Classification of Remote Sensing Datasets with Different Deep Learning Architectures

Remote sensing image classification has great advantages in the areas of environmental monitoring, urban planning, disaster management and many others. Unmanned Aerial Vehicles (UAVs) have revolutionized remote sensing by providing high-resolution imagery. In this context, effective image classification is crucial for extracting meaningful information from UAV-captured images. This study presents a comparison of different deep learning-based approach for supervised image classification of UAV images. We have experimented on four different CNN models like VGG 16, Alex net, Resnet50 and a deep neural network Efficient-Net-B0 on different remote sensing datasets; AID and AIDER. Multiple combinations were tried to find out which model performs better on which type of datasets. We have used pre-trained initial layers of four CNN models (AlexNet, VGG 16, Resnet50 and Efficient-Net-Bo) then last three layers of each of the selected models are removed and new layers have been added with better tuned parameters. Two different schemes were analyzed. In Scheme-1 the original AlexNet, VGG 16, Resnet50 and Efficient-Net-B0 were experimented without changing and tuning their number of parameters, while in Scheme-2 transfer learning was applied on the pre-trained models and after removing last three layers new layers were added with better tuned hyper-parameters. The evaluation of above schemes was ensured through comprehensive metrics across diverse land cover classes, four different performance evaluation matrices namely; F1 score, precision, accuracy and recall. The main focus of this research is towards transfer learning and adding new layers into pre-trained models to get better classification accuracy.