Enhancing remote sensing image analysis: optimization of a hybrid deep network through HHO algorithm

• Published in: Multimedia tools and applications Vol. 84; no. 28; pp. 33545 - 33565
• Main Authors: Digra, Monia, Dhir, Renu, Sharma, Nonita
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.08.2025; Springer Nature B.V
• Subjects: Accuracy; Artificial neural networks; Classification; Comparative analysis; Computer Communication Networks; Computer Science; Computer vision; Data Structures and Information Theory; Datasets; Deep learning; Effectiveness; Efficiency; Feature selection; Image analysis; Labels; Multimedia Information Systems; Neural networks; Optimization; Optimization algorithms; Optimization techniques; Remote sensing; Satellites; Special Purpose and Application-Based Systems; Track 9: Machine Learning in Multimedia; Image analysis; Harris Hawk Optimization; Deep Learning; Remote sensing; Image classification
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-20499-y

Remote sensing (RS) image interpretation includes multi-label picture classification, a key problem. The complicated and diverse character of many remote sensing scenes, which are produced by the spatial combination and association of various objects, makes classification difficult. In this research, an innovative approach that combines deep neural networks with bio-inspired optimization techniques to handle the challenges of classifying remote sensing images. First, the efficacy of proposed model was evaluated using the publicly accessible remote sensing scene dataset for multi-label classification. Second, this work employs the HHO (Harris Hawk optimization) model to improve classification performance by efficiently choosing features and also reduces the time complexity by 78%. The DenseNet-169 and VGGNet16 models extract the necessary properties to enhance remote sensing classification. Finally, validate the data by integrating additional validation data in the training set, which leads to more reliable and consistent accuracy and loss. The proposed VGGNet16 with Bi-LSTM and DenseNet169 with Bi-LSTM achieves accuracy on four distinct datasets 98.83% and 99.69% on UCM, 97.77% and 97.98% on AID, 95.06% and 96.84% on NWPU, whereas RSI-CB256 dataset achieved 98.5% and 98.91% accuracy . The hybrid fusion model distinguishes important from irrelevant data. To demonstrate its efficacy, authors conducted extensive tests on various distinct datasets such as UCM, AID, NWPU, RSICB256 and have ascertained that this approach exhibits superiority over existing state-of-the-art techniques currently employed in the field.