Moderate Resolution Imaging spectrometer is a kind of new weather satellite data. Few weather satellite images obtained are all clear sky and they are always influenced by cloud more or less. Cloud is a large obstacle to remote sensing image processing and analysis all the while. In order to extract objective information more effective, cloud should be removed from the remote sensing images, which is an essential sector in the image preprocessing. Cloud detection is the most important processing before removing cloud. Taking it into account that MODIS data includes thirty six bands,
Especially the infrared channels subdivided, it has realized cloud detection in MODIS images by multispectral synthesis method and cloud detection index in this paper.owing to the limitation to the certainty of the above methods, an automatic detection algorithm is applied based on the spatial texture analysis and neural network in this research. At last the cloud detection results gained by different methods are testified each other and analyzed by comparison. It found that the results are detected successfully. It not ony lays a good foundation or the cloud removing, but also can remove the precision of remote sensing image recognition, classification and inverse in this study.
Introduction
One of the most difficulty in remote sensing image processing is cloud detecting and removing. There have been many studies on cloud detecting and removing before. But most of them make use visible/infrared spectral thresholds to detect cloud according to the characteristic of cloud with high reflectance and low temperature. To enhance the difference between the cloud and the land surface, people adopt many methods to remove cloud, such as homomorphic filter, template synthesis method. Although visible or near infrared spectral threshold is simple, it is difficult to differentiate between cloud and land surface because of the similarity between them when the land surface is covered with ice, snow and sand or when cloud is thin cirrus, stratus or small cumulus. In addition, NOAA data with five bands are main data sources in the previous studies. The above advantage way has been developed or multi spectral cloud detection and proceeding with appearance of MODIS image.
Data sources
MODIS is of 36 bands within the spectral range from 0.4 to 14 micrometer, 20 visible or
near IR bands and 16 thermal infrared bands. MODIS data can be accepted twice a day, which is the real data source for resource and environment remote sensing monitoring in the regional scale.
This study selects seven bands sensitive to cloud, water vapor, temperature and aerosol,
band1(0.620-0.670micro meter),
band2(0.841-0.876),band18(0.91-0.941),band26(1.36-1.39),band29(8.40-8.70),band31(10.78-11.28),band32(11.77-12.27).In view of high reflectance of visible/infrared(near) band and low temperature of different thermal infrared band, MODIS data provide bands with strong aims for cloud detection. Especially 1.38 micrometer is the own band of MODIS, which makes it easy to recognize middle and low cloud.
Multi-spectral synthesis method.
In the study area land covers are singularity, including sand vegetation and water. Their spectrums are typical which makes it easy to recognize cloud. The brightness temperature is showed in the thermal infrared bands, but they are insensitive of thin cloud. Sometimes the temperature of thin cloud is equal to that of the land surface, while high reflectance character of thin cloud in visible band can be clearly presented. In consideration of the complementarity of cloud character between visible and thermal infrared bands, this method selects four bands, band1 (0.66micrometer), band2 (0.87), band31(11), band32(12).Cloud can be detected by means of reflectance in visible/near IR band and brightness temperature in thermal IR band.
To distinguish cloud and land surface, it is necessary to statistically analyze the eigen value of each kind of land cover and in the image in order to know the distribution character of values. The maximum, minimum mean and variance of each kind of land cover and cloud in reflection and emissive and the static range of temperature difference, RVI(ratio vegetation index) and NDVI(normalizes differenced vegetation index). It can be seen the minimum reflectance value of cloud in 0.66 micro meter is 0.30 and maximum reflectance value of land surface is 0.66micro meter is 0.29, which shows that the reflectance of cloud is higher than that of land surface. The minimum brightness temperature of land surface is higher than the maximum of cloud. Each kind of land cover is different obviously from mean and variance analysis.
N view of analysis of the spectrum of vegetation and water is different from
that of cloud, but it is difficult to distinguish sand and cloud, especially in
near IR bands. In addition, the reflectance of cloud varies with the kind,
height and thickness of the cloud. Only if all the conditions above are met with, a cloud pixel could be determined. Temperature difference of cloud between 11micro meter and 12 micro meter is lower than that of land cover. RVI of cloud is higher than that of water, but lower than that of vegetation and sand. The character of NDVI is similar to that of RVI. This method determines the thresholds by man-machine interaction according to the histogram, the static character of cloud and land surface. The thresholds of visible and thermal IR bands are P0.66>0.30 and BT12<282.5K.The reflectance character is different in 0.66micro meter.
Cloud shows high reflectance in 0.66 micro meter and vegetation shows low reflectance, while they are similar in 0.87micro meter, high reflectance. Sand shows low reflectance in 0.66micrometer and high reflectance in 0.87 micrometer. In addition water vapor is the main component of cloud, so the ratio of two bands can eliminate its influence n the transmittance. The thresholds are 0.8<P0.87/P0.66<1.05. NDVI of cloud is higher than that of water and temperature difference of cloud between 11micro meter and 12 micrometer is lower than that of water. Therefore, in consideration of NDVI and brightness temperature, the thresholds are BT11-BT12<0.30,0.02<NDVI<0.08.
Cloud detection index
Multi-spectral synthesis method is using threshold to detect cloud. Cloud shows high reflectance at red band(0.66micro meter) and the band is better to distinguish the edge between land and cloud. Cloud spectral character at near IR band(0.936micrometer) is related to the vapor water content, so it mainly shows vapor characteristics which shows it is vapor absorption band. In view of the obvious spectrum difference between the two bands (0.66micro meter and 0.936micro meter) it can not only extrude the cloud information, but also eliminate the affection of sun altitude, satellite scanning angle and atmospheric path radiance. Cloud detection index can be expressed as follows
CDI =P0.66-P0.936
P0.66-P0.936
Where, CDI is cloud detection index. P is the reflectance.
Cloud automatic detection method.
All the above methods are based on the reflect character of visible/near IR band and emissivity of thermal IR band and the cloud pixels are determined by thresholds. The thresholds are determined by man-machine interaction, which is of subjectivity and time/region limitation. The thresholds of cloud detection are related to time, region and sensor. In consideration of these methods, this paper adopts an automatic detection used on spatial texture and neural network method.
Firstly, cloud edge is determined by spatial filter analysis on the basis of static character. Secondly, cloud pixels are recognized using neural networks method combined spatial texture and spectra character, seven bands as input data, classification of a pixel as output data . Lastly, samples are selected to train neural and every pixel can be classified with the trained network I order to extract cloud pixels. Other pixels are classified to only one type. Thick loud, thin cloud and their edges can be determined by this way in order to make cloud detection more detailed, intuitive and objective. The effect is better than that with only spectrum. The light areas shown in the fig. are thin cloud and the dark areas is thick cloud. It can be seen that the boundaries between cloud and land surface is obvious.
Precision assessment
The cloud pixels are detected y multi-spectral synthesis method, but he cloud level are not divided. Thin and thick cloud is divided by near infrared method. In addition, cloud is recognized by automatic detection method, which is not dependent on the thresholds, so the detection results are more objective.
In order to asses cloud detection precision, the detection results are tested with measured data. First this paper selects the measured of grassland, sand and water to test the land surface information extracted from the image. Then fifty cloud samples are selected to test pixel by pixel, the detection precision is up to 95%. The test shows that cloud pixels account to ten percent by three methods. It proves that the method adopted in this study are feasible and the detection result is believable.
Conclusion
MODIS data are of high spectral resolution and ands with strong aims, which provide more advantaged way for cloud detection using multi-spectral technique. In view of the character of the MODIS data, cloud pixels are detected by means multi-spectral synthesis method and cloud detection index based on the spectrum analysis. The detection results are ideal, but they are dependent on the spectrum and thresholds are subjective. Therefore an automatic detection is studied on the basis of the spatial texture and neural network.
At last the detection results are tested and the precision is up to 95 percent. It proved that cloud detection results are consistent and believable after comparison analysis. The automatic detection method adapts to cloud detection in large scale, especially clouds pixels with little area and uneasy to be distinguished with naked eye. It is in favor of study in the regional scale.
