A CNN-Transformer Dual-Stream Architecture for Enhanced Hyperspectral Image Categorisation

The usage of hyper spectrum images (HSIs) in satellite imagery cataloguing is imperative because it enables correct credentials of materials and land cover categories. While conventional CNNs can be good at identifying localized spatial features, they do not perform as well when dealing with the longrange spectral relationships found in HSIs. On the other hand, Transformers can find long range spectral relationships but tend to ignore small-scale spatial features. Therefore, this research developed DeepsNet, an innovative dual-branch CNNTransformer architecture designed to utilize both the spectral and spatial characteristics of HSIs to create highly accurate HSI classifications. One branch uses a CNN to identify high resolution spatial characteristics of the HSI while the other branch utilizes a Transformer to find the long-range spectral relationships between the bands. Dual-branch architectures such as DeepsNet provide the means to precisely monitor environments, map land use and manage resources, thus contributing to SDG aligned sustainable development and environmental protection initiatives. The complementary representations from each branch are then fused together using a methodically designed fusion process to improve classification accuracy and discriminability. Experiments conducted on several commonly referenced HSI datasets demonstrated that DeepsNet produces higher levels of accuracy than the state-of-the-art methods and performs better under conditions where there are limited amounts of training data available. Furthermore, the experiments showed that DeepsNet is supplementary efficient than the current state-of-the-art devices when processing large volumes of data. Total, the investigational results demonstrate the effectiveness of the dual-branch design for multifaceted HSIs and provide a solid source for upcoming research studies into their likely.