Sharpend Cosine Similarity based Neural Network for Hyperspectral Image Classification

TL;DR

Hyperspectral Image Classification is hindered by strong inter- and intra-class similarity and texture information, making conventional CNNs computationally intensive. The authors introduce Sharpened Cosine Similarity (SCS) as a convolution substitute that normalizes inputs, applies a sharpening exponent $p$, and uses Abs MaxPool, trained with Adam; the spectral-spatial patches are reduced via PCA to 15 bands. Key contributions include a mathematically defined SCS operator $SCS(k, x_i) = \text{sign}(k \cdot x_i) \left| \frac{k \cdot x_i}{(||k|| + q)(||x_i + q||)} \right|^{p}$ and an end-to-end pipeline that achieves competitive accuracy on Indian Pines and Salinas with only ~5.6k parameters versus ~127k for 3D/Hybrid CNNs. This results in a lightweight HSIC backbone suitable for resource-constrained environments, enabling faster training and deployment without sacrificing performance.

Abstract

Hyperspectral Image Classification (HSIC) is a difficult task due to high inter and intra-class similarity and variability, nested regions, and overlapping. 2D Convolutional Neural Networks (CNN) emerged as a viable network whereas, 3D CNNs are a better alternative due to accurate classification. However, 3D CNNs are highly computationally complex due to their volume and spectral dimensions. Moreover, down-sampling and hierarchical filtering (high frequency) i.e., texture features need to be smoothed during the forward pass which is crucial for accurate HSIC. Furthermore, CNN requires tons of tuning parameters which increases the training time. Therefore, to overcome the aforesaid issues, Sharpened Cosine Similarity (SCS) concept as an alternative to convolutions in a Neural Network for HSIC is introduced. SCS is exceptionally parameter efficient due to skipping the non-linear activation layers, normalization, and dropout after the SCS layer. Use of MaxAbsPool instead of MaxPool which selects the element with the highest magnitude of activity, even if it's negative. Experimental results on publicly available HSI datasets proved the performance of SCS as compared to the convolutions in Neural Networks.

Figures (3)

• Figure 1: Workflow of SCS pipeline.
• Figure 2: Training and Validation Loss and Accuracy of SCS network along with 3D-CNN and Hybrid CNN.
• Figure 3: Geographical maps for Indian Pines (IP) and Salinas (SA) datasets. Per-class results are presented in Table \ref{['Tab4']}.