In this study, a DJI Phantom 4 RTK drone was used for data acquisition and captured UAV remote sensing images in R, G, and B bands with a size of 5472*3648 pixels. The drone images were captured at a relative flight height of 120 m, with a Ground Sample Distance (GSD) of 3.5 cm/px. The images were taken on 21 July 2023, between 10 a.m. and 4 p.m. In order to illustrate the validity and generalizability of the experimental results, we selected six strongly representative UAV aerial images of riparian zones with different surface features and water conditions for the experimental study, as shown in Figure 1. For example, the water surface is calm, the water body is uniform in color, and the riverbank is almost completely covered by vegetation (Figure 1A). The water body has different depths, resulting in the uneven color of the water body and sunlight reflection, and there are some floating objects on the water surface and some bare soil and roads in the vegetation-covered area on the river bank (Figure 1B). There is visible floating debris and trash on the water, shadows obscure part of the water surface, and power lines are passing over the water surface, causing obscuration (Figure 1C). Vegetation is abundant, and sunlight reflects off the water surface, creating pronounced ripples (Figure 1D). The water surface exhibites ripple distribution and the water body’s depth varies (Figure 1E). The presence of algae causes uneven green coloration on the water surface (Figure 1F).

In Figure 1, the green vegetation information of each region appears to be highly similar. To select an optimal threshold for vegetation recognition, we plotted the histogram corresponding to the VDVI and identified the lowest value of 0.11 between the double peaks. This threshold has proven to be more accurate in extracting vegetation information from our images. Consequently, we have chosen to use this threshold consistently throughout our subsequent VDVI-based vegetation extraction processes.

The results of threshold segmentation after calculating the VDVI vegetation index for each subplot in Figure 1 are shown in Figure 3. As can be seen from the figure, all subplots have water bodies that are difficult to completely separate from the vegetation. Because the color of the water body is similar to the vegetation, setting a larger segmentation threshold can further separate the water body. However, it will also lead to a large amount of vegetation loss.

Figure 4 shows the threshold segmentation results of the Canny edge detection for each subplot shown in Figure 1. It can be seen that all subplots can effectively separate water bodies, but they cannot completely exclude the non-vegetated surface cover. Additionally, the most serious issue is that a large number of voids appear within the vegetation cover. The Otsu algorithm is utilized to automatically determine the threshold value in this method. While setting a larger segmentation threshold can help in further separating the non-vegetated surface cover, it comes at the cost of potentially losing a significant amount of vegetation and creating larger voids within the vegetation cover.

The results of extracting the vegetation cover of each subplot in Figure 1 using the method proposed in this paper are shown in Figure 5. It can be seen from the figure that all the water bodies, which can easily be confused with vegetation, have been excluded completely in all the subplots, while the vegetation has been retained to the maximum extent.

To quantitatively assess the accuracy of vegetation extraction, water bodies were manually labeled and removed from each subplot in Figure 1. The VDVI vegetation index was then calculated, and the vegetation was extracted. The extraction results were compared with those obtained by the method in this paper, and the confusion matrix was calculated to obtain the overall accuracy, producer accuracy, and user accuracy, as shown in Table 1. The table shows that the overall accuracy of all regions is above 95% except for region (c), where the overall accuracy is slightly below 95%. The user accuracy is very high in all regions, while the producer accuracy is slightly lower. Among them, the producer accuracy of region (c) is slightly lower than 90% because there is relatively less vegetation cover in the region (c), and more dark shadows cover the vegetation. This causes the edge detection algorithm to have lower values in these areas, which are swamped and marked as the foreground in the watershed algorithm, resulting in non-vegetated areas.

The Urban Drone Dataset (UDD) (Chen et al., 2018) was collected using the DJI Phantom 4 at heights ranging from 60 to 100 m. The image size is 3000 × 4000 or 4096 × 2160. This dataset is divided into three types: UDD3, UDD5, and UDD6, each representing different scenarios. To validate the accuracy of our method, we selected three images (Figures 6A–C) from UDD5 that resemble the riparian zone environment, and their corresponding ground truth labels were provided (Figures 6D–F). The results of extracting the vegetation cover of each subplot in Figures 6A–C using the method proposed in this paper are shown in Figures 6G–I. The extraction results are compared with the true values of the images, and the extraction accuracy is calculated, as shown in Table 2.

From Figure 6, it can be seen that the original image in UDD5 is darker and has more shadows and reflections. On the other hand, from Table 2, the overall accuracy and kappa coefficient of classification of this paper’s method for Figures 6A–C are 90.17%, 92.31%, 81.65%, and 0.76, 0.69, 0.61, respectively. However, the producer accuracy is lower, 70.36%, 57.83%, and 59.7% in Figures 6A–C, respectively. Figure 6B has the highest Overall Accuracy, but it has a lower Kappa than Figure 6A due to its uneven distribution of error pixels. Specifically, compared to Figures 6A,B has fewer errors in recognizing water body pixels as vegetation pixels, leading to a slightly higher Overall Accuracy. The uniformity of the error pixel distribution directly influences the Kappa coefficient. Figure 6B has fewer error pixels, and the distribution location of these pixels is not uniform enough, resulting in a lower Kappa coefficient for Figure 6B than for Figure 6A.

The Canny and Sobel operators are widely used algorithms for image edge detection. The Canny operator determines image edges by considering the direction and magnitude of the gradient, while the Sobel operator calculates the gradient using pixel value differences (simple subtraction of two pixels). As a result, the Canny operator can preserve edge continuity more effectively during edge detection, resulting in more detailed and clearer detected edges, especially in scenes that require high-precision edge detection, such as image recognition. On the other hand, the Sobel operator may introduce significant errors when processing right-angled edges. However, it is faster than the Canny operator, making it more suitable for real-time applications requiring higher speed, such as video surveillance.

In order to find edge detection operators suitable for the characteristics of riparian zone images, in this study, we considered the Canny operator and Sobel operator, which are outstanding in edge recognition accuracy and noise point sensitivity. We converted Figures 1A–C from RGB images to grayscale images and then utilized Canny and Sobel edge detection operators and marked the areas in the images where segmenting the boundaries along the water body was challenging using red boxes, as illustrated in Figure 7. These challenging regions exhibit characteristics such as shallow water bodies near the shore, with a color that appears closer to the shore, and contain various undesirable elements, including scum, garbage, and power lines. All these interferences directly affect the edge detection effect of the two operators. However, it is evident that Canny detection retains more edge information than Sobel detection and provides a more refined detection. Therefore, we chose the Canny operator as the pre-extraction segmentation algorithm before the detection of the watershed algorithm.

The watershed algorithm was used to segment the results of Canny and Sobel edge detection for Figures 1A–C, as shown in Figure 8. It can be seen from the images that the segmentation results based on the Canny operator are more accurate. As indicated by the yellow boxes in the images, we can observe that for Figure 1B, the segmentation result based on the Canny operator successfully separated non-water areas locally (Figure 8B), while the result from the Sobel operator did not (Figure 8E). Similarly, for Figure 1C, the segmentation result based on the Canny operator segmented water areas locally (Figure 8C), while the result from the Sobel operator did not (Figure 8F). Comparing Figure 1 and Figure 8, we can see that although Figure 8 shows some oversegmentation, water bodies and potential vegetation were separated into different colour areas. Therefore, after extracting the potential vegetation areas, vegetation extraction can be achieved using vegetation indices.

Several visible light vegetation indices are currently proposed, depending on the purpose of the application. Each of these indices has advantages and disadvantages in distinguishing land cover types. In order to select the most appropriate index, six visible light indices were selected for testing in this paper. First, a UDD5 scene image (Figure 6A) was selected. After removing the water body interference using the method in this paper, the potential vegetation areas were extracted by vegetation segmentation using VDVI, DEVI, RGBVI, MGRVI, NGRDI, and NGBDI, respectively. Following the vegetation extraction approach used in Figure 6A, we also processed Figure6B,C. As shown in Table 3, for different vegetation indices, we calculated the average of each evaluation metric across the three images.

As can be seen from Table 3, VDVI has the highest accuracy, with the overall accuracy and kappa coefficient reaching 86.75% and 0.6423, respectively. Secondly, for the sub-map with relatively complex land cover types (Figure 1C), Figure 9 presents the histograms of the six visible vegetation indices of the potentially vegetated areas after removing the water body using this method.

As can be seen in Figure 9, although the histogram distributions of the different vegetation indices varied greatly, they basically maintained a bimodal distribution. This ensures the correctness of the subsequent segmentation of vegetation and non-vegetation using the OTSU method. The optimal segmentation thresholds for subplots from (a) to (f) in Figure 9 are 0.11, 0.88, 0.07, 0.13, 0.2, and 0.06, respectively. Among them, the histogram distribution of the VDVI indices satisfies both the undulating and smooth requirements, leading to more accurate segmentation results.

In Figure 2, the maximum value resulting from the Distance Transformation is multiplied by a percentage value (denoted as P), and this product is set as a threshold. This threshold is then used for threshold segmentation of the Distance Transformation image I ₃ to obtain the Seed Areas image I ₄. By adjusting the value of P, we can influence the number of Seed Areas and, consequently, the segmentation results that follow. The marked watershed algorithm utilizes the P parameter value to regulate the smoothness of boundaries and the fineness of segmentation results. Higher p values produce smoother boundaries and more consistent area segmentation, while lower p values generate more detailed but less consistent boundaries. As a key factor in determining the efficacy and outcome of the algorithm, the P parameter value significantly impacts the overall performance. For instance, Figures 10A–C presents an example where P is systematically increased at 0.2 intervals from 0.3 for Figure 1B. The figures clearly illustrate the decrease in the number of seed areas with an increase in P. However, too few seed areas may result in certain water bodies being overlooked in the segmentation process, leading to their inclusion in the potential vegetation areas. This, in turn, may cause errors in the final vegetation segmentation, as depicted in the red-boxed portion of Figures 10D–F.

The segmentation process, which involves distinguishing water body areas from potential vegetation areas, relies heavily on the disparity between the textures of water bodies and vegetation. Consequently, areas with water bodies exhibiting similar textures to vegetation textures will likely be erroneously classified as potential vegetation areas during the segmentation. Additionally, if such areas also exhibit high values for the VDVI index, it becomes challenging to differentiate them from actual vegetation using the approach outlined in this paper. This challenge is demonstrated by the red and yellow boxed sections in Figures 11A, B and the white boxes in Figures 11C, D. However, if the water body areas do not exhibit high VDVI values, the method outlined in this paper can still successfully remove these areas at the OTSU segmentation stage, as shown in the blue box portion of Figures 11E, F.

Table 2 demonstrates that the method proposed in this paper achieves a high overall accuracy and a high user accuracy, implying that our approach has fewer instances of misclassification. However, the lower producer accuracy suggests that the method occasionally fails to correctly classify certain categories, as illustrated in Figure 6. This discrepancy can be attributed primarily to the time and angle at which the image was captured. As can be seen in Figure 6, the images of the UDD5 dataset were basically taken in the evening with insufficient light, while the sun was tilted at a large angle. The low illumination condition (C) in Figure 6 was utilized for vegetation extraction with the method described in this paper, and the results are depicted in Figure 12, this lighting condition results in darker shadows on vegetation (indicated by the red box) and shaded water bodies (indicated by the yellow box). These darker shadows cause a loss of textural information in the areas where the watershed algorithm cannot flood, and as such, they cannot be excluded from the segmentation process (as indicated by the yellow box). Additionally, these darker shadows significantly lower VDVI values, resulting in a greater loss of vegetation categories (as indicated by the red box). However, these problems can be easily overcome by selecting a clear and cloudless midday period for data acquisition, which avoids low light and strong shadows, thus achieving high segmentation accuracy.