Reducing the Burden of Aerial Image Labelling Through Human-in-the-Loop Machine Learning Methods

Razzak, Muhammed T

Reducing the Burden of Aerial Image Labelling Through Human-in-the-Loop Machine Learning Methods

Master Thesis

2021

Abstract

This dissertation presents an introduction to human-in-the-loop deep learning methods for remote sensing applications. It is motivated by the need to decrease the time spent by volunteers on semantic segmentation of remote sensing imagery. We look at two human-in-the-loop approaches of speeding up the labelling of the remote sensing data: interactive segmentation and active learning. We develop these methods specifically in response to the needs of the disaster relief organisations who require accurately labelled maps of disaster-stricken regions quickly, in order to respond to the needs of the affected communities. To begin, we survey the current approaches used within the field. We analyse the shortcomings of these models which include outputs ill-suited for uploading to mapping databases, and an inability to label new regions well, when the new regions differ from the regions trained on. The methods developed then look at addressing these shortcomings. We first develop an interactive segmentation algorithm. Interactive segmentation aims to segment objects with a supervisory signal from a user to assist the model. Work within interactive segmentation has focused largely on segmenting one or few objects within an image. We make a few adaptions to allow an existing method to scale to remote sensing applications where there are tens of objects within a single image that needs to be segmented. We show a quantitative improvements of up to 18% in mean intersection over union, as well as qualitative improvements. The algorithm works well when labelling new regions, and the qualitative improvements show outputs more suitable for uploading to mapping databases. We then investigate active learning in the context of remote sensing. Active learning looks at reducing the number of labelled samples required by a model to achieve an acceptable performance level. Within the context of deep learning, the utility of the various active learning strategies developed is uncertain, with conflicting results within the literature. We evaluate and compare a variety of sample acquisition strategies on the semantic segmentation tasks in scenarios relevant to disaster relief mapping. Our results show that all active learning strategies evaluated provide minimal performance increases over a simple random sample acquisition strategy. However, we present analysis of the results illustrating how the various strategies work and intuition of when certain active learning strategies might be preferred. This analysis could be used to inform future research. We conclude by providing examples of the synergies of these two approaches, and indicate how this work, on reducing the burden of aerial image labelling for the disaster relief mapping community, can be further extended.

Keywords

Electrical Engineering

Reference:

Collections

Masters

Full item page