Adaptive Channel Attention Dropout (ACAD)-CNN technique for hyperspectral image-based Pichavaram mangrove species classification

Mangroves serve as habitats for diverse species. Accurate classification of mangrove species is important for their conservation and management. However, the spectral similarity among species poses significant challenges in traditional classification models. Hyperspectral Imagei (HSI) provides rich spectral information that can differentiate between species through spectral signatures. State-of-the-art (SOTA) convolutional neural networks (CNNs) and machine learning (ML) models struggle to utilize the redundant spectral information of HSI, leading to misclassification. Hence, this study proposes an Adaptive Channel Attention Dropout (ACAD)-CNN for the classification of 16 mangrove species, including both true mangrove and mangrove associate species in the Pichavaram mangrove forest, by utilizing EO1-Hyperion and AVIRIS-NG datasets processed in Google Colab. The ACAD-CNN performs inter-channel spectral correlation and assigns importance weights to feature maps for the extraction of the most discriminative spectral features. Moreover, the adaptive dropout mechanism selectively suppresses weakly contributing channels. Performance metrics demonstrated substantial improvements over the SOTA CNN model, Swin transformer, XGBoost and Naive Bayes algorithms. The proposed ACAD-CNN achieved an accuracy of 96.92% for AVIRIS-NG and 95.28% for Hyperion, with corresponding Kappa coefficients of 0.9627 and 0.9416. These results demonstrate that the ACAD-CNN effectively enhances mangrove species classification.