Comparison of Deep Learning Backbone Frameworks for Remote Sensing Image Classification

Classification of land cover in remote sensing images is a very important area of research as it has various applications like mapping of land cover and land use, detecting changes on the specified area over a period, disaster management, city planning, etc. Remote sensing image classification is a complex task because of the presence of various artifacts in RS images. In recent years deep convolution networks have outperformed other approaches in computer vision Recently, fully convolutional networks-based techniques have reached state-of-the-art performance in semantic segmentation and made dense pixel-wise classification possible, with the U-Net architecture being one of the most prominent models. This paper compares Unet, ResNet34, Efficientnetb7, DenseNet201. Dataset of aerial images of Dubai captured by MBRSC satellites are used for model training and testing. Images are of different sizes hence the images are cropped to the nearest size divisible by 256. The models were compared in terms of accuracy, Jacard coefficient, IOU and training validation loss. Total params and training params of all models are also analyzed.