Multi-algorithm comparison to predict soil organic matter and soil moisture content from cell phone images

Abstract

An accurate assessment of soil organic matter (SOM) and soil moisture content (SMC) is critical for applications in the fields of agriculture, environment, and engineering. However, characterization and measurement of these properties is costly, time-consuming, and labor-intensive. Research has demonstrated that soil spectral reflectance characteristics can be associated with various soil properties providing an indirect way of measurement. With advancements in technological and computational facilities, high resolution digital images and computer vision algorithms have shown potential to provide rapid and nondestructive characterization of soil properties. Additionally, acceptance of cell phones in everyday life made the digital photography easier and accessible. The objective of this study was to develop and compare various regression and machine learning algorithms to estimate SOM and SMC from cell phone images. A cell phone (LG G5 model) was used to capture images of 25 soil samples from two agricultural fields with highly variable SOM at 6 different soil moisture levels from over-dry to saturated. The images were preprocessed using contrast enhancement and segmentation techniques to deal with illumination inconsistencies and remove non-soil parts of the image including black cracks, leaf residues and specular reflection. A total of 22 color and texture features were extracted from images and predictive relationships were developed against laboratory measured soil properties. A set of 24 supervised regression and machine learning prediction models including six Linear Regression Models, three Decision/Regression Trees, six Support Vector Machines (SVM), four Gaussian Process Regression (GPR) Models, four Ensembles of Trees including random forest and cubist, and other models including Artificial Neural Network (ANN) were compared in this study to predict SOM and SMC. A z-score was used to identify a set of six optimum predictors (subset of 22). Exponential GPR and Cubist model performed the best for SMC prediction, with coefficients of determination (R2) values of 0.84 and 0.86, and RMSE of 10.18% and 10.43%, respectively, (internal validation with 10-fold cross-validation) when all (22) and a subset of 6 predictors were used. For SOM, ANN and Cubist produced satisfactory prediction accuracy with R2 values of 0.91 and 0.72 and RMSE values of 5.45% and 9.90%, respectively, when 22 and 6 predictors were used. The external validation results exhibited reasonable predictive ability with Exponential GPR and Matern 5/2 GPR producing R2 values of 0.92 and 0.95 and RMSE of 5.79% and 5.04%, respectively using 22 and 6 predictors for SMC. Medium Gaussian SVM and Squared Exponential GPR produced R2 values of 0.56 and 0.53 and RMSE of 8.59% and 8.27%, respectively using 22 and 6 predictors for SOM. This shows potential in fabricating an efficient proximal soil sensor using computer vision and machine learning which can be used to provide quick, accurate and nondestructive predictions of soil properties.