Optimizing Remote Sensing Image Classification with Spectral Indices and Convolutional Neural Networks

This paper analyzes the EuroSAT dataset, utilizing spectral channels of images from the dataset to generate spectral indices. We save these indices as new spectral channels of the images to create new datasets. The created datasets are evaluated in conjunction with ResNet-based deep learning models (ResNet-50 and ResNet-152) to assess their impact on the classification accuracy of 27,000 images across ten distinct classes. Our findings demonstrate that spectral indices can significantly enhance object classification accuracy in remote sensing imagery, particularly when used with less complex convolutional neural network models. This research offers valuable insights, highlighting the potential of spectral indices for improved image analysis in earth observation. The study encourages future investigations into additional spectral indices and hybrid approaches, combining spectral indices with visible spectrum (Red Green Blue) or multi-spectral datasets, to advance remote sensing image analysis and broaden its practical applications.