4.3. Experimental Results
Due to the small spectral difference between the pixels in the boundary part, the separability of red tide and seawater is weak, and it is difficult to accurately distinguish the complex boundary of red tide with small scale and dispersed distribution. However, red tides respond strongly in the red and near-infrared bands, and NDVI is introduced to enhance the characteristic information of red tides so that it is easier to distinguish the small-scale and dispersed banded boundary red tides. Red tide classification was compared between the original six-band GOCI data and the data with NDVI feature enhancement based on the original six-band data on the basis of the basic U-Net model. Use the training and test sets made in
Section 2.2.
Table 2 summarizes the detection accuracy of red tide with different characteristic data. As shown in
Table 2, after the introduction of NDVI feature enhancement on the basis of the original six-band data, the detection accuracy of each test set has been improved, in which the accuracy, precision, recall, and F1-score have increased by 1.52%, 2.19%, 2.19%, and 0.03, respectively. The experimental results also verified the difference between red tide NDVI and seawater NDVI. It has a significant detection effect on small-scale and dispersed banded red tides.
When the red tide breaks out, the blue-green band data in the remote sensing image will also have a relatively large change. On the basis of the traditional U-Net, the introduction of the ECA channel attention mechanism can effectively realize cross-channel interaction and give different weights according to the influence degree of different bands so that the model can pay maximum attention to the channel, which it needs to pay attention to, and fully explore the spectral characteristics of different channels to further extract red tide characteristics. It has a good recognition effect on the detection of complex boundary red tides with small-scale and dispersed strips, which can improve the classification accuracy of the detection model.
Table 3 shows the experimental results of feature enhancement with the introduction of NDVI on the basis of the original six-band features and summarizes the accuracy of red tide detection with the introduction of ECA channel attention mechanism and Atrous Spatial Pyramid Convolution (ASPC) on the basis of traditional U-Net. As shown in
Table 3, after the introduction of ECA on the basis of traditional U-Net, the detection accuracy of each test set is improved, among which the accuracy, recall, and F1-score are increased by 5.08%, 17.32%, and 0.12, respectively. The experimental results show that the attention mechanism of the ECA channel can effectively improve the detection effect of small-scale and dispersed banded red tide.
Atrous Spatial Pyramid Convolution (ASPC) is a multi-channel convolution with different expansion rates to expand the sensitivity field of the convolution kernel, aggregate the red tide feature information in a multi-scale context under multiple sampling rates, enhance the ability of the model to extract features of different scales, and better detect small and dispersed strips of complex boundary red tides, thus improving the detection effect of the model. On the basis of introducing the ECA channel attention mechanism into traditional U-Net, Atrous Spatial Pyramid Convolution (ASPC) is added. The common convolution blocks of the first two layers in the U-Net structure diagram are replaced with the ASPC-3 module, and the common convolution blocks of the third layer are replaced with the ASPC-2 module. Red tide pixels can be identified better via the effective fusion of spatial features of different scales, and higher detection accuracy is obtained. As shown in
Table 3, the accuracy of each test set has been improved, in which the accuracy, precision, recall, and F1-score have increased by 4.54%, 1.46%, 16.92%, and 0.10, respectively. The experimental results show that cavity convolution can significantly improve the detection efficiency of small-scale and dispersed banded red tide.
The improved U-Net method is compared with the basic U-Net method, full convolutional neural network (FCN), and support vector machine (SVM) by introducing NDVI feature-enhanced data from the original six-band data. The parameters of the full convolutional neural network used in this experiment are the same as those of the improved U-Net.
Table 4 summarizes the accuracy of these methods for red tide detection. Compared with the shallow feature extraction of SVM and FCN–8s, the encoding and decoding structure of U-Net’s jump connection can integrate features of different levels and obtain higher segmentation accuracy by classifying each pixel point. As shown in
Table 4, with the same data set and parameters, U-Net outperformed SVM and FCN–8s with an accuracy of 86.28%. The improved U-Net method introduces the ECA module and gives different weights according to the influence degree of different bands to distinguish and utilize the characteristic information of different bands conducive to red tide detection to the greatest extent. The common convolution blocks of the first three layers in the U-Net structure diagram are replaced with Atrous Spatial Pyramid Convolution (ASPC). The characteristic information of red tide in a multi-scale context is effectively fused under the condition of multiple sampling rates, and the detection accuracy of red tide is higher. As shown in
Table 4, the performance of the improved U-Net model is better than that of the basic U-Net model. The accuracy, precision, recall, and F1-score of the test set are increased by 9.62%, 1.48%, 34.24%, and 0.22, respectively. The experimental results show that the improved U-Net method is more suitable for red tide detection.
Figure 8 shows the visualization of L2A chlorophyll product data in GOCI data, the result of the original six band data based on the basic U-Net model, the result of NDVI feature enhancement based on the basic U-Net model, the result of NDVI feature enhancement and the result of the improved U-Net model (U-Net + ECA + ASPC). The visual detection results shown in the figure further validate that the improved U-Net model can effectively improve the detection effect of red tide, especially for the red tide boundary regions that are difficult to accurately detect (for example, the blue and red marked areas in the figure).
4.4. Method Applicability Analysis
In order to explore the applicability of the enhanced NDVI data and the improved U-Net method for red tide detection in other regions, the method was applied to Region 2. The experimental environment and parameter settings in Region 2 are the same as those in Region 1.
Table 5 shows the red tide detection accuracy of the basic U-Net model based on the original six-band features of GOCI and the introduction of NDVI feature enhancement based on the original six-band features. As shown in 4–4, on the basis of the original six-band data, the separation of spectral information of red tide and seawater in turbidity water is improved after the introduction of NDVI feature enhancement, and high detection accuracy is achieved, in which accuracy, precision, recall, and F1-score are increased by 1.45%, 18.54%, 4.90%, and 0.11, respectively. The experimental results also verified the difference between red tide NDVI and seawater NDVI and also had a significant effect on the detection of red tide in a large range of concentrated patches, which proved that NDVI features had strong applicability to the detection of different types of red tide.
Table 6 shows the comparative analysis results of the improved U-Net method with the basic U-Net method, FCN, and SVM after feature enhancement by NDVI in Region 2. As shown in
Table 6, with the same data set and parameters, U-Net outperformed SVM and FCN–8s with an accuracy of 87.56%. The performance of the improved U-Net model is better than that of the basic U-NET model, and the accuracy, precision, recall, and F1-score of the test set are increased by 4.49%, 2.07%, 8.34%, and 0.06, respectively. The experimental results show that compared with other methods, the improved U-Net method shows the best red tide detection effect, and the method is also suitable for the detection of large-scale block-distributed red tides in other areas.
Based on the traditional U-Net, the improved U-Net method introduces an ECA module to maximize the network’s attention to the channels to which it needs to pay attention, fully mining the spectral characteristics of different channels to further extract red tide characteristics. At the same time, the common convolution blocks of the first three layers in the U-Net structure diagram are replaced with Atrous Spatial Pyramid Convolution (ASPC). Under the condition of multiple sampling rates, the contextual red tide feature information is generated, and the red tide with complex boundaries can be better detected through the fusion of deep and shallow features and multi-scale spatial features.
Figure 9 shows the visualization of the actual red tide occurrence area, the predicted result with original six-band data, the predicted result with enhanced features, and that with the improved U-Net, respectively. This is further validated by the visual detection results shown, the improved U-Net can effectively detect the red tide, regardless of the concentrated distribution of red tide areas or the scattered distribution of red tide areas (such as the red marked area in region 2).