The use of multispectral images and deep learning models for agriculture: the application on Agave

Abstract

Agave is an important plant for Mexico, country considered as center of biological diversity of agave, in addition, one variety is used for production of tequila, an important product that brings money to the country. Demand of product has led farmers to pay more attention to plantation and to reduce quality. We can find several solutions regarding agricultural filed such as identification of weed and classification of species implementing aerial imagery along with machine and deep learning reaching good results. However, there are few solutions applied directly on agaves to monitor they health. Moreover, there is not a public dataset about agaves for the purpose of this work, for this reason we have worked to collect data using a drone equipped with a multispectral camera capable to capture five different channels of a different wavelength of the light spectrum. This dataset contains 7ha of agave information into five channels provided by the multispectral camera as well as three Vegetation Indices that were computed from the multispectral bands. In this work, we explore the use of recent deep learning (DL) algorithms as well as traditional machine learning (ML) algorithms to segment agaves based on health using aerial multispectral images. On the experiments we found out that ML algorithms were able to segment just one of the two classes defined for agaves. On the experiments of DL models we could define the size of the images we wanted to train where a size of 500x500 was the best for this problem. Experiments for both types of algorithms were done using many combinations of channels such as use just vegetation indices or using all available bands on the dataset. On the other hand, Vision Transformer (ViT) Segmenter model reached an accuracy of 92.96% using vegetation indices data while the best ML algorithm was Random Forest using the five bands captured by the drone reaching 88.06% accuracy. We also test the models using traditional RGB images to compare against multispectral images and see if there is an actual advantage on the use of this type of technology. Results show us that when we introduce the variable of health into agaves, i.e. we have two classes of agaves, models that have additional bands can get better results. Thus, the use of multispectal images actually increase the performance of all models, including ML and DL, for identification of more than one class of agave.