Hyper-Fairy Wombat Dense Classification Network for Enhanced Land Cover Classification in Polsar Images

In remote sensing, accurately classifying land cover types from Polarimetric Synthetic Aperture Radar (PolSAR) images presents significant challenges due to dynamic environmental factors and complex backscatter behaviours. To address these challenges, this paper proposes a Hyper-Fairy Wombat Dense Classification Network for effective land cover classification. Moreover, the PolSAR images are affected by the elevated HV and HH returns, which reduce the foliage in the winter and complicate the discriminative patterns. To overcome this, a Hybrid Log-Kalman Hyperspherical Pre-processing is introduced, which reduces speckle noise and stabilizes polarimetric feature distributions across seasons. However, the inadequate temporal and spatial context integration in traditional feature extraction methods leads to poor adaptability to dynamic environmental conditions. Therefore, the Optimized Time-Aware ConvLSTM Fairy-Wren Filter is employed to isolate the trend, seasonal, and residual components. Additionally, selecting relevant features from redundancy and correlation results in multicollinearity and increases computational burden. Thus, Mutual Particle Component Selection is introduced, which reduces redundancy and multicollinearity, preserving the most informative components. Moreover, classification between land cover types with similar polarimetric signatures often leads to misclassification. To tackle this, the Wombat-Kernelized DenseNet Classifier is utilized, which establishes the nonlinear decision boundaries and effectively discriminates classes with overlapping polarimetric signatures. The proposed algorithm’s performance is validated through comparative evaluation against existing state-of-the-art approaches, demonstrating superior accuracy and robustness.