Details

Autor(en) / Beteiligte

• Rohith, G.
• Kumar, Lakshmi Sutha

Titel

Design of Deep Convolution Neural Networks for categorical signature classification of raw panchromatic satellite images

Ist Teil von

• Multimedia tools and applications, 2022-08, Vol.81 (20), p.28367-28404

Ort / Verlag

New York: Springer US

Erscheinungsjahr

2022

Quelle

SpringerLink

Beschreibungen/Notizen

• Remote Sensing categorical signature classification has gained significant implications on spatial resolution image analysis due to differences in the sensors’ spatial response and surface variations. As a consequence, the grey level co-occurrence of probability texture features for the classification task is crucial. Traditionally, deep learning-based Convolution Neural Network (CNN) classifiers for spectral/spatially scaled signatures (Hyperspectral or Multispectral images) extract deep features and accurately classify remote sensing scenes into appropriate labels/categories. While dealing with raw panchromatic images, the spatial with varied angular signatures will have untrained grey scale patterns, translational and rotational variations. It is still a bottleneck to label and classify data using pre-trained models from two separate sources based on its spatial structural characteristics. In this paper, a thirteen-layer deep CNN model is designed for categorical signature classification of the raw panchromatic satellite dataset. The design is carried out in three stages- Firstly, the method extracts the global content and meanings of remote sensing images at the scene level. Then, it cross compares with training and testing of identified complex remote sensing signatures in raw inter dataset images with large inter and intra-class variations. Finally, the validation of the 70:30 training-testing set is done to classify a batch of images into the respective labeled signatures (Land and sea) with an accuracy of 88.9%. The modified versions of five state-of-the-art pre-trained classifiers are tested to check the efficacy of the proposed approach.