CN104966298A - Night low-cloud heavy-mist monitoring method based on low-light cloud image data - Google Patents

Abstract

Provided is a night low-cloud heavy-mist monitoring method based on low-light cloud image data. The method is characterized by, selecting night visible light and near infrared and far infrared waveband satellite remote sensing data to carry out default value and image matching pretreatment; by utilizing the characteristic of large difference of cloud and mist albedo data and underlying surface albedo data in the visible light waveband and by utilizing an OTSU method, determining the optimum image segmentation value automatically and separating the cloud and mist and an underlying surface; by utilizing the difference of cloud and mist reflection moonlight energy data received by a low-light sensor in the night visible light waveband and radiation energy data of light sources of city lamplight and the like, drawing a grey level histogram of a target area, and manually selecting an appropriate gray level threshold value according to the distribution information of the histogram to separate cloud and mist and city lamplight light sources; and by utilizing the difference of ice and snow albedo data and the cloud and mist albedo data in near-infrared waveband, comprehensively utilizing normalized difference snow index (NDSI) and normalized difference vegetation index (NDVI) to separate the cloud and mist and ice and snow.

Description

A kind of low clouds Monitoring Fog method at night based on low-light cloud atlas data

Technical field

The invention belongs to remote sensing technology field, is a kind of low clouds Monitoring Fog method at night fully utilizing multiband satellite remote sensing date under poor light condition.

Background technology

Satellite Remote Sensing has wide coverage, and temporal resolution is high, and information source is reliable and cost is lower waits many advantages, and utilizing Value of Remote Sensing Data to carry out night low clouds Monitoring Fog is effective method and technological approaches.Current method mainly contains Box-counting technique, based on gray scale connected domain weighted score dimension method, two infrared brightness temperature approach method and two waveband threshold method.Wherein, first two method mainly utilizes the difference of the fractal characteristic of cloud and mist on texture, the low clouds dense fog region that texture is relatively uniform is identified from infrared remote sensing image, it can distinguish the background such as cloud and mist and underlying surface preferably, but because high cloud in some is similar with low clouds dense fog on texture structure, often mixing in its recognition result has high cloud in part.Two infrared brightness temperature approach method is low clouds Monitoring Fog method at current most widely used night.The method utilizes low clouds dense fog different in the bright temperature difference of middle-infrared band (3.7 μm) and far infrared band (10.8 μm), chooses suitable bright temperature difference threshold value (being approximately 2K) and identifies low clouds dense fog region.But only use the data of two infrared bands due to the method, there is a lot of limitation in its identification for low clouds dense fog at night, the situation that cloud and mist layer thickness is less than 100m is not suitable for as the method, on two infrared band, the soil of some type is comparatively similar to low clouds dense fog in without cloud and mist situation, stratiform clouds on the middle and senior level such as also usually to obscure mutually at [the Ellrod G P with low clouds dense fog, Advances in the detection and analysis of fog at night usinggoes multispectral infrared imagery.Weather and Forecasting, 1995, 10:606-619.].Two waveband threshold method is attempted in conjunction with visible light wave range and far infrared wave segment data at night first, manually selected threshold realizes low clouds Monitoring Fog at night, achieve certain effect [Zhou little Ke, Yan Wei, Bai Heng etc., based on the low clouds Monitoring Fog technical research at night of DMSP/OLS data. sensor information, 2012,27 (6): 86-90.].But because the method uses wave band less, it is difficult to low clouds dense fog and bright earth's surface (as the light sources such as urban lighting, snow and ice cover region) to distinguish.At present, under comprehensive utilization poor light condition, the low clouds Monitoring Fog method at night of multiband satellite remote sensing date have not been reported.Therefore, need to find a kind of low clouds Monitoring Fog method at night based on low-light cloud atlas data.

Summary of the invention

The object of the invention is to, provide a kind of low clouds Monitoring Fog method at night based on low-light cloud atlas data, comprehensive utilization multiband satellite remote sensing date is monitored low clouds dense fog at night.

To achieve these goals, the present invention takes following technical scheme: a kind of low clouds Monitoring Fog method at night based on low-light cloud atlas data, is characterized in that, said method comprising the steps of:

Step 1, choose visible ray at night, near infrared and far infrared band satellite remote sensing date respectively, carry out the pre-service of default value and images match;

Step 2, the feature utilizing cloud and mist albedo data and albedo of underlying surface data (earth's surface, water body, vegetation etc.) to differ greatly in visible light wave range albedo, adopt maximum variance between clusters automatically to determine optimized image segmentation threshold, the underlying surfaces such as cloud and mist and earth's surface, water body and vegetation are separated;

Step 3, utilize low-light level sensor in the difference of light source self emittance such as cloud and mist that visible light wave range receives reflection at night moonlight energy datum and urban lighting etc., make the grey level histogram of target area, manually choose suitable gray threshold according to histogrammic distribution situation, cloud and mist is separated with light sources such as urban lightings;

Step 4, the albedo difference utilizing Snow and Ice Albedo data and cloud and mist albedo data to show at near-infrared band (especially 1.61 μm near), comprehensively adopt Normalized difference snow index (NDSI) to be separated with ice and snow by cloud and mist with normalized differential vegetation index (NDVI);

Step 5, the bright temperature difference utilizing low clouds dense fog, underlying surface and middle high cloud to show at far infrared band are different, make the bright temperature histogram of target area, choose bright temperature threshold value according to histogrammic distribution situation, low clouds dense fog, underlying surface are separated with middle high cloud;

The common region of step 6, combining step 2,3,4,5 acquired results, obtains the related data of cloud and mist, namely preliminary low clouds Monitoring Fog result;

Step 7, the characteristic of continuous uniform that should have according to low clouds dense fog, homogeneity detection is successively carried out, cancelling noise in every 3 × 3 pixel regions in the preliminary monitoring result obtain step 5, and then obtains final low clouds Monitoring Fog result.

Formula for the maximum variance between clusters determining Iamge Segmentation optimal threshold in step 2 is as follows:

Suppose that image size is M × N, gray level is L, and in image, gray scale is the pixel count of i (0≤i≤L-1) is N _i, then the probability of gray scale i is optimal threshold is:

G=arg max _0≤g≤L-1[P _a(w _a-w ₀) ²+ P _b(w _b-w ₀) ²] g is gray-scale value (0≤g≤L-1) in formula, all the other parameters are defined as follows:

$P_a={\Σ}_{i=0}^g{P}_i,P_b={\Σ}_{i=g+1}^{L-1}{P}_i,w_a={\Σ}_{i=0}^{g}i\frac{P_i}{P_a}$

$w_b={\Σ}_{i=g+1}^{L-1}i\frac{P_i}{P_b},w_0={\Σ}_{i=0}^{L-1}{\mathrm{iP}}_i$

In step 4, the formula of Normalized difference snow index NDSI, normalized differential vegetation index NDVI is as follows:

$NDSI=\frac{R_{I1}-R_{I3}}{R_{I1}+R_{I3}}$

$NDVI=\frac{R_{I2}-R_{I1}}{R_{I2}+R_{I1}}$

R in formula _i1, R _i2, R _i3represent the albedo of object at I1 (0.64 μm), I2 (0.865 μm), I3 (1.61 μm) wave band respectively.

The formula calculating 3 × 3 pixel area uniformity in step 5 is as follows:

$SH=\frac{\mathrm{\μ}}{3\mathrm{\σ}}$

In formula, SH represents the homogeneity size of target area, and μ, σ represent average gray and the standard deviation of target area respectively.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2,3 are the present invention to night low clouds dense fog monitoring result and checking situation.

Wherein: Fig. 2 is algorithm monitors result and checking situation (i.e. embodiment 1), in Fig. 2 (a, b, c) DNB visible cloud image, I5 infrared cloud image, low clouds Monitoring Fog result is respectively; (d, e) visibility measured result, cloud base height measured result is respectively in Fig. 2; Fig. 3 algorithm monitors result and checking situation (i.e. embodiment 2), a, b in Fig. 3, c) be respectively DNB low-light cloud atlas, I5 infrared cloud image, low clouds Monitoring Fog result; In Fig. 3, (d), (e) are respectively visibility measured result, cloud base height measured result.

Embodiment

The present invention fully utilizes multiband satellite data under poor light condition, proposes the multiband threshold method being applicable to low clouds Monitoring Fog at night.The method obtains the preliminary monitoring result of low clouds dense fog by implementing a series of detachment process, carries out homogeneity detection on this basis, finally extracts the low clouds dense fog region in low-light cloud atlas.Composition graphs 1, the method is specially:

1 data prediction

Choose visible ray at night, near infrared and far infrared band satellite remote sensing date respectively, carry out the data prediction of default value and images match.Wherein, go default value process to refer to the Banded improvement of rejecting in remote sensing images, conventional method has: overall line detector matching method, non linear probe matching method, statistics amendment method and spatial filter mask method.Images match process refers to mates the Pixel domain position of different-waveband remote sensing images, and conventional method has nearest neighbor method, bilinear interpolation method, cubic convolution interpolation method.

2 clouds and mists are separated with earth's surface, vegetation, water body

At visible light wave range, the emittance that satellite received at night mainly with albedo and the moon phase angle relevant (except urban lighting etc. self radiating light source) of object.According to Mie-scattering lidar, the scattering process of cloud and mist is obvious, its albedo will apparently higher than earth's surface, vegetation, the underlying surface (except ice and snow) such as water body.Therefore, on night visible cloud image, the brightness of cloud and mist is more bright compared to the underlying surface of these types.For visible cloud image at night, by the method setting gray threshold, cloud and mist and earth's surface, vegetation, water body are separated.And for the low-light cloud atlas of different time, different geographical, obviously can not use fixing threshold value, therefore the determination of dynamic threshold is the key point of this detachment process.Otsu [Otsu N, A threshold selection method from gray-level histograms.Automatica, 1975,11 (285-296): 23-27.] a kind of adaptive image segmentation threshold defining method is proposed, i.e. maximum variance between clusters.Image to be split is regarded as and is made up of background and target two class by this algorithm, weighs the difference between target and background with inter-class variance, makes the maximum gray-scale value of the inter-class variance of target and background namely think optimal threshold.Suppose that low-light cloud atlas pixel number is M × N, gray level is L, and in image, gray-scale value is the pixel count of i (0≤i≤L-1) is N _i, then the accumulative probability of occurrence of gray-scale value i is then optimal threshold G can be obtained by following formulae discovery:

G＝arg max _0≤g≤L-1[P _a(w _a-w ₀) ²+P _b(w _b-w ₀) ²]

In formula, g is gray-scale value (0≤g≤L-1), and all the other parameters are defined as follows:

$P_a={\Σ}_{i=0}^g{P}_i,P_b={\Σ}_{i=g+1}^{L-1}{P}_i,w_a={\Σ}_{i=0}^{g}i\frac{P_i}{P_a}$

$w_b={\Σ}_{i=g+1}^{L-1}i\frac{P_i}{P_b},w_0={\Σ}_{i=0}^{L-1}{\mathrm{iP}}_i$

3 clouds and mists are separated with light sources such as urban lightings

At night, the low-light level sensor that satellite carries not only can detect the characteristic body of the reflection moonlight such as cloud and mist, also can detect the light sources such as urban lighting, fire, fishing boat.In order to get rid of the interference of these light sources to low clouds Monitoring Fog result, need to reject light source from low-light cloud atlas.The characteristic body such as cloud and mist reflection at night moonlight radiation scope is roughly 1.0 × 10 ^-12wcm ^-2sr ^-1~ 1.0 × 10 ^-8wcm ^-2sr ^-1, and typical night lights radiation scope is roughly 1.0 × 10 ^-9wcm ^-2sr ^-1~ 3.0 × 10 ^-7wcm ^-2sr ^-1.Therefore, on low-light cloud atlas, the light sources such as urban lighting are brighter than characteristic bodies such as clouds and mists when crescent (be especially close this month light and shade difference more obvious) generally.By adding up the grey level histogram of target area, manually choosing suitable gray threshold according to histogrammic distribution situation, the light sources such as cloud and mist and urban lighting can be separated.Manually choose suitable gray threshold according to histogrammic distribution situation, cloud and mist can be separated with light sources such as urban lightings.

4 clouds and mists are separated with ice and snow

By at near-infrared band (1.0 ~ 3.0 μm), because the scattering particle in cloud and mist is less, its energy absorbed is less, and ice and snow is relatively more obvious in the absorption of this wave band.Therefore at near-infrared band, especially near 1.61 mu m wavebands, cloud and mist albedo still keeps higher, and the albedo of ice and snow then has and significantly declines.

The albedo difference utilizing ice and snow and cloud and mist to show at visible ray and near-infrared band, Normalized difference snow index (NDSI) can be adopted to be separated cloud and mist and ice and snow, and Normalized difference snow index is defined as:

$NDSI=\frac{R_{I1}-R_{I3}}{R_{I1}+R_{I3}}$

R in formula _i1represent the albedo at I1 wave band (0.64 μm), represent the albedo at I3 wave band (1.61 μm).Generally the NDSI value of pure ice and snow is very high, if pixel meets NDSI>=0.4 and R _i2>=0.11, so this pixel is just judged as ice and snow.But when there being other mixed characteristics in ice and snow pixel, its NDSI value can reduce to some extent.In forest zone, many ice and snow pixels are due to the covering of forest, and its NDSI value is less than 0.4, can be identified as and not have ice and snow.In order to eliminate this erroneous judgement, need NDSI and normalized differential vegetation index (NDVI) to combine use.Normalized differential vegetation index is defined as:

$NDVI=\frac{R_{I2}-R_{I1}}{R_{I2}+R_{I1}}$

Wherein R _i2represent the albedo at I2 wave band (0.865 μm).Meet the prerequisite of 0.1 < NDSI < 0.4 at pixel under, if its NDVI value meets a certain threshold range, then this pixel should be identified as the mixed pixel of the impurity such as ice and snow and vegetation.According to [Klein A G such as Klein, Hall D K, and Riggs G, Improving snow-covermapping in forests through the use of a canopy reflectance model [J] .HydrologicalProcesses, 1998,12 (10-11): 1723-1744.] research, this threshold value can be expressed as the function of pixel NDSI, and relational expression is as follows:

NDVI_1＝a ₁+a ₂*NDSI

NDVI_2＝b ₁+b ₂*NDSI+b ₃*NDSI ²+b ₄*NDSI ³

Coefficient a in formula ₁, a ₂, b ₁, b ₂, b ₃be adjustable parameter, NDVI_1, NDVI_2 are respectively the upper and lower limit threshold value of NDVI.Design parameter chooses the scheme adopting Klein etc.

5 low clouds dense fogs are separated with middle high cloud

At night, at long wave infrared region (8 ~ 14um), the emitted radiation of underlying surface and cloud and mist self is main source radiation, and satellite is main relevant with the bright temperature of object in the emittance of this band reception.Compared to low clouds dense fog and underlying surface (deicing is outer), the height of middle high cloud is higher, its Black body temperature comparatively low clouds dense fog and underlying surface low.And low clouds dense fog is close to earth's surface, generally its bright temperature is slightly lower than underlying surface, but apparently higher than middle high cloud.The bright temperature difference utilizing low clouds dense fog, underlying surface and middle high cloud to show in long wave infrared region is different, low clouds dense fog, underlying surface can be separated with middle high cloud by choosing suitable bright temperature threshold value.Due to the bright temperature of object in time, the difference of region and changing, in infrared cloud image, the determination of bright temperature threshold value also should be a dynamic process.By adding up the bright temperature distribution situation (making bright temperature histogram) in a certain concrete moment infrared cloud image, manually choosing suitable bright temperature threshold value, low clouds dense fog can be separated with middle high cloud.

6 merging draw preliminary monitoring result

The separating resulting obtained in 2,3,4,5 is merged, the preliminary monitoring result of low clouds dense fog can be obtained.

7 homogeneitys detect

Low clouds dense fog has the feature of continuous uniform.And in actual separation process, the preliminary monitoring result obtained is often doped with isolated noise.In order to reject these noises, needing that preliminary monitoring result is obtained to separation and carrying out homogeneity detection.For the doubtful low clouds dense fog of each in preliminary monitoring result pixel, its gray scale is put 1 (background gray scale sets to 0), calculate the homogeneity in 3 × 3 pixel regions centered by it, formula is as follows:

$SH=\frac{\mathrm{\&mu;}}{3\mathrm{\&sigma;}}$

In formula, SH represents the homogeneity size of target area, and μ, σ represent average gray and the standard deviation of target area respectively.For all doubtful low clouds dense fog pixel in target area, if its SH<0.22 (namely also having a low clouds dense fog pixel at most in 3 × 3 pixel regions near this doubtful low clouds dense fog pixel), so this pixel is judged as noise (not detected by homogeneity).Otherwise this pixel is judged as low clouds dense fog pixel.

Embodiment:

NPP (National Polar-orbiting Parternership) satellite launched on October 28th, 2011, the AVHRR of VIIRS (Visible Infrared Imaging Radiometer Suite) the sensor collection NOAA that it carries, the MODIS of EOS and the OLS feature of DMSP DMSP are, it is the Scanning Imaging Radiometer onboard that comprehensive survey ability is the strongest in the world at present, comprise 22 wave bands within the scope of 0.4 μm ~ 12 μm altogether, swath width is 3000km, I 1-I5 passage horizontal resolution is 375m, M1-M16 passage horizontal resolution is 750m, DNB (Day NightBand) passage full resolution is 742m.Wherein DNB wave band adopt three gains arrange, there is accurate radiation calibration and higher spatial and temporal resolution, can by day, morning and evening even night realize to earth observation, be one of main characteristics of this detector.

We adopt two the low clouds dense fog Typical Cases occurring in regional for 2012 to carry out check analysis.Wherein, satellite remote sensing date chooses the NPP/VIIRS data (respectively corresponding instance 1 and example 2) in two moment of UTC2012/08/30/18:24 and UTC2012/12/02/19:04.Concrete wave band chooses situation as table 1.

Table 1 VIIRS band selection

And for the measured data of surface weather observation website, due to by every 3 hours one time integral point record, according to the immediate principle of observation time, choose the data in two moment of UTC2012/08/30/18:00 and UTC 2012/12/02/18:00 respectively.In surface weather observation station data using visibility, relative humidity and cloud base the high discrimination standard as mist and low clouds.Visibility is less than 10km and the website that relative humidity is more than or equal to 95% has been judged to be mist, and cloud base higher primary school has been judged to be low clouds in the website of 2500m.

For the remotely-sensed data chosen, utilize algorithm proposed by the invention progressively to process, obtain the final monitoring result of low clouds dense fog.Surface weather observation website then utilizes visibility, relative humidity and cloud base height aggregation of data to judge whether this website has low clouds dense fog.Algorithm monitors result and checking situation are as shown in Figure 2,3.

In Fig. 2,3 (c), white portion is the low clouds dense fog that Satellite Remote Sensing arrives; Fig. 2,3 (d, e) represents the measured result of Ground Meteorological website.Wherein, Fig. 2 (c) be shown as Northeast China, North China one with and adjacent Korea, Korea S remote sensing monitoring result.Use alphabetical A (Liaoning Middle), B (near border) in figure respectively, C (Eastern Shandong) marked the low clouds dense fog region monitored, corresponding same area is verified in ground actual measurement fog-zone Fig. 2 (d).As can be seen from remote sensing monitoring result Fig. 3 (c), SOUTHERN CHINA has large-scale low clouds dense fog region, mainly concentrates on the ground such as A (STRUCTURES IN EAST SICHUAN), B (Guizhou), C (Yunnan), D (Guangxi).Fig. 3 (d) shows STRUCTURES IN EAST SICHUAN on the same day, Guangxi and really there is large-area fog-zone, Fig. 3 (e) show Guizhou, Yunnan then mainly low clouds cover, ground measured result and remote sensing monitoring result substantially identical.

Find out from the result of above two examples, the low clouds dense fog region that the present invention monitors is roughly the same with ground measured result, and algorithm effect is better.But for multi layer cloud region (in Fig. 2,3 pentagram mark position), this algorithm there is no the low clouds dense fog that method monitors bottom.

Claims (4)

1., based on a low clouds Monitoring Fog method at night for low-light cloud atlas data, it is characterized in that, said method comprising the steps of:

Step 1, choose visible ray at night, near infrared and far infrared band satellite remote sensing date respectively, carry out the pre-service of default value and images match;

Step 2, utilize cloud and mist albedo data and albedo of underlying surface data, in the feature that visible light wave range differs greatly, adopt maximum variance between clusters automatically to determine optimized image segmentation threshold, the underlying surfaces such as cloud and mist and earth's surface, water body and vegetation are separated;

Step 3, utilize low-light level sensor in the difference of light source self emittance data such as cloud and mist that visible light wave range receives reflection at night moonlight energy datum and urban lighting etc., make the grey level histogram of target area, manually choose suitable gray threshold according to histogrammic distribution situation, cloud and mist is separated with light sources such as urban lightings;

Step 4, utilize Snow and Ice Albedo data and cloud and mist albedo data at near-infrared band, the difference especially shown near 1.61 μm, comprehensively adopts Normalized difference snow index (NDSI) to be separated with ice and snow by cloud and mist with normalized differential vegetation index (NDVI);

Step 5, the bright temperature difference utilizing low clouds dense fog, underlying surface and middle high cloud to show at far infrared band are different, make the bright temperature histogram of target area, manually choose suitable bright temperature threshold value according to histogrammic distribution situation, low clouds dense fog, underlying surface are separated with middle high cloud;

The common region of step 6, combining step 2,3,4,5 acquired results, obtains the related data of low clouds dense fog, namely preliminary low clouds Monitoring Fog result;

Step 7, the characteristic of continuous uniform that should have according to low clouds dense fog, homogeneity detection is successively carried out, cancelling noise in every 3 × 3 pixel regions in the preliminary monitoring result obtain step 5, and then obtains final low clouds Monitoring Fog result.

2. a kind of low clouds Monitoring Fog method at night based on low-light cloud atlas data according to claim 1, it is characterized in that, the formula for the maximum variance between clusters determining Iamge Segmentation optimal threshold in step 2 is as follows:

Suppose that image size is M × N, gray level is L, and in image, gray scale is i(0≤i≤L-1) pixel count be N _i, then the probability of gray scale i is optimal threshold is:

G＝arg max _0≤g≤L-1[P _a(w _a-w ₀) ²+P _b(w _b-w ₀) ²]

In formula, g is gray-scale value (0≤g≤L-1), and all the other parameters are defined as follows:

。

3. a kind of low clouds Monitoring Fog method at night based on low-light cloud atlas data according to claim 1, it is characterized in that, in step 3, the formula of Normalized difference snow index NDSI, normalized differential vegetation index NDVI is as follows:

R in formula _i1, R _i2, R _i3represent the albedo of object at I1 (0.64 μm), I2 (0.865 μm), I3 (1.61 μm) wave band respectively.

4. a kind of low clouds Monitoring Fog method at night based on low-light cloud atlas data according to claim 1, is characterized in that, the formula calculating 3 × 3 pixel area uniformity in step 5 is as follows:

In formula, SH represents the homogeneity size of target area, and μ, σ represent average gray and the standard deviation of target area respectively.