Testing a New Ensemble Vegetation Classification Method Based on Deep Learning and Machine Learning Methods Using Aerial Photogrammetric Images

Abstract

The objective of this research is to report results from a new ensemble method for vegetation classification that uses deep learning (DL) and machine learning (ML) techniques. Deep learning and machine learning architectures have recently been used in methods for vegetation classification, proving their efficacy in several scientific investigations. However, some limitations have been highlighted in the literature, such as insufficient model variance and restricted generalization capabilities. Ensemble DL and ML models has often been recommended as a feasible method to overcome these constraints. A considerable increase in classification accuracy for vegetation classification was achieved by growing an ensemble of decision trees and allowing them to vote for the most popular class. An ensemble DL and ML architecture is presented in this study to increase the prediction capability of individual DL and ML models. Three DL and ML models, namely Convolutional Neural Network (CNN), Random Forest (RF), and biased Support vector machine (B-SVM), are used to classify vegetation in the Eastern part of Serbia, together with their ensemble form (CNN-RF-BSVM). The suggested DL and ML ensemble architecture achieved the best modeling results with overall accuracy values (0.93), followed by CNN (0.90), RF (0.91), and B-SVM (0.88). The results showed that the suggested ensemble model outperformed the DL and ML models in terms of overall accuracy by up to 5%, which was validated by the Wilcoxon signed-rank test. According to this research, RF classifiers require fewer and easier-to-define user-defined parameters than B-SVMs and CNN methods. According to overall accuracy analysis, the proposed ensemble technique CNN-RF-BSVM also significantly improved classification accuracy (by 4%).