Classifying Remote Sensing Data Through Advanced Dimensionality Reduction Approaches

One major area of interest in the field of urban studies has been the geospatial analysis of urban structures through the classification of remote sensing data. To improve the classification process, it is recommended to use supplementary data in addition to the original satellite bands. Mathematical Morphology provides a variety of tools for the generation of spatial features such as Morphological Profiles, which are employed for the optimal discrimination of pixels present in satellite images. However, these features can be redundant or irrelevant, necessitating their elimination through dimensionality reduction. This preprocessing step is crucial in hyperspectral and multi spectral image classification to remove unnecessary bands or generated features while preserving essential information, thereby improving classifier performance. The Improved F-score technique (IFS) is a feature selection method that evaluates and retains the most relevant bands/features based on computed scores. Bands/features with scores below a calculated threshold score are discarded. A significant challenge with this method is finding an appropriate threshold score. This paper proposes an improvement to the IFS technique by introducing clustering strategies. These strategies aim to separate relevant from irrelevant information without predefining a threshold score based on the obtained scores. The proposed approaches refine the selection process of the IFS while boosting classification results. The proposed methodology was evaluated using multispectral Sentinel-2 data, demonstrating its effectiveness in enhancing classification accuracy. Among the tested clustering strategies, Agglomerative Hierarchical Clustering with Average-linkage achieved the highest performance, with an overall accuracy of 98.46% and a Kappa coefficient of 0.97. These results highlight the superiority of clustering-based approaches over traditional thresholding methods for feature selection in terms of accuracy.