Empirical models for estimating the suspended sediment concentration in Amazonian white water rivers using Landsat 5/TM

Highlights

• Top of atmosphere reflectance values from LANDSAT/TM are useful for SSC estimation.

• SWIR band 5 was a very important variable for all models.

• The regionalization of the model provided an increase in the estimation accuracy.

• As the levels of SSC increased, the estimation errors increased too.

Abstract

Suspended sediment yield is a very important environmental indicator within Amazonian fluvial systems, especially for rivers dominated by inorganic particles, referred to as white water rivers. For vast portions of Amazonian rivers, suspended sediment concentration (SSC) is measured infrequently or not at all. However, remote sensing techniques have been used to estimate water quality parameters worldwide, from which data for suspended matter is the most successfully retrieved. This paper presents empirical models for SSC retrieval in Amazonian white water rivers using reflectance data derived from Landsat 5/TM. The models use multiple regression for both the entire dataset (global model, N = 504) and for five segmented datasets (regional models) defined by general geological features of drainage basins. The models use VNIR bands, band ratios, and the SWIR band 5 as input. For the global model, the adjusted R² is 0.76, while the adjusted R² values for regional models vary from 0.77 to 0.89, all significant (p-value < 0.0001). The regional models are subject to the leave-one-out cross validation technique, which presents robust results. The findings show that both the average error of estimation and the standard deviation increase as the SSC range increases. Regional models were more accurate when compared with the global model, suggesting changes in optical proprieties of water sampled at different sampling stations. Results confirm the potential for the estimation of SSC from Landsat/TM historical series data for the 1980s and 1990s, for which the in situ database is scarce. Such estimates supplement the SSC temporal series, providing a more comprehensive SSC temporal series which may show environmental dynamics yet unknown.

Introduction

Suspended sediments affect the biological processes of aquatic ecosystems (Sioli, 1984, Kerr, 1995, Wetzel, 2001), floodplain evolution (Nanson and Croke, 1992, Walling and He, 1998, Pierce and King, 2008), and siltation of reservoirs, among other engineering problems (Thornton et al., 1981, Thornton, 1990).

In the fluvial system of the Amazon River, suspended sediments are important regulators of ecological processes in rivers and floodplains (McClain and Naiman, 2008, Junk et al., 2011). As indicators of geomorphological processes in the drainage basins, particularly in the Andes (Aalto et al., 2006, McClain and Naiman, 2008, Baby et al., 2009), suspended sediments may be possible indicators of disruption of the balance of drainage basins subject to deforestation.

Traditional systems for monitoring the suspended load in rivers are based on water samples collected in situ, from predetermined sections, and subsequent laboratory analysis. Despite the accuracy of these methods, the cost and time of acquisition of these samples are high, and the necessary logistics are quite complex, which results in a sample distribution that is limited in time and space (Ritchie et al., 1987; Ritchie and Shiebe, 2000; Costa et al., 2012). Furthermore, in remote areas such as Amazonian rivers, sampling locations are often chosen on the basis of ease of access, and not by spatial representativeness of the sample. This method of data acquisition, however, is not always adequate to answer questions involving the spatial and temporal variability in the suspended sediment concentration (SSC) in the rivers.

Therefore, the SSC database of rivers of the Amazon Basin is characterized by low spatial and temporal sampling frequency. The public domain data provided by ANA (Brazilian National Water Agency; available at: www.ana.gov.br) exemplify this: there are 97 sampling stations throughout the Amazon basin, with 2.72 samples per year on average for each station. Data provided by the HYBAM program (http://www.ore-hybam.org) have higher sampling frequency, but include narrower time coverage and are limited to larger Amazonian rivers. Beyond these limitations, in situ sampling may contain errors related to local effects, as noted by Espinoza Villar et al. (2012).

The literature validates the application of radiometric data from satellites to estimate SSC (Curran and Novo, 1988, Mertes, 2002, Wang and Lu, 2010) and show many studies based on Landsat images (Khorram, 1985, Aranuvachapun and Walling, 1988, Mertes et al., 1993, Mertes, 1994, Dekker et al., 2002, Islam et al., 2004, Wang et al., 2009). Most of the studies in Case 2 waters, however, refer to reservoirs, estuarine, lacustrine and coastal environments (Wang et al., 2009). The few studies in rivers are restricted to small sections usually covering only one scene, and involving a limited number of dates. Notable exceptions are Martinez et al. (2009), Mangiarotti et al. (2013) and Espinoza Villar et al. (2013).

Data from the Landsat program ensures a wider range of spatial and temporal scales for evaluating SSC, given the far-reaching historic series of well calibrated images (Markham and Helder, 2012), as well as free access (Wulder et al., 2012, www.dgi.inpe.br). Currently there are good prospects for the extension of the Landsat program (Irons et al., 2012, Loveland and Dwyer, 2012), confirming the importance of the new techniques for the application of Landsat/TM data. Taking into account the limitations of the in situ monitoring for the Amazon region, this paper presents a series of empirical models to estimate SSC in Amazonian white water rivers, developed from the integration of in situ data and radiometric data extracted from the Landsat 5/TM sensor.