Developing an automated monitoring system for fast and accurate prediction of soil texture using an image-based deep learning network and machine vision system. (28th February 2022)

• Record Type: Journal Article
• Authors: Azadnia, Rahim; Jahanbakhshi, Ahmad; Rashidi, Shima; khajehzadeh, Mohammad; Bazyar, Pourya
• Is Part Of: Measurement. Volume 190(2022)
• Publication Date: 2022-02-28
• Subjects: Precision agriculture -- Soil texture -- Classification -- Convolutional Neural Network -- Image preprocessing; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.110669
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A deep learning model was implemented for fast and accurate prediction of soil texture. The proposed CNN model was made of two blocks 1) convolutional and 2) classifier blocks. A portable imaging box was constructed based on smartphone to capture images on the field. A user-friendly graphical user interface was implemented based on suggested model. The accuracy of presented model was achieved up to 99% for training and testing images to predict soil texture. Abstract: To guarantee proper seedbed preparation, it is important to assess and control soil aggregate size in tillage operations. Doing so would lead to higher crop yield and more efficient resource use. This study proposes a portable smartphone-based machine vision system using convolutional neural network (CNN) for the classification of soil texture images taken from 20, 40 and 60 cm heights. The proposed CNN model consists of two blocks with several different layers. The first block (feature extraction) includes Conv, Max-pooling, drop out and batch normalization layers. The second block (classifier) consists of fully connected layers, flatten and SVM classifier. Also in this study, ANN, SVM, RF and KNN algorithms were used to compare the proposed CNN results with other classifiers. The proposed CNN model was able to successfully predict soil images in distances of 20, 40 and 60 cm with the accuracies of 99.89, 99.81 and 99.58%, respectively. The results showed that the best performance was obtained when using fully preprocessed images at the height of 20 cm. Ultimately, a graphical user interface was designed in form of a user-friendly software to predict soil texture based on CNN model. The results revealed the proposed CNN method could quickly and accurately predict the type of soil texture on large scale farms and thus be a good alternative to the costly and time-consuming laboratory methods.