A comparison between different segmentation methods

NI Yu

Guide: Introduction Methodology Results Conclusion References

Introduction

Upland heather moors are of vital importance in terms of economic and aesthetic values. For better management of heather moors, a large amount of knowledge is required. Remote sensing offers an alternative approach to replace traditional manual surveys which are time-consuming. Therefore, image analysis techniques are used to interpret the imagery from a remote sensor. Object-based image analysis has its advantages over pixel-based image analysis. This research focuses on the object-based image analysis techniques, which makes image segmentation the first step for further analysis such as classification.

Four segmentation techniques were studied in this project:

1. Multiresolution segmentation (eCognition Professional 4.0);
2. Edge-based segmentation (python);
3. Marker-controlled segmentation (python);
4. Efficient graph-based segmentation (python).

A comparison between eCognition segmentation and python segmentation was made to see:

1. How are the performance of each method;
2. Whether python is able to produce better result than eCognition software which is expensive and the algorithm within it remains unpublished.

Methodology

A trial-and-error (Mathieu & Aryal, 2007) iterative procedure was used to get the best results of each method. This procedure was used to test the parameters in each algorithm: 1. Scale parameter, shape and compactness of multiresolution segmentation; 2. Sigma of edge-based segmentation; 3. Marker value for background and foreground of marker-controlled segmentation and 4. Scale, sigma and minimum size (min_size) of efficient graph-based segmentation. Once the most satisfying results were produced, the comparison was made between them. Spyder and eCognition Professional 4.0 were the main software and data used was aerial RGB imagery. The study area is shown below.

Results

a) is the best result of edge-based segmentation while b) is that of marker-controlled and c) is from efficient graph-based. d) is the result of multiresolution segmentation from eCognition Professional 4.0. According to visual interpretation, d) is the best of this four approaches. However, python segmentation shows it potential in segmenting objects. For example, red box in b) has the same quality with d). Also, the purple box in c) segments the lake precisely and matches with the result in d).

Conclusion

This study was aimed to determine whether it is possible that python segmentation could outperform the multiresolution algorithm built in eCognition Professional 4.0. The result showed that the overall quality of multiresolution segmentation is better than that of python segmentation. This can be concluded from the following points: 1. The segments produced by eCognition suited the best to a visual segmentation by human eyes; 2. Edge-based segmentation could rarely segment image objects; 3. Marker controlled segmentation produced results that partially match with that of eCognition segmentation. 4. Efficient graph-based segmentation results were mostly different in size and shape with the result of eCognition. However, this does not mean that python algorithms cannot perform segmentation with good quality. The results of marker controlled and efficient graph-based segmentation showed a potential for further investigation in acquiring good segmentation output as they have both segmented some heather stands which can be compared with the same point in eCognition output.

The parameters of all algorithms are critical for segmentation. Scale parameter and the weight of colour are more influential than other three in multiresolution segmentation in eCognition. However, the characteristic of shape should still be taken into account, not only in the domain of eCognition software but also in the implementation of python segmenting algorithms. Markers for background and foreground are important. The scale parameter in efficient graph-based algorithm is the most important one, while min_size is the least, as it only refines the result.

References

Baatz, M., Benz, U., Dehghani, S., Heynen, M., Höltje, A., Hofmann, P., Lingenfelder, I., Mimler, M., Sohlbach, M., Weber, M. & Willhauck, G. (2004) eCognition Professional User’s Guide. München, Germany: Definiens Imaging GmbH.

Trimble. (2014) eCognition Developer 9.0: Reference Book. Definiens AG, München.

Van der Walt, S., Schönberger, J. L., Nunez-Inglesias, J., Boulogne, F., Warner, J. D., Yager, N., Gouillart, E. & Yu, T. (2014) scikit-image: Image processing in Python. PeerJ 2:e453.

Back to top