面向云覆盖的遥感影像时空融合深度学习方法及其应用

Method and Application of Spatial-temporal Fusion for Cloud Coverage of Satellite Images on Deep Learning

遥感影像时空融合是一种获取高时空分辨率数据的有效手段,但现有方法在选定基础数据对时要求预测时间低分辨率数据无云覆盖影响,这极大地限制了其应用潜力。为此,提出一种面向云覆盖的遥感影像时空融合方法,即在深度学习框架下,构建重建子网络恢复预测时刻云下缺失信息,将重建后的低分辨率影像与前后相邻时刻高、低分辨率数据对构建时空融合子网络,得到最终的融合影像。以安徽淮南采煤沉陷区Landsat和MODIS反射率数据为例,对预测时刻MODIS数据模拟不同缺失率的云污染;利用所提方法进行时空融合实验,进而比较深度学习与传统方法融合数据对水体信息的提取效果。结果表明:该方法融合结果各波段的RMSE和SSIM均取得较好的定量评价效果,且总体优于传统方法;沉陷区水体提取实验表明本方法水体提取结果更加接近真实观测影像。因此,该方法降低了时空融合对数据的限制要求,且具有更高的融合精度和更有效的应用性。

Spatio-temporal fusion of remote sensing images is considered as an effective way to obtain high spatio-temporal resolution data.However,the existing methods require that the low-resolution data at the predicted time is not affected by cloud cover when the basic data pairs is selected,which greatly limits the application potential of the spatio-temporal fusion method.Thus,this article proposes a spatio-temporal fusion method based cloud-covered remote sensing image.Under the deep learning framework,there are two types of remote sensing data featured by high spatial resolution but low temporal resolution(HSLT)and the other type by low spatial resolution but high temporal resolution(LSHT).The reconstruction subnetwork is constructed to repair the missing information under the cloud coveraged area of LSHT at the prediction dates,and the reconstructed LSHT image and two prior HSLT images are integrated to obtain the final fusion result on the prediction date by the constructed spatiotemporal fusion subnetwork.We take the Landsat(HSLT)and MODIS(LSHT)reflectance data in the coal mining subsidence area of Huainan City,Anhui Province as an example,simulate cloud pollution with different missing rates on the MODIS data at the prediction time,Spatial-temporal fusion experiments are conducted with the proposed method,and then compare water information extraction effects of deep learning fusion data and traditional method fusion data.The results show that the proposed method achieves a good quantitative evaluation effect on the root mean square error and the structural similarity index of the fusion results in each band,and that the fusion results are generally superior to the traditional classical method.The experiment of water extraction in subsidence area clearly shows that the water body extraction result of the proposed method is generally closer to the real observation image.Therefore,the proposed method reduces the data limitation requirements of spatio-temporal fusion,and has higher fusion accuracy and more effective application than the classic traditional method.