Remote Sensing Scene Type Classification Using Multi-Trial Vector-Based Differential Evolution Algorithm and Multi-Support Vector Machine Classifier

Introduction

In recent periods, a vast amount of remote sensing images is available to analyze the earth surface, due to the growth of earth observation platforms such as unmanned aerial vehicles, aerial systems, and satellites (Pires de Lima & Marfurt, 2020; Yu et al., 2017). Currently, the scene type classification in high resolution remote sensing images becomes an emerging research field for mapping and monitoring the land types (Zeng et al., 2018). Additionally, the high resolution remote sensing images provide better feasibility to get land use and land cover information for natural environmental monitoring (Gong et al., 2018), urban planning (Zhang et al., 2020; Song et al., 2019) and precise agriculture (Lu et al., 2019). Recently, many techniques are implemented for scene type classification such as deep learning based techniques, middle level techniques and lower level techniques (Du et al., 2019). The conventional feature extraction techniques based on pixel domain are hard to achieve high classification accuracy, due to semantic space concern (Chen et al., 2018). Also, the existing feature extraction techniques are hard to deal with the complex spatial layout of remote sensing images (Qi et al., 2018; Nogueira et al., 2017). In most of the existing research studies, Convolutional Neural Network (CNN) is used to extract feature vectors, because it has a better ability to convert raw images into high-level feature representation that bridges the semantic space through hierarchical data abstraction (Zhang et al., 2019, Wang et al., 2018; Yang et al., 2018). However, the CNN model needs a vast amount of labelled data to achieve better accuracy that increases the computational power and cost of the system (Xu et al., 2020; Kaushik et al., 2019). The processing of inadequate remote sensing data causes overfitting risks in the CNN model and it is hard to regain better classification accuracy (Dai et al., 2019; Xu et al., 2018; Kaushik, et al., 2021). To address the aforementioned concerns, a novel model is proposed in this paper for better scene type classification with limited computational time. The major contributions of this paper are listed below;

• •

  At first, the input aerial images are collected from RESISC45, AID and UC Merced datasets for experimental analysis.

• •

  Then, feature extraction is accomplished using AlexNet, VGG 19, GoogLeNet, and ResNet 18 models for extracting discriminative feature vectors from the collected aerial images. The feature extraction is an important phase in remote sensing scene type classification, where the classification performance may reduce if the feature vectors are not extracted properly.

• •

  After extracting total feature vectors, relevant or active feature vectors are selected by using

• •

  MTDE algorithm. In several image processing applications, feature selection improves the classification accuracy of the learning technique, and shorten the computational time.

• •

  The selected active feature vectors are fed to the MSVM classifier for scene classification: 21 classes in UC Merced, 45 classes in RESISC45, and 30 classes in the AID dataset.

• •

  In the resulting phase, the proposed MTDE-MSVM model performance is validated in light of sensitivity, accuracy, f-score, specificity, Matthews Correlation Coefficient (MCC), and Critical Success Factor (CSF).