Spatial Variation Sequences for Remote Sensing Applications with Small Sample Sizes

Machine learning applications in remote sensing often require a labour-intensive feature engineering step, if only a small number of samples is available and transfer learning is not applicable. Here, we are introducing the concept of Spatial Variation Sequences, which allows to apply methodologies from automated time-series feature engineering to remote sensing applications of static images. The presented example application detects swimming pools from four-channel satellite images with an \(F_{1}\) -score of 0.95, by generating spatial variation sequences from a modified swimming pool index. The automated feature engineering approach reduced the dimensionality of the classification problem by 99.7%. A more traditional approach using transfer learning on pre-trained Convolutional Neural Networks (CNN) was evaluated in parallel for comparison. The CNN approach boasted a higher performance of \(F_{1}\) -score of 0.98 but required the use of pre-trained weights. The comparable performance of the FE and CNN approach demonstrates that time-series feature extraction is a valuable alternative to traditional remote sensing methods in the presence of data scarcity or the need of significant dimensionality reduction.