Proximal sensing of soil particle sizes using a microscope-based sensor and bag of visual words model

Abstract

While soil textural information is critical for agronomic decision making and environmental applications, its characterization in the laboratory is laborious and highly time-consuming, often costly and thus, not affordable for a considerable number of samples. This study proposed the use of a microscope-based image acquisition system and Bag of Visual Words (BoVW) model to quantify soil texture ex situ and in situ. A low-cost and portable microscope was used to design and develop a prototype soil surface image acquisition system. A BoVW computer-vision algorithm was used to extract the surficial characteristics pertaining to color and roughness. Partial least squares regression (PLSR) models were applied to correlate BoVW extracted features with the sand, silt, and clay content measured by the hydrometer method. The leave-one-sample-out cross-validation (LOSOCV) showed good agreement between measured and predicted soil separates. Coefficient of determination (R2) between the laboratory measured and image (air-dried soil samples) predicted sand, silt, and clay were 0.77, 0.68, and 0.71 and root mean squared error (RMSE) were 5.92%, 6.01%, and 2.98%, respectively. Similarly, R2 were 0.78, 0.67, and 0.52 and RMSE were 5.81%, 6.06%, and 3.85%, respectively for laboratory measured and in situ moist image predicted sand, silt and clay. These results demonstrated the potential of developing a microscope-based image acquisition system as a proximal soil sensor in conjunction with a BoVW computer-vision algorithm to efficiently characterize soil texture in greater detail, in a more cost effective and time-saving way than the conventional analysis methods.