Accurate Cloud And Cloud Shadow Detection In Multispectral Remote Sensing Images

Multispectral remote sensing images,especially moderate spatial resolution images(10 to 30 meters)from Landsats 4-8(4,5,7,and 8)and Sentinel-2(2A and 2B),provide a great opportunity for a variety of remote sensing applications,such as forest disturbance,surface water dynamics,urban expansion,and agricultural practice.However,the presence of clouds and cloud shadows reduces the usability of the Landsat and Sentinel-2 imagery and their detection is often the first and important process.For decades,most of the applications were only based on a single or a few cloud free images owing to the high cost and subsequently the cloud and cloud shadow mask can be manually interpreted.Since the free and open policy of the Landsat data was implemented in 2008,the situation has been changed completely.The Sentinel-2A and Sentinel-2B(Landsat-type)were successfully launched in 2013 and 2015 respectively and the Sentinel-2 data are also free open for public.Both the Landsat and Sentinel-2 data made time series analysis even more popular,but also made the manual interpretation of clouds and cloud shadows unacceptable for a massive number of the images.Therefore,three different algorithms were developed to automatically detect clouds and cloud shadows in Landsats 4-8 and Sentinel-2 imagery.Firstly,a new algorithm called MFmask(Mountainous Fmask)was developed for automated cloud and cloud shadow detection for Landsats 4-8 images acquired in mountainous areas.The MFmask algorithm is based on the Function of mask(Fmask)3.3 version and designed specially for mountainous areas,where the Fmask algorithm presents some problems.Besides of Top Of Atmosphere(TOA)reflectance and Brightness Temperature(BT),it also requires the corresponding Digital Elevation Model(DEM)data.On one hand,compared to the original Fmask 3.3 algorithm,the MFmask is capable of separate water and land pixels more accurately by adding a slope threshold.At the same time,it linearly normalizes the BT based on the DEM data,which can reduce the temperature variations caused by elevation changes.Both of the improvements in the MFmask algorithm can generate a more accurate cloud detection result in mountainous areas.On the other hand,the MFmask algorithm also improved the cloud shadow detection in mountainous areas.It applies a double-projection approach based on DEM data to predict cloud shadow shape on slope side better.Meanwhile,the MFmask algorithm uses a topographic correction to remove terrain shadows,which is easily confused with cloud shadow,and at the same time estimates a new cloud base height with neighboring clouds.For places with large topographic gradient,both will make the determination of the cloud shadow location more accurately.To test the performance of the MFmask algorithm,we used a total of 67 Landsat images acquired in mountainous areas from different parts of the world to calculate the cloud detection accuracy,in which 15 of them have interpreted cloud shadows and were used for assessing the cloud shadow detection accuracy.Compared with the original Fmask 3.3 algorithm,the proposed MFmask algorithm can substantially improve both cloud and cloud shadow detection in mountainous areas and do not change the performance in relatively flat terrains.Secondly,a new Fmask 4.0 algorithm was developed for automated cloud and cloud shadow detection in Landsats 4-8 and Sentinel-2 imagery.We further made three major innovative improvements based on Fmask 3.3 algorithm as follows:(1)integration of auxiliary data,where Global Surface Water Occurrence(GSWO)data was used to improve the separation of land and water,and a global DEM was used to normalize thermal and cirrus bands;(2)development of new cloud probabilities,in which a Haze Optimized Transformation(HOT)-based cloud probability was designed to replace temperature probability for Sentinel-2 images,and cloud probabilities were combined and re-calibrated against a global reference dataset;and(3)utilization of spectral-contextual features,where a Spectral-Contextual Snow Index(SCSI)was created for better distinguishing snow/ice from clouds in polar regions,and a morphology-based approach was applied to reduce the false positive cloud detection errors in bright land surfaces,such as urban/built-up and mountain snow/ice.Compared to the older version of Fmask(3.3 version)and Sen2Cor(2.4 version),Fmask 4.0 achieved substantially higher overall accuracies for images from Landsats 4-8 and Sentinel-2.Thirdly,a new algorithm called TCmask(multiTemporal Cirrus mask)was developed for automated detection of thin cirrus clouds for Landsat 8 time series data.At the same time,the effect of the cirrus TOA reflectance change on the surface reflectance for Blue,Green,Red,Near Infrared(NIR),and two Shortwave Infrared(SWIR 1 and SWIR 2)bands was analyzed based on global distributions of Landsat 8 samples;and we quantitatively defined when a cirrus cloud should be detected.TCmask algorithm requires the cirrus band TOA reflectance and consists of two steps.The first step is to use a Least Absolute Shrinkage and Selection Operator(LASSO)regression method to estimate a time series model based on all observations for each pixel and initially screen most of the thin cirrus clouds.The second step is to further build a same time series model but use the remaining “non-cirrus” clouds to finally identify all thin cirrus clouds.This algorithm will not falsely identify land cover changes as clouds,as the cirrus band is not sensitive to Earth's surface.By comparing the results of TCmask with a single-date algorithm(Fmask 3.3)for different Landsat 8 images globally,significant improvements are observed for identification of thin cirrus clouds.Those newly developed cloud and cloud shadow detection algorithms can provide more accurate cloud and cloud shadow masks for remote sensing activities based on Landsats 4-8 and Sentinel-2,and also give more paths or ideas on developing cloud and cloud shadow detection algorithm for other multispectral remote sensing images.