Hyperspectral Image Classification with Markov Random Fields and a Convolutional Neural Network (1705.00727v2)

Abstract: This paper presents a new supervised classification algorithm for remotely sensed hyperspectral image (HSI) which integrates spectral and spatial information in a unified Bayesian framework. First, we formulate the HSI classification problem from a Bayesian perspective. Then, we adopt a convolutional neural network (CNN) to learn the posterior class distributions using a patch-wise training strategy to better use the spatial information. Next, spatial information is further considered by placing a spatial smoothness prior on the labels. Finally, we iteratively update the CNN parameters using stochastic gradient decent (SGD) and update the class labels of all pixel vectors using an alpha-expansion min-cut-based algorithm. Compared with other state-of-the-art methods, the proposed classification method achieves better performance on one synthetic dataset and two benchmark HSI datasets in a number of experimental settings.

• The paper proposes a novel integration of CNNs and MRFs to enhance both spectral and spatial feature extraction in hyperspectral image classification.
• It employs a Bayesian framework with iterative optimization using stochastic gradient descent and alpha-expansion min-cut for improved spatial regularization.
• Empirical evaluations on synthetic and real-world datasets demonstrate superior classification accuracy and robust performance in remote sensing applications.

Integrating Markov Random Fields with Convolutional Neural Networks for Hyperspectral Image Classification

This paper introduces a novel approach to hyperspectral image (HSI) classification, leveraging the integration of Convolutional Neural Networks (CNNs) and Markov Random Fields (MRFs) within a Bayesian framework. HSIs are crucial in remote sensing applications, providing intricate spectral and spatial information. However, their high dimensionality and complexities demand sophisticated classification techniques that can effectively utilize both spectral and spatial data.

Innovations in HSI Classification

The proposed method combines the spectral-spatial feature extraction capabilities of CNNs with the spatial smoothness enforcement of MRFs. The authors initiate the classification task by adopting a Bayesian perspective, considering a probabilistic framework that naturally incorporates both classifiers. This integration is achieved by learning the posterior class distributions via CNNs through a patch-wise training strategy, which enhances spatial information utilization.

The innovation primarily lies in the iterative model that simultaneously updates CNN parameters via stochastic gradient descent (SGD) and the class labels of pixel vectors using the $\alpha$-expansion min-cut-based algorithm, a staple in MRF-related tasks. This iterative optimization is crucial as it allows for a dynamic adjustment of both feature representation and spatial regularization, potentially leading to superior classification performance.

Empirical Evaluation and Results

The paper presents comprehensive empirical evaluations across a synthetic dataset and two benchmark real-world HSI datasets — AVIRIS Indian Pines and ROSIS Pavia University. These datasets are known for their high spectral resolution and are often used as benchmarks due to their complexity and relevance to real-world applications.

Across various experimental setups, the proposed method consistently outperformed state-of-the-art techniques. On the synthetic dataset, the approach achieved an overall accuracy of 99.55%, a notably high result reflecting the effective integration of CNNs and MRFs. Similarly, for the real-world datasets, the method showed a marked improvement in metrics such as overall accuracy (OA), average accuracy (AA), and the Kappa coefficient, compared to standalone CNNs and other deep learning models like SS-DCNN and SPP-DCNN.

Implications for Remote Sensing

The successful combination of CNNs for deep feature extraction and MRFs for spatial regularization highlights significant advancements in HSI classification. The integration of these models suggests that a balanced approach harnessing both spectral differentiation and spatial correlation can significantly enhance classification accuracy. Such improvements aid in more accurate land-cover mapping, resource management, and environmental monitoring tasks.

Prospects and Future Research Directions

The methodology introduced in this paper also opens up pathways for future research and applications in hyperspectral imaging and beyond. The integration technique could be extended to unsupervised classification settings, potentially utilizing advanced deep generative models like Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) for better adaptability and generative capabilities.

Moreover, considering dimensionality reduction techniques could further enhance computational efficiency, especially when dealing with large-scale datasets. The exploration of different regularization regimes beyond MRFs may also present additional opportunities to tailor classification models to various hyperspectral applications effectively.

In conclusion, this paper contributes significantly to the field of hyperspectral imaging by proposing a robust technique that leverages modern machine learning advances. The innovative use of CNNs and MRFs in a unified framework addresses both feature extraction and spatial consistency, offering a comprehensive solution that can be adapted and expanded in future research.