In this chapter, the authors offer an already-trained CNN model for HSI classification. By fine-tuning the parameters, the suggested DenseNet model classification accuracy is increased. In order to fine-tune the hyper-parameters, an enhanced version of the Hunter-Prey Optimisation algorithm (IHPOA) is used. The convergence of the HPO technique is sped up by the addition of adaptive inertia weights to the optimisation search phase. At the same time, the authors tweak the starting population to boost the procedure capacity to do worldwide searches. Extensive experimental findings collected on three publicly available HSI datasets show that the suggested technique may minimise computational complexity and over smoothing while maintaining competitive performance compared to numerous state-of-the-art approaches.
Top1. Introduction
When compared to traditional multispectral or RGB (red, green, and blue) pictures, hyperspectral images (HSIs) are more successful in discriminating distinct land cover types because they collect extensive spatial-spectral material (Wambugu et al., 2021). Therefore, HSIs are used for a broad variety of purposes, from catastrophe prevention and management to military surveillance and maritime monitoring. The HSI classification method (which assigns a label to each pixel) is essential for these uses (Wang et al., 2021). However, it is challenging to extract discriminative information from HSIs for classification applications (Khan et al., 2022) because too complicated noise effects, spectral fluctuation, labelled training taster deficits, and significant spectral mixing amongst materials.
In addition, HSI misclassification is common when there are few labelled training examples (Jia et al., 2021). The curse of dimensionality arises when there are many spectral bands but few training samples, which can severely impair performance in HSI, whereas the goal of feature selection is to identify the most informative spectral bands and exclude the rest [6, 7]. Several classifiers have been tried out on HSIC so far. The early stages of HSIC relied on a variety of classification techniques to extract the necessary data, such as support vector machine (SVM), and random forest (RF) (Hou et al., 2021). These approaches were influenced by the growth of machine learning technology.
However, using these approaches, spectral and spatial info are often analysed independently, and spectral properties alone cannot properly differentiate between different land-covers (Manifold et al., 2021). Hand-engineered features including Gabor filters, extended morphological profiles, wavelets, rotation were employed for HSIC to extract rich spatial information from HSIs (Khan & Paheding et al., 2021). While the performance of HSICs has been greatly enhanced by smoothing-based methods, the flexibility of machine learning approaches to accommodate land covers of varying sizes and shapes remains a challenge (Saha et al., 2021). The spatial covers, such as in a super pixel-and graph (Yuan et al., 2021), showed the underlying topologies of HSIs, allowing for additional extraction of spatial information. These techniques can enhance the effectiveness of HSIC by extracting features of possible interaction between neighbouring land coverings. The aforementioned constructed spectral-spatial approaches, on the other hand, are highly empirical and reliant on expert knowledge (Nguyen et al., 2021).
Natural image processing are only few of the areas where deep learning methods have been widely implemented as a result of the expansion of deep learning and the advancement of artificial intelligence (AI) (Yue et al., 2021). Many deep neural networks, such as stacked auto encoders (SAEs), recurrent have been explored for HSIC (Xue et al., 2021) because of their ability to automatically capture high-level features from HSI. From 1D CNNs to 3D CNNs, from single-channel CNNs to multichannel CNNs, and from shallow CNNs to deep CNNs, recent years have seen widespread use of CNNs for HSIC (Wei et al., 2021).
This work introduces a refined DenseNet model as a response to the exceptional qualities and efficient performance of DL models in computer vision. First, the features are retrieved using a CNN model, and then the suggested model is used to do the classification. Tuning hyper-parameters like momentum, learning rate, and epochs is how the suggested model is made better. With the help of random reverse learning and an adaptive weight approach, IHPO is able to complete this procedure with increased efficiency. The remainder of the paper is laid out as shadows: In Section 5.2, we define the relevant literature, and in Section 5.3, we provide a brief description of the suggested model. In Section 5.4, we label the experimental evaluation of the proposed model on three datasets, and in Section 5.5, we draw the necessary conclusions.