DELWP Schedule 22: Mapping Planned Burns using Remote Sensing Benchmarking Image Segmentation for Burn Mapping using Aerial Imagery Technical Report, April 2019

[摘要] Executive Summary: Planned Burns are used to mitigate bushfire risk by reducing available fuels (Newnham et. al 2015). Accurate mapping of the severity and extent of burns is useful to assess performance relative to the burn plan, to help quantify the level of residual risk and to assess ecological impacts. Recent work has found that high accuracy mapping of burn severity is possible through manual delineation of burned areas using high resolution visible and near-infrared aerial imagery acquired within a small number of weeks after a planned burn (McCarthy et.al. 2017). While accurate, this approach is time consuming and requires significant expertise in manual image interpretation. Image segmentation refers to a collection of automated approaches that identify regions of similar pixels in an image (Cheng 2001). Such approaches have been proposed as a method to reduce the time taken to map burn extent in aerial imagery by automating delineation of salient landscape features. Previously, we compared different image segmentation approaches for mapping imagery and identified superpixel based algorithms as the most suitable for burn mapping (Newnham et al., 2016). Specifically, we identified the Simple Linear Iterative Clustering (SLIC) segmentation algorithm (Achanta, 2012) as an open source implementation of this approach which could be readily incorporated into a semi-automated GIS workflow to assist in burn severity mapping.Here we describe a process for assessing the performance of the SLIC algorithm, applied to post-burn aerial imagery, as an initial approximation for burn severity mapping. The assessment process used multiple objective metrics to score the relative correspondence between automatically generated polygon features to manually mapped fire severity boundaries.Our validation dataset included four corresponding sets of four-band (visible and near-infrared) aerial images and manual burn maps. For each image, approximately 400 SLIC image segmentation parameters were applied and performance was assessed with three metrics: 1. boundary recall: how well the boundaries of each manually delineated burn severity polygon was traced by the segmentation algorithm, 2. boundary precision: what proportion of the boundaries produced by the segmentation algorithm corresponded to manually-mapped severity class boundaries, and 3. area overlap: how well the final polygons produced by the segmentation algorithm aligned with the manually delineated polygons. The results suggest that segmentation performance is most comparable to expert manual mapping when fire features are clearly visible in aerial imagery, while performance is worse when visual fire boundaries are difficult to distinguish. Mean boundary recall and mean boundary precision provided the most consistent metrics showing correlated results for different parameter settings. Scores for these metrics could also be compared with less skew across different imagery. The area overlap provided a less consistent metric and was more skewed across different images due to the varying size and quantity of manual mapping polygons in each corresponding manually burn severity mapping.

[发布日期] 2019-06-30 [发布机构] CSIRO

[效力级别] Environmental Management [学科分类] 地球科学(综合)