A structural similarity-based label-smoothing algorithm for the post-processing of land-cover classification

Qikai Lu, Xin Huang, Tingting Liu, Liangpei ZhangRemote 2016 · DOI

TL;DR

This paper introduces a novel algorithm for refining land-cover classifications from remote sensing data. By leveraging structural similarity between pixel neighborhoods, it adaptively smooths noisy classifications while preserving important details, outperforming traditional methods like majority voting.

What Problem Does This Paper Address?

Remote sensing is a powerful tool for observing Earth, but raw classification maps often suffer from "salt-and-pepper" noise. This means individual pixels are misclassified, creating a speckled appearance that obscures the true land cover. While post-processing techniques can clean up these maps, many existing methods, like simple majority voting, tend to over-smooth the results, blurring important boundaries and losing fine details. This paper aims to address this by developing a more sophisticated post-processing method that balances noise reduction with detail preservation.

Methods

The core of the proposed method is inspired by the Non-Local (NL)-means algorithm, originally developed for image denoising. Instead of just considering spatial proximity, the algorithm assigns weights to neighboring pixels based on the structural similarity of their surrounding neighborhoods to the central pixel's neighborhood. This means pixels that look more alike in terms of their local context are given more influence.

The authors propose two specific ways to measure this structural similarity:

  1. Consistency Criterion (SSV-C): This method compares the labels of corresponding pixels within the neighborhoods of the central pixel and its neighbors. The weight is determined by the number of matching pixel pairs.
  2. Histogram Similarity Criterion (SSV-H): This approach calculates a local histogram for each neighborhood, representing the frequency of different land-cover classes. The weight is then based on the similarity between these histograms, using the intersection of histograms as a measure.

The algorithm then uses these calculated weights to perform a weighted voting process to determine the final class label for each pixel.

Datasets

The proposed algorithm was evaluated on three distinct remote-sensing datasets:

  • AVIRIS Indian Pines: A hyperspectral image with a spatial resolution of 20m, containing 10 classes.
  • WorldView-2 Hainan: A high-resolution multispectral image with a spatial resolution of 2m, acquired over Hainan Province, China.
  • ZY-3 Wuhan: Another high-resolution multispectral image with a spatial resolution of 5.8m, acquired over Wuhan, China.

The raw classification results for these datasets were obtained using Support Vector Machines (SVM) with a Radial Basis Function (RBF) kernel.

Key Insights and Ideas

  • Structural Similarity is Key: The paper highlights that simply considering spatial distance (as in traditional filters) is insufficient. The contextual information within a pixel's neighborhood is crucial for accurate post-processing.
  • Adaptive Weighting: By adaptively calculating weights based on structural similarity, the algorithm can better distinguish between true boundaries and noisy pixels.
  • Preserving Details: The SSV-C method, in particular, shows promise in preserving narrow and elongated features (like green belts) that are often lost by traditional over-smoothing methods.
  • Robustness to Window Size: The proposed method demonstrates greater robustness to variations in the size of the processing window compared to traditional approaches.

Key Results

The experimental results on the three datasets consistently show that the proposed structural similarity-based label-smoothing algorithm (SSV-C and SSV-H) outperforms traditional methods like Majority Voting (MV) and Distance-Weighted Voting (DWV).

  • Quantitative Improvement: The overall accuracy (OA) achieved by the proposed methods is consistently higher, especially as the window size increases. For instance, on the AVIRIS Indian Pines dataset, SSV-H achieved an OA of 86.37% compared to MV's 84.85% and DWV's 86.17%. Similar improvements were observed for the other datasets.
  • Qualitative Improvement: Visual interpretation of the classification maps reveals that the proposed methods effectively reduce salt-and-pepper noise and produce smoother results. Crucially, they also preserve finer details and structural information, such as elongated green belts, which are smoothed out by traditional methods.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, the NL-means inspired approach, which compares neighborhoods, can be computationally more intensive than simpler methods.
  • Parameter Tuning: The paper mentions that optimal values for SVM parameters were obtained via cross-validation, implying that parameter tuning is a necessary step for achieving optimal results with the proposed method as well.
  • Future Work: The authors suggest that their approach is effective and provides a promising direction for future research in land-cover classification post-processing. Further exploration of different similarity measures and their impact on performance could be a natural next step.

Summary updated

Abstract

Post-processing is able to achieve a satisfactory classification performance with a low cost and simple assumption, making it widely used in the refinement of classification maps. In this study, a novel structural similarity-based label-smoothing algorithm is developed for the post-processing of land-cover classification. Inspired by the non-local (NL) means algorithm, the proposed algorithm assigns different voting weights to the neighbouring pixels for the identification of the central pixel. Here, the voting weight of a specific neighbouring pixel depends on its structural similarity to the central pixel. In this paper, two measurements are proposed to evaluate the similarity between pixels: (1) a consistency criterion; and (2) a histogram similarity criterion. The proposed algorithm was tested on three remote-sensing images. The experimental results confirm that the proposed algorithm reduces the classification noise and preserves the detail and structural information at the same time. Compared to the traditional post-processing approaches (e.g., majority voting), the proposed algorithm exhibits a more satisfactory performance.
BibTeX
@inproceedings{lu2016structural,
  author = {Qikai Lu and Xin Huang and Tingting Liu and Liangpei Zhang},
  title = {{A structural similarity-based label-smoothing algorithm for the post-processing of land-cover classification}},
  booktitle = {Remote},
  year = {2016},
  doi = {10.1080/2150704x.2016.1149252},
}
Figure