Classification of soil aggregates: A novel approach based on deep learning

Abstract

Having a powerful tool and the knowledge to classify soil aggregates, one of the most important factors in evaluating the performance of tillage implements, will result in quick and accurate classification of soil aggregates. By considering them as virtual sieve, a large part of the energy and workforce used in this sector can be reduced. In this regard, computational intelligence tools can play an important and optimal role in the evaluation of tillage quality and its real-time employment. The objective of the present study was to introduce a method known as deep learning to classify aggregates of any size in specific classes. Accordingly, stereo-pair images were used to provide multiple images simultaneously and the proper nutrition of the network. Since stereo-pair images are not dependent on changes in ambient light, imaging was done under conditions of the field with no lighting system. To train the deep models, the images of each lens were separated from each other and entered into the network. Without the extraction of the required features that is done manually in most image and vision-processing algorithms, the presented deep model began to learn to observe and could extract the required features from the lowest level to highly complex features automatically. Among the variety of neural network algorithms in deep learning, a convolutional neural network (CNN) was used in this study for its unique properties in working on images. To train the CNN, VggNet16, ResNet50, and Inception-v4 architects were used. Classification accuracy of these networks was above 95 %, but the highest accuracy achieved with ResNet50 (98.72 %). This accuracy, which was significantly different from previous studies, indicated the good performance of the deep learning method in the classification of aggregates. The results of the current study showed that the estimation of mean weight diameter (MWD) of aggregates without limitations in size and with great precision is completely practical and achievable.