Ocular diseases are a significant problem faced by 70 million people worldwide yearly. Identifying an eye disease requires a medical professional with years of specialist training. Glaucoma screening is often done manually, which is time-consuming. Automatic eye disease identification could be introduced to clinics to help speed up the detection and treatment of eye diseases. Retinal fundus images are helpful for glaucoma screening. However, the existing glaucoma identification systems trained using images from a specific camera fail to perform well with those captured from a new camera. This research aims to find solutions to address the above-unseen domain problem in machine learning-based glaucoma detection systems from retinal fundus images. The study was conducted in three phases, including an initial study that identifies the problem domain, followed by the manipulation of image preprocessing and image augmentation techniques to produce a more generalised glaucoma detection system. In the first stage, twenty-eight pre-trained deep learning models for object recognition tasks were compared as potential feature extractors for glaucoma classification from retinal fundus images of the REFUGE dataset. First, the images were automatically cropped around the optic nerve head using a template-matching algorithm. Features were extracted using the pre-trained networks from both whole and cropped images. An extended feature set was created by concatenating those two feature sets. Finally, a ten-fold cross-validation experiment was conducted to compare the performance of random forest and logistic regression classifiers against each feature set. The best setup was when features were extracted from images using the ResNet101V2 ImageNet-pre-trained neural network and classified using a random forest classifier. However, the accuracy dropped when testing it with images from a new camera. Hence, the study was directed towards the unseen domain problem in the second stage. In the study's second phase, transfer learning-based domain generalisation was applied together with multiple image preprocessing methods: input standardisation, median filtering and multi-image histogram matching for glaucoma detection using retinal fundus images from multiple cameras. The analysis included images from the RIMONEr2 and REFUGE datasets, which were captured using three camera models. A set of experiments were conducted using all possible combinations of training and testing camera devices, using the best system found in the first stage. Images were preprocessed in six different ways using either a single or a combination of three different preprocessing methods to see their effect on generalisation. The results indicated that the stylisation of test data might lead to better generalisation while reducing the retraining of an existing system. As a result, we compared multi-image histogram matching with neural style transferring to identify the classification accuracy during the testing phase of a model. We trained a random forest classifier and an XGBoost classifier with AlexNet and ResNet101V2 as feature extractors and tested the system following the same strategy as in phase two. Comparative results indicated that the neural style transferring better predicts the labels for unseen images. We continued experiments with neural style transferring to test publicly available models trained on the ACRIMA dataset. The method results better when reference images are selected from the same class. Given that the class information of real clinical data is unavailable, we suggest possible strategies for choosing better reference images. Overall, this study provides solutions to develop robust machine learning systems that require no retraining with new fundus cameras. The experimental results indicate that the proposed combination of preprocessing methods can be successfully utilised for better domain generalisation in the context of different retinal fundus camera devices. Furthermore, test-time data augmentation with neural style transferring leads to better predictions for images taken from unseen retinal fundus cameras. This reduces model retraining and increases the reusability of a pre-trained machine learning-based glaucoma detection system.