A low-cost UAV for detection of Cercospora leaf spot in okra using deep convolutional neural network

Artificial Intelligence (AI)-enabled agricultural robotics is expected to significantly disrupt the domain of agriculture with promise of profitable farming. However, significant challenges exist while deploying such technologies in economically backward countries in a cost-effective manner that can...

Full description

Saved in:

Bibliographic Details
Published in:	Multimedia tools and applications Vol. 81; no. 15; pp. 21565 - 21589
Main Authors:	Rangarajan, Aravind Krishnaswamy, Balu, Edwin Jayaraj, Boligala, Muni Sekhar, Jagannath, Arjun, Ranganathan, Badri Narayanan
Format:	Journal Article
Language:	English
Published:	New York Springer US 01-06-2022 Springer Nature B.V
Subjects:	Algorithms Artificial intelligence Artificial neural networks Blurring Computer Communication Networks Computer Science Confusion Data Structures and Information Theory Deep learning Feature extraction Image acquisition Machine learning Mathematical models Motion effects Multimedia Information Systems Neural networks Okra Parameters Profitability Robotics Special Purpose and Application-Based Systems UAV Agricultural Robotics Cercospora leaf spot Convolutional neural network Okra
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Artificial Intelligence (AI)-enabled agricultural robotics is expected to significantly disrupt the domain of agriculture with promise of profitable farming. However, significant challenges exist while deploying such technologies in economically backward countries in a cost-effective manner that can ensure profitability. This study focusses on identification of Cercospora Leaf Spot (CLS) disease in the okra plant, which is also referred to as lady’s finger or Abelmoschus esculentus L. The disease is identified by using deep learning models on images acquired using a modified cost-effective quadcopter fitted with a camera. Two deep learning models, namely SqueezeNet and ResNet-18, were used for this study with a validation accuracy of 99.1% and 99% respectively. Testing of the models with the images collected using the modified quadcopter produced an accuracy of 92.3% and 94.6% respectively. The misclassifications have been analysed using confusion matrices and possible reasons affecting the classification process are discussed. In addition, the learning process has been visualized using feature parameters from different layers with t-SNE algorithm that reduces the dimension of the input parameters. The internal representation of the last feature extraction layer has been assessed using Class Activated Mapping (CAM). Further, the effect of motion blur on disease identification has been analysed using confusion matrix and CAM. For effective disease detection, it is required that the proposed quadcopter system is restricted to operate within certain limits. Finally, an insight on the future directions for improvement has been provided.
ISSN:	1380-7501 1573-7721
DOI:	10.1007/s11042-022-12464-4