Agricultural Pest Identification Enhanced with Deep Learning Features and Machine Learning Models
Abstract
The research proposes a novel approach combining deep feature extraction using machine learning and traditional machine learning techniques to classify 12 agricultural pests. Individual features were extracted through AlexNet, GoogLeNet, and feature fusion; afterwards, they were classified using K-Nearest Neighbors, Support Vector Machine, and Random Forest. GoogLeNet achieved 86.21% accuracy with SVM, while the fused features achieved 82.03% with Random Forest. The proposed method makes good use of deep learning with feature representation and classical models for accurate and computationally efficient pest identification in agricultural applications.
Keywords
Download Options
Introduction
Agrology is recognized as essential in ensuring global food security and economic development, especially when a significant portion of the population relies on agriculture for employment opportunities in a specific region [1]. However, agricultural productivity faces a continuous threat from pest infestations, which are a major contributing factor to crop damage and yield losses. Critical pest damage and losses need timely identification along with accurate pest recognition to determine the appropriate control measure to be put in place to curtail such losses[2]. With pest recognition relying on manual examination and the expertise of specialists, such methods, while effective, tend to be labor-intensive and ineffective for large-scale employment in agricultural settings.
Recent years have seen remarkable advances with the integration of Artificial Intelligence (AI) and Computer Vision (CV) into agriculture. Crop field monitoring and pest identification using deep learning techniques have received attention due to the high level of accuracy and autonomy they offer [3][4] Image classification has seen the adoption of Convolutional Neural Networks (CNNs), which can hierarchically learn complex representations from raw data to perform advanced classification tasks. Today, numerous CV applications such as object detection, image recognition, and classification rely heavily on previously developed models such as AlexNet and GoogLeNet[9][18]
The use of deep learning methods accomplishes remarkable outcomes, though their training requires enormous data alongside computational power, resources that are difficult to obtain in pre-established agricultural settings. To counter this challenge, the use of pre-trained CNNs for feature extraction is a practical substitution. With this method, models trained on benchmark datasets like ImageNet are employed to extract information from images about a specific field, with no training required. Such features are sufficient to train simple, low-cost classifiers that, without the need for significant resources, achieve accurate performance. In this work, we propose a new approach to classifying 12 categories of agricultural pests that combines different methods. Features from two popular deep learning networks, AlexNet and GoogLeNet, have been incorporated. To this end, both individual feature extraction and feature fusion approaches have been adopted. In the first case, features were extracted independently from each model. In the second case, the features that were extracted from the individual networks were merged to create a single comprehensive feature set.
The innovation of this study is the combination of deep feature extraction with traditional machine learning techniques, thus providing better efficiency regarding computational resources while maintaining high performance. By framing the problem as pest classification without training an end-to-end deep learning model and employing pre-trained networks for feature extraction, this study proposes an effective and scalable approach. Such an approach is important for practical use in agricultural settings with limited resources.
Input Pest images Alexnet Googlenet Deep Feature extraction Classification using ML methods (KNN, SVM, RF)
Classified results FIGURE 1: Block diagram of proposed method
Conclusion
AND FUTURE WORK The study proposed a hybrid pest classification framework that integrates GoogLeNet-based deep feature extraction with a Support Vector Machine classifier, evaluated using the standard dataset [3]. The method achieved a recognition accuracy of 86.21%, outperforming several existing deep learning-based approaches. These results highlight the effectiveness of combining deep feature representations with classical machine learning techniques for accurate and resource-efficient pest identification. Future research may focus on incorporating feature selection methods to reduce feature dimensionality and improve model interpretability. Additionally, the framework could be adapted for deployment on edge devices to facilitate real-time pest monitoring in agricultural settings and expanded to include a wider range of crop-pest datasets for enhanced generalizability.
