Satellite and airborne remote sensing of gross primary productivity in boreal Alaskan lakes

The Yukon Flats Basin is an extensive, 33 000 km² complex of low-lying lakes and wetlands inside the Yukon River Basin (figure 1). Dry, cold and underlain by discontinuous permafrost (Minsley et al 2012, Rey et al 2019), the region's continental sub-arctic climate has an annual average temperature of ∼−3 °C (Chen et al 2014) with short, ice-free summers (April/May to September/October) (Gallant 1995). Over 30 000 lakes dot this semi-arid landscape (Heglund and Jones 2003). Despite abundant surface water, the area receives very little precipitation (250 mm) and lake water balances are dominated by evaporation (Anderson et al 2013, 2019). Lakes are on the whole shallow and macrophyte-rich; historical surveys demonstrated that lake bottoms are carpeted with up to 80% coverage of aquatic vegetation (Glesne 1986).

Zoom In Zoom Out Reset image size

Study lakes (n = 7) representative of the region were carefully selected based on results from historical surveys (Heglund and Jones 2003, Halm and Guldager 2012, Halm and Griffith 2014) and a 2016 field study (Bogard et al 2019) to represent a gradient taking into consideration depth, pH, dissolved organic carbon (DOC) concentration, chlorophyll-a and chromophoric dissolved organic matter (CDOM) absorption (table S1).

Lakes were sampled by boat and/or floatplane (figures 1(a) and (b)) during a 2018 field campaign conducted as part of the National Aeronautics and Space Administration's (NASA's) Arctic-Boreal Vulnerability Experiment (ABoVE) for a suite of parameters (section 2.2, text s1).

Detailed chemical and optical measurements were collected from the center of each lake as well as in situ surface reflectance (R_s). Lake GPP was estimated using an oxygen isotope (δ¹⁸O) mass balance approach adapted from (Quay et al 1995, Bocaniov et al 2012) as described in Bogard (2017) and text S1. For clarity, GPP derived from δ¹⁸O field measurements will be referred to hereafter as in situ GPP in contrast to satellite-derived GPP. Turbidity, CDOM absorbance and phytoplankton chlorophyll-a were also measured to determine their influence, if any, on lake color. CDOM is the proportion of dissolved organic matter that absorbs strongly in ultraviolet and visible wavelengths (detailed methods in text S1).

The launch of commercial small satellites with daily return times and the increasing availability of hyperspectral imagery through large-scale campaigns such as NASA ABoVE has sparked interest in finer spatiotemporal analysis. In this study, we capitalize on the strengths of three different satellites: Landsat-8, Sentinel-2 and CubeSats (Planet, Inc). The sensors on-board Landsat-8 (L8), Sentinel-2 (S2) and PlanetScope (PS) all possess bands in the visible and near-infrared (VNIR) with varying spatial and temporal resolutions (figure 2). We also incorporate hyperspectral airborne imagery from AVIRIS-NG (figure 2).

Zoom In Zoom Out Reset image size

As part of the ABoVE campaign, AVIRIS-NG imagery was acquired on July 22, 2018. Level-2 (L2) atmospherically corrected R_s (Bue et al 2015) was downloaded from https://avirisng.jpl.nasa.gov/dataportal/. Sentinel-2 acquisitions were identified and downloaded from the U.S. Geological Survey Earth Explorer (https://earthexplorer.usgs.gov/) and atmospherically corrected using the Land Surface Reflectance Code (LaSRC) (Vermote et al 2016). Sentinel-2 images were also corrected using the dark spectrum fitting (DSF) algorithm available in ACOLITE (Python Version 20 190 326.0), which is an open-source atmospheric correction tested successfully in coastal waters (Vanhellemont 2019). PlanetScope Level-2 products were identified using the Planet Labs API (Planet 2017) and a cloud-native, scalable workflow called SWEEP (John et al 2019). No cloud-free Landsat overpasses occurred during or close in time to field operations.

To utilize long-term Landsat records, we also calculated Landsat growing season composites, which is the per-pixel median greenness over the growing season (June, July and August). A common remote sensing tool used to quantify terrestrial productivity (Roy et al 2010), growing season composites provide a continuous, stable record of seasonal lake color. While Landsat overpasses are infrequent (16 d) compared to Planet (∼daily), the Landsat program has the longest global record of Earth observation with acquisitions initiated in the mid-1980s (Lymburner et al 2016), making it a powerful tool for historical analysis. For L8 composites, Collection 1 Level 2 R_s was accessed using Google Earth Engine. To generate seasonal composites, per-pixel median greenness was calculated from all cloud-free images acquired during the growing season (June 1–September 30) over the study area. With the exception of ACOLITE, all R_s products described above are derived using 6 SV-based approaches (Vermote et al 2006). For brevity only those results are shown in the main text, but ACOLITE results are given in figures S1–2 (available at stacks.iop.org/ERL/15/105001/mmedia), table S3.

In order to verify in situ lake color, R_s (350–2500 nm) was measured at each lake center in 2018 (n = 7) using an ASD FieldSpec 4 High Resolution Spectroradiometer (3 nm and 8 nm resolution in VNIR and SWIR) (text S2). Above-water, in situ lake R_s was collected 3–5 d prior to overpasses because same-day in situ validation was not possible due to weather delays.

Lake color was extracted for each sampling location from PlanetScope, Sentinel-2, Landsat-8 and AVIRIS scenes for each lake (n = 7) using Google Earth Engine (Gorelick et al 2017). A 3 × 3 pixel-wise box centered on field GPS coordinates was drawn for each lake and the median R_s within each box was extracted from the image in order to reduce bias introduced by spatial sampling (Pahlevan et al 2016). In the case of scenes overlapping above a site, the median value between scenes was extracted. Since sampling was done at lake centers, all selected pixels were open water. A restrictive F-mask filter was also used to confirm selected pixels were water (Zhu et al 2015). Lake color from each respective platform was then regressed against in situ GPP (n = 7) using a reduced major axis (RMA) type II regression (Python package plyr2, (Haëntjens 2018)). This same workflow was also used to extract lake color from Landsat seasonal composites (section 2.3) for the 2016 samples (n = 24). The regression model was used to create satellite-derived maps of GPP by taking the regression equation derived from the GPP ∼ greenness relationship observed in the 2018 field campaign and applying it on a per-pixel basis to Sentinel-2 imagery using the green band as the model input.

In order to make hyperspectral lake color comparable to multi-spectral observations, in situ and AVIRIS-NG R_s was spectrally convolved to Sentinel-2 wavelengths as in (Pahlevan et al 2017):

$$\begin{equation*}{R_{rs}}\left( {{\lambda _j}} \right) = \frac{{\mathop \sum \nolimits_{i = 1}^n {R_{rs}}\left( {{\lambda _i}} \right) X RSR \left( {{\lambda _i}} \right)}}{{\mathop \sum \nolimits_{i = 1}^n {R_{rs}}\left( {{\lambda _i}} \right)}}\end{equation*}$$

where:

${R_{rs}}\left( {{\lambda _j}} \right)$: band center wavelength

RSR: spectral response factor

n: total number of hyperspectral band centers (i)

Median absolute percent different (MAPD) was used to compare sensor R_s:

$MAPD = median \left( {abs \left( {\frac{{x - y}}{x}} \right)} \right)*100$ where x is sensor 1 and y is sensor 2. For decadal time series analysis (figure 4), the non-parametric Mann–Kendall estimator and Sen's Slope were used to determine whether there was a positive or negative trend in lake greenness and its significance. Thiel–Sen Slope was calculated using the Scipy Python-based scientific computing package (Virtanen et al 2020) and significance of the slope was determined using a Mann–Kendall test (Hirsch et al 1982).

In situ median GPP rates (table 2) were high at 9.72 mg O₂ m⁻² d⁻¹ and spanned a wide range (interquartile range = 6.25 to 10.50 mg O₂ m⁻² d⁻¹). In situ GPP is on the same order of magnitude as other free-water techniques estimating GPP (Solomon et al 2013). The rates observed here are higher than more dilute and often deeper boreal lakes (Ask et al 2012, Deininger et al 2017) and are consistent with the high productivity observed in shallow, macrophyte-rich lakes found in the northwestern Siberia lowlands (Manasypov et al 2014) and shallow lakes in the Mackenzie Delta (Squires et al 2009, Tank et al 2009).

Other influences on lake color must be ruled out in order to avoid applying this empirical method over-ambitiously. To this end, we characterized variables (table 2) known to influence lake color, including phytoplankton pigments, CDOM and turbidity (Aurin and Dierssen 2012). We observed relatively low phytoplankton biomass based on HPLC Chl-a pigment analysis (median = 3.4 m⁻¹, interquartile range = 2.4–5.9 µg l⁻¹) (table 1). Overall, this is consistent with a previous study of 129 Yukon Flats lakes (table S1) showing Chl-a was lower than expected based on nutrient availability (Heglund and Jones 2003). The relatively clear waters and high GPP observed here in the presence of low phytoplankton biomass suggest macrophyte production is driving lake GPP (Genkai-Kato et al 2012). One exception was YF17, where field observations of a bloom were confirmed by high HPLC-derived Chl-a (83.3 µg l⁻¹)

The majority of lakes were shallow (median depth of 1.6 m) (table 1), with light extending to epiphytic periphyton and aquatic macrophytes. One lake was >10 m deep; the depth, low GPP and low Chl-a (table 2) of this lake provided an end-point representative of the gradient in the area. Other studies have established that submerged and floating macrophytes reflect in the green (560 nm) and near-infrared (850–880 nm) (Silva et al 2008, Oyama et al 2015, Palmer et al 2015, Dogliotti et al 2018). The shallow depths of these relatively clear lakes, as inferred by low turbidities (<1 FNU in most cases) could allow signals from macrophytes and periphyton to reach the surface (Mobley et al 2020). Given the low Chl-a, low turbidities and the observed presence of macrophytes within the water column we suggest that overall, phytoplankton production is not a significant component of the optical signature.

A second influence on lake color is absorption of light by CDOM. CDOM can interfere with the remote sensing of GPP in two ways. CDOM absorbs green light, so high CDOM has the direct effect of masking green light reflected from photosynthetic pigments (Carder et al 1989). Indirectly, CDOM can shade out aquatic producers, thus having the ecological impact of reducing GPP (Ask et al 2012, Seekell et al 2015). In order to rule out this influence, CDOM data were collected at each site. CDOM values (table 2) in sampled lakes were low (a₄₄₀, median = 1.13 m⁻¹, interquartile range = 0.78–1.82 m⁻¹), showing that lakes are relatively clear despite their high DOC (10.2 to 25.6 mg l⁻¹) (table 2). Our findings are consistent with other studies in semi-arid aquatic ecosystems (Osburn et al 2011, Bogard et al 2019, Kellerman et al 2019) that show low CDOM despite high DOC. While coupling between DOC and CDOM has been demonstrated at global scales (Massicotte et al 2017), CDOM-DOC decoupling in these semi-arid landscapes is hypothesized to result from decreased hydrologic connectivity which reduces terrestrial organic carbon subsidies as aromatic carbon compounds to aquatic ecosystems, resulting in clearer lakes than otherwise expected (Johnston et al 2020).

Finally, lake color can be influenced by scattering from sediment during periods of hydrologic connectivity to surface waters or during lake turnover. However, turbidity across lakes was low (<1 FNU) (table 2) with the exception of YF17, where the previously noted algal bloom drove turbidity up to 35 FNU. Taken together, these results indicate the clear and shallow nature of these macrophyte-rich lakes creates favorable conditions for the remote sensing of benthic properties.

Lake color from satellite and airborne sensors was regressed against 2018 in situ GPP (mg O₂ m⁻² d⁻¹) (n = 7), revealing a strong, positive correlation between greenness and GPP (figure 3(a), p-value < 0.05). We then applied this regression pixel-wise to Sentinel-2 imagery to create spatially explicit GPP maps (figures 3(b) and (c)) for each lake. Modeled, S2 GPP ranged from 3.9 ± 1.14 mg O₂ m⁻² d⁻¹ in the deep, upland Boot Lake to 14.21 ± 6.42 mg O₂ m⁻² d⁻¹ in shallow, eutrophic YF17.

Zoom In Zoom Out Reset image size

The relationship was then evaluated using a separate dataset from a 2016 field campaign that had been withheld for independent testing (figure 4). As same-week images were not available, we instead regressed seasonal L8 composites against 2018 (n = 7) and 2016 in situ GPP (n = 17) (Bogard et al 2019). The correlation between greenness and in situ GPP was maintained across both models.

Zoom In Zoom Out Reset image size

The L8 seasonal composites produced a slightly stronger relationship (type 2 major axis regression: y = 385.81x−0.45; r² = 0.69; p = 2.60 × 10^–4, n =24) than the Sentinel-2 same-week model. This likely results from the smoothing of short-term variability from the composite and the relative stability of oxygen pools (Bogard et al 2017) in these shallow lakes on seasonal scales. The separate models provide an independent check against each other, further substantiating the robustness of this approach across sensors and years.

The use of Landsat allows us to tap into the powerful historical imagery available through the Landsat archive in order to detect decadal (Wang et al 2014, Lymburner et al 2016) and seasonal (Kallio et al 2008) trends in lake color. Lake greenness computed for Scoter Lake (figure 5) showed a decline from 1984 to 2018. Field studies suggest that declines in greenness could result from increased CDOM absorption resulting from thermokarst processes for lakes with higher absorption (Wauthy et al 2018), increases in depth (Duguay and Lafleur 2003), or increased sediment loading during periods of surface water connectivity (Vonk et al 2015). Subsequent research combining these techniques with paleolimnology (Pienitz et al 1999, Bouchard et al 2017) and radiocarbon (Rühland et al 2003) data has the potential to unveil past lake processes, although more testing is required to establish mechanistic links between these processes and lake color across more diverse ecosystems.

Zoom In Zoom Out Reset image size

Surface reflectance from different sensors showed the greatest agreement in the green band.

Using high-resolution PlanetScope and AVIRIS-NG R_s (table S2), Sentinel-2 and in situ R_s (figures 6, S1-2), we compared R_s and respective model performance using 2018 in situ GPP. Analyzing agreement between the blue, green, red and NIR bands, we found the best agreement in the green (MAPD median = 29%, interquartile range 8%–53%) (figure S1-2, text S3-4).

Zoom In Zoom Out Reset image size

Green R_s is less subject to inter-sensor differences (Pahlevan et al 2018) and calibration errors between Landsat sensors (Vogelmann et al 2016). S2's green R_s has been successfully ground-truthed over small thermokarst lakes and ponds by Freitas et al (2019) and green R_s is also less impacted than NIR by particulates (Wang and Philpot 2007). Green R_s also has the major advantage of being measured by almost every spectral remote sensing platform, dating back to early Landsat (Dwyer et al 2018) and including the upcoming Plankton Aerosol, Cloud, Ocean Ecosystem (PACE) Mission (Werdell et al 2019). Finally, previous studies have shown greenness covaries with both phytoplankton (O'Reilly et al 1998, O'Reilly and Werdell 2019) and macrophytes (Hestir et al 2015). In light of the larger disagreements between sensors detected in the NIR (text S4), these results suggest that green R_s may be a stable, sensor-agnostic quantity suitable for GPP modeling across a range of spectral, spatial and temporal resolutions.

In situ GPP (mg O₂ m⁻² d⁻¹) corresponded with greenness (table 3, figures S3–4). The strongest correlation was between in situ GPP and in situ R_s (r² = 0.68; p-value < 0.05) and CubeSat R_s (r² = 0.78; p-value < 0.05). This could be due to two factors. First, in situ R_s was collected simultaneous to in situ GPP, in contrast to the overpasses 2–5 d later. However, unlike dynamic river and coastal systems, processes in these lakes vary on seasonal and annual scales. For example, high-frequency daily observations of dissolved oxygen collected from May to September (Bogard et al 2019) show consistently positive rates of net production driven by intense plant growth was sustained throughout most of the growing season. Johnston et al (2020) also showed diel dissolved organic matter composition and optical properties changed minimally over a 3 d period. A second reason for the stronger performance of PlanetScope and the in situ sensor's performance could be because of their finer spatial resolution relative to other sensors (10–30 m).

While restricting the time difference between field and satellite observations can improve model uncertainty in dynamic river systems (Kuhn et al 2019), albeit time differences of up to 3 to 5 d have been used successfully in several lake studies (Kloiber et al 2002, Chipman et al 2004, Olmanson et al 2008, Boucher et al 2018). Shorter time differences are ideal, yet given the temporal stability of the dissolved oxygen pool and water color over several days in these lakes, we are confident that these results are representative and provide important insights, particularly given the lack of comparisons of surface reflectance over inland waters (Maciel et al 2020).

A second advantage of Sentinel-2 over Landsat-8 is the addition of red-edge bands suitable for mapping vegetation. An even stronger correlation than with greenness was observed between GPP and the red-edge band (704 nm) (table 3, S3). Models improved by an average of 16% (table 3) due to the strong red-edge reflectance of macrophytes. This suggests that, when available, red-edge bands may be the strongest candidate for modeling GPP in these shallow lake ecosystems whose GPP is driven by both macrophytes and phytoplankton. This finding has implications for future missions as red-edge bands with narrower bandpasses are increasingly being added to, for example, Planet Lab's next-generation Dove satellites. While the green band is suitable for historic analysis, these findings encourage red-edge based algorithm development due to the strong signal in these shallow arctic and boreal lakes.

The correlation between green and NIR bands with in situ GPP across sensors and sites indicates their potential as a simple GPP proxy for shallow lakes with low turbidity and CDOM. For context, global ocean models typically only explain 40% of water-column-integrated primary production (Siegel et al 2001). The strength of this relationship between lake color and in situ GPP for both 2016 and 2018 datasets suggests potential for the development of a remote sensing modeling framework for shallow lake GPP.