Land Cover Classification of Fused Hyperspectral and Multispectral Image Using In-Ception-Resnetv2

Remote sensing data from various satellites and multimodal data processing techniques have grabbed the curiosity of geoscience communities. Accurate classification of land cover from remotely sensed satellite images remains a serious challenge. The proposed work develops an Inception-ResNet v2 to classify different land covers of the Konkan region of India. Quick Atmospheric Correction (QUAC), stacking, and merging techniques are used to preprocess multispectral images. SG smoothing and merging techniques are used to pre-process hyperspectral images. The pre-processed spectral information contains multiple wavebands, resulting in a high-dimensional feature set. Therefore, Principal Component Analysis (PCA) with Fennec Fox Optimization (FFO) is applied to extract useful wavebands. Subsequently, the extracted bands are sent to the hybrid NCTCP-DTCWT data fusion algorithm. These fused images are then given to Inception-ResNet v2 for classification. Proposed deep learning model can detect different land covers such as forest region, vegetation land, water, built-up, and barren land in the given fused image. The model achieved 96% accuracy, 91% precision, and 98% specificity. Comparison of different metrics prove that our hybrid model outperformed several other machine learning models explored in earlier work. Further, the proposed model is tested with fused images and non-fused images. The study proves that fused images can generate accurate classes than non-fused images. This study examines whether optimized band selection combined with multi-resolution hyperspectral– multispectral fusion within a unified deep-learning framework can significantly improve land-cover classification in complex coastal environments.