Enhanced faster region-based convolutional neural network for oil palm tree detection

• Main Author: Liu, Xinni
• Format: Thesis
• Language: English
• Published: 2021
• Subjects: T Technology (General); TA Engineering (General). Civil engineering (General); TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://umpir.ump.edu.my/id/eprint/34675/1/Enhanced%20faster%20region-based%20convolutional%20neural%20network.ir.pdf; http://umpir.ump.edu.my/id/eprint/34675/
• Institution: Universiti Malaysia Pahang

Oil palm trees are important economic crops in Malaysia. One of the audit procedures is to count the number of oil palm trees for plantation management, which helps the manager predict the plantation yield and the amount of fertilizer and labor force needed. However, the current counting method for oil palm tree plantation is manually counting using GIS software, which is tedious and inefficient for large scale plantation. To overcome this problem, researchers proposed automatic counting methods based on machine learning and image processing. However, traditional machine learning and image processing methods used handcrafted feature extraction methods. It can only extract low-middle level features from the image and lack of generalization ability. It’s applicable only for one application and will need reprogramming for other applications. The widely used feature extraction methods are local binary patterns (LBP), scale-invariant feature transform (SIFT), and the histogram of oriented gradients (HOG), which usually achieve low accuracy because of their limited feature representation ability and without generalization capability. Hence, this research aims to close the research gaps by exploring the deep learning-based object detection algorithm and the classical convolutional neural network (CNN) to build an automatic deep learning-based oil palm tree detection and counting framework. This study proposed a new deep learning method based on Faster RCNN for oil palm tree detection and counting. To reduce the overfitting problem during the training, this study uses the image processing method to augment the training dataset by random flipping the image and to increase the data’s contrast and brightness. The transfer learning model of ResNet50 was used as the CNN backbone and the Faster RCNN network was retrained to get the weight for automatic oil palm tree counting. To improve the performance of Faster RCNN, feature concatation method was used to integrate the high-level and low-level feature from ResNet50. The proposed model validated the testing dataset of three palm tree regions with mature, young, and mixed mature and young palm trees. The detection results were compared with two machine learning methods of ANN, SVM, image processing-based TM method, and the original Faster RCNN model respectively. The proposed enhanced Faster RCNN model shows a promising result of oil palm tree detection and counting. It achieved an overall accuracy of 97% in the testing dataset, 97.2% in the mixed palm tree region, and 96.9% in the mature and young palm tree region, while the traditional ANN, SVM, and TM methods are less than 90%. The accuracy of comparison reveals that the proposed EFRCNN model outperforms the Faster RCNN and the traditional ANN, SVM, and TM methods. It has the potential to apply in counting a large area of oil palm tree plantation.