Enhancing Hyperspectral Image Analysis With Deep Learning: An Innovative Hybrid Approach for Improved Detection and Classification Using Deep-Detect

Abstract

Surface material identification using hyperspectral imaging (HSI) analysis is a crucial and challenging issue in remote sensing. Researchers widely recognize that exploiting spectral-spatial data outperforms using spectral pixel-wise methods. Most of the work focuses on machine learning techniques to address issues with HSI categorization. The semi-supervised learning model, for example, was utilized in early research to address the issue of the expenses associated with training samples during HSI classification. Nevertheless, these approaches and techniques have certain drawbacks. The suggested framework must be improved in terms of hardware and computational capacity to enhance security. While integrated patch-based convolutional neural network (CNN) can help overcome these issues, due to the complicated geographical distribution of many targets, there are some downsides, such as using spectra as odorless vectors. The supervised learning techniques categorize the inputs by using a collection of appropriate examples for each class, referred to as training samples. For this purpose, samples must be obtained, however, the process is expensive and time-consuming. Deep learning techniques for evaluating spatial spectrum data have significantly advanced, yet security challenges remain in remote sensing with HSI. In this article, we introduce a novel HSI classification method, termed Deep-Detect, which leverages deep learning models with attention mechanisms to tackle these challenges. We use two weighting factors that may be obtained flexibly during the training phase to integrate both spectral and spatial properties, whose importance differs for various objects and scenarios. Additionally, we proposed a spatial and spectral fully convolutional networks (Spe-FCNs) attention mechanism that focuses attention on the input patches’ most informative regions. To verify performance, we used the northern Italy dataset from Pavia University and compared our proposed model with state-of-the-art HSI classification models. Our findings show that the proposed Deep-Detect model enhances classification accuracy while addressing key security concerns in HSIs.

Keywords

• Deep-detect
• security
• hyperspectral imaging (HSI)
• Spa-FCNN
• Spe-FCNN
• SSFCN