A Hyperspectral Image Classification Framework with Spatial Pixel Pair Features.

Lingyan Ran,Wei Wei,Qilin Zhang,Yanning Zhang

doi:10.3390/s17102421

Abstract

During recent years, convolutional neural network (CNN)-based methods have been widely applied to hyperspectral image (HSI) classification by mostly mining the spectral variabilities. However, the spatial consistency in HSI is rarely discussed except as an extra convolutional channel. Very recently, the development of pixel pair features (PPF) for HSI classification offers a new way of incorporating spatial information. In this paper, we first propose an improved PPF-style feature, the spatial pixel pair feature (SPPF), that better exploits both the spatial/contextual information and spectral information. On top of the new SPPF, we further propose a flexible multi-stream CNN-based classification framework that is compatible with multiple in-stream sub-network designs. The proposed SPPF is different from the original PPF in its paring pixel selection strategy: only pixels immediately adjacent to the central one are eligible, therefore imposing stronger spatial regularization. Additionally, with off-the-shelf classification sub-network designs, the proposed multi-stream, late-fusion CNN-based framework outperforms competing ones without requiring extensive network configuration tuning. Experimental results on three publicly available datasets demonstrate the performance of the proposed SPPF-based HSI classification framework.

Highlights

Hyperspectral image (HSI) classification deals with the problem of pixel-wise labeling of the hyperspectral spectrum, which has historically been a heavily studied, but not yet perfectly solved problem in remote sensing
We give the detailed configuration description of the datasets we use and the models we build for analyzing the new spatial pixel pair feature (SPPF) and the proposed classification framework
Classification measurement, we choose the overall accuracy (OA) and average class accuracy (AA) as the evaluation strategy

Summary

Introduction

Hyperspectral image (HSI) classification deals with the problem of pixel-wise labeling of the hyperspectral spectrum, which has historically been a heavily studied, but not yet perfectly solved problem in remote sensing. With the recent development of hyperspectral remote capturing sensors, HSIs normally contain millions of pixels with hundreds of spectral wavelengths (channels). While more and more high-dimensional HSIs accumulate and are made public available, ground truth labels remain scarce, due to the immense manual efforts required to collect them. The generalization ability of neural networks is unsatisfactory if they are trained with insufficient labeled data, due to the curse of dimensionality [1]. Conventional feature extraction and classifier design was popular among HSI classification practitioners. Favuel et al [2] provide a detailed review of recent advances in this area

Results

Discussion

Conclusion