|Bogner, C; Trancón y Widemann, B; Lange, H: Characterising dye patterns in soils by indices of their binary images, 2nd International Conference on Hydropedology (2012)
Groundwater pollution by agrochemicals, degradation of soil quality and pollution of aquatic ecosystems by agricultural drainage waters have become an issue in the last decades. Flow processes in the vadose zone are closely related to these problems. In general, water flow in soils can be classified into two major categories: uniform and non-uniform (preferential) flow (In: U.S. National Committee for Rock Mechanics, Conceptual Models of Flow and Transport in the ractured Vadose Zone, 2001, pp.149-187). The former describes a relatively slow movement of water through the porous soil matrix and can be modelled by Richard”s equation. The latter comprises all flow types where water bypasses a portion of the soil matrix and flows through localised (i.e. preferential) paths. Unlike uniform flow, preferential flow is hardly predictable because the assumptions of Richard”s equation of a homogeneous representative elementary volume characterised by a single value of water potential, water content and hydraulic conductivity are frequently violated (Eur J Soil Sci, 2007; 58:523-546).
Because it is difficult to predict preferential flow, dye tracer studies are often done to visua- lise flow patterns in soils. A dye solution is applied onto the soil surface and after infiltration several vertical soil profiles are excavated and photographed. The images are rectified to correct any geometrical or color distortions. Thereafter, the classical image analysis consists of image classification into stained and non-stained pixels yielding binary images. The number of stained pixels per depth, the so-called dye coverage, is determined. Its shape is interpreted to identify dominant flow regimes.
Despite the many attempts to characterize dye patterns from photographs, the full potential of these images has not been exploited yet. In the present work, we propose to combine several indices that can be used to characterise dye patterns in soils in a quantitative manner. These indices are either classical ones like the dye coverage or new ones like Shannon”s entropy often used in computer sciences. A single index contains only a small amount of information about the stained patterns. However, their combination shows important characteristics of the images. Based on these indices we investigate how dye patterns change with depth because in the vadose zone, the vertical transport of water and solutes is predominant.
We use indices to cluster images in areas of similar patterns. To improve clustering we propose to incorporate prior knowledge on the flow process by identifying exemplarily rows that must fall into the same cluster and rows that cannot be linked. Therefore, we apply state-of-the-art semi-supervised clustering that proved valuable in the area of machine learning. Our approach helps to find zones of similar flow regimes that in a later step can be linked to soil properties.