A New Sentinel-2 Spectral Index for Mapping Hydrilla verticillata in Shallow Freshwater Lakes in Florida, USA

Abstract

Hydrilla, an invasive submerged macrophyte that is classified as a noxious weed in the U.S., can quickly spread into extensive monospecific infestations, excluding other native macrophytes and disrupting entire lake ecosystems. In Florida, infestation has increased tenfold in just three years, consuming over 60% of total management costs and requiring millions of dollars in annual control efforts. Traditional monitoring methods, such as field sampling, provide accurate localized assessments but are expensive and time-consuming. This study leverages Sentinel-2 satellite imagery, introducing the Submerged Aquatic Vegetation Index for Hydrilla (SVI_H), a novel three-band index utilizing the green (G, 560 nm), red-edge 1 (RE1, 705 nm), and shortwave infrared 1 (SWIR1, 1610 nm) bands to distinguish hydrilla from water and emergent aquatic vegetation (EAV) in two Florida lakes. The index, coupled with other vegetation indices, was validated using in situ measurements of hydrilla abundance levels, confirming its strong ability to accurately distinguish hydrilla. At the highest abundance level, SVI_H produced the highest Mathew correlation coefficients (MCCs), i.e., >0.86 for Lake Yale (2021), and >0.60 (2020) and >0.68 (2021) for Lake Apopka, using three thresholding methods. For Apopka (2022), other tested indices such as MFI and FAI yielded high MCC values along with high recall using incremental search threshold. However, these indices could not distinguish EAV from SAV in the eastern regions of Lakes Apopka and Yale, where EAV was dominant. These findings encourage the use of SVI_H for routine hydrilla detection and mapping, facilitating improved management, conservation efforts, and targeted herbicide applications.

1. Introduction

Florida’s freshwater systems cover over 18% of the state’s surface, including approximately 7800 lakes. Many are shallow, warm, and polymictic lakes formed from sinkholes in the peninsula’s karst topography [1,2], with some being naturally eutrophic due to phosphatic deposits, while many others are anthropogenically enriched [2,3]. Native and non-native macrophytes are dominant primary producers of these enriched systems, providing food sources and habitat structures for aquatic wildlife with a unique diversity of functional groups exhibiting traits to thrive in freshwater aquatic environments [4]. While emergent aquatic vegetation (EAV) and floating aquatic vegetation (FAV) establish foliar canopies in the ambient environment above the water surface, submerged aquatic vegetation (SAV) must endure the light-attenuated environment of the water column [5,6].

Hydrilla (Hydrilla verticillata (L. f.) Royle) is an invasive SAV that has become one of the most significant threats to freshwater ecosystems in Florida since its introduction from South Asia in 1951 via aquarium trade [7]. This species is known for its rapid proliferation and ability to form thick monotypic surface mats that effectively block sunlight, thereby outcompeting and displacing native aquatic plants [8]. The dense hydrilla canopies severely disturb the ecological balance of invaded water bodies, obstruct water navigation, disrupt recreational activities, and reduce the economic value of the affected water bodies [9,10]. Mapping the spread of hydrilla, both spatially and temporally, is vital for invasive aquatic plant control agencies to minimize removal costs and better understand its impact on the lake’s overall health.

Field lake surveys, such as rake sampling and hydroacoustics, have been fundamental to effective aquatic plant management for decades [11,12,13]. Nevertheless, these practices are resource- and time-intensive. Satellite-based remote sensing addresses these limitations by providing cost-effective and rapid large-scale monitoring of aquatic ecosystems and enabling timely detection of changes such as shifts in aquatic vegetation types and distribution as well as water clarity changes [5,14,15,16,17]. Combined with in situ monitoring, remote sensing enhances our ability to observe and characterize SAV growth across broad spatial and temporal scales. Advances in remote sensing and earth observation make high-resolution imagery widely accessible, with frequent updates allowing for regular and efficient lake management [18,19].

Satellite selection involves trade-offs in spatial, spectral, and temporal resolution [15,20]. For example, the Moderate Resolution Imaging Spectroradiometer (MODIS) offers high spectral but low spatial resolution (250–1000 m) images. Commercial satellites typically provide high spatial resolution but are costly and often do not have spectral bands available, like in coarser resolution imagery [9]. Landsat 9 (30 m) has three visible bands, that is, a single near-infrared (NIR) band and two shortwave infrared (SWIR) bands [21]. The Sentinel-2 multispectral imager (MSI) (10–20 m), however, with three red-edge bands, in addition to the visible, NIR, and SWIR bands [22], offers an efficient balance for SAV monitoring [23]. The red-edge bands of the MSI are effective for retrieving chlorophyll content in vegetation [24]. In aquatic studies, MSI was first used to distinguish water bodies using the SWIR band [25] and later used to map benthic habitats [26,27]. Several studies have used Sentinel-2 MSI to map macrophytes. Refs. [26,28] identified SAVs such as Potamogeton crispus, Ceratophyllum demersum, Myriophyllum spicatum, and Potamogeton pectinatus. Ref. [23] mapped submerged mangrove forests, while [29] tracked the spread of the invasive Spartina alterniflora in coastal wetlands. These studies demonstrate MSI’s effectiveness across various aquatic environments.

Electromagnetic radiation has general characteristics when interacting with vegetation. Photosynthetic pigments absorb light strongly but differentially (blue and red spectra are absorbed more strongly than the green spectrum) in the visible spectrum (VIS, 400–700 nm). The leaf’s cellular structure influences light scattering in the NIR region (700–1100 nm), including effects from intercellular spaces, air pockets, and air–water interfaces [30]. The water content in vegetation strongly absorbs electromagnetic energy in the SWIR, which discriminates vegetation well from other non-vegetation features [31,32]. The electromagnetic radiation reflected by submerged vegetation is inherently influenced by water. Water constituents, such as color-dissolved organic matter (CDOM) and suspended particles, influence the amount of light reaching the sensor. CDOM absorbs electromagnetic energy, especially in the ultraviolet through the visible wavelengths, resulting in the reduction of energy reaching the sensor, while turbidity due to suspended particles increases backscattering and reflection, potentially increasing the energy detected by the sensor [33,34].

Aquatic remote sensing is affected by atmospheric interference, where atmospheric path radiance contributes 80–90%, leaving only 8–10% from water and benthic reflectance [35,36,37]. Atmospheric correction removes the contribution of atmospheric interference [35,38]. The Sen2Cor atmospheric correction method, the European Space Agency (ESA) standard atmospheric correction processor for Sentinel-2 data, was initially developed for land applications. It uses look-up tables from the libRadtran radiative transfer model [39] and is built upon the dense dark vegetation approach, which assumes the presence of dark vegetation or other reference areas to estimate aerosol optical thickness (AOT) [39,40]. When applied to water surfaces, Sen2Cor uses AOT values estimated over land, but it does not account for water-specific effects such as sun and sky glint [39]. Sen2Cor further integrates aerosol and environment correction to adjust the top-of-atmosphere reflectance based on the differences between the target and surrounding pixels [35,39,40,41,42]. The ACOLITE atmospheric correction method, developed by the Royal Belgian Institute of Natural Sciences (RBINS), is designed for atmospheric correction of Landsat and Sentinel-2 satellite data, particularly for aquatic environments [41,42]. It primarily uses the dark spectrum fitting (DSF) algorithm, which estimates atmospheric path reflectance based on dark targets in the scene without predefined bands. The DSF is especially suited for clear and mixed water conditions but requires SWIR bands for turbid waters [35,43].

Vegetation indices (VIs) are mathematically derived based on the spectral characteristics of vegetation. The Normalized Difference Vegetation Index (NDVI) [44] is widely used in global vegetation studies and is commonly used as a baseline for multi-index comparison [23]. Other indices, such as the Normalized Difference Aquatic Vegetation Index (NDAVI) [45], the Floating Algae Index (FAI) [46], and the Mangrove Forest Index (MFI) [23], were introduced and used for mapping SAV. None of the above indices effectively distinguishes between EAV and SAV, nor have they been tested for hydrilla detection. In this study, we developed a new index called SVI_H for hydrilla, as hydrilla is a prominent SAV in freshwater lakes, and assessed its performance alongside existing SAV indices in two eutrophic lakes in central Florida, improving aquatic vegetation monitoring and supporting more effective management efforts.

2. Materials and Methods

2.1. Study Area

Lake Yale (28.9146°N, 81.7404°W) is a 4041-acre lake in Lake County, Florida (Figure 1), within the Ocklawaha River Watershed [47]. It connects to Lake Griffin via the Yale-Griffin Canal, which has primary land uses, including residential, natural, and agricultural areas. The climate is humid subtropical, with average temperatures of 59.7 °F and 82.4 °F in January and August, respectively. Annual rainfall averages 47.57 inches, mostly between June and October. The lake receives inflows from surrounding marshlands and drains into Lake Griffin [48]. The lake has a mean depth of 3.7 m and a maximum depth of 4.6 m [47]. Hydrilla, first detected in 1980 [49], was controlled with grass carp (Ctenopharyngodon idella) but resurged in 2019, expanding from 100 to over 3000 acres, covering over 80% of the lake area. Recent data (2017–2025) from the U.S. Fresh Water Institute showed total nitrogen at 0.23–2.68 mg/L, total phosphorus (P) at 0.07–2.6 mg/L, Secchi depth at 20–140 cm, and turbidity at 1–15.2 NTU [50].

Lake Apopka (28.6239°N, 81.6254°W) is a 30,909-acre lake in central Florida (Figure 1), with a mean depth of 1.65 m and a maximum depth of 5 m. It lies within a 485 km² watershed that includes agricultural, urban, residential, and wetland areas. Water enters the lake through rainfall, groundwater infiltration, and pumping from surrounding lands during heavy rainfall, while outflow occurs via evaporation and a single man-made canal connecting it to the Harris Chain of Lakes [51]. The climate is humid subtropical, with average temperatures of 60.6 °F and 82.6 °F in January and August, respectively. Annual rainfall averages 51.5 inches, mostly between June and September. Historically, Lake Apopka was a clear, macrophyte-dominated lake known for its recreational fishing. However, following extensive nutrient loading, it transitioned into a hypereutrophic, algal-dominated system with limited SAV and widespread cyanobacterial blooms in 1947 [52]. Over 90% of the lake bottom is covered in unconsolidated organic muck, with high phosphorus and nitrogen concentrations, and is highly susceptible to resuspension by wind and bioturbation from benthivorous fish. A mandated comprehensive restoration plan starting in the late 1980s resulted in P load reduction with concomitant reductions in chlorophyll-a concentrations and total suspended solids, further leading to increasing water clarity [51,52,53]. A critical goal of the restoration was to reestablish submerged rooted macrophytes by improving water transparency. Small, incipient populations of hydrilla (i.e., <100 acres) have been managed on Lake Apopka over the last twenty years, but in 2021, hydrilla populations expanded rapidly to over 10,000 acres, occupying nearly 30% of the lake area. Currently, the most recent data (2022–2024) show total nitrogen levels at 2.1–2.2 mg/L, total phosphorus at 0.07–0.08 mg/L, Secchi depth between 30–60 cm, and turbidity ranging from 0.1–1705 NTU [54].

2.2. Ground-Truth Surveys

Vegetation communities were mapped by point grid surveys (rake sampling) conducted by the Fish and Wildlife Research Institute as part of their long-term monitoring program [55]. At each point, vegetation was identified to the species level. EAV was visually counted within a 3 m radius of the boat. SAV was collected using a modified frotus rake with 10 cm tines, mounted on a telescopic pole that was able to extend up to 4 m and was rotated 1–2 revolutions in the sediment at the bottom, as described by [56]. Vegetation was separated by species and assessed for abundance with a rank score of 1–3, which is a visual estimate of the fullness on the rake according to [11]; a rank of 1 represents low abundance and 3 indicates high abundance.

In situ rake sampling measurements for Lake Apopka were conducted from July 29 to August 10 in 2020, August 17 to September 29 in 2021, and August 3 to September 22 in 2022. The survey was conducted for Lake Yale from September 16 to September 21, 2021.

Table 1. Data collection dates for satellite imagery and in situ measurements.

Lake &Year	Rake Sampling Date *	Water Clarity Readings at the Time of Rake Sampling				Satellite Image Acquisition Date *	Water Clarity Readings at the Time of Satellite Image Acquisition
		SDD (m)	Date *	Turbidity (NTU)	Date *		SDD (m)	Date *	Turbidity (NTU)	Date *
Apopka 2020	07/29 to 08/10	NLA: 0.4	08/12	NLA:10.03	08/12	10/30	NLA:0.38	12/15	NLA: 7.63	12/15
		CLA:0.38	08/12	CLA: 14.54	08/12		CLA: 0.4	10/12	CLA: 9.43	10/12
		SLA: 0.41	08/12	SLA: 10.42	08/12		SLA: 0.51	12/15	SLA: 6.64	12/15
Apopka 2021	08/17 to 09/29	NLA: 1.2	08/16	NLA: 2.16	08/16	10/30	NLA:1.4	10/11	NLA:2.76	10/11
		CLA:0.34	09/13	CLA: 14.33	09/13		CLA: 0.38	10/11	CLA:9.72	10/11
		SLA: 0.35	08/16	SLA: 15.61	08/16		SLA: 0.42	10/11	SLA: 7.71	10/11
Apopka 2022	08/03 to 09/22	NLA:0.65	04/14	NLA: 2.16	08/15	10/25	NLA: 0.9	10/13	NLA: 3.96	10/13
		CLA: 0.5	09/12	CLA: 10.74	09/12		CLA: 0.3	10/10	CLA:23.75	10/10
		SLA: 0.4	08/15	SLA: 11.22	08/15		SLA: 0.3	10/10	SLA:29.35	10/10
Yale 2021	09/16 to 09/21	0.4	09/28	3.77	09/21	10/30	0.7	10/26	0.5	11/30

SDD—Secchi disk depth, * represents date format (MM/DD).

lists the rake sampling acquisition dates and Lake Yale and Lake Apopka water clarity conditions during the study period. Only one water clarity measuring station exists in Lake Yale, while Lake Apopka has three stations distributed at the lake’s northern, central, and southern parts (NLA, CLA, and SLA). Readings from these stations are reported in Table 1.

2.3. Satellite Images

The Copernicus Sentinel-2 satellite constellation was launched in 2015–2016, deploying two identical satellites in polar positions within a sun-synchronous orbit. This configuration provides global coverage every five days, depending on latitude, capturing high-resolution images. The Sentinel-2 MSI sensors capture 13 bands across the visible, near-infrared, and shortwave infrared regions, offering high 12-bit radiometric resolution and ground sampling distances of 10–60 m (

Table 2. Sentinel-2 bands, spatial resolution, and central wavelength.

Band	Resolution (m)	Central Wavelength	Description
B1	60	443 nm	Coastal aerosol
B2	10	490 nm	Blue
B3	10	560 nm	Green
B4	10	665 nm	Red
B5	20	705 nm	Red-edge (RE1)
B6	20	740 nm	Red-edge (RE2)
B7	20	783 nm	Red-edge (RE3)
B8	10	842 nm	Near-infrared (NIR)
B8a	20	865 nm	Narrow NIR
B9	60	940 nm	Water vapor
B10	60	1375 nm	SWIR Cirrus
B11	20	1610 nm	Shortwave infrared (SWIR1)
B12	20	2190 nm	Shortwave infrared (SWIR2)

). Imagery is available in the public domain and accessed through different platforms such as the Copernicus EO browser (https://browser.dataspace.copernicus.eu/) (accessed on 12 November 2023).

As mentioned in the previous section, in situ rake sampling measurements were generally collected between July and September. During this period, cloud- and glint-free images were unavailable, with clear images starting to be available in October. Since hydrilla growth peaks in mid to late summer and in early fall, before going dormant in the winter [57,58], a one- or two-month gap between the in situ measurements and the image acquisition date is unlikely to result in significant discrepancies. Satellite images with no cloud cover closest to the in situ data collection dates were selected for analysis. We analyzed the data collected in 2020, 2021, and 2022 for Lake Apopka. For Lake Yale, the study was limited to a single year (2021) due to the unavailability of field survey data in 2022 and the significant time lapse in 2020 between the in situ measurements collected in late June and July and the nearest cloud-free satellite images in October.

Atmospheric corrections were applied to the Sentinel-2 L1C images using the Sen2Cor [59] and ACOLITE [41,42] methods to generate surface reflectance images. Rake sampling points, which provided hydrilla abundance levels, were spatially aligned with the corresponding satellite reflectance values. A 70 m inward buffer around the lake was applied, excluding prominently shallow areas close to the shoreline.

2.4. Submerged Aquatic Vegetation Index for Hydrilla (SVI_H)

A new three-band index named the Submerged Aquatic Vegetation Index for Hydrilla (SVI_H) was developed after analyzing the spectral profiles of EAV, hydrilla, and water from Lakes Yale and Apopka. Spectral profiles of hydrilla, water, and EAV sample locations, captured from the 2021 image (common year for Lakes Apopka and Yale, with strong hydrilla infestation), are shown in Figure 2. The profiles show that Lake Apopka water effectively absorbs longer wavelengths, particularly in the NIR and SWIR regions, while exhibiting increased reflectance across the visible spectrum, primarily due to high turbidity. Hydrilla shows relatively strong reflectance in the RE1 region of the spectrum due to the interplay between plant reflectance and water absorption characteristics [23]. SAVs typically absorb more in the SWIR band, which helps differentiate them from EAVs [5]. Based on these observations, the spectral bands G (560 nm), RE 1 (705 nm), and SWIR 1 (1610 nm) were identified as candidates for detecting SAV (hydrilla). The above discussion led to the following formula for the SVI_H index:

$${\mathrm{S}\mathrm{V}\mathrm{I}}_{\mathrm{H}}={\displaystyle \frac{{\mathsf{\rho }\mathrm{R}\mathrm{E}}_1-\mathsf{\rho }\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-{\mathsf{\rho }\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}_1}{{\mathsf{\rho }\mathrm{R}\mathrm{E}}_1+\mathsf{\rho }\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}+{\mathsf{\rho }\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}_1}}$$

where ${\mathsf{\rho }\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n},\mathsf{\rho }\mathrm{R}\mathrm{E}}_1,\mathrm{a}\mathrm{n}\mathrm{d} {\mathsf{\rho }\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}_1$ represent the spectral reflectance values of the Sentinel-2 image’s green, red-edge 1, and shortwave infrared 1 bands.

2.5. Other Vegetation Indices

Several published spectral indices used in aquatic vegetation mapping were evaluated for mapping hydrilla infestations using atmospherically corrected images. The tested indices (NDVI, NDAVI, FAI, and MFI) are listed in

Table 3. Vegetation indices used in mapping aquatic vegetation.

Name	Expression	Reference
NDVI	${\displaystyle \frac{\mathsf{\rho }\mathrm{N}\mathrm{I}\mathrm{R}-\mathsf{\rho }\mathrm{R}\mathrm{e}\mathrm{d}}{\mathsf{\rho }\mathrm{N}\mathrm{I}\mathrm{R}+\mathsf{\rho }\mathrm{R}\mathrm{e}\mathrm{d}}}$	[44]
NDAVI	${\displaystyle \frac{\mathsf{\rho }\mathrm{N}\mathrm{I}\mathrm{R}-\mathsf{\rho }\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}}{\mathsf{\rho }\mathrm{N}\mathrm{I}\mathrm{R}+\mathsf{\rho }\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}}}$	[45]
FAI	$\mathsf{\rho }\mathrm{N}\mathrm{I}\mathrm{R}-\left[\mathsf{\rho }\mathrm{R}\mathrm{e}\mathrm{d}+\left(\mathsf{\rho }\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}-\mathsf{\rho }\mathrm{R}\mathrm{e}\mathrm{d}\right)\times {\displaystyle \frac{865-665}{1610-665}}\right]$	[46]
MFI	$$\begin{gathered} \left[\right(\mathsf{\rho }\mathsf{\lambda }1-\mathsf{\rho }\mathrm{B}\mathsf{\lambda }1)+(\mathsf{\rho }\mathsf{\lambda }2-\mathsf{\rho }\mathrm{B}\mathsf{\lambda }2)+(\mathsf{\rho }\mathsf{\lambda }3-\mathsf{\rho }\mathrm{B}\mathsf{\lambda }3)+(\mathsf{\rho }\mathsf{\lambda }4-\mathsf{\rho }\mathrm{B}\mathsf{\lambda }4\left)\right]/4 \\ \mathsf{\rho }\mathrm{B}\mathsf{\lambda }\mathrm{i}=\mathsf{\rho }\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}+(\mathsf{\rho }\mathrm{R}\mathrm{e}\mathrm{d}-\mathsf{\rho }\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R})\times (2190-\mathsf{\lambda }\mathrm{i})/(2190-665) \end{gathered}$$ where the $\rho \lambda$ is the reflectance of the band centered at λ, and i ranges from 1 to 4; λ1, λ2, λ3, λ4 correspond to the central wavelengths at 705, 740, 783, and 865 nm, respectively	[23]

ρX represents the reflectance value of the band centered at X wavelength.

. All indices were computed from preprocessed images. Index values and corresponding rake sampling hydrilla abundance rankings were used to assess hydrilla detection accuracy.

2.6. Index Thresholding

The following three approaches were employed to establish thresholds for creating a binary classification that differentiates hydrilla from other features in the image (the results of the threshold application were then used to assess index performance):

(i)

Incremental searching for the optimal threshold value: This method incrementally adjusts the threshold value to determine the value that maximizes the Matthews correlation coefficient (MCC), which is a robust binary classification evaluation metric that accounts for all elements of the confusion matrix and performs reliably on imbalanced datasets [60,61].

(ii)

The Jenks natural breaks algorithm: This algorithm minimizes and maximizes within-class and between-class variance, respectively [62,63,64].

(iii)

A zero-value threshold: A threshold of zero was previously used to evaluate the MFI index [23]. This threshold was applied exclusively to MFI and SVIH index comparisons.

In this study, samples with index values above the identified threshold were classified as hydrilla, while those with values below the threshold were categorized as non-hydrilla.

2.7. Accuracy Assessment

Accuracy assessment was conducted for the following hydrilla abundance levels based on the field rake sampling measurements: (1) samples ranked as level 3, (2) combined samples ranked as levels 2 and 3, and (3) combined samples ranked as levels 1, 2, and 3. The number of hydrilla field samples (n), along with approximately equal numbers of randomly selected water samples, was used for the accuracy assessment. Due to the small EAV sample size, EAV was excluded from the quantitative assessment and only assessed visually.

2.8. Accuracy Assessment Metrics

Five commonly used accuracy assessment metrics, i.e., precision, recall, F1 score, MCC [59,63], and overall accuracy, were computed using Equations (1)–(5) and used to compare the performance of the tested indices. Visual assessments of the results were also performed.

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(1)

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(2)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(3)

$${\mathrm{F}}_1\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2·{\displaystyle \frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

(4)

$$\mathrm{M}\mathrm{C}\mathrm{C}={\displaystyle \frac{\mathrm{T}\mathrm{P}·\mathrm{T}\mathrm{N}-\mathrm{F}\mathrm{P}·\mathrm{F}\mathrm{N}}{\surd \left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)·\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)·\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)·(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N})}}$$

(5)

where TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

3. Results

Vegetation Indices and Threshold Methods

Table 4. Thresholds from the incremental search, zero and natural breaks for Lake Yale, 2021.

(a) Incremental search Threshold
Hydrilla Levels of Abundance		ACOLITE-corrected						Sen2Cor-corrected
	Index	Threshold	Accuracy	Precision	Recall	F1	MCC	Threshold	Accuracy	Precision	recall	F1	MCC
Level 3 (n = 165)	SVIH	−0.09	0.96	0.99	0.94	0.96	0.93	−0.19	0.96	0.98	0.94	0.96	0.92
	NDVI	0.16	0.62	0.86	0.29	0.43	0.32	−0.27	0.68	0.75	0.55	0.64	0.38
	NDAVI	0.03	0.66	0.80	0.42	0.55	0.36	−0.37	0.71	0.79	0.56	0.66	0.44
	FAI	0.01	0.63	0.86	0.30	0.45	0.33	0	0.66	0.86	0.39	0.54	0.39
	MFI	0	0.82	0.74	0.96	0.84	0.66	0	0.89	0.92	0.85	0.89	0.78
Levels 2 and 3 (n = 165)	SVIH	−0.09	0.95	0.99	0.92	0.95	0.91	−0.19	0.95	0.98	0.92	0.95	0.90
	NDVI	0.14	0.62	0.83	0.30	0.44	0.31	−0.29	0.69	0.74	0.58	0.65	0.39
	NDAVI	0.08	0.63	0.81	0.35	0.49	0.33	−0.36	0.71	0.81	0.56	0.66	0.45
	FAI	0.01	0.62	0.86	0.29	0.43	0.32	0	0.65	0.86	0.37	0.52	0.38
	MFI	0	0.79	0.73	0.92	0.82	0.61	0	0.85	0.91	0.78	0.84	0.72
Levels 1, 2, and 3 (n = 165)	SVIH	−0.09	0.94	0.99	0.88	0.93	0.88	−0.19	0.93	0.98	0.87	0.92	0.86
	NDVI	0.10	0.61	0.78	0.31	0.44	0.28	−0.30	0.67	0.73	0.55	0.62	0.35
	NDAVI	0.03	0.64	0.78	0.38	0.51	0.32	−0.36	0.70	0.80	0.53	0.64	0.42
	FAI	0.01	0.60	0.84	0.25	0.38	0.28	0	0.65	0.85	0.35	0.50	0.36
	MFI	0	0.80	0.74	0.93	0.82	0.62	0	0.86	0.92	0.80	0.85	0.73
(b) Zero Threshold
Level 3 (n = 165)	SVIH	0	0.95	0.99	0.91	0.95	0.91	0	0.92	0.99	0.85	0.92	0.86
	MFI	0	0.82	0.74	0.96	0.84	0.66	0	0.89	0.92	0.85	0.89	0.78
Levels 2 and 3 (n = 165)	SVIH	0	0.93	0.99	0.87	0.93	0.87	0	0.88	0.99	0.77	0.87	0.78
	MFI	0	0.79	0.73	0.92	0.82	0.61	0	0.85	0.91	0.78	0.84	0.72
Levels 1, 2, and 3 (n = 165)	SVIH	0	0.92	0.99	0.85	0.92	0.86	0	0.89	0.99	0.78	0.87	0.79
	MFI	0	0.80	0.74	0.93	0.82	0.62	0	0.86	0.92	0.80	0.85	0.73
(c) Natural breaks Threshold
Level 3 (n = 165)	SVIH	−0.02	0.95	0.99	0.92	0.95	0.91	−0.11	0.95	0.99	0.92	0.95	0.91
	NDVI	0.01	0.61	0.71	0.38	0.50	0.25	−0.04	0.65	0.86	0.36	0.50	0.37
	NDAVI	0.05	0.64	0.80	0.38	0.52	0.33	−0.10	0.66	0.86	0.38	0.53	0.39
	FAI	0.02	0.57	0.92	0.15	0.25	0.25	0.02	0.57	0.92	0.15	0.25	0.25
	MFI	0.03	0.58	0.91	0.18	0.29	0.27	0.03	0.58	0.91	0.18	0.30	0.27
Levels 2 and 3 (n = 165)	SVIH	−0.02	0.94	0.99	0.88	0.94	0.88	−0.11	0.93	0.99	0.86	0.92	0.86
	NDVI	0.01	0.61	0.70	0.38	0.49	0.25	−0.04	0.65	0.85	0.35	0.50	0.36
	NDAVI	0.05	0.63	0.79	0.36	0.49	0.31	−0.10	0.65	0.86	0.36	0.51	0.37
	FAI	0.02	0.55	0.90	0.11	0.19	0.20	0.02	0.55	0.90	0.11	0.19	0.20
	MFI	0.03	0.56	0.88	0.13	0.23	0.22	0.03	0.56	0.89	0.15	0.25	0.23
Levels 1, 2, and 3 (n = 165)	SVIH	−0.02	0.92	0.99	0.85	0.92	0.86	−0.11	0.92	0.99	0.84	0.91	0.84
	NDVI	0.01	0.60	0.69	0.35	0.47	0.22	−0.04	0.64	0.85	0.34	0.48	0.35
	NDAVI	0.05	0.64	0.80	0.38	0.52	0.33	−0.10	0.67	0.87	0.39	0.54	0.40
	FAI	0.02	0.54	0.89	0.10	0.17	0.19	0.02	0.54	0.88	0.09	0.16	0.18
	MFI	0.03	0.56	0.89	0.15	0.25	0.23	0.03	0.57	0.90	0.16	0.27	0.25

,

Table 5. Thresholds from the incremental search, zero and natural breaks for Lake Apopka, 2020.

(a) Incremental search Threshold
Hydrilla Levels of Abundance		ACOLITE-corrected						Sen2Cor-corrected
	Index	Threshold	Accuracy	Precision	Recall	F1	MCC	Threshold	Accuracy	Precision	recall	F1	MCC
Level 3 (n = 47)	SVIH	−0.03	0.81	0.97	0.64	0.77	0.66	−0.03	0.82	0.94	0.68	0.79	0.66
	NDVI	−0.44	0.77	0.82	0.68	0.74	0.54	−0.34	0.76	0.93	0.55	0.69	0.56
	NDAVI	−0.26	0.76	0.83	0.64	0.72	0.53	−0.11	0.74	1.00	0.49	0.66	0.57
	FAI	−0.01	0.78	0.72	0.91	0.80	0.58	−0.01	0.83	0.86	0.79	0.82	0.66
	MFI	−0.005	0.82	0.79	0.87	0.83	0.64	0	0.78	0.96	0.57	0.72	0.60
Levels 2 and 3 (n = 135)	SVIH	−0.03	0.76	0.96	0.53	0.69	0.57	−0.03	0.77	0.95	0.56	0.71	0.58
	NDVI	−0.43	0.65	0.71	0.50	0.59	0.31	−0.39	0.67	0.84	0.43	0.57	0.40
	NDAVI	0.03	0.60	0.86	0.24	0.37	0.29	−0.26	0.67	0.78	0.47	0.58	0.36
	FAI	−0.01	0.73	0.68	0.84	0.75	0.47	−0.01	0.78	0.82	0.73	0.77	0.57
	MFI	−0.005	0.73	0.70	0.79	0.74	0.46	0	0.71	0.94	0.46	0.62	0.50
Levels 1, 2, and 3 (n = 215)	SVIH	−0.03	0.70	0.95	0.43	0.59	0.49	−0.03	0.73	0.95	0.48	0.64	0.53
	NDVI	−0.15	0.57	0.84	0.17	0.29	0.23	−0.37	0.63	0.85	0.33	0.47	0.34
	NDAVI	0.03	0.57	0.87	0.16	0.27	0.23	−0.26	0.63	0.78	0.37	0.50	0.31
	FAI	−0.01	0.69	0.66	0.76	0.71	0.38	−0.01	0.74	0.81	0.62	0.70	0.49
	MFI	−0.005	0.69	0.68	0.73	0.71	0.39	0	0.67	0.95	0.36	0.52	0.43
(b) Zero Threshold
Level 3 (n = 47)	SVIH	0	0.80	0.97	0.62	0.75	0.64	0	0.78	0.96	0.57	0.72	0.60
	MFI	0	0.78	0.93	0.60	0.73	0.59	0	0.78	0.96	0.57	0.72	0.60
Levels 2 and 3 (n = 135)	SVIH	0	0.74	0.99	0.50	0.66	0.56	0	0.73	0.98	0.47	0.64	0.55
	MFI	0	0.70	0.86	0.47	0.61	0.45	0	0.71	0.94	0.46	0.62	0.50
Levels 1, 2, and 3 (n = 215)	SVIH	0	0.69	0.99	0.39	0.56	0.48	0	0.68	0.99	0.37	0.54	0.47
	MFI	0	0.66	0.88	0.37	0.52	0.39	0	0.67	0.95	0.36	0.52	0.43
(c) Natural breaks Threshold
Level 3 (n = 47)	SVIH	0.01	0.78	1.00	0.55	0.71	0.62	0.02	0.79	1.00	0.57	0.73	0.63
	NDVI	−0.33	0.69	0.80	0.51	0.62	0.41	−0.17	0.72	1.00	0.45	0.62	0.54
	NDAVI	−0.19	0.71	0.81	0.55	0.66	0.45	−0.06	0.71	0.95	0.45	0.61	0.50
	FAI	0.03	0.50	1.00	0.00	0.00	0.00	0.03	0.50	1.00	0.00	0.00	0.00
	MFI	0.01	0.70	1.00	0.40	0.58	0.50	0.01	0.70	1.00	0.40	0.58	0.50
Levels 2 and 3 (n = 135)	SVIH	0.01	0.73	1.00	0.45	0.62	0.54	0.02	0.73	1.00	0.45	0.62	0.54
	NDVI	−0.33	0.63	0.80	0.33	0.47	0.31	−0.17	0.61	1.00	0.22	0.36	0.35
	NDAVI	−0.19	0.64	0.79	0.37	0.51	0.32	−0.06	0.62	0.94	0.25	0.40	0.35
	FAI	0.03	0.50	1.00	0.00	0.00	0.00	0.03	0.50	1.00	0.00	0.00	0.00
	MFI	0.01	0.60	1.00	0.19	0.32	0.33	0.01	0.60	1.00	0.21	0.34	0.34
Levels 1, 2, and 3 (n = 215)	SVIH	0.01	0.68	1.00	0.35	0.52	0.46	0.02	0.67	1.00	0.35	0.52	0.46
	NDVI	−0.33	0.58	0.74	0.26	0.38	0.22	−0.17	0.57	1.00	0.15	0.26	0.28
	NDAVI	−0.19	0.59	0.72	0.28	0.41	0.22	−0.06	0.58	0.93	0.18	0.30	0.28
	FAI	0.03	0.50	1.00	0.00	0.00	0.00	0.03	0.50	1.00	0.00	0.00	0.00
	MFI	0.01	0.56	1.00	0.13	0.22	0.26	0.01	0.57	1.00	0.13	0.24	0.27

,

Table 6. Thresholds from the incremental search, zero and natural breaks for Lake Apopka, 2021.

(a) Incremental search Threshold
Hydrilla Levels of Abundance		ACOLITE-corrected						Sen2Cor-corrected
	Index	Threshold	Accuracy	Precision	Recall	F1	MCC	Threshold	Accuracy	Precision	recall	F1	MCC
Level 3 (n = 290)	SVIH	−0.05	0.88	0.99	0.76	0.86	0.77	−0.05	0.88	0.96	0.79	0.87	0.77
	NDVI	−0.44	0.81	0.84	0.77	0.81	0.63	−0.46	0.80	0.95	0.63	0.76	0.64
	NDAVI	−0.17	0.74	0.90	0.55	0.68	0.53	−0.37	0.76	0.97	0.53	0.69	0.58
	FAI	−0.005	0.81	0.91	0.69	0.79	0.64	−0.01	0.87	0.82	0.95	0.88	0.75
	MFI	0	0.83	0.97	0.69	0.80	0.69	0	0.82	0.99	0.65	0.79	0.69
Levels 2 and 3 (n = 508)	SVIH	−0.05	0.79	0.98	0.60	0.74	0.64	−0.05	0.80	0.97	0.62	0.76	0.64
	NDVI	−0.44	0.76	0.83	0.65	0.73	0.53	−0.45	0.72	0.94	0.47	0.63	0.51
	NDAVI	−0.17	0.68	0.87	0.41	0.56	0.41	−0.37	0.68	0.94	0.39	0.55	0.45
	FAI	−0.01	0.79	0.71	0.98	0.82	0.63	−0.01	0.84	0.80	0.90	0.85	0.68
	MFI	0	0.75	0.95	0.52	0.67	0.55	0	0.80	0.78	0.82	0.80	0.60
Levels 1, 2, and 3 (n = 721)	SVIH	−0.05	0.72	0.98	0.44	0.61	0.52	−0.03	0.71	0.98	0.43	0.60	0.51
	NDVI	−0.46	0.72	0.77	0.62	0.69	0.45	−0.45	0.67	0.92	0.37	0.52	0.42
	NDAVI	−0.17	0.64	0.84	0.35	0.49	0.35	−0.37	0.63	0.90	0.30	0.45	0.36
	FAI	−0.01	0.78	0.70	0.96	0.81	0.59	−0.01	0.80	0.79	0.83	0.81	0.61
	MFI	0	0.72	0.67	0.87	0.76	0.46	0	0.74	0.77	0.69	0.73	0.48
(b) Zero Threshold
Level 3 (n = 290)	SVIH	0	0.83	1.00	0.67	0.80	0.71	0	0.85	0.99	0.71	0.82	0.73
	MFI	0	0.82	0.94	0.69	0.80	0.67	0	0.81	0.96	0.65	0.78	0.66
Levels 2 and 3 (n = 508)	SVIH	0	0.75	1.00	0.50	0.67	0.57	0	0.76	0.99	0.53	0.69	0.59
	MFI	0	0.74	0.93	0.52	0.67	0.54	0	0.73	0.96	0.48	0.64	0.53
Levels 1, 2, and 3 (n = 721)	SVIH	0	0.68	1.00	0.36	0.53	0.47	0	0.69	0.99	0.39	0.55	0.48
	MFI	0	0.68	0.93	0.40	0.56	0.45	0	0.67	0.96	0.36	0.53	0.44
(c) Natural breaks Threshold
Level 3 (n = 290)	SVIH	0.02	0.81	1.00	0.63	0.77	0.68	0.03	0.81	1.00	0.63	0.77	0.68
	NDVI	−0.23	0.69	0.91	0.43	0.59	0.46	−0.25	0.69	0.95	0.40	0.57	0.47
	NDAVI	−0.07	0.70	0.90	0.46	0.60	0.46	−0.19	0.68	0.96	0.38	0.55	0.46
	FAI	0.02	0.56	1.00	0.11	0.20	0.24	0.02	0.55	1.00	0.11	0.19	0.24
	MFI	0.02	0.60	1.00	0.20	0.33	0.33	0.02	0.61	1.00	0.22	0.36	0.35
Levels 2 and 3 (n = 508)	SVIH	0.02	0.73	1.00	0.47	0.64	0.55	0.03	0.73	1.00	0.46	0.63	0.55
	NDVI	−0.23	0.63	0.86	0.30	0.45	0.33	−0.25	0.63	0.94	0.27	0.42	0.36
	NDAVI	−0.07	0.63	0.85	0.32	0.46	0.34	−0.19	0.62	0.95	0.25	0.40	0.35
	FAI	0.02	0.53	1.00	0.07	0.13	0.19	0.02	0.53	1.00	0.07	0.13	0.19
	MFI	0.02	0.56	1.00	0.13	0.22	0.26	0.02	0.57	1.00	0.14	0.25	0.27
Levels 1, 2, and 3 (n = 721)	SVIH	0.02	0.67	1.00	0.34	0.50	0.45	0.03	0.67	1.00	0.34	0.50	0.45
	NDVI	−0.23	0.60	0.85	0.24	0.38	0.28	−0.25	0.60	0.93	0.21	0.34	0.31
	NDAVI	−0.07	0.60	0.84	0.25	0.39	0.29	−0.19	0.59	0.94	0.19	0.32	0.30
	FAI	0.02	0.53	1.00	0.05	0.10	0.16	0.02	0.53	1.00	0.05	0.10	0.16
	MFI	0.02	0.55	1.00	0.09	0.17	0.22	0.02	0.55	1.00	0.10	0.19	0.23

and

Table 7. Thresholds from the incremental search, zero and natural breaks for Lake Apopka, 2022.

(a) Incremental search Threshold
Hydrilla Levels of Abundance		ACOLITE-corrected						Sen2Cor-corrected
	Index	Threshold	Accuracy	Precision	Recall	F1	MCC	Threshold	Accuracy	Precision	recall	F1	MCC
Level 3 (n = 272)	SVIH	−0.16	0.74	0.98	0.50	0.66	0.56	−0.29	0.70	0.99	0.41	0.58	0.50
	NDVI	−0.72	0.72	0.73	0.68	0.71	0.43	−0.08	0.86	0.91	0.80	0.85	0.73
	NDAVI	−0.59	0.60	0.62	0.50	0.56	0.20	−0.07	0.75	0.70	0.89	0.78	0.52
	FAI	−0.015	0.78	0.72	0.90	0.80	0.57	−0.015	0.87	0.86	0.88	0.87	0.74
	MFI	−0.01	0.81	0.77	0.88	0.82	0.62	−0.01	0.87	0.82	0.93	0.87	0.74
Levels 2 and 3 (n = 479)	SVIH	−0.16	0.65	0.96	0.31	0.47	0.41	−0.29	0.62	0.98	0.25	0.39	0.37
	NDVI	−0.73	0.67	0.69	0.61	0.65	0.34	−0.09	0.82	0.82	0.81	0.81	0.63
	NDAVI	−0.59	0.57	0.59	0.42	0.49	0.14	−0.07	0.73	0.70	0.82	0.76	0.47
	FAI	−0.015	0.74	0.71	0.80	0.75	0.48	−0.015	0.82	0.86	0.76	0.81	0.64
	MFI	−0.01	0.75	0.75	0.75	0.75	0.51	−0.01	0.83	0.83	0.82	0.83	0.66
Levels 1, 2, and 3 (n = 853)	SVIH	−0.16	0.59	0.92	0.19	0.31	0.29	−0.29	0.57	0.96	0.15	0.26	0.27
	NDVI	−0.73	0.60	0.64	0.49	0.55	0.22	−0.09	0.74	0.78	0.65	0.71	0.48
	NDAVI	−0.59	0.54	0.56	0.36	0.44	0.09	−0.07	0.67	0.66	0.70	0.68	0.34
	FAI	−0.015	0.68	0.69	0.68	0.68	0.37	−0.015	0.74	0.84	0.59	0.69	0.50
	MFI	−0.01	0.67	0.71	0.58	0.64	0.35	−0.01	0.74	0.79	0.64	0.71	0.48
(b) Zero Threshold
Level 3 (n = 272)	SVIH	0	0.60	1.00	0.19	0.32	0.33	0	0.50	1.00	0.00	0.00	0.00
	MFI	0	0.57	0.95	0.15	0.27	0.27	0	0.58	0.98	0.17	0.29	0.30
Levels 2 and 3 (n = 479)	SVIH	0	0.56	1.00	0.12	0.21	0.25	0	0.50	1.00	0.00	0.00	0.00
	MFI	0	0.54	0.96	0.09	0.17	0.21	0	0.55	0.98	0.11	0.19	0.23
Levels 1, 2, and 3 (n = 853)	SVIH	0	0.53	1.00	0.06	0.12	0.18	0	0.50	1.00	0.00	0.00	0.00
	MFI	0	0.53	0.94	0.06	0.11	0.16	0	0.53	0.96	0.06	0.12	0.17
(c) Natural breaks Threshold
Level 3 (n = 272)	SVIH	−0.11	0.70	0.99	0.40	0.57	0.49	−0.28	0.64	0.99	0.28	0.44	0.40
	NDVI	−0.58	0.62	0.86	0.28	0.42	0.32	0.02	0.50	1.00	0.01	0.01	0.06
	NDAVI	−0.47	0.60	0.80	0.26	0.40	0.27	0.00	0.51	1.00	0.03	0.05	0.11
	FAI	0.01	0.51	1.00	0.01	0.02	0.07	0.01	0.51	1.00	0.02	0.04	0.11
	MFI	−0.002	0.59	0.96	0.19	0.32	0.31	0.004	0.53	1.00	0.06	0.11	0.17
Levels 2 and 3 (n = 479)	SVIH	−0.11	0.62	0.99	0.24	0.38	0.36	−0.28	0.58	0.99	0.17	0.29	0.30
	NDVI	−0.58	0.57	0.77	0.19	0.30	0.20	0.02	0.50	1.00	0.00	0.01	0.05
	NDAVI	−0.47	0.55	0.67	0.19	0.29	0.14	0.00	0.51	1.00	0.02	0.03	0.09
	FAI	0.01	0.50	1.00	0.01	0.01	0.06	0.01	0.51	1.00	0.01	0.02	0.08
	MFI	−0.002	0.56	0.97	0.12	0.22	0.24	0.004	0.52	1.00	0.04	0.07	0.14
Levels 1, 2, and 3 (n = 853)	SVIH	−0.11	0.57	0.99	0.14	0.24	0.27	−0.28	0.55	0.99	0.10	0.18	0.22
	NDVI	−0.58	0.53	0.66	0.13	0.22	0.11	0.02	0.50	0.75	0.00	0.01	0.02
	NDAVI	−0.47	0.52	0.57	0.14	0.22	0.05	0.00	0.51	0.91	0.01	0.02	0.07
	FAI	0.01	0.50	0.80	0.00	0.01	0.03	0.01	0.50	0.88	0.01	0.02	0.05
	MFI	−0.002	0.54	0.95	0.07	0.14	0.18	0.004	0.51	0.95	0.02	0.05	0.10

summarize the hydrilla classification evaluation metrics in Lake Yale (2021) and Lake Apopka (2020–2022) using five spectral indices. The evaluation was conducted using the three threshold types described in Section 2.6. Figure 3, Figure 4, Figure 5 and Figure 6 present the natural color satellite images, the hydrilla abundance levels derived from rake sampling, and the hydrilla maps generated using the SVI_H, MFI, and FAI indices, which were identified as the best and most consistent performing indices. The hydrilla maps were created by applying the respective threshold values derived from the three thresholding methods to the index maps. For Lake Yale in 2021, Table 4 show that the SVI_H index produced similar results using both ACOLITE and Sen2Cor corrections, except when using the zero threshold, where ACOLITE yielded higher values,. Using incremental search, SVI_H showed high MCC values (>0.90) for both level 3 abundance and levels 2 and 3 combined. For the levels 1–3 class, the MCC values were 0.88 and 0.86 for ACOLITE and Sen2Cor, respectively. Other indices produced MCC values below 0.5 across all thresholding methods except for MFI. The MFI index using ACOLITE-corrected imagery achieved an MCC of 0.66, whereas Sen2Cor achieved 0.78 for level 3 abundance. However, when using the natural breaks threshold, MFI produced a much lower MCC value of 0.27.

For Lake Apopka, Table 5 present the 2020 data analysis. The table shows FAI yielding an MCC value of 0.66, similar to SVI_H for level 3 abundance using the Sen2Cor-corrected imagery, but with high recall. SVI_H results from ACOLITE and Sen2Cor imagery were generally consistent, except for levels 1–3, where the ACOLITE MCC was slightly smaller using the incremental search method. The MFI index showed moderate MCC values but performed poorly when the natural breaks threshold method was applied.

For the Lake Apopka 2021 analysis (as shown in Table 6), the SVI_H index yielded an MCC of 0.77 for level 3 abundance, 0.64 for levels 2 and 3 abundance, and 0.52 for levels 1–3 abundance when using the incremental search method. Both Sen2Cor and ACOLITE corrections produced similar SVI_H results across all three thresholding methods. FAI achieved a high MCC with high recall when using the incremental search method for levels 2 and 3 and 1–3 abundance (Sen2Cor correction); however, its performance was poor when using the natural breaks method.

Table 7 summarize the results for Lake Apopka in 2022. Using the incremental search method, both MFI and FAI achieved high MCC values with high recall accompanied by a high rate of false positives, leading to the overclassification of non-hydrilla regions, as shown in Figure 6B. Using the natural breaks threshold method, SVI_H (ACOLITE) produced higher MCC values (0.49) for level 3 abundance, 0.36 for levels 2 and 3 abundance, and 0.27 for levels 1–3 abundance, whereas MFI and FAI performed poorly in comparison. In general, there was a large discrepancy in MCC values for all indices in the 2022 dataset, highlighting potential issues related to increased turbidity and hydrilla treatment in part of Lake Apopka during this year. Figure 7 shows spectral profiles from ACOLITE- and Sen2Cor-corrected images for Lake Apopka in 2021 and 2022, highlighting major spectral differences.

As illustrated in Figure 3B, Figure 4B, Figure 5B and Figure 6B, SVI_H excluded the EAV in the eastern region of Lake Yale (highlighted by the red circle in Figure 4A) and in the northeastern regions of Lake Apopka (indicated by the red circles in Figure 4A, Figure 5A and Figure 6A).

4. Discussion

4.1. Performance of SVI_H Compared to Other Indices

Our results compared the performance of multiple indices used for aquatic vegetation mapping and introduced the SVI_H as an additional new index for detecting hydrilla, distinguishing it from other aquatic components such as water and EAV. The spectral reflectance profiles for Lake Apopka and Lake Yale, shown in Figure 2, illustrate similar patterns among EAV and hydrilla, with distinct variations across the spectral bands. EAV and hydrilla at level 3 exhibit strong reflectance in the red-edge and NIR regions; EAV reflects some SWIR, while hydrilla at level 3 shows near-complete SWIR absorption. Hydrilla level 2 exhibits a peak in the RE1 region of the spectrum reflectance influenced by plant reflectance and water absorption at the fringes of hydrilla colonies [23]. The SVI_H index takes advantage of the increased reflectance in the lower red-edge region of the spectrum, combined with the reduced visible light reflectance near the SAV. We believe the index exploits the effect of hydrilla on nearby water clarity through nutrient compartmentalization and suspended particle reduction [8]. The index excludes EAV by subtracting shortwave reflectance, which tends to be high for emergent aquatic plants.

The spectral profiles of water differ between the two lakes. Water absorbs longer wavelengths, while turbid waters reflect more light across the visible spectrum. For example, Lake Apopka’s water has higher reflectance in the visible bands, including the green band, and comparatively less reflectance in the red-edge 1 band. It should be mentioned that the authors explored other band alternatives (e.g., using the red band instead of the green band or the red-edge 2 band instead of the red-edge 1) to optimize band selection for index development. These experiments showed that some alternative bands outperformed others in a few cases. Nevertheless, regarding the three bands used in the SVI_H, while they were not top performers in all cases, they were consistent in providing adequate hydrilla detection performance comparable to the better-tested alternatives across both lakes.

The SVI_H index generally produced high precision, indicating minimal false positives and higher confidence in the identified hydrilla pixels. This could provide environmental benefits due to reduced herbicide applications. However, using other indices such as the MFI or FAI with appropriate thresholding methods could provide higher recall values (lower false negatives), facilitating wider herbicide applications that reduce the possibility of missing invasive plant instances. An algorithm that integrates the results of different indices and utilizes them in a decision support system to guide management activities is recommended for future research.

All indices performed best for high-abundance hydrilla (i.e., level 3), typically in infested areas where the hydrilla is approaching or reaching the water surface. All evaluation metrics started to decline when lower abundance levels were included, indicating that detecting less abundant hydrilla, likely further away from the water surface, is limited by remote sensing physics constraints [5]. The MFI index demonstrated the closest performance to the SVI_H index, averaging RE1, RE2, RE3, and NIR reflectance values above the Red-to-SWIR baseline, allowing it to identify submerged vascular plants. However, it could not separate SAV from EAV when compared to the SVI_H. This is due to the high NIR reflectance values across all vascular species, resulting from the scattering of photons by the air pockets in their spongy mesophyll, aiding in the identification of all vascular plants [30,32]. Additionally, indices such as FAI, NDVI, and NDAVI, which rely on NIR reflectance values, fail to distinguish between water and SAV as depth increases [5]; this is due to the absorption of the NIR band by the water surface.

Moreover, compared to the SVI_H index, FAI and MFI were more sensitive to small threshold changes, emphasizing the importance of accurately identifying their optimal thresholds. Without analyzing hydrilla maps generated using these thresholds, there is a substantial risk of overpredicting or underpredicting hydrilla presence.

4.2. Thresholding Methods and Their Implications

The incremental search method requires in situ data to identify the optimal threshold. However, it highlights the difference between the best possible results and those obtained by algorithm-based thresholds (i.e., natural breaks) or presumed thresholds (zero-value threshold). MFI and FAI misclassified water as hydrilla using the incremental search threshold (maximizing the MCC metric) with high recall (low false negatives), accompanied by low precision (high false positives), as shown in Figure 4b, Figure 5b and Figure 6b, despite the high MCC values (Table 7). This could be attributed to the index values for turbid waters, which can resemble SAV due to strong sediment backscattering in the red band [5]. It should be noted here that we tested the performance of the thresholding methods only within our dataset’s temporal and spatial limits. Further research to develop other data-driven thresholding methods is recommended. Also, research on the optimal rake sampling size that can be used in the incremental search method to explore the best threshold values is needed.

4.3. Performance Comparison of ACOLITE vs. Sen2Cor

By comparing the results of the ACOLITE and Sen2Cor atmospheric correction methods in Table 4, Table 5, Table 6 and Table 7, we found that, for most years, both Sen2Cor and ACOLITE produced similar results. The difference between the two atmospheric correction models became more noticeable in Apopka’s 2022 dataset. This discrepancy may be attributed to the increased turbidity reported at the CLA and SLA water clarity observation stations (NTU 23.75 and 29.35), close to the 2022 satellite image acquisition time (Table 1). Figure 7 shows that the spectral profiles of Sen2Cor-corrected images in 2022 displayed a smooth trend with minimal variation across bands, producing reflectance values close to zero, whereas ACOLITE-corrected spectral profiles exhibited higher reflectance across bands compared to Sen2Cor, differing from the patterns observed in 2021. This highlights an advantage of ACOLITE in turbid water correction (Figure 7). Sen2Cor, however, is widely used on ESA platforms and cloud computing engines (e.g., Google Earth Engine), ensuring broader applicability.

4.4. Challenges in Detecting Hydrilla in Deep Waters and Study Constraints

Some hydrilla instances are expected to go undetected at lower levels of abundance. The maximum depth at which hydrilla can be detected using SVI_H remains undetermined and requires further investigation. Increased depth and lower vegetation abundance reduce sensor visibility, limiting the effectiveness of remote sensing. This was observed in the southwest area of Lake Apopka in 2022, where rake sampling reported hydrilla existence while remote sensing results were not able to detect it. To the authors’ knowledge, regarding hydrilla management efforts, this area of the lake was subjected to herbicide treatment in 2022 to control hydrilla, leading to reduced abundance and subsidence below the water surface.

Turbidity had a strong effect on reflectance in all bands [37,65], further restricting hydrilla detection. This was emphasized in Lake Apopka’s 2022 dataset, with increased turbidity that was most likely caused by the runoff from Hurricane Ian (September 28, 2022), resulting in poor performance of the SVI_H index. It is important to note that the obtained results should be interpreted in the context of the specific conditions under which the tests were conducted, including the physiological and phenological statuses of the hydrilla, the prevailing water clarity conditions [37], the quality of the remote sensing images used, and the effect of active hydrilla treatment in the water bodies.

Our exploration into the validity of utilizing satellite images for spectral analysis throughout the year has revealed significant challenges in acquiring suitable images during late spring through the summer and early fall seasons. This difficulty primarily arises due to high cloud cover and extensive glare effects in Florida during these periods. Consequently, our effective window for conducting this type of analysis is from late September to early March. Additionally, water absorption and scattering limit remote sensing-based detection of SAV at greater depths, with NIR and red-edge bands particularly affected [66,67,68]. Turbidity, suspended sediments, and algae growth are other factors that make SAV mapping inherently challenging [65].

In this study, we visually inspected the satellite images to assess their quality and selected the best available images that closely corresponded to our ground surveys. Despite our efforts to minimize the time offset between image acquisition and field sampling and keep the offset relatively short, we recognize the necessity for further testing. This need is particularly highlighted by the analysis of the 2022 Apopka image, which showed significant discrepancies likely due to Hurricane Ian impacting Florida between the two events. Based on these findings, we recommend pursuing further research with shorter time offsets or, ideally, synchronous acquisition of field sampling with satellite images. This approach could help mitigate the impact of unexpected events and improve the reliability of spectral analysis using satellite imagery.

4.5. Future Research Directions

Although this study focused on hydrilla as the dominant SAV, we believe the SVI_H can be applied to detect other SAV; however, this hypothesis remains to be tested. Future research can extend SVI_H to different types of SAV by exploring how red-edge and SWIR bands interact to enhance detection. The combined effects of these bands, along with turbidity, water depth, and canopy structure, will improve SAV detection under varying conditions [5,37]. Additionally, examining seasonal variability and phenological changes will enhance monitoring over an extended period and broaden the model’s applicability.

The main purpose of this study is to develop an index that can be implemented to support management efforts. The index can easily be implemented on cloud-based geospatial analysis platforms such as Google Earth Engine or the Copernicus browser to enable spatial and temporal monitoring of extended SAV. Our future research will focus on developing robust hydrilla mapping models using machine-learning techniques that integrate individual spectral band values and vegetation indices to enhance predictive accuracy and scalability.

5. Conclusions

This study highlights the strong potential of the newly developed SVI_H, utilizing the green, red-edge 1, and shortwave infrared 1 bands of the Sentinel-2 satellite imagery to accurately detect Hydrilla verticillata across different abundance levels in two Florida lakes. Among the spectral indices evaluated, SVI_H consistently delivered high performance across multiple thresholding methods (zero, incremental search, and natural breaks) and atmospheric correction models (ACOLITE and Sen2Cor). Four other spectral indices were tested and compared with the SVI_H results, which demonstrated a clear advantage over other indices, such as MFI and FAI, by effectively distinguishing hydrilla from water and EAV, minimizing false positives, and maintaining high MCC values.

This study further emphasizes the influence of environmental factors, such as increased turbidity and hydrilla treatment, which likely affected the spectral quality of the satellite imagery and the accuracy of classification when tested across multiple years. These findings emphasize the importance of selecting appropriate spectral indices and thresholding methods, with SVI_H emerging as a robust option for hydrilla mapping under varying conditions. Its consistent performance across years, locations, and environmental conditions supports its application in targeted hydrilla management and long-term ecological monitoring. This, in turn, can enhance decision-making for invasive species control and contribute to more sustainable freshwater ecosystem management.