Abstract
Unsupervised classification is increasingly vital for remote sensing analysis as it eliminates the need for labor-intensive training samples. This is particularly relevant for mangrove ecosystems, where difficult terrain makes field measurements challenging. This paper introduces an Artificial Immune Network Clustering (AINUC) algorithm designed for mangrove mapping. By evolving a network of "antibodies" to represent data clusters, the algorithm achieves superior stability and accuracy compared to traditional methods like K-Means.
Introduction
Mangrove forests are vital ecological components, serving as coastal barriers and significant carbon sinks. Monitoring these ecosystems via remote sensing is essential but challenging. Traditional supervised classification requires extensive field-collected training data, which is difficult to obtain in muddy, intertidal mangrove environments. Unsupervised methods offer a solution, but standard algorithms like K-Means often lack the sophistication to handle the spectral complexities of coastal vegetation.
The AINUC Methodology
This paper presents the Artificial Immune Network Clustering (AINUC) algorithm, a bio-inspired unsupervised approach. AINUC mimics the adaptive behavior of an immune system to identify data clusters without prior labeling.
1. Hybrid Initialization
Standard Artificial Immune Networks (AINs) often suffer from slow convergence due to random initialization. AINUC addresses this by using Fuzzy K-Means (FCM) to "mine" the data first, establishing a set of initial "antibodies" (cluster centers) that better represent the underlying data distribution. This "Aha" moment in seeding ensures the algorithm starts with a representative "memory" of the landscape, avoiding local optima.
2. Immune Evolution Cycle
The algorithm iteratively refines these antibodies through a sophisticated evolutionary cycle:
- Recognition: Measuring the Euclidean distance between data points (antigens) and antibodies to determine affinity.
- Clone Proliferation: Multiplying the "strongest" antibodies that show high affinity to the data.
- Mutation: Slightly altering clones to explore the feature space more effectively and maintain diversity.
- Suppression: Deleting redundant or low-affinity antibodies to keep the network efficient and prevent the "overcrowding" of cluster centers.
3. Multi-Dimensional Feature Integration
To improve discrimination between mangroves and similar vegetation, the model incorporates a 12-dimensional feature vector:
- Spectral Bands: 7 bands from Landsat TM imagery.
- Vegetation Indices: NDVI to highlight chlorophyll content.
- Topographic Constraints: A Digital Elevation Model (DEM) to exclude areas above 8 meters, where mangroves cannot grow. This ecological constraint drastically reduces false positives.
- Texture Analysis: Gray-Level Co-occurrence Matrix (GLCM) features to distinguish the fine, smooth texture of mangroves from the coarser textures of upland forests.
Experimental Results and Discussion
The algorithm was evaluated using 30-meter resolution Landsat TM imagery of the Zhangjiang Estuary in China. Performance was benchmarked against the industry-standard K-Means and ISODATA algorithms, with validation performed using 200 test samples verified by high-resolution (1.8m) WorldView-2 imagery and Google Earth snapshots.
| Metric | AINUC | K-Means | ISODATA |
|---|---|---|---|
| Overall Accuracy | 81.4% | 33.1% | 35.8% |
| Kappa Coefficient | 0.778 | 0.302 | 0.315 |
The results demonstrate that AINUC significantly outperforms traditional methods. While K-Means and ISODATA frequently misclassified mangroves as agricultural land, AINUC accurately delineated the major mangrove stands and identified critical littoral zones. The immune network's inherent self-adaptation makes it much more robust to noise than rigid geometric clustering.
Conclusion
The AINUC algorithm provides a powerful tool for the unsupervised classification of remote sensing imagery in complex environments. By combining fuzzy logic for initialization with the adaptive power of immune networks, it achieves high accuracy without the need for manual training samples. Future work will explore the application of AINUC to hyperspectral data and further refine the self-adaptive mechanisms to reduce the need for manual parameter tuning.
Summary updated