Giant Radio Galaxies (GRGs) as a population are believed to probe the low-density inter-galactic medium. Since their discovery, visual inspection has been the most successful method for GRG candidate selection and radio morphology classification - when the sample size has been manageable. However, visual inspection will no longer be efficient when classifying the millions of objects expected from new generations of radio sky survey. In this case automated classification algorithms then become necessary. In this thesis I present a transfer learning approach to galaxy morphology classification between different radio surveys which results in models that achieve humancomparable classification accuracy. In addition, I find that inheriting model weights pre-trained on higher resolution survey images (FIRST) can boost model performance when re-training on lower resolution survey images (NVSS). However, the classifier performance deteriorates if this data training sequence is reversed. Consequently, I caution that applying transfer learning when working on new survey data of higher resolution should be carefully undertaken. I further develop this work to explore CNN-based GRG classification. To start with, I selected source samples from Data Release 1 of the Radio Galaxy Zoo citizen science project. During the sample selection process, I discovered five new GRGs, one of which is also the brightest cluster galaxy (BCG) in a galaxy cluster. I further identified 13 known GRGs as BCG candidates. I examined local galaxy number densities for all known BCG GRGs and found they can reside in the centres of rich galaxy clusters. The existence of this sub-population challenges the GRG formation hypothesis that these galaxies grow to such huge sizes only in low-density environments. With this data set I develop a multibranch CNN to identify GRGs. This model can learn jointly from both NVSS and FIRST survey images as well as incorporating numerical redshift information. The inclusion of multi-domain survey data improves model performance and corrects 39% of the misclassifications seen from equivalent single domain networks. The inclusion of redshift information moderately improves GRG classification.