Repercussion of Image Compression on Satellite Image Classification using Deep Learning Models

Abstract

Satellite image classification using Deep Learning (DL) algorithms is crucial for various applications such as Environmental Monitoring, Remote Sensing, and Urban Planning. The high-resolution nature of satellite images leads to large data volumes, resulting in computing challenges when classifying and transmitting the data over the internet. In this paper, we address this challenge by reducing pixel size using Bicubic and Bilinear Interpolation. These techniques are known as image compression methods, which allow us to retain a significant portion of the original image information while reducing its size. Furthermore, adjusting the pixel size enables us to implement various levels of image compression supported by different DL classifiers, catering to diverse applications in NextG wireless networks or O-RAN. This pixel size reduction optimizes data transfer and speeds up satellite image transmission for Crop Monitoring and Land Cover Classification applications. After reducing the pixel, we employ different DL models such as CNN, ResNet, and stacked ensemble model, built by stacking three other EfficientNetB0 models to classify the low pixel images and compare the results using accuracy, precision, recall, and F1-Score. Staked EfficientNetB0 is the best performer among these, with an accuracy of 89.34% using a pixel size of 512x512. Our results show that we can reach up to 88.25% accuracy while maintaining a pixel size of 64x64, as our target is to reduce the image size with maintaining a tolerable accuracy so that network load and cost can be minimized.