High-Resolution Mapping and Biomass Estimation of Suaeda salsa in Coastal Wetlands Using UAV Visible-Light Imagery and Hue Angle Inversion

Abstract

Unmanned Aerial Vehicles (UAVs) have become powerful tools for high-resolution, quantitative remote sensing in ecological and environmental studies. In this study, we present a novel approach to accurately mapping and estimating the biomass of Suaeda salsa using UAV-based visible-light imagery combined with hue angle inversion modeling. By integrating diffuse reflectance standard plates into the flight protocol, we converted RGB pixel values into reflectance and derived hue angle metrics with enhanced radiometric accuracy. A hue angle cutoff threshold of 249.01° was identified as the optimal cutoff to distinguish Suaeda salsa from the surrounding land cover types with high confidence. To estimate biomass, we developed an exponential inversion model based on hue angle data calibrated through extensive field measurements. The resulting model—Biomass = 3.57639 × 10⁻¹⁵ × e^0.12201×α—achieved exceptional performance (R² = 0.99696; MAPE = 3.616%; RMSE = 0.02183 kg/m²), indicating strong predictive accuracy and robustness. This study highlights a cost-effective, non-destructive, and scalable method for the real-time monitoring of coastal vegetation, offering a significant advancement in remote sensing applications for wetland ecosystem management.

1. Introduction

Suaeda salsa is a crucial coastal wetland species that plays a vital role in soil improvement [1], carbon sequestration [2], and ecosystem maintenance [3]. It is widely distributed in coastal and inland saline areas, including northern China, Central Asia, and Europe [4].

Coastal wetlands are vast and complex ecosystems, often difficult to access. Traditional monitoring methods, such as visual estimation, probability, and grid methods, are insufficient for comprehensive monitoring. Remote sensing technology, with its extensive coverage and continuous observation capabilities, offers a promising solution to address these challenges [5]. Mature Suaeda salsa exhibits a distinctive reddish-brown coloration in visible-light imagery, which aids in its visual identification. However, despite this visual cue, the precise extraction of its distribution from such imagery remains challenging. Beyond its visible appearance, Suaeda salsa, typical of salt-marsh vegetation, also exhibits a prominent red-edge phenomenon in its multispectral reflectance spectrum, which is crucial for detailed spectral analysis. Traditional vegetation indices, such as the NDVI, RVI, SVI, EVI, PVI, SAVI, MSAVI, and TSAVI [6,7,8,9,10,11,12], have been proposed to extract Suaeda salsa information from remote sensing data. However, accurate cutoff thresholding is crucial, as an overlap in index values between Suaeda salsa and other vegetation, particularly green vegetation, can lead to misclassification [13]. Recently, object-based classification methods have also been applied to identify Suaeda salsa and other ground features [14], but their practical applications remain relatively limited. Additionally, the low spatial resolution of satellite imagery limits the accurate assessment of areas with low-cover, edge, and mixed-pixel areas.

To enable a quantitative representation of color, the International Commission on Illumination (Commission Internationale de l’Éclairage, CIE) developed the CIE 1931 XYZ standard color system [15], which has become the foundation of modern colorimetry. This system is widely applied in diverse fields such as color measurement, color management, image processing, and printing. However, its application in ecological and environmental monitoring began relatively late. Early studies primarily relied on satellite multispectral sensors or field-measured spectral data to derive the X, Y, and Z tristimulus values of the CIE system through various processing methods. These values were then used to calculate chromaticity parameters and assess water quality status [16,17,18]. In recent years, CIE-based color parameters—such as hue angle, chromaticity coordinates, and water color indices—have been increasingly employed in remote sensing analyses of aquatic ecosystems. These include investigations of lake eutrophication [19,20], water color constituents (e.g., chlorophyll-a, suspended matter, and colored dissolved organic matter) [21,22], nearshore algal blooms [23,24,25], and green tides [26,27]. Notably, Suaeda salsa exhibits a distinct reddish-brown hue during its mature stage, which contrasts sharply with the surrounding land cover [4]. This pronounced color difference makes hue angle-derived metrics from remote sensing imagery particularly effective for its identification. Despite the clear potential and broad applicability of this approach, to our knowledge, no prior studies have reported the use of hue angle to detect and quantify Suaeda salsa.

Recent advancements in UAV remote sensing technology have enabled high-resolution (centimeter-level) imagery, making it a powerful tool for ecological and environmental monitoring [8,28,29,30,31]. Previous studies [32] have placed high performance requirements on UAV imaging payloads, typically necessitating the use of multispectral or hyperspectral sensors for aerial photography and subsequent information extraction. This has partially limited the broader application and scalability of UAV-based remote sensing techniques for Suaeda salsa. Moreover, most extraction methods have relied primarily on vegetation indices, which still exhibit certain limitations in accurately distinguishing Suaeda salsa from other vegetation types. This study aims to leverage UAV-based visible-light imagery and hue angle analysis to (1) accurately extract Suaeda salsa from complex backgrounds, which include various land cover types such as reeds and bare land in the target area, and (2) develop a robust remote sensing inversion model to estimate Suaeda salsa biomass.

2. Materials and Methods

2.1. The Experimental Area

A coastal wetland in Wafangdian City, northern China, was selected as the experimental area (Figure 1). This approximately 12,400 m² site was characterized by the presence of Suaeda salsa, reeds, mudflats, and water bodies, as well as a mixed area of Suaeda salsa and reeds.

2.2. Design of Experiments

Figure 2 outlines the experimental procedure. A DJI (Da-Jiang Innovations, Shenzhen, China) Mavic 3M multispectral UAV, equipped with a visible-light camera and a multispectral sensor (capturing green, red, red-edge, and near-infrared bands), was employed for data acquisition. Four diffuse reflectance standard plates (DRSPs) with known reflectance values (1.2%, 5%, 10%, and 25%) were used for radiometric calibration.

On 12 October 2024, a synchronous measurement experiment was conducted to acquire UAV imagery and Suaeda salsa biomass data using square quadrats (0.6 m × 0.6 m). The experimental procedure comprised the following steps:

(a)

Experimental Site Setup (Figure 3a): Prior to measurements, the area with the highest density of Suaeda salsa was identified and designated as the initial quadrat, representing the maximum biomass. The quadrat was then secured, and DRSPs were positioned in an adjacent open area for calibration purposes.

(b)

UAV Camera Parameter Configuration and Initial Imaging: To ensure consistent illumination throughout data collection, the visible-light camera’s exposure settings—including aperture, shutter speed, and ISO—were fixed before flight. A single-frame shooting mode was employed to capture high-resolution imagery of the Suaeda salsa quadrat.

(c)

Sequential UAV Imaging and Biomass Removal (Figure 3b): The UAV ascended to a fixed altitude of approximately 25 m to capture an initial image of the quadrat at its maximum Suaeda salsa biomass. Subsequently, portions of Suaeda salsa within the quadrat were evenly removed to incrementally reduce biomass. After each removal, another UAV image was captured, and the removed plant material was sealed in labeled bags for subsequent biomass analysis. This iterative process of incremental Suaeda salsa removal, UAV image acquisition, and sample storage continued until the quadrat was completely cleared.

(d)

Laboratory Processing of Biomass Samples: Upon the completion of UAV imaging, all plant samples were promptly transported to the laboratory. Roots were thoroughly rinsed with clean water to remove soil residues. Each sample was then wrapped in aluminum foil and oven-dried to a constant weight, and its dry biomass was meticulously recorded according to the sample labels.

(e)

Image Processing: For each UAV image, the quadrat-corresponding area was accurately extracted. The hue angle was subsequently calculated for this region, and the average hue angle was determined for further analysis.

(f)

Model Dataset Construction: Based on the sequential biomass removal, the Suaeda salsa biomass within the quadrat (expressed in kg/m²) was calculated for each UAV capture. This yielded a robust modeling dataset linking varying biomass levels with their corresponding mean hue angles.

Following the quadrat-based Suaeda salsa biomass experiment, a comprehensive UAV aerial survey of the entire study area was conducted on October 14, 2024. Similarly to the previous setup, DRSPs were deployed, and the camera’s exposure settings were fixed to ensure consistent lighting conditions. The acquired UAV imagery served two primary purposes: (1) to determine the optimal hue angle threshold for Suaeda salsa identification and (2) to extract Suaeda salsa biomass information across the entire study area.

2.3. Data Preprocessing

2.3.1. Reflectance Conversion Method

The color values of four DRSPs were extracted from the UAV visible-light imagery for the red, green, and blue bands. These values were plotted against their known reflectance values, which were determined in the laboratory, and fitted to a suitable model. This model, which can vary depending on lighting conditions, is used to convert pixel color values to reflectance values. It is important to note that this conversion process must be performed individually for each image.

2.3.2. Methods for Calculating Hue Angle

The CIE 1931 XYZ color system, developed by the International Commission on Illumination (CIE), quantifies color through three spectral tristimulus values: X, Y, and Z. However, these values are not readily interpretable or visually intuitive. To overcome this limitation, the CIE introduced the CIE 1931 xy chromaticity diagram (Figure 4)—a two-dimensional representation that employs chromaticity coordinates, hue angle, and other derived parameters to visually characterize color.

Firstly, the UAV imagery for RGB bands was transformed to the Commission Internationale de l’Eclairage (CIE) tristimulus values X, Y, and Z as follows [19,33]:

$$X=1.1302R\left(\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}\right)+1.7517R\left(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\right)+2.7689R\left(\mathrm{R}\mathrm{e}\mathrm{d}\right)$$

(1)

$$Y=0.0601R\left(\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}\right)+4.5907R\left(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\right)+1.0000R\left(\mathrm{R}\mathrm{e}\mathrm{d}\right)$$

(2)

$$Z=5.5934R\left(\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}\right)+0.0565R\left(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\right)+0.0000R\left(\mathrm{R}\mathrm{e}\mathrm{d}\right)$$

(3)

where R(Blue), R(Green), and R(Red) are the reflectance of the blue, green, and red bands, respectively.

And secondly, the normalized chromaticity coordinates x and y were calculated from CIE tristimulus values X, Y, and Z as follows:

$$x=X/\left(X+Y+Z\right)$$

(4)

$$y=Y/\left(X+Y+Z\right)$$

(5)

Finally, the hue angle α can be found from the chromaticity coordinates (x, y) by using the bivariate arctangent function (arctan2) and the following equation:

$$\alpha ={\displaystyle \frac{\arctan 2(x-1/3,y-1/3)·180}{\pi }}+180$$

(6)

2.4. Evaluation Method

To evaluate the fitted model, three key performance metrics were considered: the coefficient of determination (R²), the mean absolute percentage error (MAPE), and the root mean square error (RMSE).

$$R^2={\displaystyle \frac{{\sum }_{i=1}^n{\left({\mathit{Biomass}}_{\mathrm{estimated},i}-\overline{{\mathit{Biomass}}_{\mathrm{measured}}}\right)}^2}{{\sum }_{i=1}^n{\left({\mathit{Biomass}}_{\mathrm{measured},i}-\overline{{\mathit{Biomass}}_{\mathrm{measured}}}\right)}^2}}$$

(7)

$$\mathrm{MAPE}={\displaystyle \frac{{\sum }_{i=1}^n\left(\left|\frac{{\mathit{Biomass}}_{\mathrm{measured},i}-{\mathit{Biomass}}_{\mathrm{estimated},i}}{{\mathit{Biomass}}_{\mathrm{measured},i}}\right|\right)}{n}}\times 100\%$$

(8)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^{\mathrm{n}}{\left({\mathit{Biomass}}_{\mathrm{measured},i}-{\mathit{Biomass}}_{\mathrm{estimated},i}\right)}^2}{n}}}$$

(9)

where Biomass_measured,i indicates the i-th measured value; $\overline{{\mathit{Biomass}}_{\mathrm{measured}}}$ denotes the mean of the measured values; Biomass_estimated,i is the i-th model-estimated value; n is the number of samples.

3. Results

3.1. Reflectance Conversion

The mean color values of the blue, green, and red bands were calculated for DRSPs with 1.2%, 5%, 10%, and 25% reflectance. Figure 5 shows a strong positive correlation between these mean color values and their corresponding standard reflectance values, with an R² value exceeding 0.99 for an exponential fit.

However, the ratio of the mean color values corresponding to the four DRSPs deviated from the theoretical reflectance ratio of 1.2:5:10:25, indicating a non-linear relationship between color values and surface reflectance. This observation aligns with prior research findings [34]. This deviation is likely due to the camera’s sensitivity to light intensity, highlighting the inherent limitations of consumer-grade cameras compared to high-performance spectral measurement instruments. Nevertheless, this limitation can be partially addressed by calibrating the imagery using DRSPs with known and varying reflectance levels. These results emphasize the importance of employing DRSPs with a range of reflectance values to enable a more accurate conversion of pixel-level color values to reflectance, thereby improving the reliability of image-based quantitative analyses.

3.2. Hue Angle Cutoff Threshold

The hue angles of different land cover types, including Suaeda salsa, mudflats, water bodies, and green vegetation (e.g., reeds), exhibit significant differences. Suaeda salsa generally has the highest hue angle, followed by mudflats, with green vegetation and water bodies displaying lower values.

To extract Suaeda salsa pixels from UAV visible-light imagery, a hue angle cutoff threshold can be established. While Suaeda salsa typically has a higher hue angle than mudflats, low-coverage Suaeda salsa pixels may have similar values to mudflat pixels. This suggests that the lowest hue angle of low-coverage Suaeda salsa or the highest hue angle of pure mudflat pixels could serve as potential cutoff thresholds. However, accurately identifying low-coverage Suaeda salsa pixels may be challenging. Therefore, using the highest hue angle of pure mudflat pixels as a cutoff threshold may provide a more reliable approach to extracting Suaeda salsa pixel information.

To determine a suitable cutoff threshold, we randomly selected pure mudflat areas from the UAV visible imagery and analyzed the distribution of pixel hue angles (Figure 6). The cumulative proportion of pixels with a hue angle of 249.01 exceeded 99.99%. Consequently, 249.01 was adopted as the minimum cutoff threshold for Suaeda salsa pixel extraction. Pixels with a hue angle greater than this cutoff threshold were classified as Suaeda salsa.

A cutoff threshold value of 249.01 was applied to extract Suaeda salsa information from the UAV visible imagery. The resulting extraction accurately reflected the actual distribution of Suaeda salsa (Figure 7a). Notably, this method demonstrated exceptional performance in both low-density (Figure 7b) and mixed areas with reeds (Figure 7c), highlighting its robustness and versatility.

3.3. Inversion Model Construction

A strong positive correlation was observed between Suaeda salsa biomass and hue angle. To model this relationship, various functions (linear, quadratic polynomial, exponential, and power) were evaluated. The exponential and power functions demonstrated superior performance, with the latter being selected as the optimal remote sensing inversion model for Suaeda salsa biomass in the experimental area (

Table 1. Accuracy evaluation of remote sensing inversion models of Suaeda salsa biomass.

Model Type	Model Expression	R2	MAPE (%)	RMSE (kg/m2)
Linear function	Biomass = −8.87436 + 0.03512 × α	0.93650	12.985	0.09966
Quadratic polynomial function	Biomass = 136.68861 − 1.07788 × α + 0.00213 × α2	0.99124	91.346	0.55293
Exponential function	Biomass = 3.57639 × 10−15 × e0.12201×α	0.99696	3.616	0.02183
Power function	Biomass = 4.51642 × 10−79 × α32.16447	0.99674	3.554	0.02258

, Figure 8).

Using the hue angle threshold for Suaeda salsa identification (α > 249.01), the above model was subsequently applied to estimate Suaeda salsa biomass across the experimental area (Figure 9). The resulting biomass maps accurately captured the spatial distribution of Suaeda salsa, effectively reflecting both sparse and dense stands, and showed strong agreement with field observations. Therefore, the proposed method demonstrates high accuracy in both identifying Suaeda salsa and estimating its biomass.

4. Discussion

4.1. Applying Vegetation Indices to Extract Suaeda salsa from UAV Imagery

To assess the spectral separability between Suaeda salsa and background vegetation dominated by Phragmites australis (reeds), NDVI and RVI values were calculated from red and near-infrared bands acquired by a UAV-mounted multispectral sensor, while the SVI (Suaeda salsa Vegetation Index, defined as SVI = green − (blue + red)/2) was derived from RGB imagery captured by a visible-light camera. Approximately 400,000 pixels of Suaeda salsa and 600,000 pixels of reeds were randomly sampled from their respective distribution areas for spectral analysis.

As shown in Figure 10a, the NDVI values of both Suaeda salsa and reeds exhibited a clear normal distribution. However, the value ranges largely overlapped, making it difficult to reliably differentiate the two species using the NDVI alone. Specifically, NDVI values for Suaeda salsa ranged from 0.00002 to 0.44736 (mean = 0.22369), while those for reeds ranged from 0.00006 to 0.48732 (mean = 0.24369). A similar pattern was observed in the RVI results (Figure 10b), with Suaeda salsa values ranging from 1.00005 to 2.61901 (mean = 1.80953) and those for reeds ranging from 1.00013 to 2.90110 (mean = 1.95061). These results indicate that the NDVI and RVI alone are insufficient for reliably distinguishing Suaeda salsa from reeds.

The SVI, as a vegetation index specifically designed for identifying Suaeda salsa, is typically used after the NDVI has been applied to extract all vegetation pixels and serves to further separate Suaeda salsa from other vegetation such as reeds [8]. In this study, SVI values for Suaeda salsa ranged from −0.07115 to 0.00404 (mean = −0.03356), whereas those for reeds ranged from −0.03273 to 0.03903 (mean = 0.00315), clearly higher than those of Suaeda salsa (Figure 10c). While the SVI offers some ability to distinguish between the two vegetation types, the presence of overlapping regions and the proximity of peak values reduce classification accuracy.

In contrast, hue angle values exhibited a more distinct separation between Suaeda salsa and reeds (Figure 10d). The peaks were well separated, with minimal overlap, suggesting superior classification performance. The hue angle values for Suaeda salsa ranged from 217.4529 to 278.181 (mean = 247.817), while those for reeds ranged from 206.2238 to 254.7813 (mean = 230.5025). Further analysis suggests that although the selected regions had high vegetation coverage, small gaps of exposed soil were inevitable. The minor overlap observed in Figure 10d is likely due to soil pixels within both sampling regions, which can be effectively excluded during hue angle-based classification through thresholding. In contrast, the overlap observed in Figure 10c may be partially due to soil pixels but primarily results from the spectral similarity between Suaeda salsa and reeds.

In summary, the NDVI and RVI fail to provide effective differentiation between Suaeda salsa and reeds. Although the SVI demonstrates improved performance, it still suffers from overlapping distributions and close peak values. Hue angle offers the most distinct separation, with the widest gap between distribution peaks and the smallest overlapping region, making it the most effective indicator for distinguishing Suaeda salsa from reeds in this study.

4.2. Factors Influencing Quantitative Inversion Using UAV Visible Imagery

Quantitative UAV remote sensing has become increasingly prevalent in ecology and environmental studies. While hyperspectral and multispectral sensors are commonly employed, visible-light cameras, though limited in spectral range, offer unique advantages, particularly for identifying red Suaeda salsa. Given their widespread availability as standard components of UAVs, visible-light cameras represent a promising tool for quantitative inversion studies.

To ensure standardized spectral results, exposure parameters (aperture, shutter speed, and ISO) were fixed during data acquisition. Additionally, DRSPs were used to convert raw pixel values into reflectance values. However, variations in camera sensitivity among different UAVs can lead to discrepancies in reflectance and hue angle measurements. Furthermore, weak light conditions can significantly reduce the signal-to-noise ratio of visible-light imagery, increasing the error in reflectance and chromaticity angle calculations.

To mitigate these issues, it is recommended to use the same UAV and camera model for consistent data acquisition within a specific monitoring area. Optimal imaging conditions should be selected to ensure sufficient light. When monitoring areas or periods change, adjustments may be necessary to the hue angle cutoff thresholds for Suaeda salsa identification or biomass inversion models to maintain accurate results.

5. Conclusions

In this study, we propose a novel method for the remote sensing inversion of Suaeda salsa biomass based on the hue angle derived from UAV-acquired visible imagery. Owing to its distinct chromatic characteristics, Suaeda salsa can be effectively distinguished from mudflats, water bodies, and other green vegetation such as reeds, resulting in a significant spectral separation in hue angle.

To establish a quantitative relationship between hue angle and Suaeda salsa biomass, drone-based reflectance standard panels (DRSPs) were employed to convert raw pixel values into reflectance and hue angle values. A hue angle threshold of 249.01 was identified for reliably extracting Suaeda salsa pixels, demonstrating robust performance even in heterogeneous environments where the species is interspersed with reeds or present at low density.

A strong exponential relationship was established between hue angle and Suaeda salsa biomass, described by the model Biomass = 3.57639 × 10⁻¹⁵ × e^0.12201×α, with a coefficient of determination (R²) of 0.99696, a mean absolute percentage error (MAPE) of 3.616%, and a root mean square error (RMSE) of 0.02183 kg/m². The resulting biomass inversion maps exhibited a high degree of agreement with field measurements, accurately reflecting the spatial heterogeneity of biomass distribution.

As a foundational species in coastal salt marshes, Suaeda salsa plays a critical role in maintaining ecosystem stability and biodiversity. This study demonstrates the utility of low-cost UAV visible imagery in overcoming the spatial resolution and atmospheric constraints associated with satellite-based remote sensing, significantly improving the timeliness and precision of wetland monitoring. In addressing the limitations of conventional vegetation indices such as the NDVI and RVI for species-level classification, we developed and validated a hue angle-based identification method with superior discrimination capability and quantitative accuracy. This approach offers key technical support for the development of high-resolution, intelligent monitoring systems.

Future research will focus on expanding the application of UAV-based visible imagery in ecological monitoring and advancing toward a standardized and scalable framework for supporting ecosystem assessment, management, and conservation.