研究目的
To propose and evaluate a new spectral angle and vector mapping (SAVM) classification method for improving the classification accuracy of hyperspectral remote sensing imagery in urban areas by incorporating both spectral angle and spectral vector length differences.
研究成果
The SAVM classification method, which considers both spectral angle and vector length differences, significantly improves classification accuracy over the traditional SAM method for urban hyperspectral imagery, with an overall accuracy increase of 16.77%. This makes SAVM more suitable for complex urban land use mapping, though further refinements are needed for specific classes like water and shadow.
研究不足
The SAVM method requires further improvement to better differentiate between water and shadow areas due to their spectral similarities. Additionally, it may not fully address the classification of roofs made from different materials, indicating a need for more nuanced feature extraction or additional variables in urban hyperspectral classification.
1：Experimental Design and Method Selection:
The study compares the traditional Spectral Angle Mapping (SAM) method with the newly proposed Spectral Angle and Vector Mapping (SAVM) method. SAM uses only the spectral angle between vectors, while SAVM incorporates both the spectral angle and the difference in spectral vector lengths.
2：Sample Selection and Data Sources:
Hyperspectral imagery from the Hyperspectral Digital Imagery Collection Experiment (HYDICE) for the business district of Washington, DC, USA, was used. The data had 1280 rows, 307 columns, 210 wave bands (reduced to 191 bands after excluding poor-quality bands), and a spatial resolution of
3：8 m. Training and test samples were collected from on-site sampling and high-resolution Google Earth images, covering seven land use classes:
roof, street, path, grass, tree, water, and shadow.
4：List of Experimental Equipment and Materials:
The HYDICE sensor was used for data acquisition. Software tools for data processing and classification were employed, but specific names or models are not mentioned in the paper.
5：Experimental Procedures and Operational Workflow:
The imagery was preprocessed by excluding certain wavelength bands. Reference spectra were extracted from training samples. SAM and SAVM algorithms were applied to classify the imagery. For SAM, the spectral angle was calculated using Eq. (1). For SAVM, separability was calculated using Eq. (2), which includes both angle and vector length differences. Classification results were validated using test samples to compute accuracy metrics.
6：1). For SAVM, separability was calculated using Eq. (2), which includes both angle and vector length differences. Classification results were validated using test samples to compute accuracy metrics. Data Analysis Methods:
5. Data Analysis Methods: Confusion matrices were generated to calculate overall accuracy, producer's accuracy, consumer's accuracy, and Kappa coefficient for both SAM and SAVM methods. Statistical comparison of accuracies was performed to evaluate the improvement offered by SAVM.
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
