Land Use Classification With Engineered Features

We report on a methodological framework that analyzes land-use images with engineered (manually designed) features. As older feature engineering methods suffered from excessive computation of their features, we therefore, introduce techniques that are faster and also more elaborate than any previous approach. Feature extraction and description is based on contour and region information. The contour analysis comprises the detection of ridge, river, and edge contours and is based on a technique of minimal complexity (without requiring costly multiplicative operations). The traced contour segments are then partitioned and abstracted; then they are clustered to form group descriptors. The region analysis consists of the detection of brighter, darker, and flatter regions, as well as regions obtained from clustering; the clustering is carried out with minimal complexity using a hierarchical analysis. Region segments are partitioned and abstracted using a fast implementation of the symmetric-axis transform. A total of ca. 70 parameters is developed and land-use classification experiments are performed. On the UC Merced (UCMD) collection, the classification accuracy of Deep Nets is reached; on the NWPU-RESISC45 collection the accuracy still lags somewhat.