3. Results and Analyses

3.1 Selection of feature bands

The band importance index was extracted by the random forest algorithm (Fig. 2). Sentinel-2 images have approximately four to five bands of higher importance in all three months, particularly, the red-edge band (B06, B07) and near-infrared band (B08A) are more important. Compared to Sentinel-2 images, Landsat-8 images have approximately two to three bands of higher importance in all months, especially the red-band (B04) and near-infrared band (B05). The importance of the near-infrared band in May is particularly prominent.
The importance index of the band can only determine the contribution of different bands. In order to reduce the noise generated by the band with a lower contribution rate, we determined for each image if there are any important bands for species extraction. The bands of the top n importance at different months were selected as the basis for separating different species. If the importance of a band at the given time is ranked as the top one, but its separability is poor, other bands from the same image with a lower importance index can be considered as a noise band that should be discarded. The bands that are selected with higher separability are those that will be used for final extraction.
Figure 3 shows the spectral separability of the n significant bands of Sentinel-2 images for different months. For example, during the corresponding phenological stage, the red edge band (B06, B07) and the near-infrared band (B08A) of the vegetation can be used to effectively distinguish between S. salsa , P. australis , and S. alterniflora. Among them, 09_B06, 09_b07, 09_B08A, and 10_B08A can basically achieve separation between the three species. We selected multi-temporal bands with good separability from the Sentinel-2 source data (05_B06, 05_B07, 05_B08A, 05_B12, 09_B08A, 09_B07, 09_06, 09_02, 09_B04, 10_B06, 10_B08A, 10_B12, 10_B11, 10_B04, 10_02, and 10_07) to extract the native and invasive species of the study area.
The Landsat-8 image spectral separability band is mainly concentrated in the red band (B04) and the near-infrared band (B05) (Fig. 4). From the importance and separability analysis of the bands, it was found that the Landsat-8 image in September had the best separability and the most bands (09_B04, 09_B05, 09_B01, 09_B02, and 10_B04) for S. salsa ; however, there were only two separable bands for P. australis (05_B05, 05_B07) and one for S. alterniflora(10_B05).
The above analysis shows that the multi-temporal Landsat-8 and Sentinel-2 images can distinguish S. salsa , P. australis , and S. alterniflora . The spectral separability bands were mainly concentrated in the red band (630 nm-680 nm, Landsat-8), near-infrared band (845 nm-885 nm, 848 nm-880 nm, Landsat-8, Sentinel-2), and vegetation red edge band (739 nm-749 nm, 768 nm-796 nm, Sentinel-2). Besides, some studies have shown that the multi-temporal vegetation index (such as NDVI) of the Landsat-5 image can also distinguish between the three types of vegetation (Wang et al. 2013).

3.2 Classification and accuracy analysis

Figure 5 shows the native and invasive species extracted from Landsat-8 images and Sentinel-2 images. It can be seen from the figure that the distribution patterns of S. salsa , P. australis , andS. alterniflora shown in the two images are almost the same.P. australis is predominant on the beaches on both sides of the Yellow River and the artificially restored recovery areas. S. salsa is mainly distributed in the beaches and areas of high salinity in the Yellow River Delta. It is a pioneer plant in muddy tidal flats and heavy saline-alkaline areas. S. alterniflora mainly grows in the lower part of the intertidal zone to the lower part of the middle tide zone.
The overall accuracy of data extraction based on multi-temporal Sentinel-2 is 82.86%, with a Kappa coefficient of 0.79; the overall accuracy of data extraction based on multi-temporal Landsat-8 is 78.77%, with a Kappa coefficient of 0.74 (Tab. 2). Regardless of the image, the producer accuracy and user accuracy of S. salsa andS. alterniflora are generally higher than those of P. australis , which is caused by many factors, such as plant height, plant density, and the type of underlying surface. However, P. australis and S. alterniflora exhibit a mixed phenomenon, and the extraction accuracy is low in the areas with more human activities. When the Landsat-8 image is used for classification, the P. australis growth is misclassified into that of S. alterniflorain the marginal areas where the reed grows, but this problem does not occur in Sentinel-2 images.

3.3 Analysis of landscape dynamics related to native and invasive species

The results of the extraction of native and invasive species in the Yellow River Delta wetlands for 2010, 2013, 2016, and 2018 are shown in Figure 7. The growth area for P. australis is 66.96 km2, 63.84 km2, 76.26 km2, and 72.73 km2 for 2010, 2013, 2016, and 2018, respectively, and the area change shows a steady upward trend. The area where S. salsa is found covers 47.6 km2, 27.48 km2, 39.61 km2, and 35.04 km2 in the same four time periods, respectively, which shows a downward trend. S. alterniflora greatly expanded its range from 2010 to 2016, rapidly increasing from 1.14 km2 to 30.77 km2. The expansion trend slowed from 2016 to 2018, but the area still reached 40.53 km2 in 2018.
Due to the influence of tidal currents, large areas of beaches are flooded during the high tide, which results in the death of some earlier growth ofS. salsa . After the ebb tide, the seeds ofS. salsa sprout, and new seedlings are re-grown. The S. salsacommunity in the adjacent part of the sea reciprocates with the tide period, and poor growth is exhibited mainly in the decreased density and short size of the plants. Therefore, the patch density and splitting index of S. salsa were significantly higher than that of P. australis and S. alterniflora , and the aggregation index is the lowest (Fig. 2), which indicates that the spatial distribution ofS. salsa is irregular and discrete. With the passage of time, the patch density of S. salsa is decreased, but the splitting index is always at a high level, which indicates that the area of S. salsa in the study location is gradually decreasing, and the poor growth conditions lead to severe fragmentation of the spatial distribution.
Over the past 10 years, the landscape indices of theP. australis community have been relatively stable, without a sharp change. The patch density was low, the largest patch index was the highest among the three vegetation communities, and the aggregation index was also stable at approximately 90. This occurred because a large number of wetland ecological restoration projects have been carried out in the Yellow River Delta region, and P. australis gradually formed a contiguous distribution pattern based on its rapidly expanding reproductive capacity, becoming an emergent aquatic plant community of a single dominant species.
Growth of S. alterniflora was recorded in the estuary of the Yellow River at the beginning of the 21st century. Due to the impact of the tidal runoff and the Yellow River diversion, initially, there was no large-scale expansion ofS. alterniflora . At present, under the influence of water and sediment adjustment in the upper reaches of the Yellow River, the area of the S. alternifloracommunity has been rapidly expanding. Since 2010, the patch density, largest patch index, and aggregation index of S. alterniflorahave shown an upward trend, with the splitting index showing a significant downward trend, which highlights that the area of occupation for S. alterniflora is gradually increasing. The spatial distribution of the community gradually shows a pattern of continuous, regularized, and integrated growth. A number of previous researchers have reported that S. alterniflora currently appears only on both sides of the Yellow River estuary. An area of S. alternifloragrowth was found in the southern tidal flat for the first time in this study. There was only a sporadic distribution in the 2017, but a pattern of small-scale contiguous distribution was observed in 2018 (Fig. 7).

3.4 Analysis of expansion of S. alterniflora community

3.4.1 Expansion direction analysis

To study expansion direction of S. alterniflora from 2010 to 2018, the relevant parameters of the standard deviation ellipse ofS. alterniflora were obtained. There were great changes in the spatial distribution pattern of S. alterniflora on the north bank of the Yellow River estuary. The movement path underwent a change consisting of “Northwest (2010-2013)—Northeast (2013-2016)—Northwest (2016-2018),” which generally shows a trend of moving to the northwest. The standard deviation of the elliptical distribution range of S. alterniflora showed an expanding trend from 2010 to 2018. The long semi-axis and the short semi-axis increased from 0.52 km and 0.12 km in 2010 to 83.86 km and 19.17 km in 2018, respectively. Combining the change of the long and short half-axis and the distribution pattern of the standard deviation ellipse shows that the distribution of S. alterniflora on the north shore is dominated by movement in the east-west direction. During the study period, the standard deviation of the elliptical deflection angle of the north shore ranged from 89.35° to 108.81°.
The center of gravity of S. alterniflora on the south bank of the Yellow River is mainly north-south, with small movements from east to west. It moved mainly to the south in 2010–2013 and moved northward in 2013–2018. The long semi-axis and the short half-axis increased from 0.66 km and 0.11 km in 2010 to 39.01 km and 15.66 km in 2018, respectively. The distribution of the S. alterniflora community on the south bank is mainly in the north-south direction. Compared with the north shore, the deflection angle of the south bank has been decreasing, from 75.81° in 2010 to 2.78° in 2018, which indicates that the S. alterniflora community on the south bank is mainly oriented towards land expansion.

3.4.2 Analysis of Expansion mode

Zhang et al.(2018) studied the diffusion mechanism of S. alterniflora from the perspective of genetic evolution. The study showed that all the S. alterniflora communities in this study area are from the No. 5 pile area, rather than gradually expanding from the north bank of the Yellow River estuary to the south bank, which indicates that the genetic distance ofS. alterniflora has no relationship with the geographical distance. The seeds of S. alterniflora can be transported over long distances by seawater for plant propagation. After determining the source of the expansion ofS. alterniflora , subsequent studies herein mainly analyzed the different expansion patterns after S. alterniflora colonization.
The multi-year expansion pattern obtained by using the LEI index shows that from the expansion area, the main expansion mode of the S. alterniflora community in the past 10 years is edge expansion. This mode of expansion is characterized by the outward expansion of the marginal region of the patches extracted from the previous phase, which is adjacent to the native patches. The total area of edge expansion is 33.97 km2, and the expansion area is larger than the original block area (0≤LEI<1). The external expansion area is small, with a total area of 5.42 km2, and the newS. alterniflora patches are not adjacent to other S. alterniflora patches, which demonstrates a discrete distribution state. However, from the number of patches, the number of edge expansion plaques in the S. alterniflora community is slightly less, only 73, and the total number of external expansion patches is 339, which also indicates that the growth scale of S. alterniflora community is continuous, regularized, and integrated.