Abstract:Hyperspectral images are increasingly being used in classification and identification. Data users prefer hyperspectral imagery with high spatial resolution, finer spectral resolution, and high signal-to-noise ratio (SNR). However, tradeoffs exist in these core parameters in imagery acquired by different hyperspectral sensor systems. Data users may find it difficult to utilize all the advantages of hyperspectral imagery. How to select hyperspectral data with optimal parameter configuration has been one of the essential issues for data users, which also affects the back-end applications. With advancements in computer science, various artificial intelligence algorithms from conventional machine learning to deep learning have been utilized for hyperspectral images classification and identification. Few researchers study the mechanism between the core parameters of hyperspectral imaging spectrometers and advanced artificial intelligence algorithms, which affects the application efficiency and accuracy. In this paper, we delved into the evolution of machine learning and deep learning models applied to imagery acquired by different hyperspectral sensor systems having different SNR, spectral, and spatial resolutions. Additionally, we also considered the tradeoffs among the core parameters of hyperspectral imagers. We used two conventional machine learning models, including the classification and regression tree (CART) and random forest (RF), two deep learning methods based on convolution neural network architectures-3D convolutional neural network (3D-CNN) and hamida, and two deep learning methods based on vision transformers architectures-transformer models vision transformer (VIT) and robust vision transformer (RVT), to compare the characteristics of different algorithms. In addition, five hyperspectral datasets with different species categories and scene distributions and aggregated datasets with different spatial resolutions, spectral resolutions, and SNRs were used to validate our study. The experimental results indicate that: (1) The overall accuracy (OA) using CART, RF, 3D-CNN, and VIT models decreased with coarser spectral resolution, but almost remained unchanged using the RVT classifier. The number of class and classification species affect the results. (2) The influence of spatial resolution on classification accuracy is related to the scene complexity, target size, and classification purpose. The coarser spatial resolution can achieve higher OA than the original spatial resolution for the uniform scene distribution. For the datasets with small objects and intersections of different species, OA first increased, plateaued, and then decreased with coarser spatial resolution. (3) The SNR has an obvious impact on OA for the CART and RF classifiers, and the impact decreased for deep learning models, especially for the VIT and RVT models, which were almost unaffected by SNR. Additionally, slight variations in experimental results were observed for datasets with different scene distributions and categories. Furthermore, we conducted a detailed analysis of the role of traditional machine learning and deep learning models in the experimental outcomes. The study can provide insights into understanding the relationship between the core parameters of hyperspectral imager and the artificial intelligence algorithms used for hyperspectral classification. It serves to bridge the

Effects of Different Loss Functions on the Classification Performance of Hybridsn Networks

Classification Based On Deep Convolutional Neural Networks With Hyperspectral Image

A Hybrid 3D–2D Feature Hierarchy CNN with Focal Loss for Hyperspectral Image Classification

Superpixel-Based Long-Range Dependent Network for High-Resolution Remote-Sensing Image Classification

A Comparative Study of Deep Learning Loss Functions for Multi-Label Remote Sensing Image Classification

Effects of Loss Function Choice on One-Shot HSI Target Detection With Paired Neural Networks

Classification of Hybrid Convolution Hyperspectral Images Based on Attention Mechanism

Influence of Different Optimizers on Classification Results in Deep Learning

The Effect of Artificial Intelligence Evolving on Hyperspectral Imagery with Different Signal-to-noise Ratio, Spectral and Spatial Resolutions

Hyperspectral Images Classification Based on Multi-scale Residual Network

NeighborLoss: A Loss Function Considering Spatial Correlation for Semantic Segmentation of Remote Sensing Image

Self-Attention Network with Joint Loss for Remote Sensing Image Scene Classification

Statistical Loss and Analysis for Deep Learning in Hyperspectral Image Classification

Hybrid Dense Network with Dual Attention for Hyperspectral Image Classification

An optimal model using hybrid LCNN- GRU for efficient hyperspectral image classification

Classification of Hyperspectral Image Based on Hybrid Neural Networks

Classification of Hybrid Seeds Using Near-Infrared Hyperspectral Imaging Technology Combined with Deep Learning

A Hybrid convolution neural network for the classification of tree species using hyperspectral imagery

Minimum Noise Fraction and Long Short-Term Memory Model for Hyperspectral Imaging

Hyperspectral Classification Based on Spectral-Spatial Convolutional Neural Networks

A Novel Spectral-Spatial Multi-Scale Network for Hyperspectral Image Classification with the Res2Net Block