Study Of Thin Cloud Removal Method For Multi-spectral Remote Sensing Images

The space-borne remote sensing technique plays an important role in geoscience research and application.However,data acquired by an optical remote sensing sensor are easily affected by clouds in the atmosphere.The existence of clouds adversely affects the quality and usage of the optical remotely sensed data.Therefore,how to remove or reduce the influence of clouds is an on-going research theme.In this thesis study,the thin cloud removal method for multispectral remote sensing images is investigated.Firstly,an introduction and a summary of the empirical and radiative transfer model-based(RTM-based)algorithm are given.The algorithm uses the statistical linear relationships of spectral response between bands and the radiative transfer model to model the cloud to achieve cloud removal.However,the algorithm performs not well when complicated ground features exist.Then,a revised algorithm is sought.The improvement takes advantage of the available cirrus band of Landsat-8 to improve the accuracy level of the modeled thin cloud reflectance.Two experiments with different land use and land cover(LULC)types of Landsat-8 OLI datasets were conducted to validate the improved algorithm.Results from the improved algorithm were better than those using the original algorithm visually.Quantitatively and band-by-band,the spatial correlation coefficients of the reference image and the image after the improved algorithm were between 0.919 and 0.956,which was better than the coefficients of0.717-0.912 from the RTM-based algorithm.Since the improved algorithm,as well as the RTM-based algorithm,is only able to remove thin clouds in the visible and near-infrared spectral region,a novel cloud removal approach to remove the thin clouds from the visible to mid-infrared bands is developed with the slow feature analysis(SFA).The algorithm takes the spectral bands and cirrus band as the input into the SFA and outputs the separated components from the slowest to the fastest.The slowest or first component is considered the"cloud component".Then,the calculation of the cloud in each spectral band using the"cloud component"is done so that the removal of the cloud can then be achieved.The SFA-based algorithm has been validated with simulated and acquired clouded Landsat-8datasets.R~2 values between the reference image and the image after the cloud removal algorithm were 0.9 or higher.RMSE(root-mean-squared error)values were close to 0.Finally,to investigate the applicability and robustness of the SFA-based algorithm,we applied the algorithm to Landsat-8 datasets of locations with various LULC types,and Sentinel-2A data.Satisfactory results were obtained.