UAS Remote Sensing for Coastal Wetland Vegetation Biomass Estimation: A Destructive vs. Non-Destructive Sampling Experiment

Abstract

Coastal wetlands are critical ecosystems that require effective monitoring to support conservation and restoration efforts. This study evaluates the use of small unmanned aerial systems (sUAS) and multispectral imagery to estimate aboveground biomass (AGB) in tidal marshes, comparing models calibrated with destructive versus non-destructive in situ sampling methods. Imagery was collected over South Carolina’s North Inlet-Winyah Bay National Estuarine Research Reserve, and vegetation indices (VIs) were derived from sUAS imagery to model biomass. Stepwise linear regression was used to develop and validate models based on both sampling approaches. Destructive sampling models, particularly those using the Normalized Difference Vegetation Index (NDVI) and Difference Vegetation Index (DVI), achieved the lowest root mean square error (RMSE) values (as low as 70.91 g/m²), indicating higher predictive accuracy. Non-destructive models, while less accurate (minimum RMSE of 214.86 g/m²), demonstrated higher R² values (0.44 and 0.61), suggesting the potential for broader application with further refinement. These findings highlight the trade-offs between ecological impact and model performance, and support the viability of non-destructive methods for biomass estimation in sensitive wetland environments. Future work should explore machine learning approaches and improved temporal alignment of data collection to enhance model robustness.

1. Introduction

Coastal wetlands are a vital component of healthy coastal ecosystems, providing critical services such as aquatic habitats, carbon sequestration, shoreline stabilization, and flood mitigation [1]. These ecosystems play an irreplaceable role in maintaining biodiversity, protecting coastal areas from erosion, and buffering climate change impacts [2]. As climate change intensifies, it is becoming crucial to prioritize conservation, preservation, and restoration projects that aim to safeguard and enhance these fragile yet essential landforms. A major component of such projects is often monitoring vegetation health and biomass, especially looking for changes over time [3,4,5].

Various methods are employed to measure the health and well-being of marsh environments and vegetation. Transect mapping is a common method used to examine vegetation types within defined study areas along a path [6,7]. This approach facilitates vegetation measurement and usually is non-destructive in nature, meaning vegetation is not removed from its habitat. This is especially important for sensitive ecosystems where vegetation plays important roles or is being negatively impacted already through anthropogenic disturbances. While effective, transect mapping has limitations in particularly sensitive ecosystems, where its potential to cause disturbance underscores the need for alternative non-invasive methods [7]. Furthermore, it can be time-consuming for projects covering large geographic extents.

Remote sensing has emerged as a powerful tool for studying vegetation and ecosystems. High-resolution satellite and aerial imagery and those from small unmanned aerial systems (sUAS) are increasingly utilized to collect data, create models, and improve the accuracy of restoration and conservation efforts. Among the most popular applications is the use of machine learning technologies combined with remote sensing to produce classification maps of different species [8,9,10,11,12,13,14,15,16]. Remote sensing techniques have also been used to estimate aboveground biomass (AGB) from multispectral imagery [7,17,18,19] and to measure fractional vegetation cover [20,21,22]. Additionally, hyperspectral imagery has demonstrated high accuracy in classifications, up to 97% [13], while the integration of LiDAR has enhanced vegetation classification and modeling capabilities [6,12,23,24].

Biomass modeling studies have focused on building the relationships between in situ sampling and spectral properties on imagery. The most recent studies focus on smaller scale remote sensing applications such as Sentinel-2 or Landsat 8 imagery [25,26,27]. Although freely available, the large areas that the images cover, especially at medium to large resolution sizes, make it difficult to produce an accurate model that is applicable to wetland areas with high vegetation [26]. Higher resolution satellite platforms with multispectral sensors could be used, but temporal visitation can be unideal, and monetary costs can be very expensive [28]. Due to the noted drawbacks of satellite remote sensing platforms, small unmanned aerial systems (sUAS) have risen as a reliable method to produce AGB estimation models over small areas of heterogeneous wetlands [18]. Data gathered from sUAS can provide cm level spatial resolution when studying highly heterogeneous wetland vegetation at finer scales. The combination of high spatial resolution, a multispectral sensor, and reliable in situ data can make for a powerful combination for modeling tidal marsh biomass [17,18,29]. While results varied, progress has been made in developing better models for biomass estimation.

Regardless of the remote sensing platform used, in situ destructive sampling is a critical component of modeling using remote sensing data, and it provides ground truthing data for model calibration and validation. This involves the permanent removal of vegetation to directly measure dry biomass, which serves as ground truth for developing and testing predictive models [17,19,30,31,32]. In short, the destructive measurement samples are used to show a relationship with vegetation indices, which is used to model biomass estimates for the rest of the marsh. Destructive samples are typically very accurate because they are measured from actual vegetation taken from the area of interest. However, Shew et al. evaluated multiple destructive methods and found that four of the five biomass estimation techniques tested produced either under or overestimates [31]. Furthermore, the destructive nature of this sampling technique results in the modification of the environment and is quite intrusive.

Non-destructive biomass sampling brings the idea of ground truth sampling via allometric equations based on biophysical or empirical models instead of in situ vegetation mass clipping. It has been adapted in field ecology [33,34,35,36]. For example, Morris et al. [37] adapted [38]’s foundational work on using allometric equations for biomass estimation to a study site in coastal marshes with S. alterniflora to estimate biomass over a long period of time and the results have been enlightening. However, it has not been extensively used to inform remote sensing-based vegetation biomass estimation modeling [29]. Non-destructive biomass modeling is possible through the use of techniques, but this is usually limited to small areas [33,34,35,36]. Non-destructive biomass samples are created through allometric equations based on vegetation height. Sampling can be performed in quadrats, like with destructive sampling, but it is far less intrusive since there is no cutting of the plants within the plot.

This research explores the potential of non-destructive sampling for remote sensing biomass modeling in the literature by directly comparing the calibration accuracy of destructive and non-destructive sampling for AGB estimation models based on vegetation indices derived from sUAS multispectral imagery. We propose that the sUAS-based non-destructive sampling can assist accurate biomass estimation in vast geographical areas of coastal environments.

2. Materials and Methods

2.1. Study Area

The North Inlet-Winyah Bay National Estuarine Research Reserve (NERR), situated on the Atlantic coastline just north of Georgetown, SC, USA, is known as a well-protected estuary without much anthropogenic influence (Figure 1). Specifically, the North Inlet Estuary is approximately 7655 Ha of pristine tidal marsh wetlands. The estuary is dominated by smooth cordgrass, also known as Sporobolus alterniflora, growing in its intertidal zone. The reserve is unique due to limited anthropogenic influence and its placement relative to the inlet. It provides important feeding and breeding grounds for threatened several species, including sea turtles, sturgeons, least terns, and wood storks. Two study sites within the high and low marsh regions were selected due to their involvement in previous studies and their spatial proximity to NSF-funded Long Term Research plots where biomass data are collected monthly. The sites are known as the Goat Island (GI) and Oyster Landing (OL) study sites. Both sites included regions of high marsh where plant material height is shorter, and low marsh where plant material is much taller and thicker. The diversity of each area represents well the biological makeup of a coastal estuary along the South Carolina coast.

2.2. Data Collection

2.2.1. sUAS Data Collection

sUAS imagery data were collected using a DJI Matrice 100 (SZ DJI Technology Co., Ltd., Shenzhen, China) with two batteries and a Micasense Red Edge-M sensor with 5 bands: blue (475 nm), green (560 nm), red (668 nm), red edge (717 nm), and near-infrared (842 nm) on 22 September 2022. The sensor was calibrated using a reflectance panel before and after data collection. The surface reflectance images captured from calibration were later used in imagery processing. The sUAS came equipped with a Global Navigation Satellite System (GNSS) receiver.

Each site required a variable flight path and time, though altitude was held constant at 80 m above ground with a 5 m/s flight speed. The altitude and data processing resulted in a ground sampling distance (GSD) of 5 cm on the orthoimage. The sUAS captured 516 images at the OL site, and 550 images at the GI site. The front overlap was extended to 85% while the sidelap was at 50% by default to ensure that orthomosaics and point clouds could be computed using the structure from motion (SfM) algorithm. Fights at both sites were launched on 22 September 2022. The first flight was conducted from 11:56 a.m. to 12:18 p.m. EST at the GI site, and the second flight was conducted from 1:01 p.m. to 1:18 p.m. EST at the OL site. Both were as close to low tide (11:49 a.m.) as was logistically possible on that day. The weather during each flight was sunny, without a significant impact of wind on the performance of the drone. With low tide, we visibly verified sun glint was not a significant influence on the data collection of marsh fields.

Ground control points (GCPs) were collected using an Emlid (Esztergomi út 31-39. HUB3. ép. 5. em., Budapest, Hungary) Reach RS2 RTK GNSS bundle (base station and rover) unit at the two sites. Local NOAA benchmark survey markers were used as a base station location to ensure accurate GNSS data collection. Ground control points were placed along walkways previously built in the fields. A few GCPs were placed in areas of high marsh, very close to shore, to expand GCP coverage around the study areas. Seven GCPs at each site were collected in this study.

2.2.2. Non-Destructive Biomass Data Collection

Non-Destructive Biomass of S. alterniflora in the NIWB estuary has been recorded using allometric equations since 1984 as part of a Long-term Research in Environmental Biology (LTREB) project [39,40]. The biomass measurements are determined using measured vegetation stem heights. While not every stalk is measured within each one meter by one meter plot, two subplots measuring 10 cm × 15 cm have the vegetation labeled with bird id bands and are measured every month (Figure 2). There are a total of 30 subplots used and collected. Dickerman, Stewart, and Wetzel [38] developed the original allometric equations for biomass estimation, which were later applied to the LTREB study site by Morris and Haskin [37]. The measurements are inputs into the allometric equation, resulting in a biomass measurement for the sub plot. It is important to note that six of the plots have been fertilized by researchers upkeeping the project, including with phosphorus (15 mol P/m²/y) and nitrogen (30 mol N/m²/y) each year, where vegetation is more dominant than in the remaining 24 plots.

In the time span of this study, non-destructive biomass data were collected on 23 August 2022 and 3 October 2022. The sUAS data were not able to be collected during the same time as data collection. The 3 October data were collected only 12 days following drone flights, but Hurricane Ian made landfall nearby on 30 September 2022, influencing the region and destroying some vegetation. The decision was made to use the data from 23 August since the vegetation had no large disturbance events in the 29 days between biomass measurements and sUAS flight. Despite 29 days being a nearly month-long gap, S. alternifora biomass is near its peak in both months and varies very little, on average, between August and September before a larger difference in October [29]. Therefore, the time difference should not impact our measurements and modeling significantly. During the field experiments, the same Emlid Reach RS2 RTK GNSS base and rover bundle unit was used to collect GCPs at the centroid for each of the 30 subplots of interest with centimeter-level accuracy.

2.2.3. Destructive Biomass Data Collection

A total of 65 destructive biomass samples were collected during a multi-day field study at both sites from September 21 to September 24. Each sample was randomly determined in a general u-shaped pattern by walking through both the high and low marsh sections of the study sites (Figure 3). This was performed to provide spatially diverse samples of both low and high marsh vegetation. A 0.5 × 0.5 m quadrat was placed in the field for each selected site. All marsh plants within the quadrat were clipped, bagged, labeled and then thoroughly cleaned before further processing. The biomass of each location was determined by weighing the vegetation after drying it in a large-cabinet oven at 70 °C for one week. The sample weights were recorded and added to a database for model calibration and validation.

2.3. Approaches

2.3.1. sUAS Data Processing

sUAS data was processed using Pix4D Mapper 4.6.2. Data were processed for each site separately into orthomosaics, point clouds, and digital terrain models (DTM) by using all flight photos as input into the Pix4D software. Pix4D recognizes the camera type, parameters, and other default information from the images to process them appropriately. GCPs were used to accurately georeference the SfM derived products. Each GCP was identified on at least 6 images before processing using the Pix4D software to enhance georectification. Images taken of a reflectance panel with known reflectance were used as an input to the software package for each wavelength to create percent reflectance maps and calibrate the sUAS imagery. No additional processing was needed for generating corrected reflectance maps. Each dataset’s DTM was derived from the Pix4D software as well, without any external computations. Processing was completed on a Dell Inspiron 5680 with 16 GB ram and 4 gb ddr4 video card. The processing took 4.5 h for each orthomosaic to be produced. Red, green, blue, red edge, and near-infrared bands were all included in the final orthomosaics. Each band was then used to create VI maps. The resulting data products were used in each step as described in Figure 4 and further represented in the rest of Section 2.3.

2.3.2. sUAS-Based Biomass Modeling

Vegetation indices (VIs) for this study were computed from the calibrated RedEdge-M bands of surface reflectance.

Table 1. List of vegetation indices.

Index	Equation	Sources
Normalized Difference Vegetation Index	${\displaystyle \frac{NIR-RED}{NIR+RED}}$	[30,44,45]
Visible Atmospherically Resistant Index 2	${\displaystyle \frac{GREEN-RED}{GREEN+RED-BLUE}}$	[46,47]
Soil-Adjusted Vegetation Index 1	${\displaystyle \frac{NIR-RED}{NIR+RED+L}}\ast \left(1+L\right)$	[13,45]
Simple Ratio	Ratio between any two bands. Example: ${\displaystyle \frac{REDEDGE}{BLUE}}$	[48,49]
Chlorophyll Index Green	${\displaystyle \frac{NIR}{GREEN}}-1$	[30]
Chlorophyll Index RedEdge	${\displaystyle \frac{NIR}{REDEDGE}}-1$	[30]
Difference Vegetation Index	$NIR-\left(0.96916 \ast RED\right)$	[45]
Enhanced Vegetation Index 2	$2.5 \ast {\displaystyle \frac{NIR-RED}{NIR+2.4 \ast RED+1}}$	[30]
Excess Green Index 2	$2 \ast GREEN-\left(RED+BLUE\right)$	[47]
Relative Green Index 2	${\displaystyle \frac{GREEN}{RED+GREEN+BLUE}}$	[47]
Green Normalized Difference Vegetation Index	${\displaystyle \frac{NIR-GREEN}{NIR+GREEN}}$	[30]
Green-Red Vegetation Index 2	${\displaystyle \frac{GREEN-RED}{GREEN+RED}}$	[47]
Modified Normalized Difference Aquatic Vegetation Index	${\displaystyle \frac{REDEDGE-BLUE}{REDEDGE+BLUE}}$	[49]
Modified Normalized Difference Vegetation Index	${\displaystyle \frac{REDEDGE-RED}{REDEDGE+RED}}$	[49]
Modified Water-Adjusted Vegetation Index 1	$\left(1+L\right) \ast {\displaystyle \frac{REDEDGE-BLUE}{REDEDGE+BLUE+L}}$	[49]
Normalized Difference Green Index	${\displaystyle \frac{NIR-GREEN}{NIR+GREEN}}$	[48]
Normalized Difference Red Edge Index	${\displaystyle \frac{NIR-REDEDGE}{NIR+REDEDGE}}$	[48]
Ratio Vegetation Index	${\displaystyle \frac{NIR}{RED}}$	[45]
Triangular Greenness Index 2	$GREEN-\left(0.39 \ast RED\right)-\left(0.61 \ast BLUE\right)$	[46]

¹ $L=0.5$ ² RGB only indices.

reveals the list of indices computed and explored for biomass estimation in this study. Each index was generated in ArcGIS Pro 3.2 using the equations listed in Table 1, appropriate image bands and the raster calculator tool. Both red, green, and blue- (RGB) based VIs and multispectral VIs were considered for each model. RGB-based VIs are being used more often in response to the availability of off-the-shelf sUAS without multispectral sensors capturing in the red edge or NIR wavelengths [41]. However, they have not yet shown to be as effective for biomass estimations as studies with destructive samples [42,43]. The zonal statistics as table tool was used with 1 m × 1 m square buffers to extract the VI value around each plot location.

After extracting the VIs to a database, a correlation matrix was created to determine which indices could potentially contribute unique information to a biomass model. Investigation of single index scatter plots with the biomass data revealed several VIs with an observed positive relationship between VI and the measured biomass. This was performed for both the destructive and non-destructive biomass datasets. Only 20 random data points of the 30 non-destructive sites were used for training and data analysis, and 43 out of 65 of the destructive biomass samples were used from across both study sites. The remaining data points were used for validation.

After the initial exploration of correlation, all variables were input into a stepwise linear model that removed underperforming variables to create the best performing models in IBM SPSS software 31. To accomplish this, the extracted biomass values and vegetation index numbers were compiled into columns and pasted into SPSS software. We performed the stepwise linear model with default parameters using the stepwise linear model function in SPSS. The results provided the model equation and best performing factors. For both non-destructive and destructive biomass sample dataset models, the best performing models with a single VI and with the two best performing variables were investigated further. The best performing models were determined by the coefficient of determination (R²) and statistical significance of the variables.

Once the best fit models were generated, they were applied to the remote sensing imagery using raster calculator in ArcGIS Pro to create biomass maps. These maps estimate the amount of biomass using the remotely sensed imagery. Validation was performed using the remaining samples and estimates of biomass from the biomass maps. The zonal statistics as table tool in ArcGIS Pro and 1 m × 1 m square polygons around the validation sample locations were used to extract the modeled biomass values. These values were compared to the sampled values using the root mean squared error (RMSE) metric commonly used to assess error in remote sensing and modeling analysis:

$$RMSE=\sqrt{{\displaystyle \frac{{{\displaystyle \sum }}_{i=1}^n{\left(x_i-{\widehat{x}}_i\right)}^2}{n}} }$$

where n = the number of validation biomass points, x_i = the ground truth, known non-destructive or destructive biomass values, and ${\widehat{x}}_i$ = the estimated biomass from the generated biomass maps. A smaller RMSE value represents a better agreement between the modeled and ground surveyed biomass. RMSE was calculated from the validation dataset only.

3. Results

3.1. Biomass Values and Correlations for Destructive and Non-Destructive Samples

Upon investigation, biomass values for the destructive samples and non-destructive samples were within similar ranges, though the non-destructive biomass values were higher (Figure 5). The majority of non-destructive biomass values were between 300 g/m² and 900 g/m², while the majority of destructive biomass ground truth values were between 300 and 650 g/m². The mean of the destructive biomass values was 421.99 g/m² and the mean of the non-destructive biomass values was 654.91 g/m². The higher mean biomass values for the non-destructive areas are due to being in predominately low marsh areas with taller plant material, as well as the fertilization in six of the plots.

Interestingly, of the several VIs explored for destructive and non-destructive sampling, not many of the indices indicated a strong positive correlation with above-ground biomass of S. alterniflora. The lack of strong, positive correlations could be the result of numerous factors, including seasonality and timing of data collection, environmental factors, and vegetation types. Several of the variables were strongly correlated with each other, and this was noted in the later stages of stepwise linear modeling as many indices correlated strongly with biomass were also strongly correlated with each other.

There was little overlap between indices strongly correlated with destructive biomass measurements and non-destructive biomass measurements. For destructive sampling, the normalized difference vegetation index (NDVI; 0.61) provided the strongest positive correlation, and several others such as the relative green index (RGI; 0.60) and the ratio vegetation index (RVI; 0.57) were close behind. It was a surprise to see RGB VIs and NIR/RE VIs both showing fairly strong correlations (Figure 6).

The non-destructive samples had far fewer indices with strong positive correlations, with only 4 showing any sort of use. The highest performing VIs were NIR-based indices, with RGB indices close behind in some regard, like the triangular greenness index (TGI; 0.49) (Figure 7). The highest performing NIR-based VIs were the Difference Vegetation Index (DVI; 0.66), Enhanced Vegetation Index 2 (EVI2; 0.63), and the soil adjusted vegetation index (SAVI; 0.62). These four were also fairly strongly correlated. Only one of the stronger positive correlation indices—SAVI—was used in other studies for biomass of S. alterniflora [50].

3.2. Models and Maps for Destructive and Non-Destructive Samples

The eventual models used the strongest correlated variables as determined by their influence in the stepwise linear model (

Table 2. Model structures with variables, equations, and R² values.

Models	Variables	Equation	R2
Destructive Samples Model 1	NDVI	AGB = −115.094 + 1136.393(NDVI)	0.38
Destructive Samples Model 2	NDVI, DVI	AGB = −147.314 + 1716.872(NDVI) − 3759.493(DVI)	0.46
Non-Destructive Samples Model 1	DVI	AGB = −395.775 + 25792.653(DVI)	0.44
Non-Destructive Samples Model 2	DVI, CIG	AGB = −862.319 + 61084.777(DVI) − 748.207(CIG)	0.61

). In Table 2, the best performing models, including variables and equations for both destructive and non-destructive samples, are displayed with their R² values. A best performing model with only one variable and with two variables were both used to show the influence of having both a parsimonious model and one with multiple factors. The non-destructive models had the highest R² values, with the single variable model reaching R² = 0.44 and the two-variable model reaching R² = 0.61. The destructive models were lower, with the one-variable model reaching R² = 0.38 and the two variable model reaching R² = 0.46 (Figure 8). No other variables were considered because of how correlated the other variables were with each other.

Biomass maps for each of the sites from the best performing models for destructive and non-destructive samples are shown below in Figure 9 and Figure 10. Biomass values are within ranges seen among other studies [19,32,51]. However, biomass estimates are much lower in the biomass maps derived from the models designed from the non-destructive models, as seen with the warmer colors. Biomass plots that have been fertilized as part of the longitudinal study in the region are well distinguished from the spartina alterniflora grasses around them, even in the non-destructive model.

The destructive models each show spatial patterns that we would expect. That is, higher biomass to the northwest, a small decrease as you move further into the marsh where high marsh vegetation is found, and then a significant increase as you move closer to the low marsh to the southeast of the map. This pattern is not strongly represented in the map derived from the non-destructive model.

3.3. Model Validation

Validation for each of the models was performed using the remaining samples not used in training. The RMSE values for each of the four best performing models, one with a single variable and one with two for both destructive and non-destructive samples, are shown in

Table 3. Models with variables and associated RMSE values.

Models	Variables	Validation RMSE (g/m2)
Destructive Samples Model 1	NDVI	73.17
Destructive Samples Model 2	NDVI, DVI	70.91
Non-Destructive Samples Model 1	DVI	344.73
Non-Destructive Samples Model 2	DVI, CIG	214.86

. The lowest RMSE values (indicating a more effective model) came from the destructive sample-based models and reached as small as 70.91 g/m². The non-destructive models were able to achieve an RMSE as low as 214.86 g/m².

4. Discussion

This study evaluated biomass estimation in a coastal marsh using VIs derived from sUAS imagery and modeled against both destructive and non-destructive sampling methods. Among all models, the two destructive methods yielded the best performance by way of the validation measurement chosen. The simplest of these, using NDVI alone, resulted in an RMSE of 73.17 g/m². Model performance improved modestly when the difference vegetation index (DVI) was added, lowering the RMSE to 70.91 g/m². However, despite lower RMSE values, both destructive models exhibited relatively low R² values: 0.38 for the NDVI-only model and 0.46 for the NDVI + DVI model.

NDVI is widely accepted in the remote sensing and ecological literature as a robust indicator of vegetation greenness, which is often strongly correlated with AGB [17]. Doughty and Cavanaugh found that NDVI-based models performed best across both seasonal and annual analyses, with the lowest seasonal RMSE reported at 344.3 g/m² and the annual model at 495.9 g/m² [17]. This relationship likely explains the strong performance of the destructive models, which were calibrated using direct biomass measurements. In contrast, the DVI and DVI + chlorophyll index green (CIG) models calibrated with non-destructive AGB estimates showed higher R² values (0.44 and 0.61, respectively), but these are likely inflated due to circularity. Specifically, the non-destructive biomass values were themselves estimated through prior modeling procedures, which introduces non-independence between predictor and response variables and can lead to artificially high model fit. Furthermore, the spatial distribution of non-destructive sampling points was closer together and did not include significant amounts of low marsh (higher biomass). Low marsh vegetation is the taller vegetation with higher biomass, and it is found closer to streams and waterbodies within the marshes. The high marsh vegetation is shorter, with lower biomass, and is more likely to be found closer to shore. The constraints of the environment led most non-destructive sample points to being along a boardwalk closer to shore, with more high marsh than low marsh. This seems to have led to a model that was inclined to perform well around where the non-destructive samples were taken, while the low marsh estimates struggled.

Compared to other recent studies estimating AGB in wetland environments, both of our destructive models yielded lower RMSE values. For example, Ge et al. used stepwise linear regression with Sentinel-2 and JL-1 satellite data, achieving validation RMSEs of 216.86 g/m² and 244.96 g/m², respectively [32]. These higher RMSE values can be attributed to the use of resolutions greater than 1 m which introduces error from mixed pixel values. However, Niu et al., who combined in situ sampling with sUAS imagery, reported RMSEs of 365.67 g/m² (multivariate linear regression) and 330.51 g/m² (random forest) [19]. Even with using 2 cm resolution imagery, their RMSE values remained high because of potentially limited field sampling for model calibration. In contrast, Curcio, Barbero, and Peralta achieved an RMSE of 79.05 g/m² in a habitat-specific autumn biomass model [52]. This value is similar to the values from our models, but theirs relied on the Red-Green Ratio Index (RGRI), which is based solely on red and green bands, while our models used NDVI, incorporating near-infrared reflectance—a band more strongly correlated with vegetation structure and vigor. With these studies and our own, it has been shown that using remote sensing VIs calibrated from AGB measurements provide a suitable technique for modeling on a larger scale where traditional allometric equations prove more difficult.

We also evaluated non-destructive sampling methods based on vegetation stem height as a calibration strategy for biomass modeling. This approach was motivated by the ecological sensitivity of coastal wetlands [1], where repeated destructive sampling is neither sustainable nor practical. While the stem height measurements used in our study originate from a long-term dataset with known reliability, their use as calibration input for novel VI-based models introduced limitations. Specifically, because these non-destructive estimates were themselves derived from prior models, their use likely introduced circularity, weakening the independence of validation metrics. Nonetheless, RMSE values from the non-destructive models remained within acceptable error margins and, in some cases, outperformed those reported in studies that relied on coarser-resolution or satellite-derived inputs [19,32]. Where destructive sampling is not feasible, we recommend that future remote sensing studies explore rigorous non-destructive calibration methods that prioritize both ecological preservation and statistical validity.

Although our models achieved relatively low RMSE values, we restricted our analytical framework to stepwise linear regression. Several recent studies employing machine learning approaches such as random forests or support vector machines have reported superior performance, particularly when working with high-dimensional or nonlinear vegetation data [19]. Additionally, our data collection was temporally constrained: non-destructive stem height data were collected in August (late summer), whereas both destructive samples and sUAS flights were conducted in September (early fall). This temporal mismatch likely contributed to the stronger correlation observed between sUAS-derived indices and destructively sampled AGB, since both were captured during the same phenological stage. Future research should consider aligning sampling timelines more closely, or using time-series data, to reduce temporal bias and better evaluate seasonal dynamics in biomass accumulation [53]. Finally, although this study focused on salt marsh vegetation, the applicability of these methods to freshwater marshes remains uncertain. Expanding such models to diverse wetland types could support more comprehensive conservation and restoration efforts. Future research should explore both destructive and non-destructive biomass sampling methods, as well as drone flights and indices, to freshwater marshes.

5. Conclusions

The goal of this experiment was to develop a sUAS-based model calibrated by non-destructive in situ sampling to aid in wetland monitoring. This study evaluated the predictive accuracy of employing two distinct sampling strategies to calibrate multiple linear models based on remotely sensed VIs. NDVI in combination with data from destructive sampling produced the most accurate biomass models with the lowest RMSE across all models tested. Non-destructive methods were able to explain a greater variation in the biomass data but lacked accuracy, showing higher RMSE values.

While destructive calibration provided the lowest RMSE in validation datasets, non-destructive sampling offered a useful alternative, with performance supported by other comparable models [18,52]. Because wetland ecosystems are increasingly threatened by climate change, this study provides a viable alternative for estimating AGB without the increased time and labor costs associated with destructive methods.

While our study relied on linear regression, future work should employ machine learning and artificial neural networks to increase predictive accuracy. Additionally, accounting for phenological stage variation may improve the temporal consistency of seasonal AGB estimates. This study contributes to current conservation efforts by demonstrating that non-destructive sampling can effectively calibrate practical models in sensitive wetland environments.