COLOUR AERIAL PHOTOGRAPHY FOR RIVERBED CLASSIFICATION 
J.H. Chandler *, S. Rice "and M. Church? 
“ Dept. Civil and Building Engineering, Loughborough University, LE11 3TU, UK - J.H.Chandler@lboro.ac.uk 
P Dept. Geography, Loughborough University, LE1 1 3TU, UK — S.Rice@Iboro.ac.uk 
- Dept. Geography, The University of British Columbia, Vancouver, Canada- mchurch @geog.ubc.ca 
KEYWORDS: Hydrology, Classification, Scale, Texture. Orthoimage, Land Cover, Aerial. 
ABSTRACT: 
Colour aerial photography has the potential to identify and classify the size characteristics of exposed riverbed sediments over 
large areas efficiently. This paper describes the findings of a pilot project that sought to assess this potential and consider the 
impact of photo-scale on the accuracy of classifications obtained. 
Colour multi-scale digital imagery was acquired of a test area (120 x 80m) at scales of 10,000, 5,000, 3,000, 2.000 and 1.000 
using a helicopter and a hand-held Kodak DCS460 high-resolution digital camera. Intensive ground work obtained conventional 
grain size parameters (Wolman samples), requiring 15 person da 
a | and 2 m high resolution digital elevation model (DEM) 
ys of fieldwork. This was supplemented with the measurement of 
and photo-control, using a motorised Total Station. Conventional 
photogrammetric processing and the DEM were then used to create orthophotos of the test area at differing photo-scales. 
Supervised classification methods were adopted to classif 
y each pixel into one of five classes: sand. sandy gravel, pebble, clean 
gravel and cobbles. Comparison with the known ground truth achieved a success rate initially of only 3896 at 1:5,000 photo-scale, 
but developments enabled this to be increased to a more encouraging 49%. Similar tests were conducted using orthophotos at 
other scales (1:3,000 and 1:10,000) and similar improvements were achieved using the approaches developed. 
A key parameter that indicates bed roughness and is of significant biological interest, is the percentage content of sand. Further 
work was carried out to ascertain whether this simple parameter could be extracted from the imagery at differing photo-scales. The 
dataset derived by the supervised classification procedure was converted to percentage sand using a 5x5 convolution filter. It was 
hoped to assess the accuracy of the classification by comparing this percentage sand map with a similar map derived from the 
intensive fieldwork. However, the enormous improvement in spatial resolution demonstrated that the two datasets were not 
directly comparable. Despite this, it was evident that overall sand distribution was clearly revealed using both the 1:5,000 and 
1:10,000 scale imagery. More significantly, it was apparent that significant savings in time and effort would be accrued if the 
methods developed in the study were to be used to map and classify large areas of dry river-bed using colour aerial photography. 
1. INTRODUCTION 
Gravel-bed rivers form the upper reaches of virtually all river 
networks and are ubiquitous across the world. They are a major 
source of aggregate for construction, support important 
fisheries, and provide (and frequently threaten) crucial 
transportation corridors. The sediment particles that make up 
the surface of a gravel-bed river define hydraulic roughness, 
control bed material mobility and provide benthic habitat, but 
accurate characterisation of the surface material is notoriously 
difficult. Hydraulic sorting produces a patchwork of textures 
across the channel bed so that a large number of point samples 
is required to parameterise a reach (Wolcott & Church, 1991). 
In addition, even homogeneous sedimentary patches known as 
"facies" can exhibit a range in grain size that spans two or three 
orders of magnitude, so that precise estimates require many 
measurements | (ISO, | 1992). Conventional measurement 
techniques (Fripp & Diplas, 1993) are extremely labour 
intensive and sampling is therefore costly. This typically leads 
to inadequate characterisation of individual sites and to 
sampling programmes that compromise spatial or temporal 
representation. Moreover, conventional sampling methods are 
invasive, destroying the very substrate that is being monitored. 
Although aerial photography is widely acquired in river 
engineering and commonly used for basic mapping, the 
potential of remotely sensed imagery to provide additional 
information about river systems has not yet been fully realised 
(Gilvear, 1999: Lane, 2001). Recent work using Lidar, multi 
and hyper-spectral data, has focussed on the derivation of high 
resolution digital elevation models (DEMs), the classification of 
stream morphology and the estimation of water depth (e.g. 
Gilvear et al, 1995; Thompson et al, 1998: Marks and Bates, 
2000; Wright et al., 2000; Westaway et al., 2001). No published 
work to date has presented methods for the automated 
characterisation of river bed material facies from aerial 
photography, though this has recently been flagged as a key 
challenge (Gilvear, 2001) and Lane (2001) highlights the 
relevance of texture operators in this context. 
This paper describes a pilot study carried out on the Fraser 
River, British Columbia, Canada as part of an ongoing research 
and management project (Church, 2003). The principal aim of 
this study was to assess the potential of colour aerial 
photography for riverbed classification. Objectives were to 
assess the influence of photo-scale upon the classifications and 
to ascertain whether the large archive of grey scale imagery 
could be also used for such bed classifications. 
1079