USoil: Development of USCS-Based Soil Classifier Using Digital Image Processing and Convolutional Neural Network

Abstract

Different types of machine learning algorithms can be applied in determining the classification of soil. For the classification, algorithms involving support vector machine (SVM) and faster region-based convolutional neural network (R-CNN) are assessed in this comparative study. The proponents have come up with these two algorithms since SVM is the most accurate in recent studies and R-CNN has been the most popular deep learning. In this research, Convolutional Neural Network (CNN) along with image processing are used to classify soil samples based on Unified Soil Classification System (USCS). There are five sections in this system – jar test, image capturing, image processing, system training for CNN, and the result. The convolutional neural network is a machine learning that will lead to faster performance, accurate assessment and output of image processing. Experimental results showed that R-CNN is the best algorithm for soil classification assessment with 91.2% accuracy. The data set used is taken from 30 real soil data sets that is simulated through RapidMiner.