Multispectral versus texture features from ZiYuan-3 for recognizing on deciduous tree species with cloud and SVM...

Area of Science:

• Forestry remote sensing
• Geospatial analysis
• Ecological monitoring

Background:

• Accurate tree species recognition is vital for effective forest remote sensing and resource management.
• ZiYuan-3 satellite imagery provides multispectral and texture data crucial for detailed forest analysis.
• Phenological variations significantly influence the spectral and textural characteristics of tree species.

Purpose of the Study:

• To construct and optimize spectral and texture indices for enhanced tree species recognition.
• To evaluate the effectiveness of multidimensional cloud models and support vector machine (SVM) models for identifying Quercus acutissima and Robinia pseudoacacia.
• To compare recognition accuracies between autumn and winter phenological phases and between spectral and textural features.

Main Methods:

• Selected multispectral and texture features from ZiYuan-3 satellite images acquired in autumn and winter.
• Constructed and optimized sensitive spectral and texture indices.
• Developed multidimensional cloud models and support vector machine (SVM) models using screened indices.
• Evaluated model performance based on recognition accuracy for Quercus acutissima and Robinia pseudoacacia.

Main Results:

• Spectral indices showed stronger correlations with tree species in winter than in autumn, with band 4 being particularly effective.
• Optimal texture indices varied by species: mean, homogeneity, and contrast for Quercus acutissima; contrast, dissimilarity, and second moment for Robinia pseudoacacia.
• Spectral features yielded higher recognition accuracy than textural features for both species.
• Winter data provided superior recognition accuracy compared to autumn data, especially for Quercus acutissima.
• The one-dimensional cloud model achieved 90.57% accuracy, outperforming the multidimensional cloud model (89.98%) and a 3D SVM (84.86%).

Conclusions:

• Winter phenological phase and spectral features are more advantageous for remote sensing-based tree species recognition.
• Cloud models, particularly the one-dimensional version, demonstrate higher accuracy than SVM for this specific application.
• The findings offer valuable technical support for precise tree species identification and sustainable forestry management in mountainous regions.