Surficial Sediment Size Distribution through the Use of Digital Imaging and Autocorrelation Analysis F.K. Tedla 1, J.A.

Tuhtan2, S. Hartmann2, S. Wieprecht 2, M. Schneider 3 Fichtner Gmbh & Co. KG, Sarwaystrae 3, 70191 Stuttgart, Germany; PH +49 (0)711/89 95 0; email: fidel.tedla@fichtner.de 2 Institute of Hydraulics, University of Stuttgart, Pfaffenwaldring 61, 70569 Stuttgart, Germany; PH +49(0)711/685-66615; FAX +49(0)711/685-66600; email: jtuhtan@gmail.com 3 SJE Ecological Engineering, GmbH, Viereichenweg 12, 70569 Vaihingen, Germany; PH +49 (0)711 677-3435; FAX +49 (0)711 677-3436; email: mailbox@sjeweb.de ABSTRACT Information on grain size distribution has many applications in the field of fluvial engineering, such as sediment transport and deposition, and fluvial habitat modeling. This study aims at finding a sampling and analysis technique that achieves a satisfactory characterization of grain size while simultaneously reducing the time spent in both the field and the laboratory. During this study a new method of size distribution of river bed materials through digital imaging and autocorrelation analysis has been used. 75 % of the samples analyzed show strong correlation to the sieve analysis results. Overall average error of around 10 % is observed. Moreover, the method proves to be cheaper and faster than the traditional methods by more than 100 %. 1 INTRODUCTION Grain-size is a fundamental property of materials from which additional information can be obtained with regards to their behavior (Syvitski 1991). In particular, the dynamical mechanisms of transport and deposition of the constituent grains of sediments are usually drawn from their size distribution (McLaren 1981). Furthermore, Bridge (1981) explains how the grain-size distribution of river bed sediments can be interpreted hydraulically and used to describe sediment transport and deposition. In his work, he shows that information about the transport and deposition processes can be extracted from statistical parameters of a given grain-size distribution such as mean, median, kurtosis and skewness. Once the relevant statistical properties of a given grain-size distribution are known, the data can also be used by a wide spectrum of fields, such as hydraulic engineering, riverine ecology, and fluvial geomorphology. This work focuses primarily on the determination of grain-size distribution data which can later be used to improve the application accuracy of fluvial fish habitat models. However, the method may also prove useful in calibrating sediment transport models as well. Since fish habitat simulation models typically use a combination of the parameters water depth, flow velocity, and substrate classes to produce a fish
1

33rd IAHR Congress: Water Engineering for a Sustainable Environment Copyright c 2009 by International Association of Hydraulic Engineering & Research (IAHR) ISBN: 978-94-90365-01-1

33rd IAHR Congress: Water Engineering for a Sustainable Environment

habitat map of a given river reach (Eberstaller 2006). Although values for the flow depth and velocity can be accurately calculated using one- or two-dimensional hydraulic models, river substrate maps are produced from a field survey of the grain-size distribution of river bed materials. The accuracy of the data found on substrate maps, is usually much lower than that of the water depth and the flow velocity. This difference in accuracy comes from the fact that substrate maps are predominantly based on visual observations, and are often augmented with few direct samples because direct sampling is expensive and time-consuming. Here we describe our efforts to improve the accuracy of such substrate maps via the autocorrelation-based photosieving (ABP) method developed by Rubin (2004). Section two describes current methods used for determining the grain-size distribution, and photosieving in general. Section three provides a short overview of the theory behind the ABP procedure, along with the steps required to carry it out. Sections four and five discuss the experimental procedures and results of ABP for later application in substrate mapping. Finally, conclusions and an outline of future work are discussed. 2 TECHNIQUES USED FOR THE DETERMINATION OF GRAIN-SIZE DISTRIBUTION Traditional techniques of particle size analysis rely on mechanical sorting of the sample material. Common approaches using particle counting, sieving, and settling tubes are still in use though often using automated instruments. New ways of grain-size distribution determination via digital image analysis have also been developed (Syvitski 1991). These methods are simple, rapid and to a large extent allow the practitioner to avoid the extensive field work and the need for large samples to be transported to the laboratory for analysis. It is necessary and highly desirable that the chosen sampling and analysis technique achieve a satisfactory characterization of grain size while simultaneously reducing the time spent in both the field and the laboratory. Recent developments in digital image processing software have provided the means for the development of an alternative method for soil grain-size distribution analysis based on digital images (Ghalib 1999). A number of approaches have been introduced and can generally be categorized into two groups. The first group includes deterministic methods, which use edge detection or grayscale thresholding for component segmentation and component pixel counting. These methods successfully determine not only individual grainsizes but the size distribution as well. However, these methods have one major limitation. The soil grains have to be detached from one another, otherwise overlapping or touching grains will be interpreted as individual, large components (Shin 2004). Due to the limitations of such deterministic methods, a second group of approaches for analyzing soil images has been introduced. This second group includes statistical methods based on image texture. Texture here is defined as patterns which are repeated in an image. In general, these methods represent the image as a set of index values. The index values are more easily obtained from a transformed image than from the original. Shin (2004) points out that textural analysis (statistical) methods usually involve four main tasks:

497

33rd IAHR Congress: Water Engineering for a Sustainable Environment

1. Image transformation into a form that can be more easily interpreted such as grayscale; 2. Description of the image using index values based on textural features; 3. Size distribution from digital image analysis; and 4. Establishment of relationships between the indices and grain-size. The main advantage of textural image analysis techniques is that they overcome the problem of touching and overlapping grains. The ABP method used in this study is a textural analysis technique which makes use of the spatial autocorrelation of the intensity of pixels in an image in order to calculate the grain-sizes. 3 AUTOCORRELATION-BASED PHOTOSIEVING The method used in this work, unlike the deterministic methods previously described, considers only the statistical properties of a single image in calculating the grain-size distribution using spatial autocorrelation. Spatial autocorrelation is the correlation of a variable with itself through space, and measures the level of interdependence between the variables and the nature and strength of the interdependence. It may be classified as either positive or negative on a scale from -1 to 1. Positive values of spatial autocorrelation occur when similar values appear close together, while negative values occur when dissimilar values appear in close proximity. The ABP method was first proposed by Rubin (2004) for measuring the average grain size of bed sediment, the grain-size distribution of the bed sediment, and vertical profiles in grain-size in a cross-sectional image through a bed. The method involves the use of two algorithms to determine the grain-size distribution of a digital image taken of surficial sediment (the uppermost layer in direct contact with the flow). The first algorithm (autocorrelation algorithm) calculates the autocorrelation values of an image at different offset distances within the image and is used in the calibration as well as in the grain-size analysis stages. The second algorithm (size distribution algorithm), based on calibration data resulting from a sieve analysis combines the observed autocorrelation from the first algorithm and delivers the grain-size distribution and the mean grain-size of the sample image. Autocorrelation Algorithm The first algorithm carries out the following three steps: 1. Load an image in color and display it. 2. Convert the image to grayscale. The reason for changing the image to grayscale is that less information is needed to be provided for each pixel in the image. A 'gray' color is one which has the same intensity values for the 'red', 'green' and 'blue' components of an RGB (color) image, as a result, only a single intensity value is needed to describe each pixel in a grayscale image as opposed to the three values needed to describe a pixel in an RGB image. Here grayscale intensity is stored as an 8-bit integer giving 256 possible different shades of gray ranging from black (with a value of zero) to white (with a value of 255).

498

33rd IAHR Congress: Water Engineering for a Sustainable Environment

3. Calculate the autocorrelation curve of the image for offset distances of one pixel to a user defined maximum-offset number of pixels. The core of the algorithm is based on the idea that spatial autocorrelation in an image varies with grain-size. Rubin (2004) used the intensity of the pixels in the image as a variable in calculating the autocorrelation values. The spatial autocorrelation is calculated between two rectangular regions (called plaquettes) in an image by calculating the correlation between the differences in pixel intensity using the following equation:

(x
r=
i

x )( yi y )

(x
i

x )2

(y
i

y)2

(1) Where: r = the spatial autocorrelation between two plaquettes xi and yi = the intensities of corresponding pixels in the two plaquettes x and y = the mean intensities of pixels in the two plaquettes

Figure 1. A schematic view of the stages in calculating grain-size distribution using ABP

Figure 2. Diagram of the autocorrelation algorithm: multiple offset plaquettes moving at a single fixed offset distance of 1 pixel, each is compared to the reference plaquette An autocorrelation curve, as shown in Fig. 3, is determined from the output data of the first algorithm. When the distance between the two plaquettes is small

499

33rd IAHR Congress: Water Engineering for a Sustainable Environment

(relative to the grain-size) the autocorrelation value approaches 1.0 and as the distance increases, xi becomes increasingly unrelated to yi, and the sum of their products decreases and approaches zero.
1 0. 8 0. 315m m -1m m 0. 6 1m m -2. m 5m 2. m -6. m 5m 3m 6. m -12. m 3m 5m 0. 4 12. m -20m m 5m >20m m 0. 2

Spatalaut i ocor aton r rel i

0 0 10 20 30 40

O f setdi ance ( xel f st pi )

Figure 3. Autocorrelation curves of different grain-size fractions (calibration fractions). Pixels in larger grains tend to be more similar at longer distances than pixels in smaller grains Size Distribution Algorithm The second algorithm developed by Rubin calculates the grain-size distribution of a sample based on the autocorrelation curves of the individual grain size fractions (calibration curves) and the sample curve. The size distribution algorithm solves for the proportions from the calibration curves that collectively give the best fit to the sample curve. The algorithm uses a non-negative least-squares fit method in order to calculate the proportions of each grain-size fractions and returns the values as an output in percentage. 4 EXPERIMENTAL WORK The experiments were conducted in the laboratory of the Department of Hydraulic Engineering, University of Stuttgart. Furthermore, all field images and soil samples used in this study were collected from the Schneitbach Creek, located in the Black Forest of southern Germany. This section explains the experimental equipment and procedures used during the study. Surficial substrate was collected from Schneitbach Creek to be used for calibration and validation purposes. The experiments were carried out in two stages. In stage 1 two, samples collected for calibration and validation were dried in an oven at a temperate of 105 oC for 24 hours. Both samples were mechanically sieved using a set of 8 sieves (0.315 mm, 0.63 mm, 1.00 mm, 1.60 mm, 2.50 mm, 4.00 mm, 6.30 mm, 10.00 mm). After sieving, it was found that the material could be best categorized into following four classes: <0.315 mm (class 11), 0.315 mm 2.5 mm (class 12), 2.5 mm - 10 mm (class 13), and >10 mm (class 14). Two more samples were added to the validation sample by mixing classes 11 and 12 to form class-15 (<0.315 mm - 2.5 mm), and by mixing classes 13 and 14 to form class-16 (>2.5 mm). Each class from both samples was then scanned at resolutions of 100 dpi, 300 dpi, and 600 dpi consecutively. The autocorrelation curves for calibration were obtained by running the scanned images from the first sample and the images from the second sample were used to generate the autocorrelation curves for validation. Finally, the grain-size distribution of each class in the validation

500

33rd IAHR Congress: Water Engineering for a Sustainable Environment

sample was calculated and was compared to the results obtained by mechanical sieving. This stage was carried out to test the method under ideal conditions and look for sources of error. In stage 2, a soil sample was collected to be used for calibration. The sample was mechanically sieved and categorized into six classes: 0.315 - 1 mm (class 21), 1 mm - 2.5 mm (class 22), 2.5 mm 6.3 mm (class 23), 6.3 12.5 mm (class 24), 12.5 20 mm (class 25), >20 mm (class 26). The calibration classes were then photographed under a variety of resolutions and lighting conditions (indoor/flash/outdoor). The calibration curves were then obtained by running the autocorrelation algorithm for all the calibration images. For validation four field samples were photographed with an underwater camera at a height of 30 - 40 cm, and carefully removed (the uppermost 3-5 cm of submerged substrate) using a steel sampling box affixed with a mesh filter, allowing water to pass through the box as sampling was carried out, but which retained most of the fines. The grainsize distribution of the four underwater field images was then calculated and compared to the results obtained by mechanical sieving. Table 1 provides an overview of the various configurations used to test the ABP method. Table 1. Summary of grain-size classes, image resolutions, and lighting conditions used during the two stages of the study
Stage 1 2 Image type Scanned Camera images Grain-size class Calibration class 11 - class 14 class 21 - class 26 Validation Class 11 - class 16 Resolution 100, 300, 600 dpi Lighting condition Scanner lighting Indoor with flash, Camera placed at 30 indoor without cm, 40 cm from the flash, outdoor ground without flash Scanner lighting Natural lighting

1 2

100, 300, 600 dpi Camera placed at 30 Camera images Field sample 1 - field sample 4 cm, 40 cm from the ground

Scanned

5 RESULTS The results of the experimental work are presented in this section. All the results are shown via a comparison of a sample's grain-size distribution obtained by ABP and by mechanical sieving. Here it must be noted that because the grain-size fraction from mechanical sieving is calculated using mass percentages, and those from ABP are based on surface area, a perfect fit between calibration and validation is not possible. Complete results from all the stages and a thorough analysis of the work presented in this paper can be found in Tedla (2009). Effect of Image Resolution on Calculated Grain-Size Distribution In general, it was found that images taken at higher resolutions give better grainsize distribution results than images taken at lower resolutions. Fig. 4 compares the results of soil samples investigated during the first stage of the experiments. It can be seen for all the six soil samples investigated, images taken at a resolution of 600 dpi predicted the best size distribution results, images at 300 dpi being the second best. This is because higher image resolution means higher number of pixel representation per grain which allows the algorithm to return accurate

501

33rd IAHR Congress: Water Engineering for a Sustainable Environment

autocorrelation curves thereby predicting more accurate results.

30. 00

Abs ut M ean Er or( ) ol e r %

25. 00 20. 00 15. 00 10. 00

Si e ngl cl asses

C om posi e t cl asses

100 dpi 300 dpi 600 dpi

5. 00 0. 00 Cl ass11 ( 315) <0. Cl ass12 ( 315- 5) 0. 2. Cl s13 as ( 52. 10) Cl ass14 ( >10) Cl s15 as Cl ass16 ( 315- 5) ( 5) <0. 2. >2.

Sam pl es

Figure 4. Comparison of scanned soil samples at different resolutions (stage 1) Effect of Lighting Conditions and Camera Height on Calculated Grain-Size Distribution Fig. 5 shows the comparison of grain-size distribution of field samples predicted for different lighting conditions and camera heights from the ground. Calibration images taken at a camera height of 40 cm, without flash and indoors gave the best results for field sample 1, while those images taken at a height of 30 cm, without flash and indoors gave the best results for samples 2, 3 and 4. Field lighting condition was similar to indoors lighting condition on a typical summer day. It can be concluded from this finding that using similar lighting conditions and camera heights for both calibration and validation images give better size distribution predictions. Using flash seems to exaggerate the amount of fine grains. This might be due to small patches on coarse grains being interpreted as fine grains. Taking images of samples on a very sunny day might also have the same effect. It should, therefore, be avoided using flashes except in dark or highly shadowed lighting conditions.
60. 00

Absol e M ean Er or( ) ut r %

50. 00 40. 00 30. 00 20. 00 10. 00 0. 00

B r ghtnat al i ur lghtng i i

D ar nat al k ur lghtng i i

30cm ,Fl h, as I ndoor 30cm ,No fas l h, I ndoor 40cm ,Fl ash, I ndoor 30cm ,No fas l h, Out door 40cm ,No fas l h, Out door

Fi d sam pl 1 el e

Fi d s pl 2 el am e

Fi d s pl 3 el am e

Fi d sam pl 4 el e

Fi d Sam pl el e

Figure 5. Comparison of samples under different lighting conditions and resolutions (stage 2) 6 CONCLUSIONS Based on the objectives of this work carried out during the study period, the following conclusions can be drawn: - The method of size distribution from autocorrelation analysis of digital images is successful in predicting size distribution of a diverse range of sediment types: from sand (<0.315 mm) to coarse gravel (>20 mm).

502

33rd IAHR Congress: Water Engineering for a Sustainable Environment

- Run time of the algorithms of the autocorrelation analysis method range form <1 min to 5 minutes depending on grain size, image resolution and maximum offset distance. This means that the method is faster than mechanical sieving by more than 100%. - An overall error, from the results of mechanical sieving, of around 10% was observed. This error is acceptable for the purposes of preparing fluvial substrate maps for habitat modeling. Furthermore, taking into account that the compared values of size distribution are based on a projected 2D area for the image analysis and on weight and 3D shape for the mechanical sieve analysis, the observed error is acceptable. - A grain should be represented by 2 or more pixels to have good autocorrelation results. - The higher the resolution of the images, the better the size distribution results. AKNOWLEDGMENTS We would sincerely like to thank the German Bundesministerium fr Bildung und Forschung IPSWaT Scholarship Program for providing us with the financial support to do the work. REFERENCES Bridge, John S. (1981), Hydraulic Interpretation of Grain Size Distribution Using a Physical Model for Bedload Transport, Journal of Sedimentary Petrology, 51 (1981), 1109-1124. Ghalib, Ali (1999), Soil Particle Size Distribution by Mosaic Imaging and Watershed Analysis, Journal of Computing in Civil Engineering, 13 (1999), 80- 87. McLaren, Patrick (1981), An Interpretation of Trends in Grain Size Measures, Journal of Sedimentary Petrology, 51 (1981), 0611-0624. Rubin, David M. (2004), A Simple Autocorrelation Algorithm for Determining Grain Size from Digital Images of Sediment, Journal of Sedimentary Research, 74 (2004), 160- 165. Eberstaller, Jrgen (2006), Sustainable Sediment Management in Alpine Reservoirs Considering Ecological and Economical Aspects, Institut fr Wasserwesen - Universitt der Bundeswehr Mnchen.

503

33rd IAHR Congress: Water Engineering for a Sustainable Environment

Shin, Seungcheol (2004), Wavelet Analysis of Soil Mass Images for Particle Size Determination, Journal of Computing in Civil Engineering, 2004, 19-27. Syvitski, James P. (1991), Principles, Methods, and Application of Particle Size Analysis, Cambridge University Press. Tedla, Fidel K. (2009), Surficial Sediment Size Distribution Through The Use of Digital Imaging and Autocorrelation Analysis, Institut fr Wasserbau Universitt Stuttgart.

504