1. Introduction
Rivers are vital components of a drainage basin. They act as habitats for aquatic organisms, serve as drainage channels for surface water, and regulate the hydrological cycle. They provide water for domestic, industrial and agricultural use, power generation, and navigation [
1,
2]. Anthropogenic activities have increasingly altered rivers’ physical and chemical properties [
3,
4,
5,
6]. These activities include agricultural practices, industrial operations, and land use and land cover changes, which all tend to negatively impact surface water quality through excessive point- or nonpoint-source loadings of pollutants [
3,
7,
8,
9,
10,
11]. A major driver of this problem is human-induced sedimentation, which affects the hydrologic regime and water quality of a river and its drainage basin [
12,
13,
14]. Human activities, such as agriculture, forest operations, mining, and urbanization, increase erosion and sediment transport within a drainage basin [
15].
Under natural conditions, sediment transport in rivers helps in carrying soil towards floodplains, delivering nutrient-rich soils to improve the agricultural productivity of adjacent lands. Changes in sediment transport impact fluvial properties and processes, such as geochemical cycling, channel morphology, and delta development [
15]. In recent decades, anthropogenic activities have changed sedimentation across the globe, in both increasing and decreasing trends [
16]. Martinez et al. [
17] reported that global changes in rainfall patterns and regional land cover changes have led to an increase in sediment transport in the Amazon River. In contrast, Wang et al. [
18] reported a marked decrease in sedimentation in the Yellow River due to vegetation restoration and construction of sediment trapping systems such as dams and reservoirs. Both scenarios have been reported for the Mississippi River [
19,
20,
21,
22]. A gradual increase in sedimentation has been estimated beginning with European settlement in 1830 due to the introduction of forest clearance and tillage practices, and continuing with a sharp increase between 1940 and 1970, likely caused by the rise of agricultural and industrial activities [
19]. Holeman [
20] estimated that, during 1951–1965, the Mississippi River’s sediment discharge averaged 3.12 × 10
8 tonnes (t) annually. A subsequent report by Milliman and Meade [
21] indicated that yearly sedimentation in the Mississippi River had declined to 1.91 × 10
8 t during 1963–1979 due to reservoir and dam construction, the implementation of measures to control river bank erosion, and improved soil conservation practices within the drainage basin. A later study by Meade and Moody [
22] reported a continued trend of decreasing sedimentation in the river system, with an average annual sediment yield of 1.32 × 10
8 t between 1986 and 2007. The authors suggest that the decrease was due to the reduced sediment supply upstream owing to the installation of sediment trapping structures and soil conservation practices.
Gauging stations and hydrologic models are widely used to measure and estimate sediment transport. Gauging stations are instrumented to directly measure river properties such as water discharge and sediment concentration at a site; these properties are then used to estimate sediment transport. In the US, sediment transport is commonly reported as sediment discharge, which represents the suspended sediment load across the river and is calculated based on field-measured, time-weighted discharge and suspended sediment concentration (SSC) [
23,
24]. At the watershed or regional scale, sediment transport can be assessed using hydrologic models [
25,
26], which require information about key watershed properties, such as topography, climate, soil, and vegetation [
27,
28]. These models are highly dependent upon the availability of input data. In data-scarce regions, representation of physical properties and processes as model inputs poses a challenge.
Numerous studies have assessed hydrologic processes and water quality parameters in rivers via space-based observations. These processes and parameters include floodplain water storage in braided rivers, inundation extent, stage variation, ice cover, turbidity, chlorophyll a, and flood wave propagation [
29,
30,
31,
32,
33,
34,
35,
36]. These studies demonstrated the potential of monitoring rivers from space on a broad scale as well as in cases where the study river is in a data-scarce region. River discharge, a critical property, was also estimated recently with optical satellite data. Gleason and Smith [
37] submitted that the at-a-station hydraulic geometry (AHG) parameters (site-specific coefficients and exponents statistically relating river width to discharge) are related across all cross-sections within the same river, and therefore referred to such relationships as at-many-stations hydraulic geometry (AMHG). AMHG allows reach-averaged discharge to be estimated solely by hydraulic geometry, which is especially useful where river cross-sectional widths can be extracted from satellite data, such as Landsat, Sentinel-2, or Cubesat imagery [
38]. Hagemann et al. [
39] further developed the AMHG discharge estimation method using a Bayesian inference approach. The Bayesian-AMHG method uses prior information on AHG parameters and discharge data from various sources, e.g., literature, in situ measurements, and reanalysis datasets [
38,
39]. Given the wide coverage of remotely sensed data, there is potential for using the AMHG approach for discharge estimation across river basins, especially for regions where gauging stations are limited or non-existent.
Only a few studies have attempted to estimate sediment discharge via remotely sensed data [
17,
40,
41,
42]. These studies used field-measured water discharge and estimated the sediment concentration using remotely sensed water-leaving reflectance (usually with 250-m MODIS (Moderate Resolution Imaging Spectroradiometer) data) to determine the sediment discharge. A related study also explored using satellite gravimetry to estimate the sediment discharge of major rivers to the ocean [
43]. More commonly, river sedimentation estimates from remotely sensed data are reported in terms of suspended sediment concentration along a river reach, or volume of sediment deposition at the river deltas or coastal zones [
44,
45,
46,
47,
48,
49,
50,
51,
52]. In the former, SSC is estimated through empirical relationships between the remotely sensed surface reflectance (SR) of sampling areas and field-measured SSC. In the latter, the volume of sediment deposition is determined using channel change detection from digital elevation models constructed from laser altimetry and image processing.
Over the next decades, land surface changes will continue to contribute to changes in terrestrial sedimentation. Nonetheless, despite the increasing demands for sediment data [
53], there has been an evident decline and discontinuation of operational gauging stations in many areas across the US. Monitoring operations may be hampered by accuracy requirements, cost, and safety concerns over traditional labor-intensive field sampling methods. There is a need for alternative approaches for the continuous monitoring of our river systems. With recent advancements in remote sensing of rivers, the use of space-based observations could offer greater opportunities to address the need. A method using remotely sensed information also allows us to compare rivers without gauging stations and identify which rivers have higher sedimentation rates than others.
This study was undertaken to investigate the potential of estimating sediment discharge using remotely sensed data, as an alternative method for providing useful hydrological information at lower cost. Specifically, we: (1) developed a method for estimating river sediment discharge from Landsat imagery, and applied it to nine gauging sites in the Upper Mississippi River; (2) established relationships between suspended sediment concentration and Landsat surface reflectance; and (3) evaluated the predictive performance of the approach for sediment discharge estimation against field-measured sediment data from United States Geological Survey (USGS) gauging stations.
4. Discussion
Our results of estimating river discharge using Landsat imagery support the finding of Feng et al. [
38] that the use of Bayesian-AMHG provides better discharge estimates for gauged rivers than for an ungauged setting for which
Q priors are approximated from data reanalysis. We also observed that the closer the center of distribution for
Q prior was to the peak of the discharge density plot, the more likely the discharge estimate would agree with the observed value. Hence, future studies should further explore how to best approximate the center of distribution of river discharge solely from remotely sensed data.
Our results also suggest that outputs from RivWidthCloud can be effectively used for Bayesian-AMHG discharge inference. The consistency of this automated width retrieval provides a substantial advantage over manual delineation of river widths from remotely sensed imagery. The sediment discharge estimates for the three study sites farthest downstream were “satisfactory” (±0.25 relative bias and NSE > 0.50) following the model performance evaluation criteria recommended by Moriasi et al. [
63]. We posit that these results are associated with the morphologic difference between reaches of the study sites. The Rosgen Type I plan view classification of natural rivers broadly characterizes channels as relatively straight (class A), low sinuosity (class B), meandering (class C), braided (class D), anastomosed (class DA), and tortuously meandering (class E) [
64]. Study sites 1–6 possess class D channel type features (see
Figure 4): complex multiple or braided channels with wide eroding banks. In contrast, study sites 7, 8, and 9 have entrenched and stable channels, more characteristic of class A or B. Consequently, we consider that the selection of a representative reach for a river system is crucial for multi-river studies because the performance in estimating discharge from optical sensors varies widely from one reach to another.
We also observed different patterns in our developed SSC-SR retrieval models. Study sites 1, 2, 3, 4, and 6, which have nearly identical regression functions (
Table 5), were likely influenced by the low SSC levels (2–324 mg/L) and by Landsat 5 being the predominant sensor platform used to acquire their images. The lower SSC levels at these study sites may have led to a weaker correlation with the water-leaving reflectance captured in the Landsat 5 images. Further, most of these sites have a higher mean green band than red band reflectance (
Figure 6), suggesting that the river segments with low SSC levels or turbidity do not fully illuminate the brownish sediment color in Landsat imagery. This is further confirmed with the regional SSC-SR retrieval model, which shows clearly that Landsat 5 data points differ from Landsats 7 and 8 data points, while the latter two closely match each other (
Figure 9). The matching functions for study sites 5, 7, 8, and 9 may be a result of the dominance of the Landsat sensor used and to the proximity of the study sites. In most cases, these study sites were in the same Landsat image, as they share adjacent or similar Landsat acquisition paths/rows (path 24, row 33 and path 23, row 34;
Table 2). It is worth noting that SSCs at these sites are higher (12.2–2340 mg/L) than the other five sites, suggesting that sites with lower SSC levels and captured by Landsat 5 likely have a lower coefficient of determination for the SSC-SR model.
Most published SSC–SR models were based on data from a single Landsat mission (e.g., Landsat 8). Pham et al. [
51] presented SSC-SR models (
R2 = 0.75) using the red to green band ratio from Landsat 8, with 40 images and SSC levels ranging from 22.4–178 mg/L. A related study by Pereira et al. [
49] using the red to green band reflectance ratio of 26 Landsat 8 images yielded an SSC-SR model with a similarly high coefficient of determination (
R2 = 0.86) for SSC levels of 49–533 mg/L. Despite these models being developed with a smaller sample and a lower SSC range in their respective river systems, our results show that the red to green band reflectance ratio from multiple Landsat missions can be used to estimate SSCs of certain reaches in the Upper Mississippi River. This result is consistent with the findings of Markert et al. [
48] and Peterson et al. [
50], both of which reported SSC-SR models with
R2 ≈ 0.5 using SR data from multiple Landsat missions (5, 7, and 8). Nonetheless, some biases and errors in our SSC estimates were related to the low sediments in the water column captured by the Landsat 5 sensor. As such, other band combinations should be explored when using Landsat 5 surface reflectance data. Note that the errors may also be due to other variables, such as higher blue and red absorption with the presence of chlorophyll and algae during low flows, and the presence of organic materials upstream in the Upper Mississippi River [
59,
65].
Studies estimating annual sediment discharge using gauge records and composited remotely sensed data (250-m MODIS) have shown good agreement with observed data with a ±2% mean relative difference [
17,
42]. Compared with these studies, we demonstrated a “per event” or “per image” estimation approach where we approximated the sediment discharge based on gauge records and 30-m Landsat data. Extreme events, such as the “Great Flood of 1993,” a 100-year flood event, were also covered in our simulations for the nine gauge sites of the Upper Mississippi River. Given that the approach is highly dependent on the accuracy of the discharge and SSC estimates, the uncertainties in their estimates can either increase the magnitude of errors or improve the final sediment discharge estimates, as illustrated in
Figure 13. For instance, study site 5 (Grafton, IL), shows a negative relative bias (−0.19) in
Q estimates and a positive relative bias (0.35) in SSC estimates. As a result, the
Q and SSC errors offset each other to yield an unexpectedly good match between sediment discharge estimates and measurements. Unlike study site 5, site 7 (St. Louis, IL) has a minimal relative bias (<0.15) in both
Q and SSC estimates. Thus, we can reasonably expect that this site will have good overall sediment discharge estimates with a low relative bias and better model fit. For this reason, we regard the St. Louis site as the best site within the Upper Mississippi River for estimating sediment discharge using Landsat data.
Future efforts should be devoted to improving the
Q and SSC estimation to further advance the utility of Landsat data or other optical remote sensing platforms for sediment discharge estimation in river systems. A river’s
Q can be estimated based on channel width, but prior discharge information is required to perform the Bayesian-AMHG inference. Readily accessible prior mean discharge of river reaches, e.g., from Lin et al. [
66], may be examined to apply discharge estimations to ungauged rivers. Alternative approaches to estimating SSC, such as non-linear regression and machine learning estimation [
50], should be explored. Last, even without gauge records for calibrations, a method estimating water and sediment discharge in rivers using remotely sensed data would still be useful in evaluating and screening for those susceptible to extreme flooding events or exhibiting excessive sedimentation.
5. Conclusions
We explored the potential of using remotely sensed images to estimate the discharge of water and sediment in a river system. In many cases, traditional monitoring operations are hampered by accuracy requirements, cost, and safety concerns regarding their labor-intensive field sampling methods. For this reason, the number of active gauging stations in many areas across the US continues to decline, despite increasing demands to monitor flow, sediment transport, and other aspects of river health. Our alternative approach, using space-based observations to assess the status of river systems, helps to address this need.
Conclusions from this study are:
Width outputs from RivWidthCloud can be effectively used for Bayesian-AMHG inference of river discharge.
Discharge estimations are influenced by both prior information and morphologic features along the river.
Higher biases and errors in SSC estimates tended to result from Landsat 5 sensors capturing scenes with low sediment levels in the water column. This suggests that for Landsat 5 surface reflectance data, additional band combinations should be explored.
Landsat imagery-based estimates of Q and SSC can yield reasonable sediment discharge estimates. In this study of the Upper Mississippi River, estimates had a relative bias of −25.4, MAE of 6.24 × 104 t/d, RRMSE of 1.21, and NSE of 0.49.
Because sediment discharge estimates are the product of two other independent estimates (water discharge and SSC), biases and errors from these component estimates can either increase or decrease the magnitude of errors in the sediment discharge estimates.
Even without gauge records for calibrations, this method can be used to estimate water and sediment discharges of rivers, and to evaluate and screen for rivers susceptible to extreme flooding or exhibiting excessive sedimentation.
This study demonstrates the potential of estimating water and sediment discharges—crucial hydrological information—using remotely sensed data as an alternative to labor-intensive field methods. Future efforts should be devoted to refining the Q and SSC estimation to further advance the utility of Landsat data for sediment discharge estimation in river systems.