Abstract
The prevailing framework for hyperspectral image (HSI) classification has traditionally consisted of complex feature extractors followed by conventional classifiers. However, the high spatial and spectral dimensionality of HSI pixels often hinders effective classification. This paper explores the adoption of convolutional neural networks (CNNs) as an end-to-end pixelwise scheme to classify hyperspectral imagery. By treating each pixel—which contains hundreds of continuous spectral bands—as a data-rich input, CNNs can extract hierarchical features more effectively than shallow learning counterparts. Preliminary results on the University of Pavia dataset demonstrate that existing CNN models achieve promising accuracy and exhibit robust performance.
Introduction
Hyperspectral imagery provides a detailed "fingerprint" of the Earth's surface by capturing hundreds of continuous spectral bands for every pixel. While this data richness is valuable, its high dimensionality poses significant challenges for traditional classification algorithms. Historically, researchers relied on hand-crafted feature engineering paired with standard classifiers. These "shallow" models often fail to capture the abstract, hierarchical features necessary for robust land-cover classification. This study bridges the gap by applying deep learning—specifically CNNs—to HSI, treating the task as an end-to-end learning problem.
Methodology
The authors approach HSI classification as a pixelwise spectral task. To leverage standard 2D CNN architectures, they employ a preprocessing technique where the 103-dimensional spectral vector of a single pixel is transformed into a 103x103 image through simple duplication. This allows the spatial filters of a CNN to extract features from spectral signatures.
The study evaluates several architectures:
- Adapted Models: Modified versions of LeNet and AlexNet.
- Custom "Conv" Models: Three architectures (Conv_1, Conv_2, and Conv_3) designed to investigate the impact of depth on spectral feature extraction.
Key technical implementations include the use of Rectified Linear Units (ReLU) for activation and Nesterov’s Accelerated Gradient (NAG) for optimization, which proved faster than standard Stochastic Gradient Descent (SGD) in the Caffe framework.
Results and Performance
Experiments were conducted on the University of Pavia (UPavia) dataset, featuring 103 spectral bands and 9 urban land-cover classes. The models demonstrated remarkable efficiency and accuracy.
Key findings include:
- Optimal Depth: Performance peaked with two convolutional layers (Conv_2, 87.54% accuracy). Adding a third layer (Conv_3) led to performance deterioration, suggesting a "Goldilocks" zone for depth when processing purely spectral data.
- LeNet vs. AlexNet: LeNet achieved a high Overall Accuracy (OV) of 93.11%. While slightly less accurate in some runs, AlexNet was noted for its robustness across different random samplings due to its larger scale.
- Efficiency: The models reached high accuracy in just 10 minutes of training over 30 epochs, as shown in the convergence plots.
Discussion and Future Work
The results confirm that hierarchical feature extraction via CNNs is superior to simple linear filtering for HSI. However, the current pixelwise approach ignores spatial-contextual information. Future work will focus on integrating spatial-spectral data—considering a pixel's neighbors—and utilizing regularization techniques like Conditional Random Fields (CRFs) to further smooth classification maps and improve spatial consistency.
Summary updated