Improving Transfer Learning for Use in Multi-Spectral Data
2020 ii DECLARATION I certify that this dissertation which I now submit for examination for the award of MSc in Computing (Data Science), is entirely my own work and has not been taken from the work of others save and to the extent that such work has been cited and acknowledged within the test of my work. This dissertation was prepared according to the regulations for postgraduate study of the Technological University Dublin and has not been submitted in whole or part for an award in any other
... award in any other Institute or University. The work reported on in this dissertation conforms to the principles and requirements of the Institute's guidelines for ethics in research. ABSTRACT Recently Nasa as well as the European Space Agency have made observational satellites images public. The main reason behind opening it to public is to foster research among university students and corporations alike. Sentinel is a program by the European Space Agency which has plans to release a series of seven satellites in lower earth orbit for observing land and sea patterns. Recently huge datasets have been made public by the Sentinel program. Many advancements have been made in the field of computer vision in the last decade. Krizhevsky, Sutskever & Hinton, 2012, revolutionized the field of image analysis by training deep neural nets and introduced the idea of using convolutions to obtain a high accuracy value on coloured image dataset of more than one million images known as Imagenet ILSVRC. Convolutional Neural Network, or CNN architecture has undergone much improvement since then. One CNN model known as Resnet or Residual Network architecture (He, Zhang, Ren & Sun, 2015) has seen mass acceptance in particular owing to it processing speed and high accuracy. Resnet is widely used for applying features it learned in Imagenet ILSVRC tasks into other image classification or object detection tasks. This concept, in the domain of deep learning, is known as Transfer learning, where a classifier is trained on a bigger more complex task and then learning is transferred to a smaller, more specific task. Transfer learning can often lead to good performance on new smaller tasks and this approach has given state of the art results in several problem domains of image classification and even in object detection (Dai, Li, He, & Sun, 2016) . The real problem is that not all the problems in computer vision field belongs to regular RGB images or images consisting of only Red, Green, and Blue band set. For example, a field like medical image analysis has most of the images belonging to greyscale color space, while most of the Remote sensing images collected by satellites belong to multispectral bands of light. Transferring features learned from Imagenet ILSVRC tasks to these fields might give you higher accuracy than training from scratch, but it is a problem of fundamentally incorrect approach. Thus, there is a need to create network models that can learn from single channel or multispectral images iv and can transfer features seamlessly to similar domains with smaller datasets.This thesis presents a study in multispectral image analysis using multiple ways of feature transfer. In this study, Transfer Learning of features is done using a Resnet50 model which is trained on RGB images, and another Resnet50 model which is trained on Greyscale images alone. The dataset used to pretrain these models is a combination of images from ImageNet (Deng, Dong, Socher, Li, Li, & Fei-Fei, 2009 ) and Eurosat (Helber, Bischke, Dengel, & Borth. 2017) . The idea behind choosing Resnet50 is that it has been doing really well in image processing and transfer learning and has outperformed all the other traditional techniques, while still not being computationally prohibitive to train in the context of this work. An attempt is made to classify different land-cover classes in multispectral images taken up by Sentinel 2A satellite. The dataset used here has a key challenge of a smaller number of samples, which means a CNN classifier trained from scratch on these small number of samples will be highly inaccurate and overfitted. This thesis focuses on improving the accuracies of this classifier using transfer learning, and the performance is measured after fine-tuning the baseline above Resnet50 model. The experiment results show that fine-tuning the Greyscale or single channel based Resnet50 model helps in improving the accuracy a bit more than using a RGB trained Resnet50 model for fine tuning, though it haven't achieved great result due to the limitation of lesser computational power and smaller dataset to train a large computer vision network like Resnet50. This work is a contribution towards improving classification in domain of multispectral images usually taken up by satellites. There is no baseline model available right now, which can be used to transfer features to single or multispectral domains like the rest of RGB image field has. The contribution of this work is to build such a classifier for multispectral domain and to extend the state of the art in such computer vision domains.