Regional measurements and spatial/temporal analysis of CDOM in 10,000+ optically variable Minnesota lakes using Landsat 8 imagery.

Information on colored dissolved organic matter (CDOM) is essential for understanding and managing lakes but is often not available, especially in lake-rich regions where concentrations are often highly variable in time and space. We developed remote sensing methods that can use both Landsat and Sentinel satellite imagery to provide census-level CDOM measurements across the state of Minnesota, USA, a lake-rich landscape with highly varied lake, watershed, and climatic conditions. We evaluated the error of satellite derived CDOM resulting from two atmospheric correction methods with in situ data, and found that both provided substantial improvements over previous methods. We applied CDOM models to 2015 and 2016 Landsat 8 OLI imagery to create 2015 and 2016 Minnesota statewide CDOM maps (reported as absorption coefficients at 440 nm, a440) and used those maps to conduct a geospatial analysis at the ecoregion level. Large differences in a440 among ecoregions were related to predominant land cover/use; lakes in ecoregions with large areas of wetland and forest had significantly higher CDOM levels than lakes in agricultural ecoregions. We compared regional lake CDOM levels between two years with strongly contrasting precipitation (close-to-normal precipitation year in 2015 and much wetter conditions with large storm events in 2016). CDOM levels of lakes in agricultural ecoregions tended to decrease between 2015 and 2016, probably because of dilution by rainfall, and 7% of lakes in these areas decreased in a440 by ≥3 m-1. In two ecoregions with high forest and wetlands cover, a440 increased by >3 m-1 in 28 and 31% of the lakes, probably due to enhanced transport of CDOM from forested wetlands. With appropriate model tuning and validation, the approach we describe could be extended to other regions, providing a method for frequent and comprehensive measurements of CDOM, a dynamic and important variable in surface waters.

Understanding the Effects of Optimal Combination of Spectral Bands on Deep Learning Model Predictions: A Case Study Based on Permafrost Tundra Landform Mapping Using High Resolution Multispectral Satellite Imagery.

Deep learning (DL) convolutional neural networks (CNNs) have been rapidly adapted in very high spatial resolution (VHSR) satellite image analysis. DLCNN-based computer visions (CV) applications primarily aim for everyday object detection from standard red, green, blue (RGB) imagery, while earth science remote sensing applications focus on geo object detection and classification from multispectral (MS) imagery. MS imagery includes RGB and narrow spectral channels from near- and/or middle-infrared regions of reflectance spectra. The central objective of this exploratory study is to understand to what degree MS band statistics govern DLCNN model predictions. We scaffold our analysis on a case study that uses Arctic tundra permafrost landform features called ice-wedge polygons (IWPs) as candidate geo objects. We choose Mask RCNN as the DLCNN architecture to detect IWPs from eight-band Worldview-02 VHSR satellite imagery. A systematic experiment was designed to understand the impact on choosing the optimal three-band combination in model prediction. We tasked five cohorts of three-band combinations coupled with statistical measures to gauge the spectral variability of input MS bands. The candidate scenes produced high model detection accuracies for the F1 score, ranging between 0.89 to 0.95, for two different band combinations (coastal blue, blue, green (1,2,3) and green, yellow, red (3,4,5)). The mapping workflow discerned the IWPs by exhibiting low random and systematic error in the order of 0.17-0.19 and 0.20-0.21, respectively, for band combinations (1,2,3). Results suggest that the prediction accuracy of the Mask-RCNN model is significantly influenced by the input MS bands. Overall, our findings accentuate the importance of considering the image statistics of input MS bands and careful selection of optimal bands for DLCNN predictions when DLCNN architectures are restricted to three spectral channels.

Assessment of the chlorine demand and disinfection byproduct formation potential of surface waters via satellite remote sensing.

The ability of satellites to assess surface water quality indicators such as colored dissolved organic matter (CDOM) suggests that remote sensing could be a useful tool for evaluating water treatability metrics in considering potential drinking water supplies. To explore this possibility, 24 surface water samples were collected throughout Minnesota, USA with wide ranging values of CDOM (a440; 0.41-27.9 m-1), dissolved organic carbon (DOC; 5.5-47.6 mg/L) and specific ultraviolet absorbance at 254 nm (SUVA254; 1.3-5.1 L/mg-M). Laboratory experiments were performed to quantify chlorine demand and the formation of two classes of halogenated disinfection byproducts (DBPs), trihalomethanes (THMs) and haloacetic acids (HAAs), using the uniform formation conditions (UFC) test. Chlorine demand and THMUFC were linearly correlated with CDOM (R2 = 0.97 and 0.91, respectively), indicating that CDOM is a useful predictor of these parameters. On the other hand, data comparing di- and tri-HAAUFC with CDOM were better fit by a logarithmic relationship (R2 = 0.73 and 0.87, respectively), while mono-HAAUFC was linearly correlated with CDOM (R2 = 0.46) but only for low-to moderately-colored waters (a440 ≤ 11 m-1). The correlations relating chlorine demand and DBPUFC values with CDOM were coupled with satellite CDOM assessments to estimate chlorine demand and DBPUFC values for all surface waters larger than 0.05 km2 in the state of Minnesota, USA. The resulting maps suggest that only 21.8% of Minnesota lakes would meet both the THM and HAA maximum contaminant levels, but only when pre-disinfection treatment removes 75% of DBP precursors. There are limitations to determining CDOM using satellites for high color surface waters (a440 > 11 m-1), however, leading to underpredicted values for CDOM, chlorine demand, and DBPUFC. Overall, the results demonstrate the potential benefits of satellite remote sensing for assessing potential drinking water sources and water treatability metrics.

Iron influence on dissolved color in lakes of the Upper Great Lakes States.

Colored dissolved organic matter (CDOM), a major component of the dissolved organic carbon (DOC) pool in many lakes, is an important controlling factor in lake ecosystem functioning. Absorption coefficients at 440 nm (a440, m-1), a common measure of CDOM, exhibited strong associations with dissolved iron (Fediss) and DOC in 280 lakes of the Upper Great Lakes States (UGLS: Minnesota, Wisconsin, and Michigan), as has been found in Scandinavia and elsewhere. Linear regressions between the three variables on UGLS lake data typically yielded R2 values of 0.6-0.9, suggesting that some underlying common processes influence organic matter and Fediss. Statistical and experimental evidence, however, supports only a minor role for iron contributions to a440 in UGLS lakes. Although both DOC and Fediss were significant variables in linear and log-log regressions on a440, DOC was the stronger predictor; adding Fediss to the linear a440-DOC model improved the R2 only from 0.90 to 0.93. Furthermore, experimental additions of FeIII to colored lake waters had only small effects on a440 (average increase of 0.242 m-1 per 100 µg/L of added FeIII). For 136 visibly stained waters (with a440 > 3.0 m-1), where allochthonous DOM predominates, DOM accounted for 92.3 ± 5.0% of the measured a440 values, and Fediss accounted for the remainder. In 75% of the lakes, Fediss accounted for < 10% of a440, but contributions of 15-30% were observed for 7 river-influenced lakes. Contributions of Fediss in UGLS lakes to specific UV absorbance at 254 nm (SUVA254) generally were also low. Although Fediss accounted for 5-10% of measured SUVA254 in a few samples, on average, 98.1% of the SUVA254 signal was attributable to DOM and only 1.9% to Fediss. DOC predictions from measured a440 were nearly identical to those from a440 corrected to remove Fediss contributions. Overall, variations in Fediss in most UGLS lakes have very small effects on CDOM optical properties, such as a440 and SUVA254, and negligible effects on the accuracy of DOC estimated from a440, data for which can be obtained at broad regional scales by remote sensing methods.

Color, chlorophyll a, and suspended solids effects on Secchi depth in lakes: implications for trophic state assessment.

Secchi depth (SD), a primary metric to assess trophic state, is controlled in many lakes by algal densities, measured as chlorophyll-a (chl-a) concentration. Two other optically related water quality variables also directly affect SD: non-algal suspended solids (SSNA ) and colored dissolved organic matter (CDOM, expressed as the absorption coefficient at 440 nm, a440 ). Using a database of ~1,460 samples from ~625 inland lake basins in Minnesota and two other Upper Midwest states, Wisconsin and Michigan, we analyzed relationships among these variables, with special focus on CDOM levels that influence SD values and the Minnesota SD standards used to assess eutrophication impairment of lakes. Log-transformed chl-a, total suspended solids (TSS), and SD were strongly correlated with each other; log(a440 ) had major effects on log(SD) but was only weakly correlated with log(chl-a) and log(TSS). Multiple regression models for log(SD) and 1/SD based on the three driving variables (chl-a, SSNA , and CDOM) explained ~80% of the variance in SD in the whole data set, but substantial differences in the form of the best-fit relationships were found between major ecoregions. High chl-a concentrations (> 50 µg/L) and TSS (> 20 mg/L) rarely occurred in lakes with high CDOM (a440  > ~4 m-1 ), and all lakes with a440  > 8 m-1 had SD ≤ 2.0 m despite low chl-a values (<10 µg/L) in most lakes. Further statistical analyses revealed that CDOM has significant effects on SD at a440 values > ~ 4 m-1 . Thus, SD is not an accurate trophic state metric in moderately to highly colored lakes, and Minnesota's 2-m SD criterion should not be the sole metric to assess eutrophication impairment in warm/cool-water lakes of the Northern Lakes and Forest ecoregion. More generally, trophic state assessments using SD in regions with large landscape sources of CDOM need to account for effects of CDOM on SD.

Limitations on using CDOM as a proxy for DOC in temperate lakes.

Colored dissolved organic matter (CDOM) has been widely studied as part of efforts to improve understanding of the aquatic carbon cycle, by laboratory, in situ, and remote sensing methods. We studied ecoregion-scale differences in CDOM and dissolved organic carbon (DOC) to understand variability in organic matter composition and the use of CDOM as a proxy for DOC. Data from 299 lakes across the U.S. Upper Midwest showed that CDOM, measured as absorptivity at 440 nm (a440), correlated strongly with DOC (R2 = 0.81, n = 412). Colored lakes in the Northern Lakes and Forests (NLF) ecoregion drove this relationship. Lakes in the North Central Hardwood Forests (NCHF) had low color (most had a440 < 3 m-1) and weaker CDOM-DOC relationships (R2 = 0.47). Spectral slopes and specific ultraviolet absorbance (SUVA), indicated relatively low aromaticity and non-terrestrial DOM sources in low color lakes. Multiple regression analyses that included total dissolved nitrogen (TDN) and CDOM, but not chlorophyll a, improved DOC estimates in low color lakes, suggesting a dominant contribution of non-planktonic sources of low color DOM in these lakes. Our results show that CDOM is a reliable, regional proxy for DOC in lakes where forests and wetlands dominate the landscape and the DOM is primarily terrestrial in origin. Mapping of lake DOC at broad spatial scales by satellite-derived CDOM has lower accuracy in low color lakes.