Application of Hyperspectral Image for Monitoring in Coastal Area with Deep Learning: A Case Study of Green Algae on Artificial Structure

Abstract

Remote sensing is a powerful technique for classifying and quantifying objects. However, the elaborate classification of objects in coastal waters with complex structures is still challenging due to the high possibility of class mixing. The classification through the hyperspectral images can be a reasonable alternative to problems related to such precise classification work because it has high spectral resolution over a wide bandwidth. This study introduced the results of the case study using a novel method to classify green algae on an artificial structure based on hyperspectral data and deep-learning models. The spectral characteristics of the attached green algae on the artificial structure were observed using a ground-based hyperspectral camera. The observed image had a total of three classes (concrete, dense green algae, and sparse green algae). A certain area of the image was used as learning data to create classification models for three classes. The classification models were created from one machine-learning (support vector machine, SVM) and two deep-learning models (convolutional neural network, CNN; and dense convolutional network, DenseNet). As a result, the performance for the classification results of green algae predicted from two deep-learning models was higher than that of the machine-learning model. Additionally, the deep-learning model successfully classified the interface area between concrete and green algae. This study suggests that the combination of hyperspectral data and deep learning could enable more precise classification of objects in coastal areas.

1. Introduction

The monitoring or mapping of objects in coastal areas is generally performed through remote sensing, ship-based surveys, and field observations, among which remote-sensing technology is the most suitable for detecting and classifying green algae over a large area [1]. In recent decades, satellite imaging systems with multi-spectral sensors have shown a capability to monitor many types of algae in the global ocean area [1,2,3,4,5]. Rashid and Yang (2018a) used the Geostationary Ocean Color Imager with 500 m resolution to detect green tides and analyze its movement characteristics [3]. Rashid and Yang (2018b) detected macroalgae using Landsat-8 satellites with 30 m resolution [4]. Min et al. (2019) identified and classified green and golden algae using multiple satellites [5]. However, due to the relatively low resolution of those satellites, satellite-based technology is not suitable for coastal areas where precise classification is required. Kudela et al. (2015) detected a floating green tide using a MODIS/ASTER Airborne Simulator sensor with a resolution of approximately 20 m [6]. Dierssen and Russell (2015) analyzed the spectral characteristics of floating seagrass using an airborne Portable Remote Imaging Spectrometer with 1 m spatial resolution [7]. Recently, studies on the detection and classification of various algae have been conducted in coastal waters using drones and ultra-high-resolution RGB images [8,9,10,11,12,13]. Oh et al. (2016) used a drone-mounted RGB camera to detect red tides around an ocean farm [8]. Duffy et al. (2018) and Ventura et al. (2018) classified seagrass meadows in coastal waters and underwater through machine-learning techniques, using RGB data [9,10]. Although the detection and classification of algae or seagrass areas have been successfully performed using cropped images of ultra-high-resolution data, the sophisticated classification of pixels is still challenging in the coastal region due to the distribution complexity of biophysical characteristics. Furthermore, most of these studies focused on the detection and classification of floating or underwater vegetation, and there is no established method to classify the distribution of vegetation on artificial structures.

Hyperspectral sensors provide hundreds of spectral bands with a wavelength interval of less than 10 nm, providing additional spectral information in every pixel than RGB or multi-spectral data [11,12]. Since the 2000s, the development of hyperspectral sensor technology onboard aircraft and satellites has led to improved spatial classification technology for land and ocean using remote-sensing techniques [13,14,15,16,17,18,19,20]. As hyperspectral images provide spectral and spatial information, they are beneficial for classifying crops on land [13,14,15]. Specifically, Han et al. (2020) proposed improved classification performance in agriculture using a novel method based on a multi-scale convolutional neural network (CNN) [16]. Rupasignhe et al. (2018) conducted a study to extract vegetation on the shoreline using airborne hyperspectral images [17]. In addition, this characteristic of hyperspectral images is used to classify the sediment types on tidal flats [18,19,20,21]. For example, Pyo et al. (2019) quantified the cyanobacteria concentration in inland water using hyperspectral images and deep learning [22]. As such, a sophisticated classification technique using hyperspectral can be used in various fields in coastal waters, and in this study, green algae attached to artificial structures were used as a target.

Various image-processing techniques, such as object-based classification methods, have been developed to process and extract useful information from a broad spectral range [21,22]. Kim and Yeom (2012) performed a land cover classification using multi-threshold and multiresolution segmentation techniques. Shin et al. (2015) conducted a study to improve land cover classification accuracy using a multiresolution segmentation method. However, object-based image segmentation is limited depending on the applied weight or threshold value by subjective decision. Therefore, in this study, the classification of hyperspectral images was performed using the deep leaning techniques, which minimize user intervention. In the case of CNN and dense convolutional network (DenseNet), it is possible to classify images more effectively than conventional methods, because image classification is performed by considering both spectral and spatial information [23,24].

In this study, we introduce a technique to classify the green algae on concrete artificial structures using hyperspectral images and the deep leaning techniques. A 2D grayscale image, which included both the spectral and spatial information of the target pixel, was generated for CNN and DenseNet model learning. The classification accuracy of both models was verified using the results of manual classification. In addition, a comparison with a support vector machine (SVM) was performed to compare the model performance.

2. Study Site

Green algae were observed near a breakwater in the southeastern part of Yeong-Do, South Korea. This region is a bay enclosed by Young-Do, Jo-Do, and the breakwater (Figure 1a,b). There is an artificial concrete staircase with nine steps that allows people to access the water for marine swimming training. The habitat characteristics of algae vary depending on the intensity of sunlight exposure. The primary forcing that determines this factor is tidal variation. Tides create sea level variations in a 14-day spring–neap cycle. Algae located in the lower part of the steps are less affected by sunlight, while those toward the upper part are more exposed to sunlight. In the study site, the artificial structure is influenced by tides with a tidal range of 0.38–1.08 m, causing the water to cover up to the fourth step during spring tide, whereas only the first step is submerged during neap tide. Consequently, red, brown, and green algae form zonation within the first-to-fourth steps due to the tidal variation (Figure 1c–f).

3. Data Description

3.1. Hyperspectral Camera

The hyperspectral camera is an instrument that effectively detects and classifies objects from remote-sensing images. This technology provides unique spectral information about the target material [11,12]. The hyperspectral camera used in this study was a PIKA-II, developed by Resonon, which obtains 240 spectral bands with a wavelength range of 400–900 nm (

Table 1. Resonon’s hyperspectral camera (PIKA-II) specifications.

Item	Specification
Wavelength range	400–900 nm
Spectral channels	240 bands
Spatial channels	640 samples
Spectral resolution	2.1 nm
Weight	1.3 kg
Dimensions	9.7 cm × 16.8 cm × 6.4 cm
Pixel size	7.4 µm

). This instrument was securely installed using additional portable devices, including a tripod, power supply, and notebook computer (Figure 2).

The hyperspectral camera was fixed to the tripod, and the lens was installed at an angle of 90° to the ground (Figure 2c). To determine the radiation correction required to calculate an irradiance from the surface, the solar irradiance was measured through a white Spectralon (Figure 2a). A dark correction was performed for the noise removal of the image by covering the camera with a lens cap (Figure 2b). Radiation correction was applied to the observed values through the data-processing program provided by Resonon. And the reflectance values were calculated from this program. No additional atmospheric corrections were performed.

The most compelling characteristic of hyperspectral image data is that there are several hundred spectral data for each pixel. This not only enables a detailed analysis of specific objects but also enables the generation of a large amount of information through a single observation. Although it is difficult to obtain the data compared to the RGB image, the classification using the hyperspectral image ensures sufficient learning data with only one observation and permit precise classification in pixel units.

3.2. Observation of Green Algae

The observation of the green algae was performed through the hyperspectral camera (Figure 2). The observed time is on April 19, 2018, at 20:26, UTC (11:26 LT). The hyperspectral image was collected at a latitude of 35° 4.493′ N and a longitude of 129° 5.015′ E, with a resolution of 640 × 500 pixels. There were no clouds, and the wind speed and direction were approximately 4.4 m/s and 201.6° (www.weather.go.kr; accessed on 11 November 2024), respectively. The first and second stairs, covered with red and brown algae, were submerged in seawater (Figure 3a), and the layers of green algae were exposed to the atmosphere; however, some areas were covered by seawater.

The image obtained from the hyperspectral camera showed a clear distinction between concrete and green algae. To generate the learning data of SVM, CNN, and DenseNet, a region of interest (ROI) image was selected from the full hyperspectral image (white dotted line in Figure 3a). The ROI image contained concrete, dense green algae mat, and sparse green algae mat (Figure 3b). The concrete was distributed on the left side of the ROI, and a dense and sparse green algae mat was distributed from the center to the right part of the image.

The three classes observed from the hyperspectral image were labeled for all pixels through manual inspection and digitization (Figure 3c,d). Manual inspection involves the classification of all pixels in Figure 3a into three classes, after which pixel coordinates are stored together by class. This process allows reflectance values to be extracted for each class. In this study, data in Figure 3c were used to evaluate the possibility of monitoring green algae using hyperspectral images.

3.3. Hyperspectral Data

To remove spatial variations in reflectance by the observational environment, the reflectance obtained from the hyperspectral camera was normalized using the following equation:

norm(x^(i,j)_k) = {x^(i,j)_k − min(x^(i,j) _k)}/{max(x^(i,j) _k) − min(x^(i,j) _k)}
(1 ≤ i ≤ R, 1 ≤ j ≤ C, 1 ≤ k ≤ L)

(1)

where R, C, and L are the number of rows, columns, and band layers of the hyperspectral data, respectively. Since the data vary from pixel to pixel, the normalization should be performed pixel by pixel.

Figure 4 shows the normalized reflectance from the PIKA-II hyperspectral camera. The reflectance of dense green algae was compared with the results of spectrometer provided by Min et al. (2019) [5]. The spectral information of dense green algae obtained from PIKA-II and the spectrometer showed a significantly similar pattern, with a simultaneous peak near the wavelength of 550 nm (green and green dash lines in Figure 4). Through these spectral characteristics obtained from the manual inspection of green algae, we confirmed two things. The first is that the observations and corrections in this experiment reflect the spectral characteristics of green algae. Furthermore, it shows that pixels containing green algae were correctly extracted through the manual inspection process. The second is that the increasing and decreasing trend of the reflectance appears distinctly in all classes. Therefore, the spectral characteristics of these three classes can be clearly distinguished in the input data of classification models.

3.4. Preprocessing for Classification Dataset

For the successful classification of three classes through SVM, CNN, and DenseNet, the appropriate production method of the input data for each class must be considered. The input data for the SVM simply requires the spectral information for the pixels in each class. However, the input data for CNN and DenseNet requires several processes to convert every pixel into a separate 2D grayscale image. The method of extracting spatial and spectral information from each pixel of the hyperspectral image and converting it into a 2D gray scale image suitable for the input data of the CNN and DenseNet is presented in detail in Figure 5.

First, hyperspectral data for each pixel were normalized using Equation (1). After normalization, neighborhood pixels around the target pixel using 5 × 5 windows were extracted for all band layers (Figure 5a,b). Through this process, a total of 240 metrics of 5 × 5 size centered on the target pixel were extracted. Then, each matrix of size 5 × 5 was changed to a 1D vector of length 1 × 25. This 1D vector was stacked in the order of L dimension (1–240 bands), resulting in creating a 1D vector of length 1 × 6000 (Figure 5c). The spatial and spectral vector information of each pixel was transposed into a 2D matrix of length 80 × 75 (Figure 5d). This 2D matrix includes three spectral bands in one column. The first column of the matrix contains values for bands 399, 401, and 403, while the last column contains values for 886, 889, and 891 nm. As shown in Figure 5d, the spectral characteristics of the dense green algae, for which reflectance increases with increasing wavelength, were expressed as a change in brightness with a grayscale. Finally, it was converted into the 32 × 32 input image of deep-learning models (CNN and DenseNet) using the MATLAB ‘imresize’ function, with the ‘bicubic’ option (Figure 5e). Each waveform image contains textural information for the spatial and spectral data of the target pixel.

Through the above process, a total of 149,818 grayscale input data were created for all pixels of the three classes included in Figure 3c. Among all input data, 17,515 data within ROI (Figure 3d) were used as learning data for the classification model. Learning data were divided into 80% training set and 20% test set. Each model created from the learned results was used to predict the classes of entire area in Figure 3c. Additionally, all input data were used as criteria evaluating the accuracy of three classification models (SVM, CNN, and DenseNet). The three classification models are described in detail in Section 4.

4. Classification Models

4.1. Support Vector Machine (SVM)

In general, since the SVM algorithm is sensitive to the size of the training set, it uses a method of training many classifiers in a small training set to consider model execution efficiency. Considering the characteristics of the high-resolution hyperspectral image, a one-versus-one multiple classifier was used rather than a simple binary classification [26]. The library used the SVM classifier provided by scikit-learn [27].

4.2. Basic Convolutional Neural Network (CNN)

The CNN is an artificial neural network with multiple convolution layers in front of a traditional neural network. It is one of the representative models of deep learning, widely used in various fields, such as object classification and detection in images [23,28,29]. CNN consists of several convolution layers, pooling layers, and fully connected layers. The convolution layer serves to extract features from an input layer. It consists of filters that extract features and an activation function that converts the value of filters into nonlinear values. When extracting the characteristics of each class for image classification, not all features can be used as classification criteria because the calculation time required for classification increases in proportion to the number of features. Therefore, it is necessary to select a few features, while minimizing the degradation of classification performance, so that max pooling is performed to reduce the extracted activation map. The feature maps for spatial–spectral characteristics are classified into each class after being inputted into a fully connected layer.

In this study, the basic CNN model, which was used to classify hyperspectral images, consisted of one input layer, two convolution layers, two max-pooling layers, two fully connected layers, and one output layer. The convolutional layers were chosen to effectively extract important spatial features from the images. The filters used in each convolutional layer learn to capture specific patterns and textures within the input data, highlighting useful information from the images. The max-pooling operation aids in reducing the dimensionality of the feature maps while preserving essential information. By applying max pooling, we compress spatial information and reduce the computational load, thus increasing the training speed of the model and helping to prevent overfitting. These two operations play a crucial role in maximizing the performance of the network and are essential for the classification of hyperspectral images. Figure 6 shows the structure of the utilized basic CNN model.

For input data with a 32 × 32 size, convolution layers use filters of 3 × 3 size to extract characteristic values. At these points, the stride values are one, but the size of the output through zero-padding is also maintained at 32 × 32. The filters of 3 × 3 size effectively capture spatial relationships between neighboring pixels, which is crucial for the classification of single-band images. Although the filters were not specifically selected based on the spectral characteristics of algae, the model architecture allows for the integration of spatial features with spectral information in subsequent layers. This design aims to ensure that the network generalizes well across different classes, including algae.

The activation function used ReLU [29], and the number of filters for each convolution layer was allocated as 32 and 64, respectively. For the two pooling layers, we chose to use a max-pooling method that takes a significant value and has a stride value of 2 with 2 × 2, so the size of the output is half that of the other convolution layers. Each pixel was then classified into four classes using the fully connected layer, based on the previous activation map. The size of the input image and the value of the hyperparameters were constrained by the specifications of CPU used in the experiment and the results of repeated experiments, respectively.

4.3. Dense Convolutional Network (DenseNet)

However, in the basic CNN model structure, as the number of layers comprising the network increases, there is a side effect that causes the problem of feature information and gradient vanishing or degrading. In ResNet, they have demonstrated that the layers close to the input layer or the output layer are more accurate and more efficient to train [30], and they proposed a shortcut connection that improves the existing model. The DenseNet was created by accepting this theory [24], and all layers of DenseNet are connected to other layers in a feed-forward format. In the case of ResNet, the previous layer results are added to the subsequent layer, while DenseNet is in the form of concatenating the previous layer results. The advantages of this are solving the problem of degradation, enhancing feature propagation, and using more effective parameters. DenseNet performs feed-forwarding by concatenation between feature maps. At this time, if the feature map size changes, calculation is impossible, and the size of the feature map decreases with each pooling layer. Therefore, in this study, the DenseNet network was composed of dense blocks made by grouping layers with the same feature map size.

Additionally, we also used the DenseNet-201 model consisting of four density blocks and transition blocks connecting density blocks to classify hyperspectral images. Figure 7 and

Table 2. DenseNet-201 architecture, including dense blocks and their sublayers.

Layers	Output Size	DenseNet-201
Convolution	16 × 16	7 × 7 conv, stride 2
Pooling	8 × 8	3 × 3 max pool, stride 2
Dense Block 1	8 × 8	$\left[\begin{array}{c}1\times 1 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\\ 3\times 3 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\end{array}\right]$ × 6
Transition Layer	8 × 8	1 × 1 conv
	4 × 4	2 × 2 average pool, stride 2
Dense Block 2	4 × 4	$\left[\begin{array}{c}1\times 1 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\\ 3\times 3 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\end{array}\right]$ × 12
Transition Layer	4 × 4	1 × 1 conv
	2 × 2	2 × 2 average pool, stride 2
Dense Block 3	2 × 2	$\left[\begin{array}{c}1\times 1 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\\ 3\times 3 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\end{array}\right]$ × 48
Transition Layer	2 × 2	1 × 1 conv
	1 × 1	2 × 2 average pool, stride 2
Dense Block 4	1 × 1	$\left[\begin{array}{c}1\times 1 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\\ 3\times 3 \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}\end{array}\right]$ × 32
Classification Layer	1 × 1	7 × 7 global average pool,
		fully connected

show the structure of the utilized model. The input data were 32 × 32 sizes. The first layer changes the size of the channel through a convolution layer that changes the size of the feature map. The composite function of DenseNet is combined with a batch normalization (BN), ReLU, and a 3 × 3 convolution layer, which increases computational complexity because feature map is accumulated. To reduce this, a 1 × 1 convolution layer called the bottleneck layer was added in front. The transition block consists of 1 × 1 convolution and average pooling to down-sample the feature. After the last dense block, the classification is performed through the FC Layer by flattening and global average pooling.

4.4. Classification Based on Deep-Leaning Models

In this study, we classified concrete, and dense and sparse green algae mat using the basic CNN and DenseNet models with hyperspectral data. Figure 8 describes the classification process of hyperspectral images using the models. Of the total 17,515 pixels for the ROI image, 14,012 (80%) pixels were used as training data, and 3503 (20%) were used as test data. In the both cases, the batch size was 1000, and if the accuracy was not improved after five epochs, the learning rate was reduced by half, and the lowest value was reduced to 0.001. When using the deep-learning model, the appropriate epoch size significantly impacts the model’s performance. For example, if the epoch is too small, the model will show a low classification performance because it is insufficiently trained. We set up a sufficiently large epoch to confirm this problem. We compared the accuracy while increasing the epoch size up to 5000, and we were able to confirm the accuracy that converged even in a relatively small epoch. As a result, 1000 epochs were repeatedly learned, and the optimizer used was Adam.

To perform a quantitative comparison of our results with a representative machine-learning technique, the SVM process was chosen. Deep-learning and SVM technologies are commonly used for the classification of sediment types in coastal water, and they have been verified for a high-accuracy classification performance of over 90% [19,22,31].

5. Results

5.1. Learning Results in ROI

Figure 9 shows the classification results of the three classes trained by SVM and DenseNet in the ROI area. Note that the training and testing results are shown together to compare the detailed classification performance of machine learning and deep learning. The basic CNN model result was very similar to that of DenseNet, so it was not shown in the figure. The concrete classes were well classified in both models. However, there were differences in performance between the two models in certain areas.

SVM classified the pixels labeled as dense green algae from manual inspection as sparse green algae in boxes F-1 and F-2 (Figure 9c). This means that the spectral characteristics of sparse green algae were partially mixed in boxes F-1 and F-2. However, SVM showed definite potential for misclassification in boxes E, G, and I. In the case of box E, the interface between concrete and dense green algae was misclassified as sparse green algae. Additionally, some pixels in box G were misclassified as concrete, and thin concrete pixels in box I were not detected at all. On the other hand, DenseNet showed a better classification performance than SVM in boxes E, G, and I (Figure 9d). DenseNet clearly classified the interface between two materials in box E. The sparse green algae were well classified in box G, and the concrete in box I was accurately detected.

5.2. Predicted Results in Entire Area

Three AI models generated through the training process for ROI area were applied to the entire area of the hyperspectral image. Figure 10 shows the classification results of the three classes predicted from the SVM and DenseNet models in the entire area. The three models predicted the overall pattern well, but the difference in classification performance between machine- and deep-learning models was more pronounced in the entire area than the difference in ROI area.

The SVM still tended to misclassify the interface between concrete and dense green algae as sparse green algae (Figure 10c). In addition, the problem of the SVM misclassifying concrete as sparse green algae in box I of Figure 9 was significantly highlighted in box J of Figure 10c. The SVM did not detect approximately 8% of the concrete areas correctly. On the other hand, DenseNet performed better than SVM in classifying concrete. DenseNet also did a better job than SVM in classifying the obvious dense green algae. For example, the SVM misclassified the distinct dense green algae at K in Figure 10c as sparse green algae. Furthermore, DenseNet detected the dense green algae well at L in Figure 10d better than SVM. Therefore, the deep-learning model was better than the machine-learning model at detecting objects with precision.

Table 3. Classification accuracy for predicted classes from three AI models.

Classification Models			Manual Inspection Class (Pixels)			Accuracy
			Concrete	Dense Green Algae	Sparse Green Algae
SVM	Predicted class (pixels)	Concrete	78,907	9	1017	90.22%
		Dense green algae	12	29,146	1168
		Sparse green algae	7274	5167	27,118
CNN		Concrete	83,133	126	1866	91.48%
		Dense green algae	302	32,660	6171
		Sparse green algae	2758	1536	21,266
DenseNet		Concrete	83,133	126	1866	92.80%
		Dense green algae	302	32,660	6171
		Sparse green algae	2758	1536	21,266

shows the accuracy of the classification results for the SVM, CNN, and DenseNet, respectively. The values in tables indicate what classes classified by manual inspection were actually classified by the AI models. For instance, in the SVM results for concrete, 78,907 pixels were detected well as concrete, but 12 and 7274 pixels were misclassified as dense and sparse green algae, respectively. In other words, the higher the matrix components that correspond to [1, 1], [2, 2], and [3, 3] in the table for each model result, the higher the accuracy.

The accuracy for the classification results of the three AI models was all calculated higher than 90%. Of these, the accuracy of SVM had the lowest at 90.22% and DenseNet had the highest at 92.80%. Overall, the models that classified concrete and dense green algae well were CNN and DenseNet. The SVM had a relatively high number of pixels (>12,000) that misclassified concrete and dense green algae as sparse green algae. However, the SVM had a high number of pixels (>27,000) that correctly classified sparse green algae than did CNN and DenseNet. Nevertheless, the percentage of pixels in concrete and dense green algae classified well by CNN and DenseNet was higher than the percentage of pixels in sparse green algae classified well by SVM. Therefore, these results showed that the accuracy of deep-learning models was higher than that of the machine-learning model.

6. Discussion

6.1. Comparison Between Machine- and Deep-Learning Models

The machine-learning and deep-learning models performed well, with classification accuracy of over 90% in detecting green algae on an artificial concrete structure. However, even though all models showed high classification accuracy, there was a clear difference in the classification ability of machine-learning and deep-learning models in some areas. These characteristics were identified in the areas marked E, F, G, I, J, K, and L in Figure 9 and Figure 10.

In the area of the interface between two materials, such as box E in Figure 9, adjacent pixels could be contaminated, and the characteristics of the spectrum could change. Additionally, some pixels in box G in Figure 9 have spectral characteristics similar to concrete due to strong reflectance from sunlight. These problems can lead to various misclassifications in the machine-learning model, such as SVM with input data of only one spectral information for one pixel. Indeed, SVM tended to misclassify the interface area between concrete and dense green algae as sparse green algae (Figure 9c and Figure 10c), and it misclassified sparse green algae as concrete in box G in Figure 9c. However, because the input data used in the DenseNet model include a spatial variation in the spectrum at the surrounding pixels, it can perform classification with high accuracy at areas such as boxes E and G (Figure 9d).

In addition to the misclassification problems mentioned above, three classes with similar spectral characteristics (Figure 4) could lead to additional misclassifications. Sparse green algae not only had a similar spectral pattern to dense green algae but also contained spectral information of concrete due to its relatively transparent characteristic. Areas marked E, F, J, and K in Figure 9 and Figure 10 illustrate these misclassifications of SVM. On the other hand, DenseNet successfully classified concrete in boxes E and J, and it classified dense and sparse green algae better than SVM in the areas marked K and L in Figure 9 and Figure 10. Therefore, the classification accuracy of deep-learning models (CNN and DenseNet) was calculated to be higher than that of the machine-learning model (SVM) due to the input data of the deep-learning model, which has the characteristic of sharing the spectral information of neighbor pixels.

6.2. Comparison Between Hyperspectral, Multispectral, and RGB Data

The most significant disadvantage of hyperspectral data in remote-sensing technology is the complexity of the data acquisition and management. The primary difference between hyperspectral, multispectral, and RGB images is the number of spectral bands. RGB images are represented using only the blue, green, and red bands (red area in Figure 4). Multispectral images include about 4-to-10 bands, incorporating the RGB bands (blue area in Figure 4). In contrast, hyperspectral images consist of hundreds of spectral bands, providing more detailed information for each pixel. As a result, object identification using hyperspectral data has a significantly higher potential for accurately classifying various objects than multispectral and RGB data.

To evaluate whether the performance of hyperspectral data is superior to that of multispectral and RGB data in classifying green algae, we resampled the bands of hyperspectral data to align with the bands of multispectral and RGB data (refer to the blue and red areas in Figure 4). Due to the inability to convert the input data for multispectral and RGB into the 2D grayscale image, we assessed performance based on the SVM algorithm. As a result, the detection accuracy of green algae predicted using hyperspectral data was approximately 2–4% higher than that predicted using multispectral and RGB data (Figure 11). These results indicate that hyperspectral data could better classify green algae on artificial structures. Therefore, this study emphasizes that using hyperspectral data can yield better results in accurately detecting and classifying green algae.

6.3. Application Plan

Recently, coastal waters have been widely used for tourism and leisure sports. Over 90% of all marine accidents in South Korea occur due to the lack of systematic safety management of accident-causing risk factors [32]. Slipping accidents are the second most common type of marine accident [32], accounting for 26.8% of accidents along shore rocks in coastal activity incident statistics in South Korea from 2011 to 2015 [33]. Moreover, the rate of deaths resulting from falls was 32.28% [34]. These statistics show that such accidents frequently result in severe wounds or even death [33]. The main cause of falling accidents is inebriation or slipping on algae attached to shore rocks or artificial structures. In the case of slippage due to algae, the risk of slipping is proportional to the distribution density of the algae. Therefore, if the algal distribution can be calculated spatially through the method proposed in this study, it could be possible to distinguish places with a high risk of slipping.

Although the accuracy difference between the results of deep learning and machine learning is not significant, this study showed that deep learning is more effective than machine learning in classifying the interface between green algae and other substances. In future applications of algae detection using hyperspectral images over large areas, it is crucial to clearly distinguish the interfaces between the two substances due to reduced spatial resolution. In such situations, object detection using deep learning could serve as an alternative for effectively separating ambiguous regions between the two substances.

The conventional method for classifying high-resolution images using deep learning depends on the cropped image size, and the novel approach proposed in this study could classify pixel units. However, the experiment was conducted only at a single site, so more observation experiments are required to apply to other coastal areas. Therefore, we plan to conduct additional observational experiments on various surfaces, such as rock and huge coastal areas. For example, the method is expected to be used in classifying the spatial distribution of green algae in a broader area, using an unmanned air vehicle (UAV) or the airborne system. Nevertheless, using the method proposed in this study is expected to be useful for the elaborate classification of objects in coastal waters with complex structures.

For our research to be practically applied in the future, several improvements are necessary. It is essential to consider and appropriately correct elements that may introduce noise into the images, such as light, shadows, and water reflectance. Additionally, to effectively apply the model in various environments, it is important to collect diverse training data, including not only concrete but also seawater, vegetation, sand, and rocky substrates. Continuous validation of the model results through real-time monitoring of green algae in port areas is also crucial for ensuring reliability. Finally, integrating the model results with an automated warning system is necessary to enable the rapid identification and response to potential hazards. Through these improvements, we anticipate that our research will significantly enhance its practical applicability in the future.

7. Conclusions

In this study, a quantitative evaluation of technology for monitoring green algae on coastal artificial structures was conducted by applying machine-learning and deep-learning models to hyperspectral images. One machine-learning model (SVM) and two deep-learning models (CNN and DenseNet) were used to classify three classes (concrete, dense green algae, and sparse green algae) observed from the hyperspectral camera. The input data for the two deep-learning models was designed to reflect the spectral and spatial characteristics of each pixel. Finally, a quantitative comparison and evaluation for the classified results obtained through machine- and deep-learning models were performed.

The most obvious difference between the results of machine-learning and deep-learning models was most pronounced at the interface of concrete and green algae. The machine learning misclassified the interface area, while the deep learning successfully classified two classes at the interface area. Deep learning also classified dense green algae and sparse green algae, which have similar spectral characteristics, more accurately than machine learning. These results stemmed from the input data of the deep-learning model, which is characterized by the target pixel sharing information with its neighboring pixels. As a result, the classification accuracy of the deep-learning model was higher than that of the machine-learning model.

Although classification through deep learning exhibited a high performance, there still exist misclassified pixels. This means that there is a need for continuous improvement in the preprocessing method of the input data. In addition, the observation in this study was conducted only at a single site, so more observations are required to apply to other coastal areas. Nevertheless, the classification technology through deep learning proposed in this study suggests that it is a novel method that can effectively classify the spatial distribution of green algae. In the future, it is expected that the likelihood of marine accidents caused by green algae will be further reduced if the above-mentioned limitations are gradually improved.