A resource limited artificial immune system algorithm for supervised classification of multi/hyper‐spectral remote sensing imagery

Liangpei Zhang, Yanfei Zhong, Bo Huang, Peng LiInternational Journal of Remote Sensing 2006 · DOI

TL;DR

This paper introduces RLCRSI, a novel artificial immune system algorithm designed for supervised classification of multi/hyper-spectral remote sensing imagery. By adapting the resource-limited artificial immune system (RLAIS) principles, RLCRSI offers a more efficient and accurate approach to classifying complex remote sensing data, outperforming traditional methods in experimental evaluations.

What Problem Does This Paper Address?

Classifying multi/hyper-spectral remote sensing imagery is crucial for various applications, from environmental monitoring to urban planning. Traditional classification algorithms, while established, often struggle with the high dimensionality and complexity of this data. Artificial Immune Systems (AIS), inspired by the human immune system, have shown promise in data analysis and clustering. However, the conventional RLAIS algorithm, despite its strengths, has been limited in its application to remote sensing due to its computational complexity and the sheer volume of data involved. This paper aims to bridge this gap by developing a resource-limited artificial immune system algorithm specifically tailored for supervised multi/hyper-spectral image classification, addressing both accuracy and computational efficiency concerns.

Methods

The core of this paper is the development of the Resource-Limited Classification of Remote Sensing Image (RLCRSI) algorithm. RLCRSI is built upon the Resource Limited Artificial Immune System (RLAIS) model, which itself is inspired by the B-cells of the human immune system. Key aspects of the RLCRSI methodology include:

  • Artificial Recognition Balls (ARBs): These are the fundamental units of the RLAIS, representing a population of B-cells with similar antibodies. In RLCRSI, ARBs are used to recognize and classify image data.
  • Resource Allocation Mechanism: ARBs compete for a finite pool of resources, with more stimulated ARBs consuming more. This competition drives an evolutionary pressure, favoring ARBs that are better at recognizing antigens (image data).
  • Memory Cell Pool: A crucial component for classification, the memory cell pool stores evolved ARBs that have demonstrated strong recognition capabilities. RLCRSI limits the size of this pool to improve efficiency.
  • Data-Driven Self-Adaptation: RLCRSI is designed to adapt to the data without requiring explicit functional or distributional assumptions about the underlying model, making it flexible for complex, real-world data.
  • Immune Model Properties: The algorithm leverages immunological properties like memory and clonal selection to model complex relationships within the data.
  • RLCRSI Specific Modifications: To address the limitations of conventional RLAIS for remote sensing, RLCRSI focuses resource allocation and stopping criteria on ARBs of the same class as the antigen, reducing computational load without sacrificing accuracy.

The algorithm's implementation involves six steps: selection of regions of interest (ROIs), normalization and initialization, ARB generation, competition for resources and nomination of candidate memory cells, promotion of candidate memory cells to the memory cell pool, and finally, classification.

Datasets

The paper evaluates the proposed RLCRSI algorithm using three distinct datasets:

  1. Wuhan TM Image: A 30-meter resolution multispectral Landsat TM image of Wuhan city acquired on October 26, 1998. This dataset was used for Experiment 1 and included four classes: water, vegetation, road, and building.
  2. Wuhan MODIS Image: A MODIS image of Wuhan acquired on April 2, 2002, featuring 500-meter reflectance data for channels 3-7. This dataset was used for Experiment 2 and comprised four classes: water, cloud, city, and vegetation.
  3. Xiaqiao PHI Image: An airborne imaging spectrometer (PHI) image from the Xiaqiao test site, a mixed agricultural area. This dataset was used for Experiment 3 and included seven classes: road, corn, vegetable, tree, grass, water, and soil.

Key Insights and Ideas

  • Immune Systems as Powerful Classifiers: The paper strongly advocates for the potential of artificial immune systems, particularly RLAIS, as robust and effective tools for complex classification tasks like remote sensing.
  • Resource Limitation is Key to Efficiency: The core innovation of RLCRSI lies in its resource-limited approach. By controlling the memory cell pool and focusing computational effort, it makes AIS practical for large-scale remote sensing data.
  • Self-Adaptation for Complex Data: The ability of RLCRSI to adapt to data without prior assumptions about its distribution is a significant advantage over many traditional methods that rely on assumptions like Gaussian distributions.
  • Immune Metaphors Translate to Performance: The paper demonstrates that concepts like "memory," "clonal selection," and "affinity" from immunology can be effectively translated into algorithmic components that yield high classification accuracy.
  • Trade-offs in Parameter Tuning: The sensitivity analysis highlights that while parameters like clonal_rate and ATS can influence accuracy, they also have a significant impact on computational time, suggesting a need for careful selection based on application requirements.

Key Results

The RLCRSI algorithm consistently outperformed traditional supervised classification algorithms (Parallelepiped, Minimum Distance, Maximum Likelihood, K-Nearest Neighbour, and Back-Propagation Neural Network) across all three experimental datasets:

  • Wuhan TM Image (Experiment 1): RLCRSI achieved the highest overall accuracy (92.33%) and Kappa coefficient (0.891), surpassing other methods.
  • Wuhan MODIS Image (Experiment 2): RLCRSI demonstrated superior performance with an overall accuracy of 93.67% and a Kappa coefficient of 0.91, while maintaining a reasonable computational time (5.4 s).
  • Xiaqiao PHI Image (Experiment 3): RLCRSI again achieved the highest overall accuracy (90.15%) and Kappa coefficient (0.88), showing its effectiveness even with hyperspectral data.

Qualitatively, RLCRSI provided visually more accurate classifications, particularly in distinguishing challenging classes like vegetation and city, where other methods struggled.

Limitations and Future Work

  • Parameter Sensitivity: The paper acknowledges that RLCRSI's performance is sensitive to user-defined parameters such as clonal_rate, stimulation_threshold, and ATS. While the paper provides insights into these sensitivities, selecting optimal parameters for diverse datasets might require further investigation.
  • Computational Time: While RLCRSI is more efficient than some traditional AIS, its computational time can be higher than some simpler methods like ML, especially with higher clonal_rate values. The authors suggest that selecting appropriate parameters can mitigate this.
  • Future Work: The authors propose further exploration of AIS and RLCRSI for more extensive remote sensing applications, indicating a belief in the broad applicability of this approach. This could include investigating its performance on even larger and more complex datasets, or exploring hybrid approaches combining RLCRSI with other machine learning techniques.

Summary updated

Abstract

The resource limited artificial immune system (RLAIS), a new computational intelligence approach, is being increasingly recognized as one of the most competitive methods for data clustering and analysis. Nevertheless, owing to the inherent complexity of the conventional RLAIS algorithm, its application to multi/hyper‐class remote sensing image classification has been considerably limited. This paper explores a novel artificial immune algorithm based on the resource limited principles for supervised multi/hyper‐spectral image classification. Three experiments with different types of images were performed to evaluate the performance of the proposed algorithm in comparison with other traditional image classification algorithms: parallelepiped, minimum distance, maximum likelihood, K‐nearest neighbour and back‐propagation neural network. The results show that the proposed algorithm consistently outperforms the traditional algorithms in all the experiments and hence provides an effective new option for processing multi/hyper spectral remote sensing images.
BibTeX
@article{zhang2006resource,
  author = {Liangpei Zhang and Yanfei Zhong and Bo Huang and Peng Li},
  title = {{A resource limited artificial immune system algorithm for supervised classification of multi/hyper‐spectral remote sensing imagery}},
  journal = {International Journal of Remote Sensing},
  year = {2006},
  doi = {10.1080/01431160600675903},
}
Figure