Abstract
Mixed pixels present a significant challenge in remote sensing image processing, as traditional unmixing methods often suffer from low accuracy when assuming a fixed number of endmembers across an entire scene. This paper introduces a novel spectral unmixing model based on the Fuzzy ARTMAP (FAM) neural network that incorporates dynamic endmember selection. By leveraging the spectral similarity between mixed pixels and pure endmembers, the model first identifies the specific types and quantities of materials present within a single pixel. Subsequently, a modified Fuzzy ARTMAP architecture is employed to calculate material abundances. Experimental results demonstrate that this selective approach significantly outperforms traditional linear mixing models and standard neural network implementations, providing results that more closely align with ground-truth observations.
Introduction
In satellite imagery, the spatial resolution often results in "mixed pixels"—single data points that represent a combination of multiple land-cover types, such as vegetation, water, and urban structures. The process of decomposing these pixels into their constituent materials and calculating their respective proportions is known as spectral unmixing.
A major limitation of conventional unmixing techniques, such as the Linear Mixing Model (LMM), is the assumption of a static set of endmembers. If a scene contains four primary materials, standard models attempt to find a non-zero proportion for all four in every pixel, leading to "spectral noise" or "phantom" signatures where materials are reported in areas they do not physically occupy. This paper addresses this issue by proposing a selective endmember method that dynamically identifies the relevant components for each pixel before quantification.
Methodology
The proposed framework consists of a two-stage process designed to improve both the qualitative and quantitative accuracy of unmixing.
1. Iterative Endmember Selection
To determine which materials are actually present in a pixel, the authors utilize Cross-Correlogram Spectral Matching (CCSM). This technique compares the spectral signature of a mixed pixel against a library of known pure endmembers. Through an iterative projection process, the system identifies the best-matching endmember, accounts for its contribution, and repeats the process on the residual signal. This stage effectively filters out irrelevant materials, ensuring that the subsequent unmixing step only considers physically present components.
2. Fuzzy ARTMAP Neural Network Unmixing
Following selection, the material proportions are calculated using a Fuzzy ARTMAP (FAM) neural network. FAM is chosen for its ability to handle non-linear relationships and uncertain boundaries, which are common in complex light-scattering environments. The innovation lies in the dynamic adjustment of the network's output layer; the architecture is pruned to match only the specific subset of endmembers identified in the first stage. This hybrid approach combines physical spectral logic with the robust processing power of neural networks.
Experimental Setup and Results
The model was validated using Enhanced Thematic Mapper (ETM) imagery from the Yangtze River Three Gorges region, captured in April 2002. Six spectral bands were analyzed to distinguish between four primary classes: the Yangtze River, Residential Land, Vegetation, and Lakes.
Quantitative Performance
The selective endmember method was compared against a standard LMM and a non-selective Fuzzy ARTMAP model. The proposed method achieved the lowest Root Mean Square Error (RMSE) across all test categories. For instance, in the "River" category, the RMSE was reduced to 0.0610, compared to 0.0811 for the linear model.
Qualitative Improvements
Visual analysis of the resulting abundance maps revealed that the selective method produced significantly "cleaner" data. Traditional models often showed low-level "gray noise" (false positives) for residential land in the middle of dense forests. In contrast, the selective method correctly identified these pixels as pure vegetation, assigning zero abundance to other classes and eliminating false detections.
Discussion and Conclusion
The study highlights that identifying the components of a pixel is as critical as measuring their abundance. By overcoming the limitations of fixed-endmember assumptions and addressing the non-linearity of spectral mixing, the proposed Fuzzy ARTMAP-based model offers a more realistic representation of land cover.
Future research will focus on automating the initial endmember extraction process and optimizing the computational efficiency of the iterative matching stage to facilitate large-scale applications. Additionally, refining the stopping criteria for iterative selection remains a key area for further development to ensure the model's robustness across diverse geographic regions.
Summary updated