Research On Quality Enhancement Of Gaofen-1 Satellite Images Spatiotemporal-Spectral Fusion

Gaofen-1 satellite is the first satellite launched for China’s high-resolution Earth observation project.It carries a 2-meter panchromatic camera,an 8-meter multispectral camera,and four 16-meter wide-field multispectral cameras,realizing key technologies such as high spatial resolution,multispectral,and wide coverage optical remote sensing.However,due to limitations in sensor technology,Gaofen-1 cannot simultaneously obtain images with high temporal,spatial,and spectral resolution,which limits the value of data usage.In addition,existing quantitative remote sensing research is generally based on foreign satellites,and there is relatively little research on applications specific to Gaofen-1 satellite,requiring a lot of basic work,such as thin cloud detection,to be designed from scratch with targeted algorithms.To address these two issues,we conducted research to enhance the data quality of Gaofen-1 satellite.(1)Traditional temporal-spatial fusion on the Gaofen-1 platform only improves the 16-meter wide-field image resolution to 8 meters,which has limited significance.In order to obtain high resolution in time,space,and spectral domains,this study combines temporal-spatial and spectral fusion to simultaneously obtain new data with2-meter spatial resolution and 4-day temporal resolution.This research is a segmented temporal-spatial spectral fusion process,which is used to verify the fusion performance boundary of the Gaofen-1 satellite.(2)After verifying the possibility of temporal-spatial spectral fusion of Gaofen-1platform data through segmented fusion,this study analyzed the mechanism of feature aggregation in the fusion process,and proposed that the above-mentioned temporalspatial spectral fusion problem can be achieved through an end-to-end model.This approach can overcome the inconvenience of segmented fusion and potentially improve the fusion quality.Using convolutional neural networks,this study designed an end-to-end generation model and improved the training process through dual discriminators division in a generative adversarial network for generating target-free images.(3)Most mainstream temporal-spatial fusion algorithms are based on the LandsatMODIS platform,but Gaofen-1 has a shorter operating period and its spectral response intervals intersect with each other.To enhance the universality of Gaofen-1 data,we proposed a method to approximate the spectral style of Gaofen-1 to the more widely researched Landsat satellite series,to leverage the rich research results of Landsat series in quantitative applications.To achieve this,we proposed a cross-radiation calibration method to obtain Landsat as the target image and constructed an image translation network to achieve spectral normalization.