Abstract

Plants are very important living organisms on earth because humans and animals depend on them for nutrition, oxygen, medicine and balance in the ecosystem. Therefore, plant species recognition is critical to the improvement of agricultural productivity, mitigation of climate change and the discovery of new medicinal plants. However, species recognition has remained a difficult task even for trained botanists, because using the traditional approaches, an expert on a specie may be unfamiliar with others. Thus, researchers and practitioners are increasingly interested in the automation of species recognition problem. Recently, deep learning algorithms such as Convolutional Neural Network (CNN) have provided huge breakthroughs in various computer vision tasks compared to their shallow predecessors. Deep learning automates features extraction by learning salient representations of the data and subsequently classifies the features using a supervised learning approach. Inspired by this capability, we leveraged on five pre-trained CNN models and Leafsnap image dataset of 185 plant species to experimentally evolve an accurate species recognition model in this study. Among the pre-trained models, MobileNetV2 with ADAM optimizer gave the highest testing accuracy of 92.33%. This result provides a basis for developing a mobile app for automated species recognition on the field. This will augment existing efforts to alleviate the difficulties of manual species recognition by botanists, farmers, biologists, nature tourists as well as conservationists.

Highlights

  • It is estimated that there are more than 450,000 plant species on earth and approximately 385,000 of these have been identified and classified (Pimm and Joppa, 2015)

  • Results from the experiment show that MobileNetV2 with Adaptive Moment (ADAM) and ResNet50 with SGDM gave the best performance and second-best validation accuracies of 92.33% and 92.13% respectively, while GoogLeNet with Root Mean Square Propagation (RMSProp) gave the poorest performance with a validation accuracy of 1.73%

  • We have presented the development of automated plant species identification model through rigorous experimentations with five state-of-the-art pretrained Convolutional Neural Network (CNN) models (i.e., AlexNet, GoogLeNet, VGG19, ResNet50 and MobileNetV2) and the Leafsnap plant image dataset

Read more

Summary

Introduction

It is estimated that there are more than 450,000 plant species on earth and approximately 385,000 of these have been identified and classified (Pimm and Joppa, 2015). Due to the shrinking number of botanists that can classify plants and the complexity of plant classification, it is necessary to design a computerized system for plant species identification. Many computer-based systems for plant identification have been designed using leaf analysis, DNA analysis, flower analysis, or a combination of these and other features (Kaur and Kaur, 2019). In current literature, leaf analysis remains the most popular approach used to identify a given plant, because leaves are visible and carry enough information to differentiate plant species. Feature extraction and selection are the most challenging and time-consuming steps when using leaf images to recognize plant species. Deep learning techniques came as a solution to achieve automatic feature extraction and selection in computer vision, especially with Convolutional Neural Networks

Methods
Results
Conclusion
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call