PaperlayerView exact revision on arXiv

Source-linked AI summary

HSI-CNN: A Novel Convolution Neural Network for Hyperspectral Image

Yanan Luo, Jie Zou, Chengfei Yao, Tao Li, Gang Bai

arXiv:1802.10478v1cs.CV

TL;DR

Limited training samples remain a bottleneck for improving HSI classification. The paper proposes HSI-CNN, which extracts spectral-spatial features, reshapes convolutional feature maps into image-like matrices, and evaluates related XGBoost and CapsNet variants. Across four datasets, HSI-CNN reports state-of-the-art accuracies of 99.28%, 99.09%, 99.42%, and 98.95%.

  • Problem

    The shortage of training samples is a bottleneck for further improving hyperspectral image classification.

  • Method

    HSI-CNN extracts spectral-spatial features from a target pixel and neighbors, stacks one-dimensional convolutional feature maps into a two-dimensional matrix, and feeds it to a standard CNN.

  • Results

    HSI-CNN reports state-of-the-art accuracies of 99.28%, 99.09%, 99.42%, and 98.95% on KSC, IP, PU, and SA, respectively.

  • Takeaways & Limitations

    The results indicate that reorganizing spectral-spatial features into image-like data is feasible for HSI classification and can achieve the paper’s best overall accuracy.

Abstract

from arXiv · show

With the development of deep learning, the performance of hyperspectral image (HSI) classification has been greatly improved in recent years. The shortage of training samples has become a bottleneck for further improvement of performance. In this paper, we propose a novel convolutional neural network framework for the characteristics of hyperspectral image data, called HSI-CNN. Firstly, the spectral-spatial feature is extracted from a target pixel and its neighbors. Then, a number of one-dimensional feature maps, obtained by convolution operation on spectral-spatial features, are stacked into a two-dimensional matrix. Finally, the two-dimensional matrix considered as an image is fed into standard CNN. This is why we call it HSI-CNN. In addition, we also implements two depth network classification models, called HSI-CNN+XGBoost and HSI-CapsNet, in order to compare the performance of our framework. Experiments show that the performance of hyperspectral image classification is improved efficiently with HSI-CNN framework. We evaluate the model's performance using four popular HSI datasets, which are the Kennedy Space Center (KSC), Indian Pines (IP), Pavia University scene (PU) and Salinas scene (SA). As far as we concerned, HSI-CNN has got the state-of-art accuracy among all methods we have known on these datasets of 99.28%, 99.09%, 99.42%, 98.95% separately.

I. INTRODUCTION

HSI classification uses rich spectral-spatial information, but limited training samples constrain further improvement. The paper proposes HSI-CNN and compares it with XGBoost and CapsNet-based variants across four HSI datasets.

  • HSI contains both spatial and spectral information and supports applications in ecological, geological, hydrological, and agricultural science.
  • Earlier HSI classification methods extracted spatial and spectral features using mathematical morphology, sparse representation, and neighborhood-based representations.
  • The paper proposes HSI-CNN as a convolutional neural network framework for hyperspectral image classification.
  • HSI-CNN is presented as a trade-off between training-sample requirements and network complexity that can overcome overfitting.
  • HSI-CNN+XGBoost substitutes XGBoost for HSI-CNN’s output layer, and experiments report that HSI-CNN is not overfitting.
  • 99.28%, 99.09%, 99.42%, and 98.95% are the reported HSI-CNN accuracies on KSC, IP, PU, and SA, respectively.

II. RELATED WORK

Related work characterizes HSI as spatial-spectral cube data and describes preprocessing strategies that incorporate neighborhood pixels around a classified target.

  • Recent HSI classification methods use strategies including mathematical morphology, sparse representation, and spectral-spatial feature extraction.
  • HSI combines spatial imaging with dozens or hundreds of narrow spectral bands for each spatial pixel, forming three-dimensional data blocks.
  • Figure 1 presents raw HSI as either three-dimensional data or cube data.
  • Using eight neighboring pixels around the central target pixel produced the best results among eight-neighbor, four-neighbor, and single-pixel strategies.

B. CNN Based Hyperspectral image Classification

Prior CNN-based HSI classifiers process spectral data directly, while cube-CNN and 3D-CNN incorporate spectral-spatial information. Capsule Networks are introduced to address CNN limitations involving spatial hierarchies and rotational invariance.

  • Hu et al.’s CNN classifies HSI directly in the spectral domain using convolution, pooling, fully connected, and output layers.
  • Figure 2 depicts the CNN structure proposed by Hu for HSI classification.
  • Spectral-only CNN processing can lose spatial information, motivating cube-CNN and 3D-CNN methods using three-dimensional kernels.
  • Capsule Networks use dynamic routing and vector outputs to address CNN limitations involving spatial hierarchies and rotational invariance.

III. HSI-CNN: A NOVEL CNN FOR HSI CLASSIFICATION

HSI-CNN uses neighborhood spectral-spatial data and a reshape layer to convert convolutional feature vectors into an image-like matrix for subsequent CNN processing. Its pipeline extracts samples, performs convolution and pooling, then trains the classifier with cross-entropy and SGD.

  • HSI-CNN comprises two convolution layers, one reshape layer, one pooling layer, and three fully connected layers, ending with category-sized softmax output.
  • HSI-CNN uses the central pixel’s 8-neighborhood cube as input, with the central pixel providing the sample label.
  • The model applies an initial 3D convolution, reshapes the resulting vectors into an image-like matrix, and then performs 2D convolution and max pooling.
  • The reshape layer stitches correlated feature vectors into a matrix so standard 2D image-classification operations can process them.
  • Training updates the network with cross-entropy loss and stochastic gradient descent until convergence.

C. HSI-CNN+XGBoost

HSI-CNN+XGBoost replaces HSI-CNN’s softmax layer with XGBoost as an alternative intended to prevent overfitting. The paper relates this design to a similar accuracy-improvement approach in handwritten-digit recognition.

  • HSI-CNN+XGBoost substitutes XGBoost for HSI-CNN’s softmax layer to prevent overfitting.
  • The authors motivate the substitution by a similar method reported to improve handwritten-digit recognition accuracy.
  • The XGBoost component is used as a classifier-layer replacement rather than as an additional preceding feature-extraction stage.

D. HSI-CapsNet

HSI-CapsNet adapts a baseline MNIST CapsNet architecture for hyperspectral image classification by retaining convolutional and fully connected layers while adding reconstruction-based regularization.

  • HSI-CapsNet is adapted from Hinton’s MNIST model for hyperspectral image classification.
  • Its architecture retains two convolutional layers and one fully connected layer, while using three fully connected layers to reconstruct inputs.
  • The reconstruction regularization uses the sum of squared differences between the final fully connected output and the input.

A. Datasets

The study evaluates HSI-CNN on four hyperspectral datasets using an 80% per-class training split and a held-out test split. The datasets span distinct sensors, locations, spatial dimensions, spectral bands, and class counts.

  • The experiments use KSC, IP, PU, and SA, with 80% of each class selected for training and the remainder used for testing.
  • KSC was acquired by NASA AVIRIS and provides 176 spectral bands across 13 classes.
  • IP contains 145 × 145 pixels, 200 spectral reflectance bands, and 16 designated classes.
  • PU contains 610 × 610 pixels, 103 spectral bands, and nine ground-truth classes.
  • SA originally has 224 bands; after discarding 20 water-absorption bands, it covers 512 lines by 217 samples and 16 classes.

B. Experiments setup

The experiments evaluate HSI-CNN and comparison models across four HSI datasets using repeated runs, fixed overall-accuracy evaluation, and specified implementation settings.

  • Experiments use an Intel Core i7 PC with 16 GB memory and an Nvidia GPU.
  • The experiments use learning rate 0.1, decay term 0.09, batch size 100, and dataset-specific Conv1 kernel counts.
  • The comparison includes HSI-CapsNet, HSI-CNN+XGBoost, and CCS.
  • HSI-CapsNet uses convolutional layers and capsule outputs, with dataset-dependent kernel and stride settings.
  • HSI-CNN+XGBoost replaces the softmax classifier with XGBoost trained on HSI-CNN FC2 features.
  • Overall accuracy evaluates each model, with normalized cubes, 10 runs per dataset and method, and fixed split ground truths.

C. Results and Analyze

HSI-CNN converges across the four datasets and achieves high classification accuracy, with balanced performance across categories and reported dataset-specific results.

  • After convergence, training accuracy approaches 100% and testing is described as fitting well.
  • The network converges after approximately iterations, with convergence speed varying by dataset size and batch count.
  • 99.52% OA is reported for PU, while 98.95% OA is reported for SA.
  • HSI-CNN reaches around 99% state-of-the-art performance across four datasets and exceeds comparison methods by 0.5%, 1.28%, 0.1%, and 3.91%.
  • Many classes exceed 99% or even 100% accuracy, while all categories exceed 94%.

V. CONCLUSIONS

The paper proposes an HSI classification model that reorganizes one-dimensional convolution results into image-like two-dimensional data, enabling deeper feature extraction. It combines HSI-CNN feature extraction with XGBoost, evaluates Capsule Networks, and reports the CNN model achieving the best overall accuracy, while identifying limited samples as the main challenge.

  • HSI-CNN model: HSI-CNN reorganizes one-dimensional convolution results into image-like two-dimensional data to deepen the network and improve feature extraction and discrimination.The model uses correlations between convolution results to splice one-dimensional data into two-dimensional data.
  • HSI-CNN model: HSI-CNN combines high-level feature extraction with XGBoost as a classifier.XGBoost is presented as a machine-learning classifier used alongside the HSI-CNN feature extractor.
  • Comparative models: The paper proposes and implements an application of hyperspectral data to a Capsule Network.This modified Capsule Network is described as an additional model in the paper.
  • Results: The CNN model achieved the best overall accuracy, supporting the feasibility of the proposed ideas.The conclusion reports this as the paper’s final performance outcome.
  • Future challenge: The biggest challenge is the lack of samples, which the paper identifies as the core issue for future work.The authors frame sample scarcity as the next major problem to solve.
Loading 1802.10478v1…