研究目的
Identifying a suitable distance function for hyperspectral images to maintain the accuracy of hyperspectral image processing results.
研究成果
The most suitable distance function for hyperspectral data, one that is a metric and responds proportionally without saturating in both cases of translation and magnitude change, is Euclidean distance of cumulative spectrum (ECS). Other metrics that saturate with preferably bigger saturation point, such as Canberra and χ2 1 distances, are also considered suitable under less strict constraints.
研究不足
The study identifies that many distance functions saturate when there is no overlapping between two spectra, limiting their effectiveness in certain scenarios. Additionally, some distance functions are unstable when dealing with noisy data.
1：Experimental Design and Method Selection:
The study compares existing distance functions and defines a set of selection criteria for evaluating their suitability for hyperspectral images. Theoretical constraints and behavior, as well as numerical tests, are proposed for the evaluation.
2：Sample Selection and Data Sources:
Simulated spectral reflectance signals using Gaussian distribution functions and real hyperspectral images of pigment patches are used for evaluation.
3：List of Experimental Equipment and Materials:
Not explicitly mentioned in the provided text.
4：Experimental Procedures and Operational Workflow:
The evaluation involves testing distance functions on simulated spectra for translation and magnitude change cases, and on real spectral data from pigment patches to assess classification performance.
5：Data Analysis Methods:
The study analyzes the theoretical properties of distance functions, including their responses to simulated and real spectral data, to determine their suitability for hyperspectral image processing.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
