High resolution mapping of development in the wildland-urban interface using object based image extraction

Abstract

The wildland-urban interface (WUI), the area where human development encroaches on undeveloped land, is expanding throughout the western United States resulting in increased wildfire risk to homes and communities. Although census based mapping efforts have provided insights into the pattern of development and expansion of the WUI at regional and national scales, these approaches do not provide sufficient detail for fine-scale fire and emergency management planning, which requires maps of individual building locations. Although fine-scale maps of the WUI have been developed, they are often limited in their spatial extent, have unknown accuracies and biases, and are costly to update over time. In this paper we assess a semi-automated Object Based Image Analysis (OBIA) approach that utilizes 4-band multispectral National Aerial Image Program (NAIP) imagery for the detection of individual buildings within the WUI. We evaluate this approach by comparing the accuracy and overall quality of extracted buildings to a building footprint control dataset. In addition, we assessed the effects of buffer distance, topographic conditions, and building characteristics on the accuracy and quality of building extraction. The overall accuracy and quality of our approach was positively related to buffer distance, with accuracies ranging from 50 to 95% for buffer distances from 0 to 100 m. Our results also indicate that building detection was sensitive to building size, with smaller outbuildings (footprints less than 75 m²) having detection rates below 80% and larger residential buildings having detection rates above 90%. These findings demonstrate that this approach can successfully identify buildings in the WUI in diverse landscapes while achieving high accuracies at buffer distances appropriate for most fire management applications while overcoming cost and time constraints associated with traditional approaches. This study is unique in that it evaluates the ability of an OBIA approach to extract highly detailed data on building locations in a WUI setting.

Keywords: Environmental science; Geography.

Fig. 1

Basemap of the study area and the ten randomly selected National Aerial Image Program (NAIP) quadrangles used in the evaluation.

Fig. 2

Extracting buildings in the wildland urban interface using an object based image analysis. The evaluated method uses an iterative process that starts with a user defined training set and user defined algorithms to produce an initial dataset of objects interpreted as buildings (pink polygons). It then uses manual intermediate steps to identify a secondary training set of correctly and incorrectly identified objects to refine object detection algorithms for subsequent outputs. Lastly, a user can conduct quality control by manually adding missed buildings or removing incorrectly identified buildings in the final dataset.

Fig. 3

3A) Separation distance between the control building footprint (solid polygon), and the extracted feature centroid (circle within hatched polygon) used to identify accurately identified features and extraction errors. 3B) Buffers around control building footprints (solid polygon) are used to assess agreement with extracted feature centroid (circle within hatched polygon) at different scales.

Fig. 4

Accuracy, omission error, and commission error between control buildings and extracted features during each iteration of the Object Based Image Analysis extraction. The error bars represent the 95th confidence interval between samples, and the composite quality index is shown in parenthesis above the error bars.

Fig. 5

Accuracy, omission error, and commission error between control buildings and extracted features produced from the final manual iteration. The error bars represent the 95th confidence interval between samples, and the composite quality index is shown in parenthesis above the error bars.

Fig. 6

Extraction accuracy for different building sizes, classified into 25 m² bins, using a representative 30 m buffer.