Red Tide Detection Method for HY − 1D Coastal Zone Imager Based on U − Net Convolutional Neural Network

: Existing red tide detection methods have mainly been developed for ocean color satellite data with low spatial resolution and high spectral resolution. Higher spatial resolution satellite images are required for red tides with ﬁne scale and scattered distribution. However, red tide detection methods for ocean color satellite data cannot be directly applied to medium–high spatial resolution satellite data owing to the shortage of red tide responsive bands. Therefore, a new red tide detection method for medium–high spatial resolution satellite data is required. This study proposes the red tide detection U − Net (RDU − Net) model by considering the HY − 1D Coastal Zone Imager (HY − 1D CZI) as an example. RDU − Net employs the channel attention model to derive the inter − channel relationship of red tide information in order to reduce the inﬂuence of the marine environment on red tide detection. Moreover, the boundary and binary cross entropy (BBCE) loss function, which incorporates the boundary loss, is used to obtain clear and accurate red tide boundaries. In addition, a multi − feature dataset including the HY − 1D CZI radiance and Normalized Difference Vegetation Index (NDVI) is employed to enhance the spectral difference between red tides and seawater and thus improve the accuracy of red tide detection. Experimental results show that RDU − Net can detect red tides accurately without a precedent threshold. Precision and Recall of 87.47% and 86.62%, respectively, are achieved, while the F1 − score and Kappa are 0.87. Compared with the existing method, the F1 − score is improved by 0.07–0.21. Furthermore, the proposed method can detect red tides accurately even under interference from clouds and fog, and it shows good performance in the case of red tide edges and scattered distribution areas. Moreover, it shows good applicability and can be successfully applied to other satellite data with high spatial resolution and large bandwidth, such as GF − 1 Wide Field of View 2 (WFV2) images.


Introduction
Red tides refer to harmful algal blooms (HABs) that constitute a marine ecological disaster resulting from excessive reproduction or accumulation of plankton, protozoa, or bacteria that cause water discoloration [1,2]. Red tide outbreaks are a threat to fisheries, marine ecosystems, and human health [3][4][5][6]. The dominant phytoplankton which caused red tide in China include dinoflagellates (such as Noctilucent scintillans [7], Prorocentrum donghaiense (P. donghaiense) [8], Alexandrium catenella [9]), and diatom (such as Skeletonema costatum (S. coatatum) [10]). In recent years, red tides have been found to occur more frequently as a result of eutrophication [11][12][13][14]. According to statistics, in 2019 alone, a total of 38 large−scale red tides occurred in China, resulting in direct economic losses of up to USD 4.86 million [15]. Automatic detection and monitoring of red tides is important for red tide prevention and reduction.
Red tide detection and monitoring methods include in situ surveys and remote sensing [16]. Owing to its wide range and short revisit period, remote sensing technology has played an important role in red tide detection and monitoring [17][18][19]. Multiple ocean color satellites, including the Coastal Zone Color Scanner (CZCS, NASA), Sea−Viewing Wide Field-of-View Sensor (SeaWiFs, NASA), Moderate Resolution Imaging Spectroradiometer (MODIS, NASA), Visible and Infrared Imager Radiometer Suite (VIIRS, NASA), Medium Resolution Imaging Spectrometer (MERIS, ESA), and Geostationary Ocean Color Imager (GOCI, South Korea), have been used for red tide detection and monitoring. However, the spatial resolution of such ocean color satellite data is too low (>250 m) to detect red tides with fine scale and scattered distribution [20]. Therefore, remote sensing images with medium-high resolution have been increasing in use in red tide detection [21]. In recent years, China has launched a series of remote sensing satellites with medium-high resolution. For example, HY−1D, China's fourth ocean optical satellite, was launched in June 2020. The CZI onboard HY−1D, with a spatial resolution of 50 m, wide swaths of 950 km, and a revisit period of 3 days, has been increasingly used for monitoring marine disasters including red tides.
Most existing red tide detection methods based on spectral features have been developed for ocean color satellites, such as the red tide index (RI) [22], P. donghaiense index (PDI), and diatom index (DI) [23], as well as a series of improved RI algorithms [24,25]. Of these, the RI is calculated by three bands at 443, 510, and 555 nm. PDI and DI utilize the 443, 510, and 555 nm bands and the 488, 531, and 555 nm bands, respectively. Studies have also attempted to use the chlorophyll−a concentration, bio−optical properties of seawater, or fluorescence line height (FLH) as alternative indicators of red tides [26][27][28]. However, due to without corresponding bands, these red tide monitoring methods cannot be directly applied to remote sensing satellite data with medium-high spatial resolution, such as HY−1D CZI images for limited bands (R, G, B, and NIR) and broad bandwidth. Some red tide detection algorithms have been developed for remote sensing satellites with mediumhigh spatial resolution. For example, the spectral shape at 490 nm from multiple sensor data, including Sentinel−2 MSI and Landsat OLI, was used by Shin et al. [29] to detect red tides of M. polykrikoides near the Korean Peninsula with high performance. Liu et al. [20] proposed a red tide detection algorithm (GF1_RI algorithm) for broad band satellite data with high spatial resolution using GF−1 WFV data, which was successfully applied to the red tide of Noctilucent scintillans detection and monitoring in Guangdong and Shandong provinces, China. These studies verified the red tide detection capability of images with medium-high spatial resolution with broad bandwidth. However, the aforementioned methods usually require a predetermined threshold to diagnose red tides effectively, and the selection of the threshold is strongly region−dependent and susceptible to the marine environment. For example, the threshold may be invalid when the image is contaminated by clouds, sun glint, or aerosols. Therefore, it is necessary to develop an automatic red tide detection method for images with medium-high spatial resolution without a threshold.
Deep learning [30] has been widely used in remote sensing applications such as remote sensing image classification [31][32][33], target recognition [34,35], and image fusion [36]. Owing to its powerful large data mining and feature extraction capabilities [37], deep learning has also been applied to red tide detection with good results. Most existing red tide detection methods based on deep learning use hyperspectral images [38,39], synthetic dataset based Inherent Optical Properties (IOP) combinations [40], and in situ data [41], which have rich red tide features and response bands. However, owing to the low spectral resolution, the spectral information of satellite data with medium-high spatial resolution is not as rich as that of hyperspectral data. As a nonlinear method with powerful data mining capability, deep learning is expected to extract more red tide features from satellite data with medium-high spatial resolution and detect red tide ranges more accurately. Considering the red tide of Noctilucent scintillans that occurred in the East China Sea in August 2020 which reported by the National Satellite Ocean Application Service (NSOAS) as an example, this study proposes an advanced red tide detection method based on the U−Net model, namely RDU−Net.
The remainder of this paper is organized as follows. Section 2 introduces the satellite data and related methodology. Section 3 presents the proposed method, namely RDU−Net, and the experimental results. Section 4 discusses the effects of the loss function coefficients and input characteristics on the proposed model. Finally, Section 5 concludes the paper.

Satellite Data
Noctilucent scintillans red tide occurred in the East China Sea in August 2020, as reported by the National Satellite Ocean Application Service (NSOAS) [42] (Figure 1a). The HY−1D CZI Level−1B data (Figure 1b) imaged during this red tide were acquired from NSOAS. The CZI characteristics are summarized in Table 1. Compared with ocean color satellites with spatial resolution >250 m, HY−1D CZI has an obvious advantage owing to its spatial resolution of 50 m. Furthermore, CZI has the advantages of wide swaths (950 km) and a short revisit period (3 days). Therefore, CZI has been widely used in marine monitoring.  To explore the applicability of the proposed method, GF−1 Wide Field of View (WFV) data were acquired from China's Center for Resource Satellite Data and Application (CRESDA) [43] in this study. As a representative Chinese high−resolution satellite, the GF−1 with four WFV cameras was launched in 2013. GF−1 WFV also has the advantage of high spatial resolution (16 m), wide scene swaths (800 km), and a short revisit period (4 days). Therefore, GF−1 WFV images are often used in marine monitoring. The sensor characteristics of GF−1 WFV [44] are summarized in Table 1. The GF−1 WFV2 Level−1 data are shown in Figure 1c. The detailed information of the two images is summarized in Table 2.

Dataset Construction
Considering the distribution characteristics of different objects in the study area, three areas were selected when constructing the training sample dataset, as shown in Figure 2. The true values ( Figure 2b) were obtained on the basis of visual interpretation. Moreover, considering the computer memory and speed, the input sample size of the model was set to 256 × 256. Therefore, the images and corresponding labels were randomly divided into sample images with a size of 256 × 256 pixels. A total of 680 samples were generated. Furthermore, three types of data augmentation were adopted to obtain sufficient training data and avoid overfitting, including horizontal flip, vertical flip, and diagonal flip. Therefore, a dataset containing 2720 samples was established and randomly divided into training (80%) and validation (20%) datasets.  [45] for the semantic segmentation of biomedical images. U−Net has been widely used in medical diagnosis [46,47], remote sensing [48], and other fields [49,50] as a pixel−wise prediction model that can be trained in an end−to−end manner. Owing to context−based learning, U−Net can propagate context information to higher resolution layers to build high level features for precise detection. Moreover, U−Net can support image segmentation with a few samples. Therefore, research on red tide detection has been carried out on the basis of the U−Net model. The structure of U−Net is mainly divided into an encoder and a decoder, which can be represented by a U−shaped diagram. Unlike a general convolutional network, U−Net does not have any fully connected layers and uses only the valid part of each convolution. The encoder path is also known as the contracting path. It is used to extract features from the input image following the typical architecture of a convolutional network. It consists of five blocks, each of which includes two convolutions (unpadded convolutions), an activation function, and a max pooling layer for down−sampling. In each block of the encoder, the receptive field size is increased, the output size is halved, and the number of feature channels is doubled. The decoder path is also known as the expansion path. It consists of up−convolutions and concatenations with features from the contracting path. Every block in the decoder path consists of an up−sampling, a concatenation with the correspondingly cropped feature map from the contracting path, and two convolutions, each followed by an activation function. The output of the decoder path consists of the category and localization, i.e., a category label (red tide or not) is assigned to each pixel.

Comparison Methods
For comparison, three convolutional neural networks (including U−Net), a machine learning method (SVM), and a spectral ratio method (GF1_RI) are selected in this study. The details of each algorithm are as follows: (1) Fully Convolutional Neural Networks (FCN) The fully convolutional neural network (FCN) [51] has achieved good results. It adopts the skip architecture to combine semantic information from a deep, coarse layer with appearance information from a shallow, fine layer in order to produce accurate and detailed segmentations. A series of networks developed on the basis of FCN have confirmed its strong generalization ability [52]. FCN includes three networks: FCN−32s, FCN−16s, and FCN−8s.
FCN−8s is an advanced version of FCN−32s and FCN−16s, and it outperforms FCN−16s and FCN−32s in image segmentation. Thus, FCN−8s was selected for comparative experiments. Furthermore, FCN−8s does not have fully connected layers. The classification at the image label is extended to the pixel label by modifying fully connected layers into convolutional layers.
(2) SegNet SegNet [53] is a deep network for image semantic segmentation, which is similar to FCN. SegNet was designed for scene segmentation with high learning speed and accuracy using a highly unbalanced dataset [46]. The dataset used in this study was also unbalanced. Hence, SegNet was chosen for comparison. The architecture of SegNet consists of encoder and decoder processes. The encoder process involves image compression and feature extraction using the rectified linear unit (ReLU) [54] during activation. Upon its completion, the decoder process restores the image. Image spatial information is maintained during the decoder process, as image restoration is performed using the same pooling layer as that in the encoder process. This feature of SegNet distinguishes it from FCN. When the image reconstruction is complete, the image is classified using a softmax function [47]. Compared with the corresponding structure of FCN, the volume of SegNet is much smaller mainly because of the operation performed by SegNet to balance the computational effort: the location information of the recorded pooling process is used instead of a direct deconvolution operation.
(3) SVM SVM is a kernel−based supervised classification algorithm proposed by Cortes and Vapnik [54]. It has been shown to be a powerful machine learning algorithm for pattern recognition and nonlinear regression, and it is mainly used for classification, regression, and time−series prediction [55,56]. The basic idea of SVM is to find the optimal hyperplane that can correctly partition the training dataset with the maximum geometric spacing, as follows: where ϕ(Xi) is the nonlinear mapping function, and Wi and b are the linear support vector regression function parameters. The common kernel functions in SVM include the linear kernel, polynomial kernel, and Gaussian kernel [57]. Among them, only the Gaussian kernel needs to adjust its parameters to achieve the best performance. The parameters of the Gaussian kernel function include the penalty factor c and kernel width γ. Smaller values of c and γ will increase the training error, while larger values will increase the complexity of the model and lead to overfitting. In this study, the optimal parameters of the model were c = 2.0 and γ = 4.0.
(4) GF1_RI GF1_RI is a red tide detection method for GF−1 WFV images proposed by Liu [20] according to the spectral characteristics of red tides. The clouds, turbid water, and construction regions in the image are first masked according to the spectrum of blue and green bands; then, the red tide area is detected using the GF1_RI index as follows: where L2, L3, and L4 represent the radiance values of the green, red, and NIR bands, respectively.

Accuracy Evaluation
The performance of the proposed method was evaluated using four criteria, including Precision, Recall, F1−score, and Kappa. Precision is the ratio of red tide pixels accurately detected to all red tide pixels identified. Recall represents the percentage of pixels classified as red tides that are correctly classified. F1−score is an indicator of the accuracy of the red tide detection method, which combines precision and recall. The kappa coefficient is a type of accuracy evaluation index used to measure the consistency of the red tide detection results with greater objectivity. The calculation methods are given below: where TP and FP denote the true positive and false positive, i.e., the number of red tide and non−red tide pixels identified as red tide, respectively; FN denotes the false negative, i.e., the number of red tide pixels identified as non−red tide; and TN denotes the true negative, i.e., the number of non−red tide pixels correctly identified as non−red tide.

RDU−Net Model Framework
Owing to limitations in terms of the bands and spectral resolution, red tide detection for satellite images with medium-high resolution requires strong feature representations, especially for red tides with fine scale and scattered distribution. In addition, U−net is not effective in extracting detailed information of red tides [58], such as the red tide boundary. Therefore, RDU−Net is proposed by incorporating channel attention and improving the loss function on the basis of traditional U−Net. Furthermore, some layers are also added in RDU−Net to improve the training performance. The batch normalization (BN) layer [59] can achieve a stable distribution of activation values throughout training, thereby accelerating the model convergence and increasing the model capacity. Moreover, to make the output image size consistent with the input image size, the same strategy is adopted to replace the valid part in the padding. In addition, as the model training in this study is based on a small training set, a dropout layer is added before up−sampling to avoid overfitting. This design retains the information lost by down−sampling to a certain extent.
The RDU−Net model framework is shown in Figure 3.

Channel Attention Model
Red tides have different spectral characteristics in different bands (channels). Mining the inter−channel features can facilitate red tide information detection. However, the convolutional layer extracts feature by blending cross−channel and spatial information together [60,61]. In recent years, the attention model has also been widely used in deep learning networks to improve the performance. Channel attention can automatically obtain valuable features by learning and exploring the inter−channel relationships. Therefore, this study uses channel attention to generate the channel attention map by exploring the characteristic relationship between channels, and greater attention is paid to the red tide formation between different channels.
The channel attention module integrates MaxPool and AvgPool to enrich the extracted high−level features and considerably improve the representation capability of the network. The channel attention framework is shown in Figure 4. The channel attention module can be expressed as follows: where F FM denotes the feature map from the convolution block and F CA denotes the channel attention map; F MP and F AP represent the feature map obtained by the MaxPool and AvgPool operations, respectively; σ denotes the activation function; and ⊕, ⊗ denote element addition and multiplication, respectively. First, two feature maps are obtained by applying the MaxPool and AvgPool operations to the input features. Second, the two maps are input to the multi−layer perceptron (MLP) with one hidden layer. Third, the output feature vectors are merged and activated by the element−wise summation and activation function. Finally, the channel attention map is obtained by multiplying the output feature vectors by the feature map obtained by the convolution block.

Boundary and Binary Cross Entropy Loss Function
For unbalanced red tide samples, i.e., when the number of positive (red tide) and negative samples (background) varies considerably, using only the loss function based on the regional integrals can affect the performance of model training and red tide boundary detection [62]. The boundary loss function can alleviate the difficulty arising from unbalanced red tide samples. Therefore, to achieve red tide edge detection with greater accuracy and clarity, the boundary and binary cross entropy (BBCE) loss function, which combines the boundary and region information, was chosen in this study.
The boundary loss function [63] is proposed by calculating the distance between the boundaries in a differentiable form: (8) where G and S are two nearby regions; ∂G and ∂S are boundaries of G and S; ∆S represents the region between ∂G and ∂S; and D G (q) denotes the distance between any point q and the nearest point on ∂G.
Subsequently, the sum of the region integrals represented by the boundary−based level set φ G is calculated, and the final loss value is obtained by outputting the binary variables S θ (q) through the softmax probability of the network as follows: (9) In this study, the BCE loss function is used to measure the loss value of the region as follows: where w is a hyper−parameter, and x and y are the predicted result and true label, respectively. By incorporating the boundary loss, the loss function L BBCE makes the training consider the accuracy of the boundary: where α is an arbitrary weight belonging to [0, 1].

Data Preprocessing
Data preprocessing includes geometric correction, feature combination, and data normalization. In this study, we used ENVI 5.3 software to perform geometric correction.
High−resolution remote sensing images contain more intricate details and texture information. However, owing to the limitation of band configuration, it is difficult to distinguish features (such as turbid water) similar to red tides. Moreover, the separability of the spectral information of red tides and seawater is weak at the red tide edges. From the spectral curves of the selected samples, it can be seen that the red tides respond strongly in the red and NIR bands. Therefore, NDVI was introduced to enhance the spectral information of red tides. NDVI is calculated as: where R(NIR) and R(R) denote the radiance values of the near−infrared and red bands, respectively. As shown in Figure 6b, the statistics results show that the NDVI of red tides is significantly different from that of other objects such as seawater. Therefore, NDVI is taken as the input feature in this study. The average NDVI values were all less than 0, among which that of seawater was the lowest with the minimum variance, followed by red tides. By using the multi−feature multispectral data, which combines the original multispectral data (four bands) and NDVI, high level information can be extracted. The spatial information of seawater is suppressed, and the spectral differences between red tides and seawater are increased. To unify the dimensions of multiple features, the maximum and minimum value normalization method [64] was adopted to normalize the gray value of the preprocessed image and samples to [0, 1]. The algorithm is given by: (13) where x * represents the normalized vector, x is the sample spectral vector, and x max and x min are the maximum and minimum values of the sample spectral vector, respectively.

Splicing Method Based on Ignoring Boundary
Splicing method based on ignoring boundary is used to avoid the obvious stitching marks. Following steps were performed to achieve the detection of the red tide.
Step 1: Calculate the step size of image cropping by setting the ratio of the splicing area to the input area. The formula is as follows. s = n × OL_ratio (14) where s represents the step size during image cropping, OL_ratio represents the ratio of the splicing area to the input area, and n represents the width of the image block. In this paper, the OL_ratio is set to 50%, and n takes 256 pixels.
Step 2: The whole image was cropped into a series of image blocks with a size of n × n according to the s. Then, a dataset is created using these image blocks. These image blocks have a specific overlapping area with adjacent images.
Step 3: Input the dataset created in Step 2 into the trained model for prediction.
Step 4: The detection results were spliced. The middle part of the adjacent predicted image blocks is spliced except for the surrounding area.

Results
The experiments were performed on a computer with an Intel Core i5−10300H CPU (2.50 GHz) and 16.0 GB of physical memory, using PyCharm 2019 (Python) as the programming environment. For the RDU−Net model, which was implemented on the basis of Keras, the input size was 256 × 256, the number of channels was 5, and the output was a single−channel classified image of the same size. The α of the loss function is 0.8, the initial learning rate of the model is 10 -4 , the batch size is 32, the optimizer uses Adam [65], and the number of iterations is 100 times.
We compared our method with the five methods introduced in Section 2.3, namely three fully convolutional neural networks (U−Net, FCN, and SegNet), the traditional machine learning method (SVM), and GF1_RI. The parameters of the fully convolutional neural networks used in this experiment were the same as those of RDU−Net. Table 3 summarizes the accuracy of these methods. As can be seen in Table 3, the proposed method achieved high detection accuracy. The Precision and Recall were greater than 86%, while the F1−score and Kappa were 0.87. This is further corroborated by the detection results shown in Figure 7. The proposed method could effectively detect the red tide pixels, whether they were in concentrated or edge areas or surrounded by scattered clouds. Compared with the other five methods, the precision and F1−score were improved by 6.14%-19.98% and 0.07-0.21, respectively. In addition, with the same training set and parameters, U−Net outperformed FCN−8s and SegNet, with Precision of 81.33%. Thus, the U−Net framework is more suitable for red tide detection. Furthermore, as can be seen from Figure 7, the proposed method has obvious advantages for red tide detection in scattered and edge distributed areas. Specifically, the six methods work well for the concentrated area of red tide distribution, such as Area−A in Figure 8. However, RDU−Net achieves the best performance in the red tide edges or areas surrounded by scattered clouds. For example, as shown in Area−B and Area−C in Figure 8, a large number of red tide pixels could not be identified in the methods used for comparison, especially SVM and GF1_RI. In addition, even though U−Net extracts more red tide information, the red tide edges are not as clear as those in the case of RDU−Net, such as Area−C in Figure 8. This is because RDU−Net includes channel attention and focuses on the boundary loss, which can train the samples more adequately and detect the red tides more accurately in the edge areas compared to the traditional U−Net method. Compared with SVM and GF1_RI, the deep learning method considerably reduces the probability of classifying clouds and seawater as red tides. However, some red tide pixels cannot be detected in the case of FCN−8s and SegNet mainly because U−Net has a large number of feature channels in the up−sampling part, which allows the network to propagate context information to higher layers. Figure 8 confirms the red tide detection accuracy of the methods listed in Table 3. These findings demonstrate the superiority of the proposed method in the red tide detection of broadband HY−1D CZI data.

Discussion
In this section, the effects of the loss function coefficients and input characteristics on the model are discussed. In addition, the red tide detection capability of the proposed method from multi−source images with high spectral resolution is demonstrated using a GF−1 WFV2 image.

Sensitivity Analysis of Loss Function Parameters
The loss function in RDU−Net is a combination of boundary loss and BCE loss, which should be set during training. Therefore, sensitivity analysis of α was carried out experimentally. Table 4 summarizes the red tide detection precision (F1−score) with α ranging from 0.1 to 1. The results show that appropriate inclusion of the boundary loss in the BBCE loss function can improve the red tide detection accuracy, and the highest F1−score is obtained when α = 0.8. However, when α is less than 0.8, the F1−score decreases with α. Therefore, 0.8 is selected to construct the loss function in this study. In addition, as can be seen in Table 4, the red tide detection will be affected when α is too small: when α is less than 0.4, the F1−score is lower than that when using the BCE loss function.

Analysis of Multi−Feature Effect on Red Tide Detection
To evaluate the effect of using NDVI in the model, two datasets, including four bands (R, G, B, and NIR) and multiple features (R, G, B, NIR, and NDVI), are used as the input of the RDU−Net network. The red tide detection accuracy using the different datasets is summarized in Table 5. The results show that the four metrics are improved using the multi−feature dataset. The Precision of the multi−feature dataset was 87.47%, i.e., 7.93% higher than that of the four−bands dataset. Compared with the multi−feature dataset, the Recall of the four−bands dataset was higher than the Precision.
The red tide detection results of the two datasets are shown in Figure 9. The detection results indicate that most of the red tides can be detected with the two datasets, especially in the concentrated area of red tide distribution. However, as shown in Figure 10 (Area−B and Area−C), the red tide range detected using the multi−feature dataset is more accurate in the case of the red tide edges or cloud−covered areas. It can be seen from the local view that, consistent with Table 5, the red tide range obtained by the multi−feature dataset is more accurate.

Applicability Analysis of Rayleigh Correction
It should be noted that the RDU−Net is designed based on the radiance value. Accurate atmospheric correction is difficult to perform on the broad band sensors, such as HY−1D CZI, because they do not have all bands needed for atmospheric correction [66]. The Rayleigh scattering by atmospheric molecules accounts for the largest proportion of the atmospheric contribution [67]. Existing research has shown that the Rayleigh−corrected reflectance can be accurately calculated. In this paper, to explore the applicability of Rayleigh correction, the Rayleigh−corrected reflectance (R rc ) was calculated using the predetermined look−up table of Tong et al. [68] (Figure 11).
As shown in Table 6, the red tide detection accuracy using the R rc image is close to that of the radiance image. Red tide detection result based on the R rc image (Figure 12b) also shows that the RDU−Net is suitable for the Rayleigh correction image. Compared with the Rayleigh correction product, the Radiance product is more readily available and simpler to handle. Therefore, RDU−Net was designed using the Radiance product in this paper.

Method Applicability Analysis
To explore the applicability of the red tide detection method to other high−resolution satellite images, the method was applied to a GF−1 WFV2 image. Red tide, seawater, and lots of clouds are also covered in GF−1 WFV2. The total number of samples of GF−1 WFV2 image after data augmentation (including horizontal flip, vertical flip, and diagonal flip) was 1940, of which 1820 were used for training and 120 were used for verification. The experiment environment and parameter setting of RDU−Net were consistent with those of the HY−1D red tide detection experiment. The α of the loss function is 0.8, the initial learning rate of the model is 10 -4 , the batch size is 32, and the number of iterations is 100 times. Table 7 shows the red tide detection accuracy for the GF−1 WF2 image. The Precision and Recall are greater than 85%, and the F1−score and Kappa are 0.86. From Figure 13, we can see that the proposed method can also effectively detect red tides in the GF−1 WFV2 image. The proposed method has been experimentally shown to have strong applicability, and it can be applied to other high−resolution broadband satellite data with the same band settings.  Unlike to the HY−1D CZI image in this work, the GF−1 WFV2 image has a large area of thin clouds and fog. The radiance and NDVI values of the thin clouds and fog are intermediate between seawater and thick clouds. Thin clouds and fog will be detected as red tide when RDU−Net trained with HY−1D CZI images is used for red tide detection on GF−1 WFV2. Therefore, it is necessary to consider the distribution of red tide in more scenarios when constructing the training sample dataset.

Conclusions
Using an HY−1D CZI image, this study developed a red tide detection framework based on RDU−Net for satellite images with medium to high spatial resolution. The RDU−Net method focuses on the red tide feature relationship between different channels using the channel attention model and detects more precise boundaries by introducing the boundary loss function. A multi−feature dataset including four bands and NDVI was selected as the input of the model to improve the separability of red tides, seawater, and clouds. The experimental results showed that the proposed method outperforms the existing methods in terms of the accuracy of red tide detection. The red tide detection accuracy of the proposed method reached 87.47%, and the F1−score was 0.87, representing improvements of 6.14%-19.98% and 0.07-0.21, respectively, over the method used for comparison. RDU−Net has obvious advantages in the detection of red tide edges and cloud−covered areas. In addition, this method can be applied to other high−resolution remote sensing images, such as GF−1 WFV images.
This study explored the application of deep learning and remote sensing with mediumhigh spatial resolution to red tide detection using broadband data. Owing to its high performance, the proposed method provides a new reference for developing red tide detection methods. As the training dataset constructed in this study only includes HY−1D data, it is necessary to reselect the sample when it is applied to other high-resolution images. In the future, a larger training dataset will be built using satellite data with medium-high resolution from multiple sources and different areas for detecting red tides in various marine environments. The proposed method is expected to yield better performance with sufficient training samples.