Original papersCrop pest classification based on deep convolutional neural network and transfer learning
Graphical abstract
Introduction
Crop pest identification and classification represent one of the major challenges in the agriculture field. Insects cause damage to crops and mainly affect the productivity of crop yield. Classification of insects is a difficult task due to the complex structure and having a high degree of similarity of the appearance between distinct species. It is necessary to recognize and classify insects in the crops at an early stage, especially to prevent the spread of insects, which cause crop diseases by selecting effective pesticides and biological control methods. Traditional manual identification of insects is typically labour-intensive, time-consuming and inefficient. The vision-based computerized system of image processing using machine learning was developed for accurate classification and identification of insects to overcome these problems in agriculture research field (Martineau et al., 2017).
Wen et al. (2009) proposed an effective local feature based insect classification for orchard insects using six machine learning algorithms. The maximum classification accuracy of 89.5% was observed and insect species are misclassified due to similar insect species, various poses of wings and bodies. An automatic insect identification system was developed by Wang et al. (2012) by defining seven geometrical features and the classification results of Artificial neural networks (ANNs) and support vector machine (SVM) provide good results only for less number of classes of insects. In machine learning, the classification accuracy mainly depends on the design of extracted features and only the best features are selected to pass over to the machine learning algorithm, which increases the computational complexity. Further, the accuracy is improved by applying deep learning, which is a branch of machine learning for classifying larger image data sets. Deep learning performs automatic feature extraction from raw data that reduces the challenges in handcrafted features and solve more complex problems, especially for image classification.
In recent years, deep learning models based on CNN are extensively used as a powerful class of models for classification of images in a variety of problems in agriculture field such as plant disease recognition, fruit classification, weed identification and crop pest classification (Kamilaris and Prenafeta-Boldu, 2018). Convolutional neural network models were developed to diagnose and identify plant diseases from the leaf images of healthy and diseased plants (Ferentinos, 2018). Rice diseases identification method was proposed by Lu et al. (2017) based on deep CNN (DCNN) techniques to identify ten common rice diseases, which increases both the convergence speed and recognition accuracy. Later, transfer learning was introduced to fine-tune the pre-trained deep networks to improve learning efficiency. Recently, Too et al., reported the analysis of state-of-the-art deep learning models for plant disease identification (Too et al., 2018). Fine-tuning is a concept of transfer learning which need a bit of learning, it is proved that much faster and more accuracy than built models (Mohanty et al., 2016). Ghazi et al., tested GoogLeNet, AlexNet, and VGGNet models using transfer learning to improve the plant species identification accuracy (Ghazi et al., 2017). Deep pre-trained models were implemented by Khan et al., to extract deep features for classifying six types of apple and banana fruit diseases with improved precision and accuracy of classification (Khan et al., 2018). In Liu et al. (2016), 8-layer CNN network was developed to learn powerful local features from the complex insect image dataset and achieved a high mean Accuracy Precision for classification of 12 important paddy insect species. Wang et al. (2017) applied LeNet-5 and AlexNet to classify crop pest images by analysing the effects of both the convolution kernel and the number of layers on the network. However, he reported only two pre-trained models to classify crop pest images. The advanced pre-trained and CNN models are needed to classify crop pest images for better accuracy.
In the present work, CNN model is proposed to provide high accuracy in insect classification task and pre-trained CNN models using transfer learning are applied for comparison of classification accuracy. Three insect datasets were selected, which are collected from different field crops. The first insect dataset is collected from NBAIR that contains 40 classes of insects from various field crops such as rice, maize, soybean, sugarcane and cotton crops (http://www.nbair.res.in/insectpests/pestsearch.php). The second (Xie1) and third insect dataset (Xie2) were adopted from (Xie et al., 2015) with 24 classes of insect pests and (Xie et al., 2018) with 40 classes of insects respectively. We explore and evaluate different DCNN models of AlexNet, ResNet, GoogLeNet and VGG with transfer learning and achieving significantly better classification performance. This work was implemented in MATLAB 2018a and utilized GPU parallelization for fast computation. Our main contributions in this paper are introducing an effective CNN model with improved performance for insect classification tasks than pre-trained models using transfer learning and investigating the effect of learning rate, the number of epochs and mini-batch size, which are important hyper parameters to achieve less classification error and avoid the model over-fitting. This proposed work used to recognize the different classes of insects in crop fields at early stage to improve the crop quality and increase the crop productivity.
Section snippets
Insect dataset collection
In the experiment, the first insect dataset is collected from NBAIR that contains 40 pest types from field crops. The 24 insect classes of Xie1 and 40 insect classes of Xie2 are used as second and third insect dataset. The performance of the insect classification task is improved by applying the image pre-processing techniques to extract the insect from the original input image automatically before input in to the deep learning models. First, RGB insect image is converted in to gray scale
Results and discussion
In this work, the experiments were conducted on three different insect datasets: NBAIR, Xie1, and Xie2. Each insect dataset is divided into two categories i.e., training and testing data sets, such that 70% of insect images in each class are used as training dataset and the remaining 30% of insect images are used as a testing set (Liu and Cocea, 2017). The framework that is used to implement deep learning models is Matlab2018a using GPU NVIDIA Quadro K2200 with 4 GB of VRAM.
Conclusion
The quality and quantity of field crops are affected by pest attacks. In this study, we proposed a CNN model and focussed the development of pre-trained models for field crop insect classification. Transfer learning approach is utilized to use the pre-trained models such as AlexNet, ResNet-50, ResNet-101, VGG-16, and VGG-19 for our insect classification problem and the performances were compared with the proposed model. The effect of hyper parameters is also investigated in this work. The
Acknowledgments
Author K. Thenmozhi acknowledge to Department of Science and Technology, India for financial support of the work under women scientist scheme B (WOS-B), grant number: DST/Disha/SoRF-PM/059/2013. The authors also gratefully acknowledge infrastructural supports from Machine Learning and Data Analytics Lab, Massively Parallel Programming Lab, Department of Computer Applications, National Institute of Technology, Tiruchirappalli, India.
References (26)
- et al.
Pest identification via deep residual learning in complex background
Comput. Electron. Agric.
(2017) Deep learning models for plant disease detection and diagnosis
Comput. Electron. Agric.
(2018)- et al.
Deep learning in agriculture: a survey
Comput. Electron. Agric.
(2018) - et al.
CCDF: Automatic system for segmentation and recognition of fruit crops diseases based on correlation coefficient and deep CNN features
Comput. Electron. Agric.
(2018) - et al.
Identification of rice diseases using deep convolutional neural networks
Neurocomputing
(2017) - et al.
Transfer learning for the classification of sugar beet and volunteer potato under field conditions
Biosyst. Eng.
(2018) - et al.
A new automatic identification system of insect images at the order level
Knowl.-Based Syst.
(2012) - et al.
Local feature-based identification and classification for orchard insects
Biosyst. Eng.
(2009) - et al.
Multi-level learning features for automatic classification of field crop pests
Comput. Electron. Agric.
(2018) - et al.
Automatic classification for field crop insects via multiple-task sparse representation and multiple-kernel learning
Comput. Electron. Agric.
(2015)
Agricultural pests tracking and identification in video surveillance based on deep learning
Plant identification using deep neural networks via optimization of transfer learning parameters
Neurocomputing
Deep residual learning for image recognition
Cited by (310)
Tea yield estimation using UAV images and deep learning
2024, Industrial Crops and ProductsNon-destructive classification of sturgeon stress using cross-modal data fusion and multi-input deep learning models
2024, Computers and Electronics in AgricultureDesign and experiment of an adaptive cruise weeding robot for paddy fields based on improved YOLOv5
2024, Computers and Electronics in AgriculturePost-flood disaster damaged houses classification based on dual-view image fusion and Concentration-Based Attention Module
2024, Sustainable Cities and SocietyEnhanced rendering-based approach for improved quality of instance segmentation in detecting green gram (Vigna Rediata) pods
2024, Smart Agricultural Technology