Human activities change suspended sediment concentration along rivers

ML is increasingly used in remote sensing of inland water quality because of its robust predictions, low computational costs, and the challenges of applying physically-based models in optically complex waters (Peterson et al 2018, Pahlevan et al 2020, Sagan et al 2020). AquaSat was designed as an inclusive database, therefore we performed additional quality control to identify matchups for algorithm development (see SI). We generated a diverse and representative training dataset across space, time, SSC values, and waterbodies (e.g. estuaries, lakes/reservoirs, and rivers). The resulting training-validation dataset had 21 671 match-ups with 2407 set aside as test hold-outs (10% of all match-ups) across the contiguous United States (CONUS) with SSC concentrations ranging from 0.1 to 2228 mg l⁻¹ (figure S1).

To reduce overfitting, spatial-temporal bias, and to develop a parsimonious model, we used forward feature selection (FFS) to select the optimal SSC predictors, combined with leave-location-leave-time out cross validation (Meyer et al 2018, 2019) within an Xgboost generalized linear regression ML algorithm using the CAST (Meyer et al 2018), CARET (Kuhn 2008), and Xgboost packages in R (Chen et al 2015). FFS found ten optimal features (figure S2). The Xgboost regression model hyper-parameters were tuned using a grid search constrained to conservative values to avoid overfitting and produce the final model (Topp et al 2021). We calculated error metrics such as root mean square error (RMSE), mean absolute error (MAE), bias, and relative bias (Dethier et al 2020). We focus on MAE and relative bias, as they are the preferred metrics for remote sensing of water quality validation, particularly since SSC spans nearly six orders of magnitude (Seegers et al 2018).

We quantified trends in mean annual SSC over time using the non-parametric Mann–Kendall test, which is robust to outliers. We analyzed trends at 28 397 river reaches with representative data, only including reaches that had more than four SSC observations per year sampled across two or more seasons for more than 15 years. Seventy-five percent of reaches had a time series of 30 years or more. Significant trends were identified if the p-value was <0.05. SSC rates of change (mg l yr⁻¹) were quantified using the Theil–Sen slope estimator (Sen 1968).

We identified longitudinal SSC patterns using the Mann–Kendall trend test as a robust and conservative test. We calculated the long-term mean SSC for reaches with representative data that had more than 50 SSC observations collected over at least 15 years. We restricted analysis to rivers with dense spatial observations and filtered to rivers that had more than 100 km of spatially continuous SSC observations, resulting in 244 unique rivers (53% of the rivers in RivSed). The SSC patterns were categorized as increasing downstream, decreasing downstream, or no pattern depending on if the slope was significantly positive, significantly negative, or not statistically significant. To test the impact of dams on longitudinal SSC patterns, we calculated metrics that represent the average and cumulative effect of dams along the 244 rivers, including median dam height (m), median dam storage (mm), and cumulative reservoir water storage (mm). Storage metrics were converted to mm by normalizing volume to basin area to enable comparisons across river basins with different sizes and climates. Dam density was calculated as the number of dams divided by the total length of the river (dams km⁻¹). Dam characteristics were derived from the National Anthropogenic Barriers Database (NABD) (Ostroff et al 2013) and US Army Corps of Engineers National Inventory of Dams (NID). We tested if rivers with longitudinal SSC increases versus rivers without longitudinal increases had differences in dam metrics using the non-parametric Kruskal–Wallis test using an alpha of 0.05.

To analyze the impact of widespread damming of US rivers over the last century (Graf 1999) on longitudinal SSC patterns, we compared our observed to modeled longitudinal SSC patterns from a global sediment transport model, WBMsed (Cohen et al 2013, 2014). To isolate the impacts of dams, we used a model simulation without dams but kept other contemporary hydrological and geomorphic conditions. Similar to observed data, we classified rivers in the 'no dam' scenario as having decreasing, increasing, or no longitudinal patterns. We then compared how the fraction of rivers with different downstream patterns changed between observed and modeled 'no dam' scenarios. For more information about the model structure, inputs, and post-processing see Cohen et al (2022).

We quantified temporal stability in SSC profiles along major rivers by comparing the shape of mean annual SSC profiles to the shape of the long-term mean SSC profile within each river. For data consistency and representativeness, we filtered RivSed to river reaches with more than four observations per year across all 34 years and used the same reaches across all years to represent mean annual SSC profiles leaving 190 unique rivers (41% of rivers in RivSed, figure S3). To isolate differences in only SSC profile shapes, we log transformed SSC because SSC can span several orders of magnitude along a river and then detrended temporal changes in SSC. The normalized distance between annual SSC profiles and the long-term mean SSC profile within each river was calculated using dynamic time warping, which is widely used to quantify the similarity between two ordered sequences with high variability, such as time series, sound waveforms, or animal movements (Giorgino 2009, Cleasby et al 2019). For clarity, we refer to normalized distance as 'stability,' where higher values represent similar profile shapes, or more stability, and low values approaching 0 represent dissimilar profile shapes, or less stability (figure S4). We then examined human and natural factors related to the SSC profile stability value for each river using dam and basin properties from NID and StreamCat (Hill et al 2016), such as drainage area (km²) at the outlet of each major river, percent cropland cover within the drainage area, and the median dam height (m) along major rivers.

Our model performed as well or better than most Landsat-based empirical SSC models (table S1) with error metrics of MAE = 12 mg l⁻¹, RMSE = 29 mg l⁻¹, and relative bias = 0.59 (figure 1). Published models use various error metrics and showed RMSE ranging from 16 to 286 mg l⁻¹, MAE 10–12 mg l⁻¹, and relative bias of 0.73. We examined error metrics stratified by latitude, longitude, and time period and found no evidence of spatial or temporal biases (figure S5). Our model differs from most published models in three major ways. First, it can predict across Landsat 5, 7, and 8 while only one other Landsat-based SSC model was found with this capability (Du et al 2020). Second, our model was trained on and can predict over a diverse set of US rivers, lakes, and estuaries, whereas most models were only trained over a small region or a particular waterbody type, except for a global river model by Dethier et al (2020) and a global, largely coastal-based model by Balasubramanian et al (2020). Third, our model is ML based, as is Peterson et al (2018), and can extract complex relationships that classic regression techniques cannot. However, ML models are limited to the range of the training data, here 0.1–2228 mg l⁻¹. Further limitations and future improvements of our SSC model are discussed in supplemental information.

Zoom In Zoom Out Reset image size

Mapping long-term trends in SSC since 1984, we found declining SSC not only at basin outlets but rather throughout entire river networks. Mean annual SSC decreased in 32% of US rivers by length, increased in 2%, and had no significant trend (p > 0.05) in the remaining 66% (figure 2). Rates of decline ranged from 0.02 to 15.6 mg l yr⁻¹ with a mean decline of 0.52 mg l yr⁻¹ (figure S6). Using the annual percent decline to estimate total change over the 34 year period showed a mean SSC reduction of 40% across all declining rivers with a range of 9%–98%. Our results agree with analysis of in-situ data on the magnitude of SSC reductions, which found a mean SSC reduction of 37% across 140 sites in the US (Murphy 2020). The similar magnitudes of SSC reduction suggest our image-based database accurately captures SSC changes. However, we found fewer declining SSC trends with 33% (n = 28 397) of river reaches declining compared to 58% (n = 140 sites) (Murphy 2020). The disparity in percent declining SSC trends suggests a small sample of in-situ observations is not fully representative of the large scale trends in SSC that we can obtain through remote sensing (Stanley et al 2019). We also attribute these differences to different record lengths, from 1992 to 2012 in Murphy (2020) compared to 1984–2018 reported here.

Zoom In Zoom Out Reset image size

The consequences of widespread SSC declines for river ecosystems have not been studied at this scale but could lead to cascading changes in biogeochemical cycles and habitat. As SSC is the main driver of light attenuation in rivers (Davies‐Colley and Nagels 2008), entire rivers are potentially becoming more clear, which might lead to increased primary productivity and photodegradation of carbon (Cory and Kling 2018). High SSC also stimulates nitrogen removal through denitrification occurring on particles, as well as production of nitrous oxide, a potent greenhouse gas (Liu et al 2013, Reisinger et al 2016), as most nitrous oxide emissions in rivers are produced in the water column (Gardner et al 2016, Wang et al 2022). Declines in SSC could therefore lead to more bioavailable nitrogen due to decreased denitrification, more algal blooms due to the combined impact of increased light and nitrogen availability, and, under extreme SSC reductions, increased vegetation or algal growth on river bottoms. Such changes would represent a major shift in the ecology of large rivers. In some cases, declines in SSC might point to successful management and reduction of soil erosion (Meade and Moody 2010). However, prior to widespread damming starting in the early 1900s, suspended sediment loads in North American rivers were 23%–167% higher on average than modern sediment loads (Dethier et al 2022, Syvitski et al 2022). SSC was likely much higher a century ago, suggesting some rivers might have already undergone ecosystem regime shifts in part due to SSC reductions.

SSC increased downstream in 117 rivers (48%), decreased downstream in 68 rivers (28%), and had no significant pattern (p > 0.05) in 59 rivers (24%). Rivers with longitudinal decreases or no longitudinal pattern had greater spatial variability in SSC on average than rivers with longitudinal SSC increases, likely due to effects of dams and flow barriers (figures 3 and S7). Dams and their reservoirs were larger on average on rivers with longitudinal decreases or no pattern compared to rivers with longitudinal SSC increases, with a 5.6 m taller median dam height of 16.6 m compared to 11 m (p = 0.009, Kruskal–Wallis test) and double the median reservoir water storage of 0.96 mm compared to 0.46 mm (p = 0.008). However, there were no differences in cumulative metrics along rivers such as total reservoir water storage (p = 0.56), and dam density was surprisingly higher in rivers with longitudinal SSC increases, 0.012 compared to 0.009 dams km⁻¹ (p = 0.02; figure S8). Our results suggest a few large storage dams, rather than many smaller dams, are linked to progressive longitudinal SSC decreases.

Zoom In Zoom Out Reset image size

A majority of rivers now have different longitudinal SSC patterns than they did before the onset of widespread river damming in the early 1900s. Modeled SSC profiles were available for 66 rivers. Of the 66 rivers, 53% changed longitudinal patterns comparing modeled no-dam profiles to contemporary observed profiles (figure 4). Notably, the Colorado, Snake, and Rio Grande Rivers showed longitudinal SSC increases before damming, but now show rapid longitudinal decreases. In contrast, the Potomac, Willamette, Cumberland Rivers showed longitudinal SSC decreases before damming, but are now increasing downstream. In addition to the observed association between large dams and longitudinal SSC decreases, shifts in longitudinal SSC patterns between no-dam simulations and current observations is further evidence that dams alter SSC patterns along entire rivers, often leading to longitudinal SSC decreases. Since the onset of widespread damming, about half of large US rivers and their biota likely encountered different longitudinal SSC patterns.

Zoom In Zoom Out Reset image size

While many rivers may have different longitudinal SSC patterns than they would without dams, the shape of observed SSC profiles are remarkably similar from year to year within a river since 1984 (figure 3). SSC profile stability values are skewed toward high, more stable values (figure 5). Less stable SSC profiles are slightly more likely to occur during the most extreme high or low flows, but all other flows showed no differences in stability (figure S9). Despite the stochastic nature of river hydrology and sediment transport, the shape of SSC profiles is consistent over time for a given river, with persistent locations of high and low SSC along rivers.

Zoom In Zoom Out Reset image size

Examining human and natural factors related to the stability of SSC profiles, we found 41% of the variation in the mean stability was explained by drainage area (p < 1 × 10⁻¹⁶), median dam height along the river (p = 0.0002), and percent cropland in the drainage area (p = 0.0008) using multiple linear regression. Larger basins with minimal agriculture and tall storage dams have the most stable SSC profiles on average (figure 5). Drainage area can be interpreted as a first order control over SSC profile stability, which is modified by secondary human influences of land use and dam size. SSC in large basins that generate larger rivers with higher flows are more resistant to changes in concentration from additional sediment inputs, therefore larger basins show greater SSC profile stability. The inverse correlation between cropland cover and stability suggests that cropland is a sediment source due to soil erosion (Walling and Webb 1996). The effect of cropland is complex, with potential threshold effects (figure 5(d)). At high cropland cover, and likely high connectivity to agricultural land, sediment is mobilized at different locations and times along a river network, leading to less stable SSC profiles from year to year. In contrast, dams and river management stabilize SSC profiles due to both reduced flow variability (Poff et al 2007) and reservoirs acting as sediment sinks, creating consistent zones of sediment erosion and deposition along rivers (Kondolf 1997, Vörösmarty et al 2003). The influence of dams and land use suggests humans are an important force in determining the shape of SSC profiles along rivers.