Morphometric classification of soil aggregates using deep learning within the concept of precision agriculture

The Within the concepts of precision agriculture, this research sought to develop and evaluate an artificial neural network model capable of identifying and classifying different morphometric classes of aggregates through images. The methodology developed used a Convolutional Neural Network, with the application of MobileNetV2 architecture transfer learning in three experiments, Training with 5 morphometric classes (prismatic, angular, subangular, rounded and rounded), Training with 4 classes (prismatic, angular, subangular and rounded) and Training with 3 classes (prismatic, angular, rounded). During network development and data_set processing, sampling and data cleaning methods were used. The MobileNetV2 performance results for the three Trainings showed an average accuracy of 79%, with the Training with 3 classes performing best with an accuracy of 87%. In all three experiments, the morphometric classes with the highest accuracy were round, rounded and angular, while for prismatic and subangular classes the network showed a lower accuracy. It can be concluded that the preparation of the data set, the preprocessing in general, is a very important phase, as it can influence and make a difference in the performance of the network. For this it is necessary to have a robust dataset, in quality and quantity of images per class, in addition to hardware and software configurations, the more optimized this set is, the greater the likelihood of improvement in the accuracy of the classification of aggregates. This research constitutes a starting point for research related to the development of technologies and innovations for soil analysis.