Deep Learning Techniques for Automation and Industrial Applications

4Classification of University of California (UC), Merced Land-Use Dataset Remote Sensing Images Using Pre-Trained Deep Learning Models

Abhishek Maurya^*, Akashdeep and Rohit Kumar

UIET, Panjab University, Chandigarh, India

Abstract

Nowadays, researchers prefer advanced deep learning-based approaches over traditional computer vision techniques for satellite image classification since traditional techniques have certain limitations such as lack of flexibility, poor response, and less efficiency. Deep learning techniques overcome these problems since the models learn features automatically and can deal with low-quality images and complex and bulky datasets. In this chapter, we compared different deep learning-based classification techniques on a remote sensing image dataset. The dataset has been taken from the UC Merced Land-Use Dataset, which contains a total of 21 classes, with every class consisting of 100 images of size 256 x 256 pixels. The models used in this study are VGG, ResNet, Inception, DenseNet, and EfficientNet, which are deep convolutional network architectures for image classification with different numbers of layers. In order to make comparisons meaningful, all models were extended by adding three layers at the end to improve their performance. The performance of the VGG19 model was found to be better and was able to classify almost all images belonging to 21 classes with an accuracy of 100% in training and 95.07% in testing data, followed by VGG16 with 93% accuracy ...