Mathematical morphology is a powerful tool for analyzing the geometrical properties of land cover classes while mitigating spectral confusion in remote sensing data, particularly in hyperspectral images. Among various approaches for modeling spatial information, hierarchical image processing via morphological multi-scale decomposition is a prominent technique for structural image characterization. A key component of this decomposition is morphological leveling, which provides a robust framework for hierarchical image analysis. This study systematically investigates and compares different morphological leveling decomposition strategies using a hyperspectral image acquired in 2002 over the University of Pavia campus in Italy. Two experiments were conducted to assess the impact of different morphological leveling configurations on classification performance. The first experiment evaluated four morphological leveling constructions, each defined by distinct reference and marker parameters combined with three filtering techniques (Gaussian Convolution, Alternate Sequential Filter, and Averaged Alternate Sequential Filter). The second experiment assessed the effectiveness of residual leveling images derived from the results of the first experiment. Outputs from both experiments were integrated with the spectral information and used as input to a two-dimensional convolutional neural network to improve classification accuracy. The analysis of overall accuracy and agreement with ground truth data shows that incorporating morphological leveling significantly enhances classification performance. One configuration achieved nearly 99% accuracy with a very high level of agreement, highlighting the potential of multi-scale spectral–spatial feature extraction. These results offer valuable insights for future applications in remote sensing and environmental monitoring.