Sebastian Böck

On the Objectivity of the Objective Function—Problems with Unsupervised Segmentation Evaluation Based on Global Score and a Possible Remedy

Image segmentation is a crucial stage at the very beginning of many geographic object-based image analysis (GEOBIA) workflows. While segmentation quality is generally deemed of great importance, selecting adequate tuning parameters for a segmentation algorithm can be tedious and subjective. Procedures to automatically choose parameters of a segmentation algorithm are meant to make the process objective and reproducible. One of those approaches, and perhaps the most frequently used unsupervised parameter optimization method in the context of GEOBIA is called the objective function, also known as Global Score. Unfortunately, the method exhibits a hitherto widely neglected, yet severe source of instability, which makes quality rankings inconsistent. We demonstrate the issue in detail and propose a modification of the Global Score to mitigate the problem. This hopefully serves as a starting point to spark further development of the popular approach.

Identifying suitable segmentation parameters for an object-based image classification

Only well-chosen segmentation parameters ensure optimum results of object-based image classifications. Manually defining suitable parameter sets can be a time-consuming approach, not necessarily leading to optimum results. Moreover, the manual approach is obviously subjective. An automated evaluation approach is needed to reduce human intervention and to provide objective criteria for ranking different parameter sets. In this work, we test three different ways to find the optimum segmentation. (i) We used a supervised approach integrating the segmentation and classification tasks. The segmentation is optimized directly with respect to the overall accuracy of the subsequent classification. (ii) Using the global score value considering withinand between segment heterogeneity, we run an unsupervised segmentation optimization to find the best segmentation parameters. (iii) Using manually delineated objects we calculate discrepancy measurements like euclidean distance between automatic generated objects and the manual delineated objects in a supervised segmentation evaluation. For all approaches, we present fully autonomous workflows for supervised and unsupervised object-based classification, combining image segmentation, segmentation evaluation and image classification. Starting from a fixed set of randomly selected and manually delineated training samples, suitable segmentation parameters are automatically identified. Finally, we compare the results of the three different approaches. ∗Corresponding author

Automated segmentation parameter selection and classification of urban scenes using open-source software

Object-based image analysis (OBIA) using high spatial resolution remote sensing imagery has been successfully applied in numerous land-use/land-cover studies. OBIA usually requires user inputs at several stages, limiting reproducibility and automation. This work focuses on developing the potential of a relatively simple and general solution to classification of very-high resolution airborne remote sensing imagery. The approach is demonstrated on the datasets made available for the ISPRS 2D Semantic Labeling Challenge and using solely freely available open-source software. In the first step, Large-Scale Mean-Shift Segmentation (LSMS) is used to obtain an initial partitioning of the test images, using the spectral component as the only input to the algorithm. On the output of LSMS, a in-house implementation of Spectral Difference Segmentation (SDS) is performed. Hundreds of candidate segmentation results are produced over a range of scales, by altering input parameters of the segmentation algorithms. A modified version of a well-known unsupervised segmentation evaluation method based on a combination of global inter-segment and intra- segment heterogeneity is employed to objectively rank and select potentially superior segmentation results for subsequent random forest (RF) classification. Statistical features are calculated for each segment based on its pixel values, using spectral (visible and NIR), height (nDSM) and textural (wavelet) data as input. RF classifiers are trained using subsets of the reference data available and their performance is assessed using held-out test data for validation. Results demonstrate the potential of unsupervised segmentation evaluation to reduce user intervention in OBIA and the ability of freely available software to produce high-quality classification results.