Characterizing soil particle sizes using wavelet analysis of microscope images

Soil texture (relative proportions of soil particles of varied sizes) is a fundamental soil physical property affecting almost all other soil physical properties and processes of agricultural, environmental and engineering importance. However, characterization of particle sizes in the laboratory presents a range of challenges in terms of the time, labor, difficulty and/or cost involved with the analysis. Continuous wavelet transform (CWT) has been used in characterizing scale-specific variations in spatial or temporal domain as well as in image analysis. The objective of this study was to develop a CWT-based computer vision algorithm to characterize soil particle sizes from digital images captured with a microscope. A cheap portable microscope with 5 MP camera and maximum magnification of 200× was used to develop an image acquisition system. Three images of air-dried, ground (2 mm) soil samples were captured in laboratory conditions for each soil sample (56 + 67 = 123) collected from two agricultural fields (Field26 and Field86) with highly variable soils. Triplicate in-situ images were also collected from 67 locations from Field86 after scrapping off surface residues. The color images were transferred to grey-scale images and the CWT was performed along the 20 equally-spaced rows and columns. The total area under the average global wavelet spectrum represented the total variation in any image. Two fractions of particle sizes; 'coarse' (diameter between 2.0 mm and 0.05 mm) and 'fine' (diameter < 0.05 mm), respectively, representing sand and sum of silt and clay were calculated based on the area under the curve and compared with lab-measured particle sizes using the hydrometer method. The lab-measured coarse- and fine-fractions showed strong agreement with the predicted (from image) fractions. The regression relationship showed the prediction capability of 87% and 88% for coarse (RMSE 44.7 g kg−1) and fine (RMSE 44.7 g kg−1) fractions, respectively for Field26 samples. A similar prediction was obtained (88% with RMSE 40.2 g kg−1 for coarse and 88% with RMSE 40.3 g kg−1 for fine) for Field86 samples. The efficiency of the wavelet algorithm shows promise in determining the particle sizes from an image and the portable nature of the image acquisition system results in a good proximal soil sensor. In contrast to the laboratory images, a weak prediction (48% for coarse and 56% for fine) was observed for the images taken in-situ mainly due to the quality of the images as they were affected by various field conditions; this requires further research.