Applicability of low-cost cameras for monitoring suspended sediment in rivers through close-range remote sensing

Suspended sediment in rivers is a major problem globally. Monitoring of water turbidity and suspended sediment concentration (SSC) using satellites and in-situ sampling has been used widely to assess fine sediment pollution. However, due to low image resolution, application of satellite remote sensing is limited to only large water bodies, while in-situ sampling does not provide the continuous spatial data that are needed to address certain scientific questions or management problems. This research aimed to understand the potential of using low-cost cameras to estimate SSC in smaller rivers and streams and produce reach scale ‘maps’ of SSC. The study consists of development and testing of statistical models to predict SSC from pixel information contained in digital images, and validation of these models through field tests. An overarching goal was to assess the transferability of models between rivers and the effects of different camera sensors on SSC predictions. Laboratory experiments developed predictive models for two cameras (Vivo V9 smartphone and DJI Mavic Pro drone). Experiments involved manipulation of SSC in a water filled tank, with images taken with each camera and over a different coloured bed at each controlled sediment concentration. Digital Number (DN) values for each bed colour, camera and colour channel combination was extracted, with Generalised Additive Models fitted to Red, Blue and Green (R, G, B) colour bands. In general, there were significant relations between SSC and the mean DN values, with G and B most frequently providing the best fits. Relations differed appreciably depending on bed characteristics, as a function of the relative colour of the bed and the material in suspension; some relations were direct (positive) and some indirect (negative). Thus, laboratory tests indicated that predictive relations need to be developed on a river-by-river basis due to differences in bed characteristics. There were some subtle differences between the two cameras, but in general both yielded images from which SSC could be predicted reliably in laboratory conditions. However, almost all relations broke down at very high SSCs depending on the bed colour, camera and colour channel combination; once the amount of fine material in suspension exceeded a certain threshold, SSC could not be predicted reliably from DN values. The field tests demonstrated that it is possible to produce accurate maps of SSC using an orthomosaic developed directly using DN values. These involved developing a calibration relationship for SSC v DN from images collected from drone flights at 30 m height above a reach of the Semenyih River, Malaysia. This relationship successfully predicted SSC, with the B colour band providing the best fit (R2 >0.86 for the observed v predicted). The SSC map was able to shed light on the influence of a tributary on main stem SSCs and patterns of mixing of the fine sediment delivered by the tributary. Such fine scale spatial patterns (1cm2/pixel) are evident neither from satellite data nor in-situ monitoring. The methods presented here are applicable to a variety of questions and contexts, from understanding downstream changes in SSC in glacial rivers to assessing effects of forest loss on SSC in tropical systems