Advanced hyperspectral image classification via adaptive triplet networks and chaotic quasi oppositional optimization

• Published in: Optical and quantum electronics Vol. 56; no. 7
• Main Authors: Rose, J. T. Anita, Daniel, Jesline, Chandrasekar, A.
• Format: Journal Article
• Language: English
• Published: New York Springer US 05.06.2024; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Characterization and Evaluation of Materials; Classification; Computer Communication Networks; Computer vision; Data mining; Deep learning; Electrical Engineering; Hyperspectral imaging; Image classification; Lasers; Machine learning; Optical Devices; Optics; Optimization; Photonics; Physics; Physics and Astronomy; Structural models; Feature extraction; Triplet network; CNN; CQFFA; Shot learning; Hyperspectral image classification
• ISSN: 1572-817X; 0306-8919; 1572-817X
• DOI: 10.1007/s11082-024-06753-5

In recent years, deep learning networks have been utilized in computer vision applications to validate the depth of the model and also significantly improve data mining. However, it improved the performance of the hyperspectral image classification process in different structural models. However, due to the presence of uncleared data and unclear images, the efficiency of the model is diminished. Also, there is a possibility of generating overfitting issues due to slow processing and this reduces the classification accuracy. Therefore, a novel Adaptive Triplet Network (ATN) based Hyperspectral Image Classification method is proposed to overcome the above-mentioned issues while classifying the features. This method is generated by integrating the Triplet network and Chaotic Quasi Oppositional Farmland Fertility Algorithm (CQFFA) that determines an efficient hyperspectral image classification process with few-shot learning. Mostly the CNN classifier is utilized for classification and the determination of triplet network classifies multiple data but, in such situations, the classification is not accurate due to overfitting of images. Hence the CQFFA optimization method is determined to solve the overfitting issue and extract multiple features simultaneously for better classification. The efficiency of the proposed method is evaluated by USGS and ICVL-HSI datasets as well as the metrics namely accuracy, precision, recall, F1-score, specificity, ROC, and time complexity. The experimentation result revealed that the proposed method has a better hyperspectral image classification accuracy of 98.03% than state-of-the-art methods.