Xcep-Dense: a novel lightweight extreme inception model for hyperspectral image classification

EXNet: (2+1)D Extreme Xception Net for Hyperspectral Image Classification

“…The Xception network proposed in this study is based on the idea that "mapping spatial and spectral correlations can be completely decoupled" which can help to reduce the number of trainable parameters [26], [55], [58]. As compared to the literature, this work proposes the use of (2+1)D convolutions instead of 3D convolutions to achieve two-fold advantages.…”

“…Nevertheless, these handcrafted features often possess limited representational capacity, resulting in suboptimal classification performance [25]. In contrast to TML, Convolutional Neural Networks (CNNs) have gained widespread use and demonstrated impressive outcomes [26]. In the realm of CNNs for HSIC, three main architectural categories emerge 1D, 2D, and 3D CNNs.…”

“…In the realm of CNNs for HSIC, three main architectural categories emerge 1D, 2D, and 3D CNNs. Numerous studies in the literature have explored the application of these CNN variants for HSIC [26]- [28]. CNNs offer several advantages, including weight-sharing capabilities and minimal data preprocessing requirements, which contribute to the acquisition of robust features for the ultimate goal of classification [29], [30].…”

“…For such reasons, CNNs might not be inherently adept at capturing fine-grained spectral variations, which are critical in HSIC, where materials can exhibit subtle spectral differences [38]. Deep CNNs can be prone to overfitting, and for such reasons, regularization techniques are often required to mitigate this issue [26], [39]- [42]. Training Deep CNNs is computationally expensive, and the time required can be prohibitive, especially for large datasets [43].…”

“…However, there are certain issues, for instance, Xception net aimed to obtain a convolutional structure that can handle translation invariance [52] and introduced sparsity using the idea of the network-in-network approach [50]. This enabled such networks to reduce the dimensions of input data, thereby, reducing the operations performed by the network while still being able to capture the non-linearity between data [26].…”