研究目的
To address the limitation of current MSE-focused multispectral pan-sharpening with machine learning by shifting the learning loss from pixel-wise error to a higher-level feature loss.
研究成果
The proposed feature-level loss function, based on SSIM and SAM, allows the pan-sharpening model to preserve more desirable spatial and spectral information. Extensive experiments on WV2 multispectral images and a building segmentation task confirm the superiority of the model over state-of-the-art algorithms.
研究不足
The study focuses on the limitations of pixel-wise MSE in pan-sharpening algorithms, which may not fully capture the spatial and spectral distortions in the images. The proposed feature-level loss function aims to address these limitations but may require further optimization for different types of satellite images.
1：Experimental Design and Method Selection:
The study employs a deep convolutional neural network based learning architecture to model the mapping function from LRMS and PAN images to HRMS images. A new feature-level loss function is designed based on spatial structural similarity (SSIM) and spectral angle mapping (SAM).
2：Sample Selection and Data Sources:
200 sets of multispectral image patches of size 320×320 from the Worldview2 (WV2) satellite are used, split into 85%/15% for training and testing.
3：List of Experimental Equipment and Materials:
NVIDIA Titan X Pascal GPU for training the network.
4：Experimental Procedures and Operational Workflow:
The network is trained using stochastic gradient descent (SGD) with a batch size of 16 to minimize the proposed feature-level loss function. The initial learning rate is set to 10^-5, and divided by 10 every 2×10^5 iterations.
5：Data Analysis Methods:
The performance is evaluated using four metrics: MSE, SSIM, SAM, and correlation coefficients (CC). Additionally, the pan-sharpening algorithms are applied in a building extraction task to assess their performance in real applications.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
