Automated Pixel Purification for Delineating Pervious and Impervious Surfaces in a City Using Advanced Hyperspectral Imagery Techniques

Conventional urbanization transforms natural into paved landscapes, posing a significant environmental challenge. Detecting the changes in (im)pervious surfaces in cities, where patches are small and intermingled, is particularly challenging. This study introduces a novel approach to these changes by integrating Coupled Non-negative Matrix Factorization (CNMF) image fusion with an automatic pixel purification algorithm. By fusing low-resolution hyperspectral (30m) with high-resolution panchromatic (5m) PRISMA imagery, we achieved enhanced spatial resolution, crucial for accurate land use and land cover (LULC) classification. We introduced automatic pixel purification as a key innovation method to improve LULC mapping accuracy, sensitive to training pixel selection and mixed pixels. This method, which is tested in Dublin City area, enhanced/ refined spectral signatures and clarity across major LULC classes including bare soil, industrial roofs, grasslands, trees, residential roofs/asphalts, and water bodies, significantly improving classification accuracy by removing outliers and ensuring spectral consistency. The Random Forest (RF) algorithm, applied before and after pixel purification, showed substantial increases in overall accuracy (from 94.04% to 96.69%,) and Kappa coefficient (from 92.60% to 95.91%) for 2021, with similar improvements in 2022. This method enabled accurate differential analysis of (im)pervious surfaces, revealing a 4.08% decrease in pervious (from 33.29 km2 to 28.08 km2) and a 4.09% increase in impervious surfaces (from 79.96 km2 to 82.92 km2) over one year, highlighting the rapid urbanization’s impact on Dublin’s landscape permeability. This study significantly advances LULC classification and urban monitoring, offering valuable insights for sustainable urban development and advocating for its integration into future remote sensing and urban planning initiatives.