Hyperspectral Classification Based on Lightweight 3-D-CNN With Transfer Learning (2012.03439v1)

Abstract: Recently, hyperspectral image (HSI) classification approaches based on deep learning (DL) models have been proposed and shown promising performance. However, because of very limited available training samples and massive model parameters, DL methods may suffer from overfitting. In this paper, we propose an end-to-end 3-D lightweight convolutional neural network (CNN) (abbreviated as 3-D-LWNet) for limited samples-based HSI classification. Compared with conventional 3-D-CNN models, the proposed 3-D-LWNet has a deeper network structure, less parameters, and lower computation cost, resulting in better classification performance. To further alleviate the small sample problem, we also propose two transfer learning strategies: 1) cross-sensor strategy, in which we pretrain a 3-D model in the source HSI data sets containing a greater number of labeled samples and then transfer it to the target HSI data sets and 2) cross-modal strategy, in which we pretrain a 3-D model in the 2-D RGB image data sets containing a large number of samples and then transfer it to the target HSI data sets. In contrast to previous approaches, we do not impose restrictions over the source data sets, in which they do not have to be collected by the same sensors as the target data sets. Experiments on three public HSI data sets captured by different sensors demonstrate that our model achieves competitive performance for HSI classification compared to several state-of-the-art methods

• The paper proposes 3-D-LWNet, a novel lightweight 3D-CNN architecture using depthwise separable convolutions to significantly reduce parameters and computation for hyperspectral image classification.
• Two transfer learning strategies, cross-sensor and cross-modal, are introduced to leverage abundant data from diverse sources and improve classification accuracy, especially with limited labeled HSI data.
• Experimental results demonstrate that the lightweight model and transfer learning strategies achieve competitive classification performance, particularly benefiting from pretraining on diverse datasets for practical remote sensing applications.

Hyperspectral Classification Based on Lightweight 3-D-CNN With Transfer Learning

The paper "Hyperspectral Classification Based on Lightweight 3-D-CNN With Transfer Learning" presents an innovative approach to hyperspectral image (HSI) classification employing deep learning (DL), specifically tailored to address challenges associated with limited labeled data and computational resource constraints. Proposed is a novel lightweight 3-D convolutional neural network, 3-D-LWNet, which ingeniously minimizes parameter size while achieving competitive classification accuracy. The framework integrates this model with transfer learning strategies, offering new avenues for improving HSI classification performance.

Technical Advancements

The major technical advancement in the paper is the design and implementation of a lightweight 3-D-CNN architecture, termed 3-D-LWNet. The architecture diverges from conventional 3-D-CNNs by increasing the network depth and reducing both the computational cost and the parameter count. This is accomplished through the novel Lightweight Unit (LW Unit) which utilizes depthwise separable convolutions and pointwise convolutions followed by batch normalization and activation functions. The reduction in computational complexity is particularly beneficial for scenarios with limited training samples, effectively mitigating issues such as overfitting.

Transfer Learning Strategies

Furthermore, the paper introduces two distinct transfer learning strategies to address the small sample problem prevalent in HSI datasets:

1. Cross-Sensor Strategy: Pretraining the 3-D-LWNet on a source HSI dataset with abundant labeled samples, irrespective of the sensor used for acquiring the data, and transferring to a target dataset.
2. Cross-Modal Strategy: Utilizing abundant labeled data from 2-D RGB datasets to pretrain an inflated 3-D model, which is then fine-tuned for HSI data, effectively utilizing more widely available labeled RGB data.

These transfer learning strategies are pivotal as they leverage pretrained models to improve classification accuracy, demonstrating a significant performance enhancement in experiments.

Experimental Results

The experimental validation on several HSI datasets demonstrates that the proposed methodology competes well with the state-of-the-art approaches. Notably, performance improvements are observed even when dealing with heterogeneous datasets captured by different sensors. The transfer learning strategies distinctly show that employing datasets with higher diversity and a greater number of classes for pretraining imposes substantial improvements in classification accuracy.

Implications and Future Directions

The research holds practical implications, primarily in remote sensing applications where rapid classification of complex hyperspectral data is required. The reduced computational complexity and increased classification accuracy make the proposed framework a viable solution for real-time applications. Theoretically, by eliminating constraints on source datasets, researchers can broaden the utility of pretrained models across different domains, potentially leading to advancements in cross-domain transfer learning techniques.

The exploration of enhanced CNN architecture design potentially sets the stage for future research in automated architecture optimization and intelligent model design, aiming for an optimal balance between accuracy, model size, and training efficiency.

In conclusion, this paper significantly contributes to the field of hyperspectral image classification by providing a sophisticated yet efficient model coupled with advanced transfer learning strategies, facilitating improved accuracy in contexts constrained by limited training, and sensor heterogeneity.