1. Introduction
Coastal wetlands are characterized as remarkable and crucial systems located in the transition zone between terrestrial land and ocean [
1,
2]. They perform a wide range of ecological functions including reducing coastal erosion, preserving environment health, removing nutrient and suspended solid, adjusting local climate, providing habitats and refuges for fish and wildlife [
3,
4,
5]. Unfortunately, dramatic degradation in the area of coverage, ecological value and biodiversity has occurred in the worldwide coastal wetlands over the past decades, which severely increases the vulnerability of coastal wetlands [
6,
7]. The main reasons come from global sea-level rise [
8,
9] and inappropriate anthropogenic activities [
10]. Therefore, it is an urgent work to map spatial distribution and dynamic changes of land covers in coastal wetlands, in order to restore and protect natural resources and ecological services in coastal environments.
The complicated and vulnerable ecosystems of coastal wetlands make the traditional manual investigation could not work in realistic applications, because many coastal wetlands are inaccessible for humans and field measurements are difficult and time consuming [
11,
12]. Satellite remote sensing techniques provide an important and effective data source for mapping and monitoring coastal wetlands, because of its unique characteristics in an easy data acquisition, spatially continuous coverage and short revisiting periods [
4,
13,
14]. During the last several decades, the fast development of satellite remote sensing sensors has greatly enhanced our capability to map coastal wetlands, that is, panchromatic sensors [
13,
15], Synthetic Aperture Radar (SAR) [
16], multispectral sensors [
17,
18] and even hyperspectral imagers [
19,
20]. Conventional panchromatic sensors could not discriminate different species in coastal wetlands (e.g., vegetation) because of their single limited band information [
21]. SAR has strong penetration but its data processing work is complicated and its inversion of biomass is more limited than optical remote sensing [
5]. Satellite multispectral images like Landsat TM/ETM+, Gaofen-1 WFV, Sentential-2A and advanced spaceborne Thermal Emission and Reflection Radiometer (ASTER), have more advantages in monitoring coastal wetland habitats, because of their moderate spatial resolution, wide spatial coverage and multiple optical bands [
22,
23]. They could identify main types of land covers, that is, water, farmland and woodland but could not distinguish vegetation communities, for example, different subtypes of swamps. Accordingly, it is a big challenge to achieve a fine classification map of coastal wetlands. Combing multispectral sensors with other sensors like SAR and Lidar might integrate respective advantages of these sensors and it has shown some promising results in mapping coastal wetlands [
24,
25,
26]. On the other hand, the imaging spectrometer could provide numerous narrow spectral bands at each pixel [
27]. The high spectral resolution (<10 nm) of hyperspectral images renders that it owns greater potentials in identifying different ground objects of coastal wetlands with subtle spectral divergence [
28,
29]. Some hyperspectral sensors have been utilized in mapping coastal wetlands, for example, Earth Observing-1 (EO-1) Hyperion [
20], Hyperspectral InfraRed Image (HyspIRI) [
19], Compact High Resolution Imaging Spectrometer (CHRIS) [
30] and Hyperspectral Imager for the Coastal Ocean (HICO) [
31].
The divergences in spectral characteristics within multispectral and hyperspectral images could reflect different ground objects [
32]. Accordingly, researchers tried to utilize the spectral information to identify different land covers of coastal wetlands. For example, the work of spectral analysis on five wetland classes by Meisam Amani show that Near-Infrared (NIR), Red Edge (RE) and red bands are top three most useful bands for separating different wetland classes, whereas Shortwave Infrared (SWIR) and green bands have poorer separability [
33]. Elhadi Adam analyzed the ASD field spectrometer data and found that there are statistically significant differences in spectral reflectance between papyrus and its co-existing species on 412 wavelengths located in different portions of the electromagnetic spectrum [
34]. René R. Colditz explored 14 spectral indices and their appropriate thresholds for water mapping in coastal wetlands and told us that MODIS green (band 4) and short wave infrared (band 6) “MNDWI6” perform the best of all [
35]. Moreover, supervised classifiers are usually employed to make the pixel-based classification on spectral signatures, for example, Statistical classifiers like maximum likelihood classifier (MLC) [
13] and machine learning like support vector machine (SVM) [
36] and random forest (RF) [
37]. Reschke combined multi-temporal Landsat imagery with high resolution satellite data and adopted the RF algorithm to extract subpixel information of coastal wetland classes [
38]. Petropoulos proposed a semi-automatic classification based on SVM to classify different spectral responses of Landsat TM images and quantify the deposition and erosion in two Mediterranean wetland deltas [
39]. Pierre Dubeau investigated the potential of optical remote sensing images for mapping the primary vegetation groups in the Dabus Wetlands using the non-parametric RF classifier [
40]. However, the pixel-based classification result is always limited by the complicated landscape with diverse land-cover types and their small patch sizes [
41]. Moreover, spectral variations in spatial locations make that different species in coastal wetlands might have similar spectral signatures [
27]. The single source of spectral information accordingly could not distinguish mixed species composition of wetlands and often leads to comparatively low mapping accuracies and class confusions of specific classes, for example, paddy and swamp, croplands and shrublands.
The idea of object-based image analysis was then introduced into classifying remote sensing images of coastal wetlands [
42]. It considers the spatial location and context of homogeneous pixels and integrates multiple spatial features to promote the classification accuracy. The object-based classifier groups objects that have similar spatial features, for example, a similar pixel shape, color or texture and classify them based on the object features. Berberoglu extracted the vegetation information using Grey Level Co-occurrence Matrix (GLCM) and explored the land cover change dynamics of a Mediterranean coastal wetland [
43]. Jia adopted an object-oriented classification method on Landsat data to map the spatial distribution of mangrove areas in coastal wetlands of China and obtained accurate results [
44]. Wang explored the potentials of completed local binary patterns (CLBP) in extracting texture feature and applied it to classify vegetation in Yancheng coastal wetlands of China [
45]. Unfortunately, determining proper spatial-object features is time consuming and the performance of object-based classifier might diverge greatly across different features selected.
Furthermore, it will be natural to wonder whether one can combine spectral information with spatial-object information to upgrade the classification result, because they are complementary to each other. The idea of spatial-spectral based classification has been investigated in remote sensing image processing, where spatial features are extracted from the original images and stacked with spectral signatures to carry out classification. Many popular algorithms can be used to extract spatial features, for example, extended morphological profiles (EMP) [
46], attribute profiles (AP) [
47] and Markov random fields (MRF) [
48]. However, the above schemes treated remote sensing data as a regular image and the good performance highly relies on the complicated parameter tuning work of spatial operators. Moreover, they did not involve the essential expert rules extracted from spectral variables of specific land cover types. For example, the brightness index (BI) from Tasseled Cap transformation can be well distinguish built-up areas from non-built up areas.
Inspired from the decision tree, the purpose of our study is to present a simple and effective framework for mapping coastal wetlands from multispectral and hyperspectral images. The decision tree does not require assumptions concerning the frequency distribution of spectral data in each land cover type [
49]. We then would like to involve the information of expert rules of spectral variables and spatial-object information into the framework and promote the classification of mapping coastal wetlands. We implement multispectral data from Landsat TM/ETM+/OLI sensors because of their typical features of free to access, long archived history and widely used status [
50]. On 9 May 2018, China successfully launched the Gaofen-5 (GF-5) satellite, which takes an advanced hyperspectral sensor. The GF-5 advanced hyperspectral imaged (AHSI) has 330 spectral bands ranging from 400 to 2500 nm and is designed to provide users hyperspectral data with higher spectral resolution and larger swath width. Unfortunately, mapping capacity of GF-5 hyperspectral data has not been investigated in classifying coastal wetlands. Therefore, our study focuses on the following two key objectives: (1) to develop a hierarchical classification framework (HCF) for mapping coastal wetlands from satellite multispectral and hyperspectral images; and (2) to investigate the mapping capabilities of Landsat and GF-5 remote sensing data in classifying different land covers of coastal wetlands.
Compared to previous work, our paper favors three scientific contributions. First, the idea of two-level HCF is different and advantageous to current methods. In the first level, the implemented expert rules of spectral indices for 4 rough classes and the structure of decision tree are specifically designed and different from current methods. In the second level, the spectral, geometric and texture features are carefully chosen for coastal wetland mapping. Second, our proposed HCF is easy to implement and flexible in mapping different land covers in coastal wetlands, which works well on either multispectral or hyperspectral images. Third, we are the first to demonstrate that the GF-5 hyperspectral data have clear advantages in coastal mapping than multispectral images like Landsat and Sentinel data.
4. Experimental Results
We design two groups of experiments on two study areas in order to comprehensively testify the classification performance of proposed HCF. The first group is to verify the classification behaviors of HCF on the multispectral data by comparing it with two-state-of-the art methods SVM and RF. The second group is to investigate the performance difference of HCF in mapping coastal wetlands with different source data, that is, GF-5 hyperspectral, Sentinel-2 and Landsat multispectral data. Classification accuracy is quantified as kappa coefficient (KC), overall accuracy (OA), producer’s accuracy (PA) and user’s overall (UA).
4.1. Classification Comparison between SVM, RF and HCF on Landsat Data
In this experiment, we use three images (Landsat-5, Landsat-7 and Landsat-8) of both Yellow River Estuary and Yancheng coastal wetland to testify the classification performance of HCF. The classification results of HCF is compared with those of SVM and RF on the two study areas. To make a fair comparison, the training and testing samples of each Landsat data are independently selected.
4.1.1. Yellow River Estuary
The Yellow river estuary comprises of 13 land cover types including sea, river, built-up, dry land, paddy field, ditch, aquaculture, pond, mudflat, marsh, vegetation, bare land and salt marsh. Using the field survey with global positioning system (GPS) and with the help of high spatial resolution images, we obtain the region of interesting (ROI) of ground truth for all land cover types. The training and testing samples are then randomly selected from the ROI and their detailed information is listed in
Table 3. To ensure a fair sampling condition, the number of training samples is almost three times that of testing samples.
Table 4 shows the OA and KC of three classification methods on the Yellow River Estuary Landsat data. The SVM on three periods has OA of 88.7%, 84.33% and 81.37% and KC of 0.85, 0.8 and 0.76, respectively. The RF has OA of 90.46%, 88.77% and 94.19% and KC of 0.82, 0.85 and 0.93, respectively. The OAs of HCF in the three periods is 93.63%, 93.01% and 95.97% and the KCs are 0.91, 0.87 and 0.94, respectively. HCF behaves better than SVM and RF, having higher OA and KC in the three periods of Landsat data. Meanwhile,
Figure 3 illustrates the UA and PA from each method. More than eight classes of HCF has over 85% PA and UA in Landsat-7 data, the RF has eight classes exceed 85% and the SVM only has six classes with over 85% PA. The HCF, SVM and RF classification maps of landsat-5 have slightly worse UA and PA than those of the Landsat-7 data. Only seven classes of HCF have the PA and UA over 85% on the Landsat-5 data, RF has six classes with over 85% PA and only four classes (sea, aquaculture, salt marsh) has higher PA and UA in the SVM classification map. The HCF performs best on the Landsat-8 data, regardless of UA or PA and ten classes of HCF has PA over 90%. The next is RF and eight classes have over 85% PA and UA in Landsat-8 classification map, whereas there were only five classes in the SVM classification map. Misclassifications exist mostly in dry land of 2005, the paddy field of 2010 and the ditch of 2015.
Moreover,
Figure 4 shows the classification maps of HCF, SVM and RF in the three periods. Land cover types change significantly over time and a lot of land was developed and utilized. The classification results are consistent with the original image. In general, the visual performance of the two methods was roughly the same in each period. A lot of coastal land was developed and utilized from 2005 to 2015. The visual behaviors of classification maps of HCF, SVM and RF are roughly the same in each period. Both methods misclassify salt marshes into marsh on the Landsat-7 data. For the Landsat-5 data, the SVM and RF misclassify salt marsh into river and HCF misclassifies pond into river. Both methods misclassify part of sea into ditch and the SVM brings about a higher misclassification rate on Landsat-8 data. Overall, the HCF is superior to SVM and RF in classification on the Yellow River Estuary Landsat data. And HCF achieves the best result on the Landsat-8 data, followed by Landsat-7 data.
4.1.2. Yancheng Coastal Wetland
Yancheng coastal wetland consists of 12 land cover types including sea, river, buildings, dry land, ditch, aquaculture, pond, mudflat, marsh, vegetation and fallow land.
Table 5 lists the training and testing sample information of Landsat data in the Yancheng coastal wetland. To ensure a fair sampling condition, the number of training samples is almost three times that of testing samples.
Table 6 shows the OA and KC of both methods on the Landsat data of Yancheng coastal wetland. The RF on three periods has OA of 78.76%, 78.39% and 72.41% and KC of 0.75, 0.72 and 0.65, respectively. The SVM has OA of 83.42%, 78.81% and 89.9% and KC of 0.77, 0.73 and 0.86, respectively. The OAs of HCF is 93.46%, 90.1% and 93% and KCs are 0.92, 0.86 and 0.91, respectively. The HCF performs better than SVM and RF on all three Landsat datasets and achieves the best classification accuracy on the Landsat-7 data. The PA and UA of HCF in
Figure 5 also supports the above observations. Specifically, HCF has six classes of land covers with over 85% PA and UA in Landsat-7 data, whereas RF and SVM have only four classes. In contrast with Landsat-7 data, all the three methods have slightly worse PA and UA on the Landsat-5 data. The HCF has only four classes with over 85% PA and UA in Landsat-5 data, whereas RF has four classes and SVM has three classes. The HCF classification accuracies on Landsat-8 data perform best of all. It has seven classes with over 90% PA and UA. Moreover,
Figure 6 shows the classification maps of HCF, SVM and RF on Yancheng coastal wetland. Land covers changed greatly in the past decade (2001–2015), especially in the first five years. The HCF behaves better than SVM and RF in distinguishing paddy field from dry land and it could identify buildings more accurately, especially on the Landsat-8 data. Therefore, HCF achieves better classification accuracy than SVM and RF in mapping coastal wetlands. HCF achieves the best results on Landsat-8 data, followed by Landsat-7 data, which is consistent with the results in the Yellow River Estuary area.
4.2. Classification Comparison between GF-5, Landsat-8 and Sentinel-2 Data
This experiment investigates the difference in classification performance of HCF on GF-5 AHSI, Landsat-8 and Sentinel-2 data. To better demonstrate the capability of different data, we carry out two classification experiments on the Yellow River Estuary and identify 13 coarse classes and 21 fine classes of land covers respectively.
Table 7 shows the OA and KC of 13 land cover types, which include marsh, paddy, dry land, bare land, aquaculture, river, ditch, sea, vegetation, pond, building, mudflat and salt marsh. The OA and KC of the GF-5 AHSI data are 94.98% and 0.93, the Sentinel-2 data are 93.4% and 0.92 and Landsat-8 data are 92.34%and 0.90. GF-5 AHSI data have higher OA and KC than Sentinel-2 and Landsat-8 and Landsat-8 is the worst of three datasets.
Figure 7 plots the PA and UA of HCF on GF-5 hyperspectral data, Landsat-8 and Sentinel-2 multispectral data. The Landsat-8 and Sentinel-2 have more than eight classes with over 75% UA, whereas GF-5 has more than ten classes.
Figure 8 shows classification map of HCF on the three datasets. The classification behavior of three data is close to each other, where GF-5 has slightly better classification results. In the Landsat-8 classification map, HCF misclassifies sea into mudflat and paddy into dry land. Dry land and buildings, aquaculture and marsh, pond and aquaculture and marsh and vegetation are confused in the classification map, particularly vegetation are easily misclassified into paddy and marsh. The Sentinel-2 data misclassifies mudflat into buildings. GF-5 data misclassifies sea into pond and paddy into vegetation but it has less misclassification rate than Landsat-8 and Sentinel-2 data. In general, the GF-5 data are slightly more advantageous than Landsat-8 and Sentinel-2 data in classifying 13 land cover types.
To further explore different capabilities of GF-5, Landsat-8 and Sentinel-2 data in mapping coastal wetlands, we increase land cover types into 21 classes on the Yellow River estuary data. The 21 classes are shallow sea, deep sea, river, buildings, locust, maize, broomcorn, soybean, rice, flood plain, mudflat, aquaculture, pond, reed, suaeda salsa, spartina, tamarix chinensis, intertidal saltwater marshes, freshwater herbaceous marshes, emergent vegetation and salt marsh. As illustrated in
Table 7, OA and KC of GF-5 AHSI data are 97.38% and 0.97, the Sentinel-2 are 86.15% and 0.84 and Landsat-8 are 77.81 and 0.74. GF-5 AHSI data has higher OA and KC than Landsat-8 and Sentinel-2 in classifying 21 classes of land cover types. The PA and UA plots in
Figure 7 tell that Landsat-8 data has twelve and seven classes with over 80% of UA and PA, Sentinel-2 data has twelve and fourteen, whereas GF-5 has twenty-one and nineteen classes. Especially, 95% of land cover types have over 90% UA and PA in the GF-5 classification map. Moreover,
Figure 9 shows classification maps of GF-5, Landsat-8 and Sentinel-2 data. The shallow sea and deep sea, broomcorn and soybean is confused in Landsat-8 classification map. Large areas of Mudflat are misclassified into ponds, spartina, intertidal saltwater marshes and aquaculture are misclassified into ponds. In contrast, fewer misclassification cases exist in GF-5 classification map, with only shallow sea partly misclassified into ponds. In general, GF-5 classification map is the best in both PA and UA. Therefore, we conclude that the GF-5 data has great advantages in fine classification for mapping coastal wetlands.
4.3. Discussions
Considering existing problems in classifying and mapping coastal wetlands, we propose the HCF and design two groups of experiments on two study areas of coastal wetlands. The first group of experiments compares the classification performance of HCF with those of SVM and RF on the Yellow River estuary and Yancheng Landsat images. The HCF has OA above 90% on all Landsat datasets, while the OA of SVM and RF is below 90%. The HCF has lower misclassification rate than of SVM and RF and its UA and PA are obviously superior to those of two comparison methods. Therefore, that HCF achieves better classification accuracies than SVM and RF methods. The SVM and RF are two typical classifiers of machine learning and it owns excellent performance in land cover classification. Unfortunately, land covers in coastal wetlands are complicated and diverse and many different land covers have similar spectral responses with each other. Two classifiers take spectral signals of all samples into the training process and the similar spectrum of different land covers might negatively impact the result of trained classifier. In contrast, our proposed HCF is a hierarchical framework. The first level distinguishes four main classes of land covers using the designed decision tree and the second level combines spectral features, texture features and geometric features and implements SVM to further divide four main classes into fine subtypes of land covers. The first level of HCF adopts expert rules of spectral variables into the decision tree and could reduce confusion or misclassification between different land covers in main classes. The second level adopts the spatial-object information and could well identify different subtypes of land covers in the same main classes. Particularly, in the second level, the classification of each main class is an independent and non-interfering process and it does not require too many training samples, thus reducing the sample asymmetry and promoting the distinguishability among different main classes.
The second group of experiments investigates the classification performance of GF-5 AHSI data, Landsat-8 and Sentinel-2 multispectral data in mapping coastal wetlands using our HCF. When having 13 land cover types, the OA and KC of Landsat-8 are 92.34% and 0.9, Sentinel-2 are 93.4% and 0.92, while those of GF-5 are 94.98% and 0.93. When having 21 land cover types for fine classification, Landsat-8 and Sentinel-2 data have poor behaviors and their OA and KC are reduced to 77.81% and 0.74, 86.15% and 0.84 respectively. In contrast, the OA and KC of GF-5 data do not degrade but increase significantly into 97.38% and 0.97. Therefore, the proposed HCF is also suitable in hyperspectral data and GF-5 data is much advantageous than Sentinel-2 and Landsat-8 in mapping coastal wetlands. Hyperspectral data has the characteristics of numerous bands and narrow spectral range. It accordingly provides approximately continuous spectral information of ground objects and greatly improves the capability in identifying different land covers. GF-5 AHSI data have a spatial resolution of 30 m, same with the Landsat data but it has clear spectral advantages to the Landsat-8 data. When the number of land cover types for classification is small, the spectrum width of Sentinel-2 and Landsat-8 multispectral data could separate different classes. It accordingly has similar classification performance to GF-5 data in classifying the 13 classes of the Yellow River estuary. But when having too many land cover types, the spectral information of Sentinel-2 and Landsat-8 data is too coarse and could not identify some specific similar land cover subtypes (e.g., shallow sea and deep sea). But narrow bands of GF-5 data provide sufficient spectral information, which ensures that it could work well in identifying similar land cover types with subtle spectral divergences. Therefore, compared with Landsat and Sentinel data, the GF-5 data own great advantages in finely mapping coastal wetlands.
Our work in this paper still has some drawbacks and needs further improvements. First, we do not carefully investigate the explanation of selected thresholds in the first level of HCF. Whether the thresholds in the decision tree is reasonable is closely related to the classification accuracy. In the future work, we will try to theoretically explain the selected threshold. Second, the quality of samples is an important factor affecting the classification accuracy. The classification maps still have mixed and misclassified land cover types. Therefore, how to optimize the sample selection is another issue worth exploring in the future. Finally, more satellite remote sensing datasets and deep learning classifiers like cube-pair-based convolutional neural networks (CNN) [
71] and CNN-Based Spatial Feature Fusion Algorithm (CSFF) [
72] will be adopted in our framework to further promote the classification behaviors of our HCF.