Semi-Automated Detection of Fragmented Rectangular Structures in High Resolution Remote Sensing Images with Application in Archaeology

Abstract

Automated visual analysis has substantially advanced in recent years, allowing a variety of targets to be automatically detected. Remarkably successful algorithms and technologies have been developed, e.g., for face detection and for obstacle detection for autonomous car navigation. In archaeology, however, remote sensing images are still analyzed in the traditional way, by a visual inspection. Such a visual inspection is performed prior to a field survey in order to identify potential sites that may guide later fieldwork. Though this approach saves fieldwork time, visual inspection remains very time consuming and requires the highly concentrated attention of an expert. Due to human fatigue, this approach might be unreliable. Moreover, the visual inspection of image data over vast unexplored areas is not feasible at all. This is especially frustrating, since a large amount of high resolution image data has become available due to recent developments of satellite technology. It is, therefore, very appealing to automate screening of large datasets of remote sensing images.An automated screening does not aim at replacing visual image interpretation by an Archaeologist. Though machine vision algorithms have already become quite powerful at some visual tasks, human vision and its ability to interpret scenes and recognize objects is clearly superior in general. On the other hand, in contrast to a human expert, machine vision algorithms are capable of routinely screening a large amount of imagery and generating plausible candidate locations. These findings can be verified easily and timely by an archaeologist, who can also estimate their potential significance. Such a semi-automated approach can increase the efficiency of an archaeological survey of vast unexplored areas. In this thesis we develop a semi-automated methodology for detecting unknown rectangular structures, such as archaeological remains of livestock enclosures (LSE), in wide alpine areas covered by high resolution remote sensing images (HRRSI). The LSE structures were of a special interest in several recent archaeological studies, because such architectural remains offer important insights into the origins and historical development of alpine pasture economy. The LSEs have varying sizes and aspect ratios, may be heavily ruined, and may have spectral properties similar to the surrounding terrain and rocks. They appear in HRRSI images as faint fragmented usually approximately rectangular contours on a complex background. As a part of our methodology we introduce particular image analysis algorithms that are briefly outlined below. We introduce an approach for segmenting out large contextually inappropriate regions of high texture contrast, such as urban areas, forests, or rocky areas. This approach is shown to be superior to other methods in terms of accuracy of segmentation at texture borders and ability to distinguish texture details from individual features, which might be a part of the structures we are searching for. We also introduce a complementary method that extracts individual image features, i.e. linear segments corresponding to walls of ruined LSEs, while suppressing background texture. A quantitative comparison with alternative approaches is also provided. We propose a method for fast detection of initial candidates. It generates sparse locations that are, at least partially, enclosed by a structures of an arbitrary shape. Image patches at candidate locations are further analyzed based on dedicated rectangularity-size features introduced in the thesis. These features allow capturing rectangular enclosures, even if distorted, incomplete, or fragmented. On the other hand, they are not sensitive to a variety of irrelevant structures, such as isolated corners, line intersections, parallel curves, etc. The LSE structures are detected using a linear classifier fed with the rectangularity-size features. We have designed a dedicated linear classifier that is not prone to overfitting the data even in our case of extremely unbalanced data with only a few positive and a large number of negative examples. We quantitatively compare the effectiveness of the rectangularity-size features for our detection task with other handcrafted features and with state-of-the-art pre-trained deep CNN features. The flow of the image analysis algorithms, which automatically generates detections and their confidence, is followed by a visual inspection by means of a specially designed graphical user interface (GUI). The GUI allows quick and convenient validation of true detections and rejection of falsely detected sites. We demonstrated the feasibility of our methodology by applying it to two large alpine regions. We were able to detect the LSE structures of interest, some of which were hitherto unknown. Although, we developed the algorithms for processing HRRSI with a particular archaeological application in mind, they can be used in different domains and for different purposes. We have, for example, briefly discussed the application of some of the algorithms to detecting individual buildings in rural or mountainous areas.