Ensemble Representation Learning for Land Use Land Cover Classification

Representation learning techniques are used to extract meaningful feature set from raw data for classification or analysis. This study examines the possibility of various traditional and deep-learning-based Dimensionality Reduction Mechanisms such as Linear Discriminant Analysis (LDA), Principal Component Analysis (PCA), and Auto-encoder, for Land-Use-Land-Cover (LULC) classification in the field of Remote Sensing. Comparison on performance of individual and ensemble methods is examined by using a Logistic Regression Classifier as a baseline model for a 10-class classification problem. The results show that the combination of LDA and Auto-encoder gives superior performance in results with reduced running time in classification. The study also explores future research directions of simultaneous feature extraction methods and multimodal representation learning in Remote Sensing. Representation Learning techniques can enhance the performance of algorithms in real world scenarios involving high resolution spatial band data. The study can be utilized as a base for integrating multiple representation learning techniques to efficiently classify observations in Remote Sensing.