could be possibly because of similar datasets used. However, there were some underestimation in some regions from the GMW compared to this study, which may have been due to unavailability of ground control points. We also showed a trend in Nipa Palm invasion in the Calabar estuary (Figure 8c) by comparing the global maps from different years. The maps shows a clear replacement of mangrove forests by Nipa Palm.

Discussion
We produced the first LC classification of the Niger Delta based on comprehensive ground data, to estimate mangrove and Nipa Palm cover and recent changes. We used both radar and optical imagery [39] and compared the accuracy of different types of classification algorithms to effectively classify these land cover types. Our results show that the Niger Delta has nearly a million hectares of mangroves, itself more than any other country in Africa [82], and previous global studies may have underestimated mangrove forest area in Nigeria. However, we also found rapid changes are occurring, with a decrease in mangrove area and likely rapid increase in the invasive Nipa Palm from 2007 to 2017.

Comparison of Classifier Performance
The classification results showed the SVM has a better ability to classify LC in this region than LMC using these datasets. This is not surprising, as the SVM classifier has been known to outperform ML [83] in other areas because it allows a more complex separation plane between classes [43]. SVM creates a multi-dimensional plane which allows for a separation boundary to be established amongst classes [73]. Huang (2002) [43] reported that SVM had a more stable overall accuracy over ML, ANN, and DT [43]. We observed similar trend where ML had the lower classification overall accuracy and for the individual classes (Table 3). SVM has been reportedly used in the classification of mangrove forests [84] and invasive species [85]. The danger of this more sophisticated type of classifier is that it can over fit, by learning the characteristics of the training data to increase apparent accuracies, but perform worse against test data.
The urban area and water classes had uniformly high accuracy. This easy detection could be due to spectral nature of the classes (quite different to the vegetation classes of the rest of the images), and the inclusion of DEM which takes into account the height of these classes above sea level. We found the lowest accuracy was the Nipa Palm, which again is not surprising. Nipa Palm is spectrally quite similar to forest and mangrove, but further occurs as a thin fringe along mangrove forests and hence its spectral qualities are mixed with those of mangroves at the 30 m resolution of our classification. Difficulty in classifying vegetation ecotones was reported by [86] when classifying vegetation in muddy tidal flats [86]. Hence, importance in detecting ecotones can improve accuracy of classifications. Field surveys showed mature mixed stands of mangrove and Nipa Palm, causing even more difficulties for the classifier ( Figure 9). As there is no clear hard border between Nipa Palm and mangroves, trying to produce a binary classification into two classes will always cause errors. Secondly, the fringe nature of Nipa Palm created a large percentage of confusion with surface water (Tables 4 and 5). This confusion is because of the tidal nature of coastal vegetation which can interfere with both optical and radar satellite data. In order to improve our classification accuracy, we applied texture measures on the radar and optical data, theorizing that the uniform height and composition of dense Nipa Palm stands compared to mangrove forests would have smoother textural characteristics than mangroves [87]. The combination of multiple explanatory bands, which explained Nipa Palm structure and difference in canopy structure to mangrove forests, improved the classification accuracy. However, significant errors remain likely due in part because the textural windows (7 × 7 pixels) are larger than the normal patch size of the vegetation types, meaning pure statistics were rare.
A further issue peculiar to coastal regions is the tide: changing water levels can cause misclassification of Nipa Palm, mangrove and surface waters as the water level differs at the time of remote sensing data acquisition. This issue is compounded by different sources of satellite imagery with different sensing times, which can increase uncertainties in spatial characteristics and result in biased maps [88]. Ideally, all imagery should be collected at a particular point in the tide cycle (e.g., low tide), but the lack of data (both radar and cloud-free optical) made this selection impossible: we had to use what was available.

Current Mangrove and Nipa Palm Extent
Mangrove and Nipa Palm area from this study are comparable to past studies. Our estimates of mangroves area in the Niger Delta (801,774 ha) is~14% higher than those reported by Bunting et al., (2018) [11]. The relatively large variation in reports of mangrove area regionally and globally are a result of the methodology used in land cover classification ( Table 7). The various reports of mangrove area regionally and globally will naturally contain errors in this area as they are based on global algorithms and ground data. We believe the differences in area was particularly caused by the landward extent of mangrove forests from the coast, which is high in Nigeria, possibly higher than average: this is an important threshold in mangrove classification that varies regionally. Our results show similarities with the land ward extent of mangroves from Spalding et al., (1997) [9] especially in the central Niger Delta [9]. We therefore conclude that mangrove area in Nigeria (and also the rate of Nigeria's mangrove area loss) has been underestimated by global datasets.
The global mangrove maps mirrored the spatial extent of mangrove forests in the Niger Delta (Figure 7). We observed that more than 90% of Nigerian mangroves lie within the core Niger Delta states of Delta, Bayelsa and Rivers states. This coverage was also reported by Fatoyinbo and Simard, (2013) reporting about 80% of mangroves lie within the Niger Delta swamps [50]. The ability of regional maps to estimate the presence of mangrove patches could also be a reason of the differences in globally reported mangrove area. The utilization of >500 GCPs during our study aided in the identification of mangrove patches which may have been undetected due to generalization. The landward extent of mangrove forests is an important threshold in mangrove classification as they vary regionally; GCPs are likely important in fixing this boundary. Thomas et al. (2018) [14] reported a 25.9% larger estimate in Riau region, Indonesia, due to a greater estimate on mangrove landward margin. They also attributed the higher estimate to mangrove gain [14]. Landward expansion has also been reported by Suyadi et al. (2019) [89] in Auckland, New Zealand, sometimes at the expense of salt marshes [89].
Nipa Palm distribution in Nigeria has a long and interesting historical record. Spalding et al., (1997) [9] reported that in the Palaeocene era, Nypa pollen had fossil records in both Nigeria and Brazil. Hence, Nipa Palm has always been associated with the Atlantic coasts. However, none was likely there in 1906 when it was introduced exotically for beautification and beach erosion [23]. We recorded Nipa Palm cover of about 11,000 ha, though with high uncertainty, with the largest recorded in both the Calabar estuary and the eastern Niger Delta. These two regions have had Nipa Palm introduced in 1906 [23] at Calabar estuary and in the 1960s at the Imo River estuary [90]. Our results show a similar trend with reducing Nipa Palm cover along an east to west gradient. We recorded lower Nipa Palm cover than those recorded by Isebor et al. (2003) [26], recording a 82,100 ha cover in 2003 [26]. Despite being an underestimate of previous reports and wide range of uncertainty, more research is needed to have a more accurate estimate to facilitate management of Nipa Palm. Understanding the trend in invasive species spread can help manage the conservation of native species in an ecosystem.

Change Detection
Urbanization, oil pollution and unsustainable exploitation of wood products are the major causes of mangrove deforestation in Nigeria. We recorded a 12% loss of mangrove and tropical forests over the decade, which could have been because of population dependent factors such as pollution, urbanization, agricultural expansion and Nipa Palm invasion [21]. This conversion rate shows the influence of population growth on forest resources in the Niger Delta, with pressure on land areas and food resources increasing. The Niger Delta region had a population of 30 million in 2005 but this is projected to almost double, currently estimated with a population of about 45 million in 2020 [16]. The increasing population has been described as an ecological time bomb especially in Rivers state [91]. Hence, there will be increasing the demand for available land for development and agriculture. The increasing urbanization in the Delta could have possible effects on mangrove cover. Infrastructural development at the expense of mangrove regions in the delta threatens the biodiversity of this carbon rich ecosystem [92]. The replacement of mangrove regions in the Niger Delta could also be because of hydrographic modifications in the Niger delta due to dredging activities [49]. The widening and deepening of canals in creeks can modify the estuarine properties in which mangrove forests thrive.
We reported a likely 10-fold increase in Nipa Palm from 1441 ha in 2007 to 11,444 ha in 2017 with a range of a gain of 1641 ha to a loss of 18,787 ha. Nipa proliferation was greater in mangrove-cleared areas. The likely large increase in Nipa Palm was mostly at the expense of mangrove forest:~50% of the Nipa Palm area in our analysis was from the loss of mangrove cover in the region, especially in Akwa Ibom state. This non-native invasive species colonize structurally matured native species through disturbance or a pattern of penetrating mangrove stands (Figure 9). Our confidence intervals also show a possible reduction in Nipa Palm area (29%; 1641 ha) and possibly no change. These possible error in change detection could be a baseline for further investigation of invasive species in Niger Delta mangrove forests.

Caveats and Limitations
We encountered the problem of creating training classes for Nipa Palm classification, because Nipa Palm occurs in a strip along mangrove fringes, which only go some, meters inland. However, the fine resolution of the sensors (30 m) assisted as this could account for a larger coverage especially in areas of heavy colonization. Future research can take into account finer scale remote sensing products. Our results also show low accuracy in classifying Nipa Palm evident from the low Producer and User's accuracies (Table 3). Care should be taken when interpreting these results.
Cloud cover was also a limitation in using Landsat imagery for land cover analysis. Cloudiness had an effect on the clarity of optical data which is visible in our classification results over the Niger Delta (Figure 10), and scan line (SLC-off) error from the Landsat equipment movement affecting some of the regions during the classification [11]. Over the mangrove forests estimated in both years, there was no cloud influence in 2007 and 1.8% in 2017. Future regional classification of mangrove forests can make use of air-borne instruments on airplanes while Nipa Palm differentiation can make use of drones. These instruments will be able to account for finer optical characteristics of coastal vegetation that can improve accuracy.
We used image texture measures to increase the identification of major land cover classes especially agriculture lands and forests that spans large areas. Non-inclusion of these texture measures resulted in lower classification accuracies (Table 3). However, the use of texture measures probably reduced the map resolution, further hindering the accuracy of Nipa Palm classification which occurs as thin strips. Future land cover classification for identification of invasive species occupying thin strips could make use of finer scale resolution, which will improve the effectiveness of image texture.

Conclusions
Using a combination of EO data and ground control points, mangrove forests of the Niger Delta can be effectively monitored. We showed that mangrove forests in the Niger Delta cover a larger area, and stretch further inland, than was known previously. However, we also show that the area of mangrove forests is decreasing fast, through conversion to agriculture or urban areas, and through the proliferation of the invasive Nipa Palm.
Our best results involved direct observations and texture data from optical and SAR data, combined with a DEM, and analyzed using the SVM method. In general, our 95% confidence intervals, based on independent test data, were narrow enough for confident conclusions on area change to be drawn. However, classifying Nipa Palm poses a challenge, with no method producing reliable classifications with accuracies for this class over 32%. Our results therefore provide a foundation for monitoring the vegetation of the Niger Delta, but monitoring of Nipa Palm might require a higher resolution method.
The results have shown opportunities for monitoring and management of mangrove forests in the Niger Delta. Nipa Palm in the Niger Delta is a non-resourceful invasive species, which, if not checked, will eradicate the valuable mangroves. Improved mapping precision can target areas with high incidence of logging, population growth and economic activity in order to generate a mangrove vulnerability map in the Niger Delta. Our results show that Nigeria is one of the top five ranked countries in the world in terms of mangrove forest area, and the first ranked in Africa, but this may be rapidly changing due to adverse effects from development, logging and sea level rise. Utilization of remote sensing products, which can provide a baseline in spatial and temporal analysis, can be the first step in national mangrove forests monitoring, protection and restoration plans.