Because this study is the rst of its kind, it serves to not only identify abandoned buildings in the city of Guelph, but also to test the TensorView software and develop associated workflows. With this in mind, the study was divided into 7 distinct phases. Developing the tool itself is not considered to be part of the study, as it was effectively a one-off work effort that can be reused in other contexts. The TensorView tool itself can be found at its GitHub repository.
Phase 1: Training Set Development
A neural network is as effective as the training set that it is given. In the case of InceptionV3, the training set consists of a series of images, as well as the single category to which each image belongs. We developed the training set for this survey using Google Image Search - search terms corresponding to different building types and elements of the streetscape were provided, and the resulting images were bulk-downloaded, deduplicated, and manually sorted. For this proof-of-concept survey, this yielded 2,885 images across 18 categories. Because of the particular subject matter being studied, it was not feasible to train the network using Street View images of abandoned buildings - manually locating enough buildings would have required more time than was available, and likely not have significantly improved accuracy. Refining the training set was a process that required approximately 8 weeks of part-time work. With additional resources brought to bear, this could likely be accomplished significantly faster, and "real-world" training sets would tend to be reusable - unlike the comparatively naïve set used for this study.
Phase 2: Survey Area Selection and Preparation
Due to time and budget constraints, it was decided to focus the survey on a few comparatively small sub-areas of Guelph in which abandoned buildings were considered likely to exist, as well as some in which they were not. In order to generate the points le provided as input to TensorView, the OpenStreetMap shapefile dataset for Ontario was used as a baseline, with further operations performed on it using ArcGIS (OpenStreetMap Foundation, 2017). In this case, these steps were:
Masking to Guelph Only: The OpenStreetMap roads dataset is broken down by national-level political subdivisions, with the smallest being those at the level of a province or state. In order to save processor power during the remaining operations, the city of Guelph was clipped out of the Ontario roads dataset by manually drawing a box around the city and applying the clip operation.
Trimming Non-Roads: The OpenStreetView dataset contains paths that are not roads, but instead footpaths, parking lot lanes, etc.. Since these are not surveyed for Street View imagery, they were removed. This was done by selecting any road that was not labelled as a freeway, arterial, secondary, or tertiary road, and deleting them. This yielded the overall road layer that served as a basis for creating the survey areas.
Drawing Survey Sub-Areas: Because of limited time and budget, several sub-areas of the city of Guelph were chosen, the majority of which were considered likely to contain abandoned buildings based on the authors' local knowledge. For comparison purposes, some areas that were considered unlikely to contain abandoned buildings were also included. Bounding boxes for these areas were manually drawn, and the roads contained within them were clipped from the Guelph street grid. An overview of these study areas can be found in Figure 1.
Buffering Around Roads: Because TensorView employs a "guess and check" method for locating panoramas, it is important that the road dataset it uses corresponds to that used by Google. Because the OpenStreetMap road dataset does not precisely correspond to Google's, especially around major intersections, it was decided to create a 5m buffer around the roads within the survey areas within which to search for panoramas.
Search Grid Generation: In order to allow proper "guess and check" searching for panoramas, a search grid within the buffered zone was created. This consisted of generating an approximately 1m by 1m point grid within the buffered areas, with the hope that each panorama would have at least one point associated with it. The Street View API uses a "nearest point" system for determining whether there is a panorama at a specific location, so this was considered to have a high chance of success (Google Inc., n.d.).
The points shapefile resulting from the final step consisted of 19,655 points across approximately 17.3km of roads, and was provided as the input to the TensorView tool. This task required approximately half an hour to complete.
Phase 3: Automated Survey
This step consisted of providing the survey-area shapefile and trained neural network to the TensorView tool and allowing it to run. Because buildings only appear at the side of the road, the automated survey was set to only download and process two slices per panorama - one each to the immediate left and right of the car, with a 45-degree field of view. The survey area, under these parameters, required approximately two hours to download and classify all of the images.
Phase 4: Manual Review of Identified Images
After running, it was found that TensorView had identified approximately 433 (approximately 12.9%) of the 3,358 images supplied to it as being of abandoned buildings. Once the automated survey had concluded, the images that had been identified were manually reviewed, in order to determine which ones were actually of abandoned buildings, and which were false positives. This process required approximately one hour of manual work, and the 433 tagged images yielded 58 images of 18 locations considered promising enough to field-survey.
Phase 5: Field Survey
In order to verify the results of the survey, and account for Street View's low temporal resolution, a field survey based on the identified images was conducted. This consisted of driving to the locations identified by the images, and photographing the current state of the structure in question. Occasionally, this involved photographing otherwise vacant lots or empty fields where the structures had been demolished in the intervening time. The field survey required approximately 1 1/2 hours of work spread over two days due to snow and cold, which significantly hampered photographic fieldwork.
Phase 6: Archival and News Research
In order to provide a more complete picture of the status and value of each site, a search was made of local newspaper articles and the City of Guelph and Wellington County online archives. Due to time constraints, it was not possible to perform a search of land ownership at the land registry office.
Phase 7: Recommendation Development
Once all of the data had been compiled, recommended courses of action were developed based on the results of the field survey and archival research. From these, and neighbouring land uses noted during the field survey, recommendations as to what course of action the city and private developers should pursue were created for inclusion with this report.