Landcover class extraction from Prisma hyperspectral data using Fuzzy Machine Learning Techniques

Abstract. Hyperspectral remote sensing enables highly detailed spectral discrimination between landcover types. However, conventional classification methods often struggle with spectral variability, mixed pixels, heterogeneous landscapes or when limited training data is provided. This study presents a fuzzy machine learning framework for extracting landcover classes from Prisma hyperspectral data, using the strength of soft classification to address classification related challenges. The spaceborne hyperspectral dataset, with its rich spectral resolution spanning across VNIR to SWIR region, serves as an ideal dataset for detecting subtle spectral differences between surface materials/features. By implementing fuzzy machine learning approach, the study moves beyond the conventional binary/hard classification, enabling the identification of partial class memberships and improving the mapping accuracy in heterogeneous and spectrally overlapping landcover types. After pre-processing and MNF dimensionality reduction, different fuzzy techniques (Fuzzy C-Means, Possibilistic C-Means and Modified Possibilistic C-Means) were applied to different vegetation types, built-up area and water body. The model was trained with limited ground truth, and results show that fuzzy techniques achieve higher class precision and spatial consistency particularly in urban and vegetation-based landscape. The study highlights the potential of fuzzy based machine learning as a robust soft classification technique for hyperspectral data achieving a classification accuracy of 87.01% (MPCM-HSI) and 92.38% (MPCM-HSI+PAN). This demonstrates, how the spectral resolution of Prisma hyperspectral sensor can be efficiently used for enhanced landcover mapping in real-world situations where only partial class knowledge is known.