Modelado de una red neuronal artificial para reconocimiento de suelos utilizando deep learning

Abstract

Currently, digital technology is being applied in all areas, medicine, economy, education, agriculture, among others, where innovation aims to facilitate processes and find increasingly efficient and autonomous solutions to problems or needs. Artificial intelligence through Deep Learning is a clear example, with its application made possible thanks to the development of computer tools and the continuous improvement of technological equipment. This has allowed innovation in traditional techniques, where ICTs are used as alternative or complementary tools in the study or research, in the search or analysis of data, it is even capable of completely automating certain tasks, without the need for human intervention, e.g., the Internet of things. Today, technology is also being applied in agriculture, there are many studies that reveal strong achievements in the field, bringing innovation and significant improvements, with which the aim is to achieve, intelligent, and precision agriculture to generate competitive advantages, resource savings and efficiency in the field. Therefore, in search of innovation, the development of this technological project was proposed, to develop an artificial neural network modeling for soil recognition, which will use color as a recognition criterion. Soil color is a very important physical property in soil classification and allows to infer certain information such as type and nature, very useful when there is no time or tools available for an accurate laboratory analysis. Thus, the prototype can be used as an alternative method to determine soil color; or, in case a previous analysis already exists, as a confirmation and comparison of the results. It is further emphasized that, being a computerized tool, it would eliminate the problem of color ambiguity, which occurs when soil color is determined by the traditional method of direct observation with Munsell tables. Therefore, an artificial neural network modeling for soil recognition was developed, which will use images for the supervised learning process. So, a convolutional neural network architecture was chosen, as this is the most widely used architecture with the highest overall performance in image processing and feature extraction. As programming language, Python was used in conjunction with Keras and Tensorflow Deep Learning libraries; with Visual Code as code editor and runtime environment.