1.

Introduction

River dikes are substantial river structures constructed to safeguard people and assets from flooding. Such dikes are typically managed by local river offices established across Japan by the Ministry of Land, Infrastructure, Transport and Tourism.¹ The herbaceous vegetation growing on these dikes vitally prevents soil erosion,² which could lead to dike failure and severe flooding with failure. Maintaining ideal vegetation conditions, such as the presence of native grass species and ensuring that the dikes are visible for safety checks, is essential. However, the spread of invasive alien species complicates dike vegetation management. Solidago altissima and Sorghum halepense are the dominant alien species in this study. S. altissima, the tall goldenrod, dominates several plant communities via allelopathy and reduces biodiversity.³ S. halepense competes with crops and grass species in cropland, grassland, and urban environments, increasing invasion risk.⁴ Accurately identifying areas on the dikes where native grasses and alien species grow would streamline dike vegetation management. Compared with traditional field surveys, which are geographically limited and require significant time and resources, remote sensing offers a cost-effective, large-scale, and long-term monitoring solution, systematically tracking species distribution over vast areas.⁵ However, classifying herbaceous vegetation on river dikes using remote sensing is complicated; it demands a high level of spatial resolution and accuracy to identify small, complex-shaped vegetation on the steep slopes of the dikes.⁶

Satellite remote sensing has become increasingly popular as a feasible alternative to map alien species. Unlike traditional methods, this approach can be applied to large and remote areas using satellite data and techniques.⁷ High spatial resolution multispectral sensors, such as IKONOS, WorldView-2/3, and Quickbird, advance broadband sensor technology for improved alien invasive plant (AIP) detection and mapping. These sensors have enhanced their ability to distinguish among different types of vegetation compared with the limited spatial resolution of multispectral sensors.⁸ For instance, Ngubane et al.⁹ reported a higher mapping accuracy (91.67%) for “Bracken” fern in Durban, South Africa, using the high spatial resolution WorldView-2, compared with the medium spatial resolution SPOT-5 (72.22%). The large pixel sizes of satellite remote sensing restrict their application to homogeneous landscapes, leading to the mixed pixel issue. The mixed pixel problem complicates invasive plant differentiation from a mixture of different plants and influences classification precision in remotely sensed images.¹⁰ Conflict occurs in labeling mixed pixels that encompass spectral responses of different land cover on the ground element, leading to confused classification. Therefore, an innovative method is required for large-scale mapping of invasive plants using satellite remote sensing. One technique for enhancing images involves the sharpening of multispectral (MS) bands with a panchromatic (pan) band. This pan-sharpening process aims to increase spatial resolution and reduce mixed pixels.¹¹ For instance, Ai et al.¹² evaluated five pan-sharpening algorithms to enhance AIP mapping accuracy. Their study demonstrated that pan-sharpening could improve AIP mapping; however, their results still included around 10% of accuracy misclassification errors. Therefore, future research should focus on integrating various data sources to enhance AIP mapping.

Drones or unmanned aerial vehicles (UAVs) have been used to identify AIPs^13–15 as these vehicles can capture high-resolution images at low altitudes. For example, Akandil et al.¹⁶ developed a method using UAVs equipped with multispectral sensors to accurately detect and assess giant goldenrod coverage. Their study yielded a kappa coefficient of 0.902 and an overall accuracy (OA) of 92.12%. These study findings highlight the effectiveness of UAV remote sensing in generating high-resolution maps with high accuracy. However, remote-sensing UAVs have limited flight operation, owing to legislation regarding air traffic and the safety and privacy of society.¹⁷ Considering the area coverage, satellite remote sensing has the advantage of being able to cover a larger area than drones. Thus, the combination of data from satellites and drones might allow for the acquisition of broad coverage from satellites and detailed information from drones. Additional data from drones might enhance wide-area satellite-based AIP mapping.¹⁸

The synergy between satellite and UAV data is essential to study the dynamics of Earth’s surfaces.¹⁹ Data fusion combines data from two or more sensors to create a new, comprehensive dataset. Although it can involve any type of sensor, this study concerns merging data from satellites and UAVs, to fully exploit the characteristics of each data source, resulting in a robust synergy between satellite and UAV data. Satellite and drone data are typically used separately, despite their complementary nature and strong potential synergy. Data fusion is a well-known technique for exploiting multisource synergy. However, in practice, the synergy between satellite and UAV data is specific, not well-established, and requires formalization.²⁰ Moreover, a review by Royimani et al.⁸ highlighted several research gaps in long-term and large-scale AIP mapping. One such shortcoming was exploring the benefits of multisource datasets, or data fusion, for mapping AIPs. Data fusion or multisource datasets are increasingly used for mapping AIPs using remote sensing approaches. However, the use of data fusion for mapping AIPs has been dominated by the integration of multispectral satellite datasets with light detection and ranging (LiDAR) data, not with drone data. Hence, the full potential of data fusion between satellite and drone datasets for optimal AIP detection and mapping has not yet been fully explored.

This study aims to improve the mapping accuracy of the AIP, specifically, S. altissima and S. halepense, on river dikes along the Tone River using fused WorldView-3 satellite and drone imagery. The goal of this study is to explore the potential of the fusion data between high-resolution satellites and drones to accurately map the AIP on wide-scale river dikes to aid effective vegetation management.

2.

Materials and Methods

2.1.

Study Area

The river dike environment (Fig. 1) is in the middle stream of Tone River in Ibaraki prefecture, Japan. The environment contains nine river dikes. Dike 3, for example, is $\sim 1\mathrm{km}$ long with a width of 40 m and a height of 7 m.¹⁸ These dikes are crucial for protecting human lives and urban properties from potential flooding and overflow incidents from the Tone River. The structural maintenance of these dikes is vital, which requires efficient vegetation management along their slopes. The vegetation management on river dikes helps conserve native ecosystems and prevents soil erosion along the slopes of dikes. In addition, dikes are surrounded by paddy fields, farmlands, and other agricultural zones.¹⁸ Thus, it is essential to avoid AIP spreading from the dikes to the surrounding areas. Invasive plants pose a threat to crops and other agricultural products cultivated by local farmers.

Fig. 1

Study area of the river dike environment.

River dikes are predominantly covered by herbaceous vegetation including a variety of plant species. Among these are native grass species and alien species, of which S. altissima and S. halepense are the two dominant alien species¹⁸ (Fig. 2). S. altissima grows over 1 m high and has broad leaves that block sunlight from reaching native species, physically obstructing the visibility of the river dike.²¹ S. halepense has strong competitive abilities and allelopathic potential, affecting the growth and development of nearby plants, causing agricultural losses and negatively impacting natural biodiversity.²²

Fig. 2

Target alien species. Solidago altissima (a). Sorghum halepense (b).

2.3.

Image Analysis

2.3.1.

Image pre-processing

Drone images were pre-processed using the Pix4Dmapper software. For the sample dike, RGB spectral bands with 0.1-m ground sampling distance (GSD) and an orthomosaic with 0.01-m GSD were generated. For the representative dike, only an orthomosaic with 0.01-m GSD was generated. In addition, WorldView-3 satellite bands and drone bands underwent pre-processing such as georeferencing, resampling, and pixel matching. The summary of pre-processed imagery used in this study is shown in Table 3.

Table 3

Summary of pre-processed imagery.

Target area	Pre-processed image	Information
Study area	WorldView-3 Satellite	1.6-m RGB bands
		0.4-m panchromatic band
Sample dike (dike 3)	Drone	0.1-m GSD RGB bands
		0.01-m GSD orthomosaic
Representative dike (dike 5)	Drone	0.01 m GSD orthomosaic

2.3.2.

Image pan-sharpening

Pan-sharpening is a technique used to generate high-resolution multispectral images, which is achieved by merging low-resolution multispectral images with panchromatic ones.²³ The process involves the fusion of a high-resolution panchromatic raster band with a lower-resolution multiband raster dataset. The outcome is a multiband raster dataset that matches the resolution of the panchromatic raster in the overlapping areas. Numerous pan-sharpening methods have been explored over the past several years. We used the Gram–Schmidt pan-sharpen method to enhance the 1.6-m resolution RGB bands with the 0.4-m panchromatic band of the WorldView-3 satellite image. Since its introduction in 1998, the Gram–Schmidt pan-sharpen method has gained popularity as one of the leading algorithms for pan-sharpening multispectral imagery and excels in enhancing image sharpness and reducing color distortion, outperforming most other pan-sharpened methods.²⁴

2.3.3.

Image fusion

Image fusion was performed on an RGB-based band-by-band fusion between the pan-sharpened RGB bands of the satellite and the RGB bands of the drone on a sample dike. The “+” operator of the raster calculator in ArcGIS Pro was used to merge each corresponding band from the 0.4 m RGB pan-sharpened satellite and the 0.1-m GSD RGB drone. The statistical range of a pan-sharpened satellite raster was between 0 and 2047, whereas that of the drone raster was between 0 and 255.

For instance, a fused band is created using a mathematical expression denoted as Eq. (1) and expressed in Fig. 3. In this equation, the “+” operator signifies the pixel-wise addition of values from two raster datasets. The raster calculator extracts the pixel values from the band of the pan-sharpened satellite and adds them to the corresponding pixel values from the band of the drone. The outcome of this addition is stored in a new raster layer, termed the “fused band.” Consequently, each pixel in the “fused band” raster possesses a value equivalent to the sum of the corresponding pixel values from the bands of the pan-sharpened satellite and drone.

Fig. 3

Band fusion.

Mathematically, for each pixel ($i,j$), the operation is represented as follows:

Eq. (1)

$$\mathrm{Pan}\text{-}\text{sharpened satellite Band}(i,j)+\text{Drone Band}(i,j)=\text{Fused Band}(i,j).$$

2.3.4.

Vegetation index image

Vegetation indices, which are linear combinations of bands, are highly effective in mapping AIPs.^25,26 In this study, we used the Green–Red Vegetation Index (GRVI) as the vegetation index, as only RGB bands were used.²⁷ The GRVI was computed according to Eq. (2) and was independently extracted from the original satellite, pan-sharpened satellite, and fused satellite–drone image

Eq. (2)

$$\mathrm{GRVI}=\frac{\text{Green band}-\text{Red band}}{\text{Green band}+\text{Red band}}.$$

2.3.5.

Object-based image analysis

Beyond the conventional pixel-based classification techniques, object-based image analysis (OBIA) categorizes imagery into objects using perception-based and meaningful knowledge.²⁸ OBIA presents several benefits, encompassing the addition of multiparameter classification, improved performance with high-resolution data, and sharp images facilitating advanced image classification.^29,30 Therefore, this study adopted OBIA for the classification of alien species distribution.

Using the Definiens Professional 5.0 software, images were segmented into objects according to the configuration¹⁸ as detailed in Table 4. The choice of scale parameter in the segmentation process is a key factor in determining the quality and accuracy of the results in OBIA and influences the size of image segments, landscape representation, and the success of the classification process. In this study, a 5.0-scale parameter was selected, which resulted in an average size of $19.28{\mathrm{m}}^2$ per object. According to Table 5, a scale parameter of 5.0 provides a balance between detail and generalization. This parameter is more detailed than a scale parameter of 10.0 and 20.0, which results in larger objects, but less detailed than a scale parameter of 1.0, which results in the smallest objects and is equivalent to pixel size. Moreover, the vegetation patches in the study area were $˜20{\mathrm{m}}^2$ in size on average; thus, a scale parameter of 5.0 would be suitable for this study.

Table 4

Image segmentation setting.

Segmentation mode	Multiresolution segmentation
Scale parameter	5.0
Color	0.9
Shape	0.1
Compactness	0.5
Smoothness	0.5

Table 5

Scale parameters comparison.

Scale parameter	Average size of objects (m2)
1.0	2.56
5.0	19.28
10.0	86.13
20.0	348.80

After the image segmentation process, statistical data were extracted from each band across all images, according to each object of the segmented image. The data extraction used the zonal statistic function in quantum GIS (QGIS). The common types of extracted statistical data are mean, median, standard deviation (stdev), minimum (min), and maximum (max) values. Following the data extraction, the dataset was prepared so that each image of each dike was represented by the created dataset. A sample of the dataset is shown in Table 6. Consequently, the sample dike (dike 3) was represented by three datasets: original satellite, pan-sharpened satellite, and fused satellite–drone dataset. In contrast, the other river dikes were represented by only two datasets: the original satellite datasets and the pan-sharpened satellite datasets.

Table 6

Sample dataset.

Object ID	Red band	Green band	Blue band	GRVI band
First object	Mean, median, stdev, min, and max	Mean, median, stdev, min, and max	Mean, median, stdev, min, and max	Mean, median, stdev, min, and max
.	.	.	.	.
.	.	.	.	.
.	.	.	.	.
Last object	Mean, median, stdev, min, and max	Mean, median, stdev, min, and max	Mean, median, stdev, min, and max	Mean, median, stdev, min, and max

2.4.

Ground Truth

Drones play three important roles: explanation, validation, and completion of satellite data. The validation role involves using drone data as a “ground truth” (or “drone truth”), relying on the capability of the drone to clarify satellite data.²⁰ In this study, the ground truth data were based on the 0.01-m GSD drone orthomosaic,¹⁸ which was acquired for both the sample dike and the representative dike. Using these resources, the location and proportion of the target AIP were determined by visually constructing representative polygons, as depicted in Fig. 4. These polygons were then overlaid with the segmented image to determine the percentage of target alien species within each object. The percentages of target alien species within each object were classified into five categories,¹⁸ as illustrated in Table 7. These classes served as a benchmark for supervising the classification models and assessing the accuracy of mapping performances.

Fig. 4

Alien species location on sample dike (dike 3) and representative dike (dike 5).

Table 7

Classes of alien species.

Class of target alien species	Description of class
Class 0	0% of species
Class 1	0% to 25% of species
Class 2	25% to 50% of species
Class 3	50% to 75% of species
Class 4	75% to 100% of species

3.

Results

The outcomes of the regression analysis are depicted in Fig. 5. The results revealed that six variables were linked to the regression model with an $R^2$ value of $<0.5$. Consequently, these variables were omitted in the following steps. However, regression models of the 14 remaining variables were used to predict the fused satellite–drone datasets of the other dikes by applying these models to the pan-sharpened satellite datasets.

Fig. 5

Regression results ($n=1263$).

Three classification models were developed from the sample dike to classify the distribution of each target alien species, namely, S. altissima and S. halepense. For S. altissima species (refer to Table 9), the training results demonstrated that the fused satellite–drone dataset achieved the highest classification performance in almost all evaluation metrics, followed by the pan-sharpened satellite and the original satellite datasets. A similar pattern was observed for S. halepense species (see Table 10), where the fused satellite–drone dataset again achieved the highest classification performance in almost all evaluation metrics, followed by the pan-sharpened satellite and the original satellite datasets. From these results, it appears that the fused satellite–drone dataset consistently outperforms the other two datasets in classifying the distribution of both S. altissima and S. halepense.

Table 9

Evaluation of classification models for Solidago altissima species classification.

Table 10

Evaluation of classification models for Sorghum halepense species classification.

After the mapping was completed, the accuracy of the mapping performance was evaluated on the representative dike by comparing the results of the alien species mapping with the ground truth data. For S. altissima species (refer to Table 11), the mapping results derived from the fused satellite–drone dataset were the most alike across all classes compared with the ground truth. Similarly, for S. halepense species (see Table 12), the mapping results obtained from the fused satellite–drone dataset were the most comparable across all classes compared with the ground truth, although class 0 was predominantly represented. These comparisons indicate that the fused satellite–drone dataset demonstrates superior performance in AIP mapping.

Table 11

Comparison of Solidago altissima species mapping results on representative dike (n=3602).

Table 12

Comparison of Sorghum halepense species mapping results on representative dike (n = 3602).

The accuracy of the target alien species mapping was determined based on OA and the kappa coefficient. The results highlighted that the use of the fused satellite–drone dataset to predict the distribution of S. altissima and S. halepense species resulted in a high mapping accuracy. For the mapping of S. altissima species (refer to Fig. 6), the use of the fused satellite–drone dataset achieved the highest OA of 98.39% and a kappa coefficient of 0.976. This was $\sim 9\%$ higher in OA and 0.14 higher in the kappa coefficient compared with the pan-sharpened satellite dataset and $\sim 11\%$ higher in OA and 0.17 higher in the kappa coefficient compared with the original satellite dataset. In the case of mapping S. halepense species (see Fig. 7), the use of the fused satellite–drone dataset achieved the highest OA of 97.78% and a kappa coefficient of 0.936. This was $\sim 4\%$ higher in OA and 0.12 higher in the kappa coefficient compared with the pan-sharpened satellite dataset and $\sim 5\%$ higher in OA and 0.14 higher in the kappa coefficient compared with the original satellite dataset. These results further validate the effectiveness of the fused satellite–drone dataset in improving AIP mapping accuracy.

Fig. 6

Accuracy assessment of Solidago altissima species mapping on representative dike.

Fig. 7

Accuracy assessment of Sorghum halepense species mapping on representative dike.

Finally, based on mapping results from the fused satellite–drone datasets, ArcGIS Pro 3.2 was used to create predictive maps to provide a visual representation of the distribution of the target AIP, S. altissima (refer to Fig. 8) and S. halepense (see Fig. 9). These maps highlight the large coverage of the S. altissima species within the study area, whereas the growth of the S. halepense species was relatively limited. These maps can be crucial for environmental management strategies, as they provide insights into the spatial distribution of these AIPs. This information can guide efforts to control the spread of these species and mitigate their potential impacts on local ecosystems.

Fig. 8

Predicted map of Solidago altissima species distribution.

Fig. 9

Predicted map of Sorghum halepense species distribution.

4.

Discussion

Improved mapping accuracy was achieved using a fused satellite–drone dataset to predict the coverage of S. altissima and S. halepense species and can be attributed to the complementary strengths of high-resolution satellite and drone imagery. WorldView-3 satellite imagery provides a broad landscape overview, capturing large-scale patterns and trends. However, its resolution may not be sufficient to detect fine-scale features, such as individual plant species on a river dike. In contrast, drone imagery, with its 0.1-m GSD, can capture these fine-scale features in notable detail. However, owing to the limited flight operation of drones, it may not be feasible to use drone imagery for mapping vast river dike environments. By fusing these two types of imagery, this study leverages the strength of both the large-scale coverage of satellite imagery and the fine-scale detail of drone imagery. This fusion results in a dataset that is both comprehensive in coverage and rich in detail, thereby enhancing the accuracy of mapping the coverage of S. altissima and S. halepense species. The pan-sharpened satellite dataset, while an improvement over the original satellite dataset in terms of resolution, still lacks the fine-scale detail provided by drone imagery. As such, the fused satellite–drone dataset outperforms the pan-sharpened satellite dataset. The original satellite dataset, with its lower resolution, is less capable of accurately mapping the coverage of AIPs, resulting in lower overall accuracy.

A previous study³⁴ has acknowledged the cost and time-effectiveness of remote sensing as a technique for identifying alien vegetation. Furthermore, this study used not only high-resolution satellite imagery but also fused it with drone imagery to improve the AIP mapping accuracy. Fusing these data sources provides a comprehensive and detailed dataset, which leads to higher overall accuracy. Moreover, this study contributes to the growing body of literature³⁵ supporting the use of advanced remote sensing techniques and machine learning models in improving the mapping accuracy of alien species distribution. The successful application of the RF classification model in this study adds to this narrative.

The research method exhibits strengths contributing to the findings of this study. Ai et al.¹² described a method to improve the mapping accuracy of an invasive plant (Spartina alterniflora), which is based on the integration of pan-sharpening and classifier ensemble techniques. The parallel observations among our studies underscore the consistent benefits of pan-sharpening techniques in remote sensing applications, in the context of discriminating among closely related species. Moreover, the findings of Elkind et al.³⁶ are closely related to the findings of this study, which underscored the importance of integrating DigitalGlobe WorldView-2 satellite imagery with UAV data for mapping non-native buffelgrass (Pennisetum ciliare). By leveraging the capabilities of both data sources, RF classification models for buffelgrass detection over larger areas achieved an average overall accuracy of 93%. Furthermore, OBIA methods improve classification performance when used with machine learning algorithms.³⁷ Therefore, OBIA provided a systematic and effective way to process and analyze the high-resolution satellite and drone imagery, leading to accurate AIP mapping in this study. In addition, RF has been successfully applied to this study to map the distribution of alien species on river dikes. Meyer et al.³⁸ used five supervised classification algorithms to estimate the total areas covered by the invasive dune species Carpobrotus edulis. Among these algorithms, the RF algorithm emerged as the most effective, demonstrating superior performance across various accuracy metrics. This highlighted the efficacy of the RF in accurately identifying and classifying invasive species, even in complex environments such as dune ecosystems. Therefore, our findings contribute to the growing body of knowledge on the application of RF in ecological monitoring and invasive species management.

In summary, this study not only aligns with the existing studies but also provides new insights and advancements in remote sensing for mapping alien species distribution. It highlights the potential of fusing high-resolution satellite and drone imagery for enhancing mapping accuracy, thereby contributing to effective vegetation management on river dikes, particularly in terms of the conservation of native species and the control of invasive alien species. However, in the context of generalization, the methodology and findings of this study are so far limited to the river dikes along Tone River and two alien species, S. altissima and S. halepense. The effectiveness and reliability of the methodology and application of this study may vary in different contexts, landscapes, and species.

5.

Conclusion

We demonstrate the significant potential of fusing high-resolution satellite and drone imagery for mapping AIP distribution, namely, S. altissima and S. halepense, on the river dikes environment along Tone River. We used a combination of various processes, including pan-sharpening of WorldView-3 satellite images, fusion with drone images, object-based image analysis, regression analysis, and the RF classification model. The key findings of this study underscore the significant contribution of these techniques in enhancing the mapping accuracy of AIP distribution. The fused satellite–drone dataset outperformed the pan-sharpened satellite dataset and the original satellite dataset in terms of overall accuracy and kappa coefficients for both species. This indicates that the fusion of high-resolution satellite and drone imagery provides more accurate and reliable data for mapping the distribution of alien species on river dikes. These findings have important implications for the management of herbaceous vegetation on river dikes. With improved mapping accuracy, it is possible to obtain a comprehensive understanding of the distribution of native grasses and AIPs along expansive river dikes. This can significantly influence the conservation of native species and the maintenance of river dike visibility for safety inspection.