Tracking Phytoplankton Biomass Amid Wildfire Smoke Interference Using Landsat 8 OLI

Abstract

This study investigates the escalating impact of wildfire smoke on the remote sensing of phytoplankton biomass in freshwater systems. Wildfire smoke disrupts the accuracy of Chlorophyll-a (Chl-a) retrieval models, with Chl-a often used as a proxy for quantifying phytoplankton biomass. Given the increasing frequency and intensity of wildfires, there is a need for the development and refinement of remote sensing methodologies to effectively monitor phytoplankton dynamics under wildfire-impacted conditions. Here we developed a novel approach using Landsat’s coastal/aerosol band (B1) to screen for and categorize levels of wildfire smoke interference. By excluding high-interference data (B1 reflectance > 0.07) from the calibration set, Chl-a retrieval model performance using different Landsat band formulas improved significantly, with R² increasing from 0.55 to as high as 0.80. Our findings demonstrate that Rayleigh-corrected reflectance, combined with B1 screening, provides a robust method for monitoring phytoplankton biomass even under moderate smoke interference, outperforming full atmospheric correction methods. This approach enhances the reliability of remote sensing in the face of increasing wildfire events, offering a valuable tool for the effective management of aquatic environments.

1. Introduction

Since the late 20th century, freshwater environments have experienced a troubling rise in phytoplankton blooms, marked by increasing frequency, intensity, and shifts in community composition [1]. Although phytoplankton populations typically exhibit seasonal patterns, anthropogenic stressors, such as nutrient loading and climate change, can disrupt these natural cycles, leading to harmful algal blooms [2,3]. These blooms degrade water quality, causing issues including toxin production, disruption of food webs, and anoxia, which in turn create substantial social, economic, and ecological challenges [4,5]. Given their responsiveness to environmental changes, phytoplankton are crucial indicators of water quality and are routinely monitored to assess the health of water bodies globally [6,7]. Developing tools to monitor phytoplankton dynamics within freshwater lakes is essential for enabling lake managers and practitioners to make informed decisions [8].

Chlorophyll-a concentration (Chl-a), commonly used as a proxy for phytoplankton biomass [9], is measured through both field-based and remote sensing methods. Traditional field sampling provides high-resolution data at specific locations but lacks comprehensive spatial coverage and is resource intensive [10,11]. The time-intensive nature of these methods can delay water management responses, which are critical for addressing phytoplankton-related threats [12]. In contrast, remote sensing offers a time-sensitive and cost-efficient technique for monitoring phytoplankton biomass across large areas [7,13,14,15,16,17,18,19]. This approach enables near-real-time, quantitative, and qualitative analyses of phytoplankton biomass by capturing optical data on water bodies [17,20]. As clouds, smoke, and haze can significantly degrade image quality, optical remote sensing is dependent on cloud-free conditions and minimal atmospheric interference [6,21].

One emerging concern is the increasing prevalence of wildfires, which produce smoke that obstructs remote sensing capabilities. Recent years have seen a global rise in wildfire frequency and intensity, exacerbated by climate change and historical fire suppression practices [22,23] (Figure 1A). For example, Canada experienced its worst wildfire season on record in 2023, with atmospheric carbon emissions four times higher than in previous years (Figure 1B). Wildfires not only impact water quality through ash deposition and light inhibition [23,24,25,26,27] but also degrade the quality of remote sensing observations. Elevated aerosol levels can be partially mitigated through atmospheric correction in remote sensing products like Landsat Level 2 or Sentinel-2A [28,29]. For aquatic applications that often use partial atmospheric corrections, wildfire smoke remains a challenge.

This study utilizes a Landsat 8 coastal/aerosol band (B1) to categorize the level of wildfire smoke interference in remote sensing images. Methodologies to classify and mitigate smoke effects in satellite imagery are crucial as wildfire activity increases globally. Such methodologies enable the exclusion of smoke-impacted observations from analysis, thereby enhancing the reliability of water quality models in wildfire-affected areas. We assessed the impacts of excluding varying degrees of smoke interference on the performance of Chl-a retrieval models. To validate our approach, we compared Rayleigh-corrected reflectance (i.e., partially corrected reflectance) data filtered by the aerosol band with fully corrected, analysis-ready Landsat 8 Level 2 products. Developing robust methodologies to mitigate wildfire smoke interference can equip lake managers with greater confidence in water quality monitoring. By defining thresholds for wildfire smoke below which model performance is reliable, these methodologies help ensure the accuracy of remote sensing products in the face of increasing wildfire activity.

2. Materials and Methods

2.1. Ground-Based Dataset

Chl-a samples were collected from 108 lakes near and within the Lake Winnipeg watershed of Canada, spanning Alberta, Saskatchewan, and Manitoba (Figure 2). Water samples were obtained using a 1 m integrated sampler at the center or deepest point of each lake, as determined by bathymetric data. Phytoplankton samples were taken during the peak growing season, from mid-July to the end of August [30,31]. To concentrate phytoplankton for Chl-a analysis, lake water was filtered through Whatman GF/F filters. The volume of water filtered ranged from 30 to 500 mL, depending on the algae concentration. The filters were subsequently frozen at −20 °C to preserve pigments. Chl-a was extracted using 90% (v/v) acetone under dim light, and the filters were mechanically disrupted with a bead beater (3 × 10-s cycles) containing 0.1 mm zirconia/silica beads [32]. The mixture was stored at −20 °C for 4 h, followed by centrifugation (6000× g for 5 min) for an initial clarification. The supernatant was further clarified by filtering through a 0.22 µm filter and then measured at 664 nm using a spectrophotometer (1 cm cuvette) following the Jeffrey and Humphrey method [33]. Chl-a concentrations were reported in µg/L and used as a proxy for phytoplankton biomass [9].

2.2. Landsat Image Acquisition, Processing, and Analysis

Landsat 8 Operational Land Imager (OLI) data were acquired, with the first six spectral bands extracted for each study lake (

Table 1. Summary of Landsat OLI bands and bandwidths used in this research.

Sensor	Bands	Wavelengths (μm)	Resolution (m)
Landsat 8 OLI	Band 1 (B1): Coastal/aerosol	0.43–0.45	30
	Band 2 (B2): Blue	0.45–0.51	30
	Band 3 (B3): Green	0.53–0.59	30
	Band 4 (B4): Red	0.64–0.67	30

). Data acquisition and radiometric and atmospheric corrections were carried out using Google Earth Engine (GEE), a cloud-based platform that facilitates the processing of large geospatial datasets, including remote sensing data [34,35]. GEE has become a widely utilized tool for monitoring water quality in both inland and marine environments [36,37,38,39,40].

2.2.1. In Situ and Satellite Match Up Considerations

Ground-based Chl-a measurements were matched to corresponding satellite observations at the point of sampling within a temporal window of ±1, 3, and 7 days of satellite image capture. These temporal windows were selected based on considerations of the sample size and the correlations between satellite band formulas and in situ Chl-a (transformed using the natural logarithm of Chl-a (ln(Chl-a)).

2.2.2. Atmospheric Correction

Atmospheric correction is essential for accurately mapping water optical properties to mitigate atmospheric interferences, such as aerosol and Rayleigh scattering effects, in satellite imagery [41,42]. Despite its importance, inland water research still lacks a standardized atmospheric correction algorithm [43,44]. Many studies (e.g., [17,44,45,46,47,48,49,50]) recommend partial atmospheric correction, particularly the removal of Rayleigh scattering effects, due to the difficulty in accurately estimating aerosol correction over water bodies, which often leads to an overestimation of atmospheric radiance contributions.

In this study, partial atmospheric correction was performed as follows:

1.

Conversion to at-sensor spectral radiance: Landsat 8 OLI Level 1 data, comprising raw digital numbers (DN) ranging from 0 to 65,000, were used. Converting DN values to a common radiometric scale by calculating radiance is the first step in analyzing images from different sensors and platforms [51,52]. DN values for each band (λ) were converted to top-of-atmosphere (TOA) radiance (L_λ) using Equation (1) (

Table 2. Table of formulas used to calculate Rayleigh-corrected bottom-of-atmosphere (BOA) reflectance (Rrc) (Equations (1)–(7)) and model performance metrics (Equations (8)–(12)).

${\mathrm{L}}_{\mathsf{\lambda }}=\left({\mathrm{D}\mathrm{N}}_{\mathsf{\lambda }}{\mathrm{g}\mathrm{a}\mathrm{i}\mathrm{n}}_{\mathsf{\lambda }}\right)+{\mathrm{b}\mathrm{i}\mathrm{a}\mathrm{s}}_{\mathsf{\lambda }}$	(1)
where, ${\mathrm{L}}_{\mathsf{\lambda }}$ is the TOA radiance for band $\mathsf{\lambda }$; ${\mathrm{D}\mathrm{N}}_{\mathsf{\lambda }}$ is the raw digital number (DN) for band $\mathsf{\lambda }$; ${\mathrm{g}\mathrm{a}\mathrm{i}\mathrm{n}}_{\mathsf{\lambda }}$ is the multiplicative rescaling factor for band $\mathsf{\lambda };$ and ${\mathrm{b}\mathrm{i}\mathrm{a}\mathrm{s}}_{\mathsf{\lambda }}$ is the additive rescaling factor for band $\mathsf{\lambda }$.
${\mathrm{L}}_{\mathrm{r}\left(\mathsf{\lambda }\right)}=\left({\displaystyle \frac{{\mathrm{E}\mathrm{S}\mathrm{U}\mathrm{N}}_{\mathsf{\lambda }}\times \cos {\mathsf{\theta }}_{\mathrm{s}}{\mathrm{P}}_{\mathrm{r}}}{4\pi \times \left(\cos {\mathsf{\theta }}_{\mathrm{s}}+\cos {\mathsf{\theta }}_{\mathrm{v}}\right)}}\right)\times \left((1-\exp -{\mathsf{\tau }}_{\mathrm{r}\left(\mathsf{\lambda }\right)}\times \left(\left({\displaystyle \frac{1}{\mathrm{c}\mathrm{o}\mathrm{s}{\mathsf{\theta }}_{\mathrm{s}}}}\right)+\left({\displaystyle \frac{1}{\mathrm{c}\mathrm{o}\mathrm{s}{\mathsf{\theta }}_{\mathrm{v}}}}\right)\right)\right)\times {\mathsf{\tau }}_{\mathrm{o}\mathrm{z}\left(\mathsf{\lambda }\right)}$	(2)
where, ${\mathrm{L}}_{\mathrm{r}}$ is the Rayleigh path radiance; ${\mathrm{E}\mathrm{S}\mathrm{U}\mathrm{N}}_{\mathsf{\lambda }}$ is the exo-atmospheric solar irradiance constant for band $\mathsf{\lambda }$; ${\mathsf{\theta }}_{\mathrm{s}}$ is the solar zenith angle in degrees; ${\mathsf{\theta }}_v$ is the satellite view angle in degrees; ${\mathrm{P}}_{\mathrm{r}}$ is the Rayleigh phase function (Equation (3)); ${\mathsf{\tau }}_{\mathrm{r}\left(\mathsf{\lambda }\right)}$ is the Rayleigh optical thickness for band $\mathsf{\lambda }$ (Equation (4)); and ${\mathrm{t}}_{\mathrm{o}\mathrm{z}\left(\mathsf{\lambda }\right)}$ is the ozone transmittance for band $\mathsf{\lambda }$ (Equation (5)).
${\mathrm{P}}_{\mathrm{r}}={\displaystyle \frac{3}{4}}\times {\displaystyle \frac{1-\mathsf{\gamma }}{1+2\mathsf{\gamma }}}\times \left(1+{\mathrm{c}\mathrm{o}\mathrm{s}}^2\Phi \right)+{\displaystyle \frac{3\mathsf{\gamma }}{1+2\mathsf{\gamma }}}$	(3)
where, $\mathsf{\gamma }$ is obtained from $\mathsf{\gamma }={\displaystyle \frac{\mathsf{\rho }\mathrm{n}}{2-\mathsf{\rho }\mathrm{n}}}$ in which $\mathsf{\rho }\mathrm{n}$ is the depolarization factor, and $\Phi$ is the scattering angle in degrees (180 − ${\mathsf{\theta }}_{\mathrm{s}}$).
${\mathsf{\tau }}_{\mathrm{r}\left(\mathsf{\lambda }\right)}=0.008569{\mathsf{\lambda }}^{-4}\times \left(1+0.0113{\mathsf{\lambda }}^{-2}+0.00013{\mathsf{\lambda }}^{-4}\right)$	(4)
${\mathrm{t}}_{\mathrm{o}\mathrm{z}}\left(\mathsf{\lambda }\right)=\mathrm{e}\mathrm{x}\mathrm{p}\left({-\mathsf{\tau }}_{\mathrm{o}\mathrm{z}}\right)\times \mathrm{e}\mathrm{x}\mathrm{p}\left({\displaystyle \frac{-{\mathsf{\tau }}_{\mathrm{o}\mathrm{z}}}{\cos \mathsf{\theta }\mathrm{s}}}\right)$	(5)
where, ${\mathsf{\tau }}_{\mathrm{o}\mathrm{z}\left(\mathsf{\lambda }\right)}$ is the ozone optical thickness for band $\mathsf{\lambda }$.
${\widehat{\mathrm{L}}}_{\lambda }={\mathrm{L}}_{\mathsf{\lambda }}-{\mathrm{L}}_{\mathrm{r}}$	(6)
where, ${\widehat{\mathrm{L}}}_{\lambda }$ is the Rayleigh-corrected radiance for band $\mathsf{\lambda }$.
${\mathrm{R}\mathrm{r}\mathrm{c}}_{\mathsf{\lambda }}=\left({\displaystyle \frac{\pi {\widehat{\mathrm{L}}}_{\mathsf{\lambda }}\times {\mathrm{d}}^2}{{\mathrm{E}\mathrm{S}\mathrm{U}\mathrm{N}}_{\mathsf{\lambda }}\times \cos {\mathsf{\theta }}_{\mathrm{s}}}}\right)$	(7)
where, $\mathrm{R}\mathrm{r}\mathrm{c}$ is the partially corrected BOA reflectance for band $\mathsf{\lambda }$; and d is the Earth–Sun distance in astronomical units.
${\mathrm{R}}^2=1-{\displaystyle \frac{\mathrm{R}\mathrm{S}\mathrm{S}}{\mathrm{T}\mathrm{S}\mathrm{S}}}$	(8)
where, ${\mathrm{R}}^2$ is the coefficient of determination; $\mathrm{R}\mathrm{S}\mathrm{S}$ is the residual sum of square; and $\mathrm{T}\mathrm{S}\mathrm{S}$ is the total sum of square.
$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\displaystyle \frac{{\left({\widehat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}{\mathrm{n}}}}$	(9)
where, $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$ is Root Mean Square Error; ${\widehat{\mathrm{y}}}_i$ is the predicted value; ${\mathrm{y}}_{\mathrm{i}}$ is the observed value; andn is the number of samples.
$\mathrm{N}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}{\mathsf{\sigma }}}$	(10)
where, $\mathrm{N}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$ is Normalized Root Mean Square Error; $\mathsf{\sigma }$ is the standard deviation of the observed values.
$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}|{\widehat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}|}{\mathrm{n}}}$	(11)
where, $MAE$ is Mean Absolute Error.
$Bias={\displaystyle \frac{{\sum }_{i=1}^n({\widehat{\mathrm{y}}}_i-{\mathrm{y}}_{\mathrm{i}})}{\mathrm{n}}}$	(12)

; [51]).

2.

Calculation of Rayleigh scattering effects on at-sensor spectral radiance: To remove Rayleigh scattering from the TOA radiance, the Rayleigh path radiance (L_r) for each band was calculated using Equation (2) (Table 2; [53]). The Rayleigh pressure (P_r) was calculated using Equation (3) (Table 2; [54,55]). The Rayleigh optical thickness (τ_r) for each band was calculated using Equation (4) (Table 2; [53]), and ozone transmittance (τ_oz) for each band was calculated using Equation (5) (Table 2; [56,57]). Finally, Rayleigh-corrected TOA radiance ($\widehat{\mathrm{L}}$) for each band was obtained by subtracting L_r from L_λ using Equation (6) (Table 2).

3.

Conversion of Rayleigh-corrected radiances to partially-corrected BOA reflectance: The Rayleigh-corrected TOA radiance was then converted to Rayleigh-corrected bottom-of-atmosphere (BOA) reflectance (Rrc) for each band using Equation (7) (Table 2); Rrc_λ was then used to develop Chl-a retrieval models. This step offers multiple advantages: it eliminates the impact of varying solar zenith angles due to different image acquisition times, accounts for differences in Earth–Sun distances, and compensates for varying exo-atmospheric solar irradiances [52].

2.3. Wildfire Correction

While wildfires are a frequent occurrence in Canada, the 2023 wildfire season was unprecedented, marking the worst on record (Figure 1B). The resulting smoke significantly degraded the quality of satellite images. To quantify the extent of smoke interference, we used the Landsat coastal/aerosol band (B1 = 0.43–0.45 μm), which is particularly effective for monitoring atmospheric aerosol properties [58]. It is crucial to first address the influence of Rayleigh scattering before classifying observations based on their aerosol levels (B1) e.g., in [59]. Rayleigh scattering, which arises from the diffusion of very small atmospheric particles (particle scale < 1/10 wavelength), is particularly pronounced in shorter wavelength bands, such as deep blue or aerosol bands. Therefore, the magnitude of this effect is significant for visible wavelengths and must be corrected prior to analyzing the aerosol band, which is most affected by this scattering.

A k-means clustering approach (k = 3) was applied to the B1 data, allowing us to classify the lakes based on the severity of smoke interference into three categories: low (cluster 1), moderate (cluster 2), and high (cluster 3). Classification assesses the impact of smoke-affected observations on retrieval model accuracy, identifies the threshold beyond which wildfire smoke impairs data reliability, and determines conditions where the method remains viable for tracking phytoplankton dynamics.

2.4. Chl-a Retrieval Model Development

For the Chl-a retrieval model development, we tested three band formulas that we selected based on the physical reflectance and absorption properties of Chl-a and common usage. Blue (B2) and red (B4) bands were selected for their alignment with the maximum absorption of Chl-a, while the green band (B3) was chosen for its alignment with the maximum reflectance (i.e., minimal absorption) of Chl-a. These bands were incorporated into the development of band formulas to create a physically based approach for Chl-a retrieval from remote sensing images. The performances of three band formulas—B2/B3, B2/B4, and (B2−B4)/B3—were evaluated using 3-fold cross-validation. The B2/B3 [14,45] and B2/B4 [14,45,60,61] ratios are single-band formulas commonly used for mapping Chl-a in Landsat products. The third formula, (B2 − B4)/B3, has been widely adopted in Chl-a mapping on Landsat products (e.g., [17,19,60,62,63,64,65,66,67]).

2.5. Chl-a Retrieval Model Evaluation

To determine the optimal relationship between the band formulas and ln(Ch-a), we employed 3-fold cross-validation. The 3-fold cross-validation process involves dividing the data into three sets for training and testing, enhancing the generalizability and predictive accuracy of the models. The performance of each model was assessed using multiple metrics, including the coefficient of determination (R²) (Equation (8); Table 2), Root Mean Square Error (RMSE) (Equation (9); Table 2), Normalized Root Mean Square Error (NRMSE) (Equation (10); Table 2), Mean Absolute Error (MAE) (Equation (11); Table 2), and Bias (Equation (12); Table 2).

3. Results

3.1. Landsat Image Acquisition, Processing, and Analysis

The selected temporal window for matchups between ground-based Chl-a measurements and corresponding satellite observations at the point of sampling was ±3 days of satellite image capture (Figure 3). This temporal window achieved a balance between sample size and the correlations (represented by the Pearson correlation coefficient r) between satellite band formula and in situ ln(Chl-a). The ±1 day window had the smallest number of samples but the highest correlations with in situ ln(Chl-a), while the ±7 day window had the largest number of samples but smaller correlation values with in situ ln(Chl-a) compared to ±1 and ±3 days.

3.2. Clustering the Impact of Wildfires on Remote Sensing Imagery

To assess wildfire effects on remote sensing imagery, B1 Rrc values at the in situ sampling locations were extracted and subjected to k-means clustering analysis (k = 3) (Figure 4). The first cluster consisted of samples with low aerosol band values (B1 < 0.05, Rrc), indicating minimal interference. The second cluster represented lakes with moderate aerosol levels (0.05 ≤ B1 ≤ 0.07, Rrc), while the third cluster consisted of observations with high aerosol levels (B1 > 0.07, Rrc). The significance of differences between these clusters was evaluated using the Kruskal–Wallis test, a nonparametric method that compares mean ranks across clusters. The results indicated that B1 Rrc values were statistically distinct between the clusters (p < 0.001).

3.3. Performance of Chl-a Retrieval Models with Partial Atmospheric Correction

To investigate the impact of incorporating samples with varying aerosol band values into the Chl-a retrieval models, three calibration datasets were established:

• Calibration set 1: Includes cluster 1 (low wildfire interference).
• Calibration set 2: Includes clusters 1 and 2 (low and moderate wildfire interference).
• Calibration set 3: Includes all clusters (low, moderate, and high wildfire interference).

A 3-fold cross-validation method was employed, with

Table 3. The linear regression models developed for the three band formulas (B2/B4, B2/B3, and (B2 − B4/B3) using a ±3 days window based on (A) partial atmospheric correction using Rayleigh-corrected reflectance and ((B) full atmospheric correction using the Landsat 8 Level 2 process ready reflectance.

A. Partial Atmospheric Correction (Rayleigh-Corrected Reflectance)
Calibration Set (Sample Size)	Band Formula	Correlation (r)	Chl-a Retrieval Model	Training					Testing
				R2	RMSE	NRMSE	Bias	MAE	R2	RMSE	NRMSE	Bias	MAE
1 (n = 9)	x = (B2/B4)	−0.90	ln(Chl-a) = −2.796x + 7.685	0.80	0.48	0.41	0	0.41	0.79	0.53	0.37	+0.07	0.45
	x = (B2/B3)	−0.93	ln(Chl-a) = −6.360x + 9.923	0.84	0.51	0.37	0	0.48	0.87	0.18	0.30	−0.03	0.17
	x = (B2−B4)/B3	−0.95	ln(Chl-a) = −5.988x + 5.610	0.83	0.43	0.37	0	0.38	0.99	0.10	0.08	+0.05	0.08
2 (n = 17)	x = (B2/B4)	−0.89	ln(Chl-a) = −2.646x + 7.342	0.78	0.52	0.45	0	0.40	0.74	0.64	0.46	−0.09	0.58
	x = (B2/B3)	−0.90	ln(Chl-a) = −4.631x + 8.017	0.80	0.45	0.42	0	0.38	0.77	0.68	0.43	−0.01	0.55
	x = (B2−B4)/B3	−0.90	ln(Chl-a) = −4.580x + 4.879	0.80	0.55	0.43	0	0.45	0.87	0.42	0.32	+0.04	0.34
3 (n = 26)	x = (B2/B4)	−0.77	ln(Chl-a) = −3.315x + 8.383	0.58	0.88	0.63	0	0.73	0.61	0.62	0.59	+0.05	0.46
	x = (B2/B3)	−0.75	ln(Chl-a) = −4.749x + 8.377	0.52	0.90	0.67	0	0.64	0.61	0.70	0.59	+0.02	0.64
	x = (B2−B4)/B3	−0.78	ln(Chl-a) = −4.685x + 5.016	0.58	0.85	0.63	0	0.60	0.68	0.63	0.53	+0.06	0.58
B. Full Atmospheric Correction (Landsat 8 Level 2 Reflectance)
Calibration Set (Sample Size)	Band Formula	Correlation (r)	Chl-aRetrieval Model	Training					Testing
				R2	RMSE	NRMSE	Bias	MAE	R2	RMSE	NRMSE	Bias	MAE
1 (n = 9)	x = (B2/B4)	−0.84	ln(Chl-a) = −2.033x + 4.637	0.71	0.65	0.49	0	0.61	0.71	0.54	0.44	+0.01	0.51
	x = (B2/B3)	−0.86	ln(Chl-a) = −3.75x + 4.695	0.72	0.64	0.48	0	0.54	0.79	0.45	0.37	−0.05	0.39
	x = (B2−B4)/B3	−0.88	ln(Chl-a) = −4.331x + 2.501	0.80	0.47	0.41	0	0.41	0.74	0.65	0.41	+0.00	0.60
2 (n = 17)	x = (B2/B4)	−0.67	ln(Chl-a) = −2.157x + 5.121	0.37	0.95	0.76	0	0.76	0.64	0.74	0.53	+0.00	0.56
	x = (B2/B3)	−0.66	ln(Chl-a) = −2.981x + 4.787	0.38	0.87	0.75	0	0.74	0.46	1.03	0.66	−0.10	0.85
	x = (B2−B4)/B3	−0.69	ln(Chl-a) = −3.515x + 2.932	0.42	0.93	0.73	0	0.75	0.63	0.69	0.54	−0.06	0.56
3 (n = 26)	x = (B2/B4)	−0.63	ln(Chl-a) = −1.923x + 5.064	0.32	1.08	0.80	0	0.80	0.53	0.78	0.65	+0.09	0.68
	x = (B2/B3)	−0.63	ln(Chl-a) = −4.003x + 5.325	0.37	1.13	0.77	0	0.90	0.53	0.57	0.65	−0.06	0.50
	x = (B2−B4)/B3	−0.63	ln(Chl-a) = −3.867x + 3.018	0.35	1.06	0.78	0	0.81	0.52	0.79	0.66	+0.09	0.59

A detailing the performance metrics of the best-performing fold based on the test fold. This table summarizes the performance of Chl-a retrieval models using the partial atmospheric correction dataset developed and using three band formulas across various calibration sets. Models based on the (B2 − B4)/B3 formula exhibited higher R² and lower MRSE, NRMSE, and MAE than the other band formulas (B2/B4 and B2/B3) in testing datasets across all calibration sets. Calibration set 1 (low wildfire interference) consistently outperformed the other calibration sets (i.e., higher R², lower RMSE, NRMSE, and MAE). Calibration set 2 (low and moderate wildfire interference) showed slightly weaker performance than calibration set 1 across all band formulas, while calibration set 3 (low to high wildfire interference) exhibited the weakest performances.

3.4. Performance of Chl-a Retrieval Models with Full Atmospheric Correction

To assess how Chl-a retrieval models developed with partial atmospheric correction compared to those with full atmospheric correction, Landsat 8 Level 2 (process-ready) products were used. A calibration dataset incorporating both in situ and satellite overpass data was employed, with a matchup window size limited to ±3 days. Table 3B presents the performance metrics of the three Chl-a retrieval models using Landsat Level 2 products for the same calibration sets. The advantage of using models based on the (B2 − B4)/B3 formula compared to the other band formulas were less distinct using these Landsat data. In calibration set 1 (low wildfire interference), (B2 − B4)/B3 outperformed the other band formulas in the training dataset but B2/B3 exhibited the strongest performance in the testing dataset. In the other calibration sets (low and moderate, and low to high wildfire interference), there were few clear distinctions between the performance of different band formulas in either training or testing datasets, although B2/B3 exhibited a markedly weaker performance in calibration set 2 (low and moderate wildfire interference). As with Chl-a retrieval models developed with partial atmospheric corrected Landsat data, calibration set 1 (low wildfire interference) consistently outperformed the other calibration sets, and calibration set 3 (low to high wildfire interference) consistently exhibited the weakest performances.

Calibration set 1 demonstrated superior performance across all band formulas compared to other calibration sets. Models based on (B2 − B4)/B3 showed the highest R² and NRMSE values for calibration sets 1 and 2. Calibration set 3 consistently exhibited the weakest performance among all calibration sets.

3.5. Comparison of Chl-a Retrieval Modeling between Partial and Full Atmospheric Correction

Figure 5A illustrates the performance of Chl-a retrieval models as measured by R² using different band formulas and in different calibration sets using the training dataset of Rayleigh-corrected reflectance. Within each calibration set, there were no significant differences in R² between band formulas (p ≥ 0.001). Calibration set 1 (low wildfire interference) exhibited the highest R² values for all band formulas. There were small declines in R² for all band formulas from calibration set 1 to calibration set 2 (low and moderate wildfire interference), although a two-way ANOVA test indicated no significant difference in R² values between these calibration sets (p ≥ 0.001). Calibration set 3 (low to high wildfire interference) consistently demonstrated significantly (p < 0.001) weaker performance.

Figure 5B illustrates the performance of Chl-a retrieval models as measured by R² using different band formulas and in different calibration sets using the training dataset of Landsat Level 2 products. As with models developed from partial atmospheric correction data, there were no significant differences in R² between band formulas within each calibration set (p ≥ 0.001). Unlike models developed from partial atmospheric correction data, the improvement in R² for all band formulas using calibration set 1 was statistically significant (p < 0.001); the differences in R² calibration sets 2 and 3 were not significantly different (p ≥ 0.001).

Observations from calibration set 1 for both Rayleigh-corrected reflectance (Figure 5A) and Landsat Level 2 products (Figure 5B) correspond to minimal aerosol disturbance. Comparing the performance of this calibration set between the two reflectance products illustrates the impact of aerosol effects on Chl-a retrieval modeling in surface waters. Specifically, the full atmospheric correction employed in the Landsat Level 2 product (Figure 5B—calibration set 1) demonstrated weaker performance compared to the partial atmospheric correction (Figure 5A—calibration set 1), with R² values for the three different band formulas decreasing by between 0.03 and 0.12. The two-way ANOVA showed that the (Figure 5A) Rayleigh-corrected reflectance and (Figure 5B) Landsat Level 2 calibration sets between both products were statistically different across all three calibration sets (p < 0.001).

3.6. Effect of Aerosol Band Filtering on Lake Chl-a Concentrations across the Lake Winnipeg Watershed

To evaluate the performance of aerosol band filtering on removing pixels with elevated aerosol levels in the mapping lake Chl-a concentrations, we employed a model based on the (B2 − B4)/B3 band formula. This model was used to map phytoplankton biomass (i.e., Chl-a concentration) across lakes in the Lake Winnipeg watershed from 2013 to 2023. The formula was derived from calibration set 2 using Rayleigh-corrected reflectance. Although calibration set 2 (low and moderate wildfire interference) did not achieve the strongest performance among all sets (R² = 0.80), its performance was only marginally lower than that of calibration set 1 (low wildfire interference; R² = 0.83), with no statistically significant difference between the two sets. Calibration set 2 included nearly twice as many samples as calibration set 1 (17 versus 9) and, thus, providing a more extensive dataset for model development. When assessing differences in lake average Chl-a using Landsat OLI data from 2013 to 2023 (Figure 6A), we observed a slight restructuring of the frequency distributions between lakes without aerosol screening (Figure 6B) and with aerosol screening (Figure 6C). For higher productivity lakes (Chl-a > 54.6 µg/L), the frequency distribution remained similar in both scenarios, with 3964 observations without screening and 4001 observations with screening. However, for lower productivity lakes (Chl-a < 7.4 µg/L), slight skewness was observed with aerosol screening, with an increase in lower productivity observations with aerosol screening (n = 13,883) compared to those without screening (n = 12,769).

4. Discussion

The rising frequency and intensity of wildfires underscore the need for advanced monitoring techniques to address the challenges that elevated haze and smoke levels pose to remote sensing methodologies [68]. Elevated haze and smoke from wildfires reduce visibility and degrade satellite imagery quality [69,70], making it difficult to extract reliable information about surface properties, such as water quality characteristics [21,71,72].

Lake managers and practitioners require guidance on whether image quality is adequate for accurately characterizing algal dynamics.

4.1. Removing Effects of Severe Smoke

In this study, we applied a three-tiered image categorization system based on aerosol band values to assess smoke severity. We found that including samples with higher aerosol band values (i.e., higher smoke levels) in the calibration dataset led to a continuous decline in model performance, as indicated by decreasing R² values for all band formulas. When B1 Rrc values exceeded a threshold of 0.07, the accuracy of Chl-a models was significantly affected, and such observations should be excluded to avoid inaccuracies. Nevertheless, even with highly smoke-affected observations (cluster 3), the models achieved R² values around 0.55 (Table 3, section A, calibration set 3), which is still considered acceptable in previous studies using Landsat 8 OLI (e.g., [64,73,74]). We further tested each cluster independently (e.g., Table S2, where cluster 2 contains only moderate-impact observations), compared to the combinatorial approach used in Table 3 (e.g., calibration set 2 contains both low- and moderate-impact observations). We observed that retrieval models trained with low-impact samples (cluster 1) performed significantly better than those trained with moderate- or high-impact samples. The model trained with only moderate-impact samples outperformed the one trained with high-impact samples (Table S2). Despite the high performance observed, the low sample size limits the model’s generalization and affects its train test validity.

Our analysis of aerosol band filtering on Chl-a mapping (Figure 6A) revealed that, for most lakes in the Lake Winnipeg watershed from 2013 to 2023, the differences between applying and not applying aerosol band filtering were minimal, ranging from −5 to +5 µg/L. Additionally, Figure 6B and C demonstrates that for phytoplankton-rich lakes with Chl-a concentrations exceeding 54.6 µg/L, both methods produced similar frequency distributions of lakes. In contrast, for lakes with Chl-a concentrations below 54.6 µg/L, the differences between the two methods became more pronounced. Specifically, for lakes with Chl-a concentrations below 7.4 µg/L, aerosol band filtering resulted in a restructuring of the frequency distribution, with a greater number of lakes showing lower Chl-a concentrations compared to when the filter was not used.

A growing number of researchers are relying on Rayleigh-corrected reflectance, which remains vulnerable to smoke interference from wildfires [59]. This underscores the need for continued research into correction algorithms that can accurately account for elevated aerosol levels from wildfires across diverse surface types. Filtering Rayleigh-corrected reflectance with the coastal/aerosol band can be useful for detecting and excluding observations with high levels of smoke in the remote sensing of water bodies. Moreover, we found that full atmospheric correction, commonly used in previous studies, is often ineffective at removing wildfire effects and can impair model performance. This underperformance may be due to the overestimation of aerosol contributions, as documented in the literature [46,75]. The partial atmospheric correction approach (Rayleigh-corrected reflectance) developed in this study was further evaluated against established full atmospheric correction models (e.g., SIAC (Sensor Invariant Atmospheric Correction) and ACOLITE (Atmospheric Correction for OLI ‘lite’)). SIAC is a generalized atmospheric correction algorithm applied to various satellite sensors and was originally developed for terrestrial applications [76] but has recently been used for water areas [77,78]. ACOLITE is a tailored algorithm for correcting atmospheric effects in imagery from Landsat 8 (OLI) and Sentinel-2 and was developed for aquatic applications [79,80]. The comparison revealed that the partial atmospheric correction outperformed both the SIAC and ACOLITE models (see Table S3). ACOLITE performed better than SIAC but fell short of the performance achieved with Rayleigh-corrected reflectance. The atmospheric correction method developed by Page et al. [81] yielded a higher R² than SIAC and ACOLITE. A key innovation of approach of Page et al. [81] is the incorporation of in situ data to adjust the atmospheric correction process, particularly for complex aquatic environments. While this atmospheric correction method yielded similar performance to the Rayleigh-corrected reflectance for calibration set 1 (low wildfire interference; Table 3 and Table S3), its performance decreased when observations with elevated aerosol levels (calibration sets 2 (low and moderate wildfire interference) and 3 (low to high wildfire interference) were included in the modeling. A comparison of Rayleigh-corrected reflectance with the three atmospheric correction methods tested shows that modeling with Rayleigh-corrected reflectance consistently yielded better results across all calibration sets in terms of R² (see Table S3).

4.2. Evaluating Chl-a Retrieval Model Performance after Removal of Smoke Effects

The study area, which covers the prairie and boreal plain ecozones, encompasses a broad range of lake productivity, from oligotrophic to eutrophic conditions [28]. In the southern prairie ecozone, lakes are often eutrophic due to the region’s phosphorus-rich geology. Eutrophic conditions have been further exacerbated by intense agricultural activity over the past century [2,31]. In contrast, lakes in the northern boreal plain ecozone are typically oligotrophic due to minimal human activity [31]. Our calibration lakes exhibited a wide range of phytoplankton biomass, from 3.7 to 344.2 µg/L Chl-a (Table S1).

All three band formulas used for Chl-a retrieval showed strong performance (R² ≥ 0.78) when using calibration sets 1 (low wildfire interference) and 2 (low and moderate wildfire interference). Regardless of the band formula, calibration set 3 (low to high wildfire interference) performed poorly (R² ≤ 0.58), illustrating the negative impact of high wildfire smoke on model performance. The (B2 − B4)/B3 ratio exhibited the best performance in modeling Chl-a across all three calibration sets. This band formula has been extensively used for mapping Chl-a across a diverse range of aquatic environments (e.g., [19,60,63,64]) and is known for its broad applicability, effectively tracking Chl-a across different trophic states [17,63].

A key consideration when applying the (B2 − B4)/B3 ratio is its susceptibility to high chromophoric dissolved organic matter (CDOM). High CDOM can limit the effectiveness of this method, as CDOM interferes with phytoplankton detection in the blue spectrum [65]. However, our calibration lakes had relatively low color (<46.6 mg L⁻¹ PtCo), supporting the suitability of this method in our analysis (Table S1).

5. Conclusions

As wildfires become more frequent and intense worldwide, the resulting smoke plumes can travel hundreds to thousands of kilometers, posing significant challenges to the accuracy of satellite sensor data, which are crucial for monitoring freshwater lakes. To maintain precision in water quality assessments under these wildfire-affected conditions, it is essential to implement tailored strategies, such as refined atmospheric corrections and well-defined cutoff thresholds. In this study, we identified a specific threshold at which wildfire smoke significantly impairs the prediction of algal productivity. By integrating this threshold with Rayleigh-corrected reflectance, we effectively assessed lakes impacted by wildfire smoke. This highlights the need for remote sensing practitioners to recognize when predictive accuracy is compromised, ensuring more reliable monitoring and assessment of wildfire-affected regions.