热红外与多光谱遥感图像的神经网络回归融合方法研究

为验证神经网络方法用于遥感图像融合的有效性，归纳了利用神经网络对遥感数据进行回归来实现融合的3种途径，并提出了一种结合图像数据回归和多光谱遥感图像锐化技术来实现热红外图像的全色锐化新方法。这种热红外图像的全色锐化方法，利用了极限学习机（ELM）这种新型神经网络算法，快速高效地由训练样本得到遥感图像数据间的回归关系；同时，方法注重图像数据本身的物理含义，以提高热红外图像数据的真实质量为目标，是一种定量化的图像融合方法。经这种方法融合得到的热红外数据也能很好地用于定量遥感的物理模型，为遥感的实际应用提供方便。该方法的有效性通过对ETM+图像进行实验得到了证明，而直接对热红外图像数据和全色图像数据进行回归的融合模式，在实验中则无法得到满意的结果。

Fusion of Thermal Infrared and Multispectral Remote Sensing Images via Neural Network Regression

The paper summarizes three modes for the use of neural network regression to fuse remote sensing images, while proposing a new pansharpening method, based on neural network regression, to fuse thermal infrared(TIR) image and the panchromatic(Pan) image, which can hardly be done using traditional image fusion techniques. Extreme learning machine algorithm is applied to obtain the regression relationship between remote sensing data, in a rapid and efficient manner, while the pansharpening for TIR focus on the internal physical relations of pixel values recorded as an image, and aiming at a real improvement of the TIR data quality rather than a visual enhancement. TIR data synthesized by this new image fusion method is qualified to be used in physical models. This provides convenience for quantitative remote sensing applications. Experiments on ETM+ images prove the effectiveness of this approach which achieves fairly accurate results, while direct fusing mode achieves dissatisfactory results.