Hybrid Taguchi-Objective Function optimization approach for automatic cave bird detection from terrestrial laser scanning intensity image
- Publisher:
- University of South Florida Board of Trustees, a public body corporate, having locations in Tampa, St. Petersburg, and Sarasota
- Publication Type:
- Journal Article
- Citation:
- International Journal of Speleology, 2016, 45, (3), pp. 289
- Issue Date:
- 2016
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
This paper proposes an optimized Taguchi-objective function segmentation-based image analysis to detect bird nests in a cave from high resolution terrestrial laser scanning intensity images. First, the Taguchi orthogonal array was used to design 25 experiments with three segmentation parameters: scale, shape, and compactness, each having five variable factor levels. Then, a plateau objective function was computed for each experiment using their respective level combinations. A merger of the factor level combination in the orthogonal array and the computed plateau objective function values was used to generate main effects and interaction plots for signal-to-noise ratios, which provided a measure of robustness for scale, shape, and compactness factors. The optimized parameters were used in the segmentation process in eCognition. The image object was then classified into nest and cave-wall on the basis of laser return intensity and area index using knowledge-based rule sets, and the detection accuracy was evaluated. The result produced area under ROC curve of 0.93 with P<0.0001 at 95% confidence level. This indicates that the proposed method is effective for distinguishing birds from cave-wall with high precision. The classification result was transferred to ArcGIS where the detected nests were counted after post-classification editing. A total number of 25,959 nests were counted from the seven scan scenes used. This shows that the fusion of Taguchi and objective function is indeed an effective method to determine optimal segmentation parameters to group image objects as small as birds within a segment. Moreover, the use of segments’ spectral intensity value and area index increased classification accuracy significantly. Further, the method was tested for reliability using six additional images. The test of heterogeneity using Cochran’s Q and Inconsistency tests produced a P value of 0.384 and I2 value of 5.10% at 95% confidence interval, respectively. This shows that the method is consistent with non-significant difference among the trials.
Please use this identifier to cite or link to this item: