Engineering orthogonal intensity and texture features for sample design, classification, regression, and clustering analyses using principal component analysis and convolution kernels

Local raster texture is an important attribute that can be used for sample design and raster-based classification, regression, and clustering. However, determining the appropriate kernel weights to quantify texture can be challenging and can vary depending on the remote sensing or computer vision task. We present an automated approach and companion Jupyter notebook that uses principal component analysis (PCA) and convolution kernels to quantify orthogonal intensity and texture metrics. Our approach linearly projects raster data across covariance axes, learns kernel cell weights, and provides a straightforward mechanism to retain as much information as possible within the raster dataset while removing redundancy. Using this PCA convolution approach, we reduced the dimensionality of a Landsat 8 scene and corresponding five by five analysis window from 175 to 10 independent intensity and texture components that maintained more than 95% of the variation within the raster data. Applying our PCA convolution methodology, we created a 10-band raster dataset that can be used to inform sample design or as orthogonal predictors in classification, regression, and clustering models.