研究目的
To address the issue of spectral aliasing in coastal wetland object types that leads to class mixing by proposing a multiobject CNN decision fusion classification method for hyperspectral images to improve classification accuracy.
研究成果
The proposed multiobject CNN decision fusion method effectively improves classification accuracy for hyperspectral coastal wetland images by inheriting the advantages of single-object feature band classification, with an overall accuracy of 82.11%, outperforming other methods by 3.33% to 6.24%. It provides a practical solution for handling class mixing in challenging conditions.
研究不足
The number of CHRIS hyperspectral image bands is limited, which may not fully demonstrate the advantage of the feature band and decision fusion approach; more data is needed for comprehensive validation.
1：Experimental Design and Method Selection:
The method uses spectral deviation to extract separable spectrum ranges for object types based on field spectral data, employs a CNN model with spatial-spectral features for classification, and applies a fuzzy membership decision fusion algorithm to combine single-object classification results.
2：Sample Selection and Data Sources:
The study area is the Yellow River Estuary coastal wetlands, using a CHRIS hyperspectral image acquired in June 2012 with 18 bands, and field spectral data collected from 2008-2016 for reed, Spartina, Tamarix, Suaeda salsa, tidal flat, and water types.
3：List of Experimental Equipment and Materials:
ASD Fieldspec Handhold 2 spectrometer for field spectral measurements, CHRIS hyperspectral image, SPOT6 high-resolution image for validation.
4：Experimental Procedures and Operational Workflow:
Preprocessing of the image (geometric correction, missing pixel filling, noise elimination), feature band selection using spectral deviation method, CNN model training and classification for single-object types, decision fusion based on fuzzy membership rules to obtain final classification.
5：Data Analysis Methods:
Classification accuracy evaluation using overall accuracy and Kappa coefficient, comparison with SVM and full-band CNN methods.
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
