| Land surface temperature,as a direct driver of heat flux exchange between long-wave radiation and ground gas turbulence,is one of the important parameters reflecting regional and even global scale water and heat balance.Timely and effective access to regional surface temperature information is of great practical significance for climate and agricultural research.Daily thermal infrared surface temperature products of pixels under the cloud inversion accuracy is seriously affected by the weather.Microwave remote sensing has the advantages of all-day,all-weather,penetrating the soil surface,is an effective means of reconstructing the surface temperature image elements under the clouds.However,the spatial resolution of land surface temperature products obtained by passive microwave remote sensing,is low and rough.Such spatial resolution cannot meet the requirements of regional land surface temperature spatio-temporal analysis and land surface model.Therefore,downscaling of passive microwave bright temperature products is an effective means to obtain daily,all-weather and high spatial resolution surface temperatures.This paper took Gurbantunggut Desert as the study area and selects the GCOM-W AMSR-2 passive microwave bright temperature product Aqua’s MODIS product as the data source which had close transit time,focusing on the theme of passive microwave sub-cloud surface temperature reconstruction and comparative research of downscaling methods.Firstly,the correlation analysis of downscaling feature factors conducted;then,five machine algorithms(Catboost,random forest,support vector machine,Cubist,and deep belief network)and five downscaling schemes of combining passive microwave features with vegetation indices(polarization difference,full vertical polarization,full horizontal polarization,physical-based 4-channel combination,and full channel separately with vegetation indices)were compared.The optimal model was obtained.Finally,a process-oriented downscaling evaluation algorithm was built to evaluate the optimal model.The main conclusions are as follows.(1)The spatially divergent characteristics of the correlation between the daily and nighttime eigenvectors(14-channels passive microwave bright temperature and two MODIS vegetation indices)and surface temperature in the Gurbantunggut Desert are more regular,showing a high correlation in the desert and a low correlation in the oasis,and a stronger daytime divergence;the overall correlation is highest at23.8 GHz and lowest at 6.9 GHz;while salt mine cover reduces the correlation between microwave and surface temperature.Correlation.The mean correlation coefficients of each feature in the region were ranked:Daytime vertical polarization,23.8GHz(0.980)>6.9GHz(0.977)=10.7GHz(0.977)>7.3GHz(0.975)>18.7GHz(0.972)>36.5GHz(0.962)>89GHz(0.936);Daytime horizontal polarization,23.8GHz(0.96)>36.5GHz(0.95)>18.7GHz(0.94)>89GHz(0.93)>10.7GHz(0.92)>7.3GHz(0.89)>6.9GHz(0.88),EVI(0.76)>NDVI(0.74).Night vertical polarization,23.8GHz(0.957)>18.7GHz(0.943)>36.5GHz(0.932)>89GHz(0.929)>10.7GHz(0.926)>6.9GHz(0.923)=7.3GHz(0.923);Night horizontal polarization,89GHz(0.934)>23.8GHz(0.921)>36.5GHz(0.919)>18.7GHz(0.856)>10.7GHz(0.703)>7.3GHz(0.611)>6.9GHz(0.595),EVI(0.802)>NDVI(0.798).(2)In this study,the all-channel combination of Ctaboost(Catboost|VH and EVI)was determined as the downscaling model for passive microwave surface temperature downscaling study and reconstruction construction of surface temperature image elements with 1 km spatial resolution under clouds.The4-channel downscaling model constructed by 5 downscaling machine learning algorithms was evaluated,and the Catboost algorithm was determined to be superior to the other 4 algorithms;the downscaling model with 5 feature combinations based on the Catboost algorithm was evaluated,and the combination of full-channel passive microwave and enhanced vegetation index with the highest accuracy was determined to be the best combination.(3)The factor importance of the Catboost all-channel model is evaluated,the downscaling accuracy is verified by MODIS clear-sky image elements,and the correlation between the downscaling results and the multilayer temperature is verified.Based on the"all-channel"passive microwave downscaling model of the Catboost algorithm,the 10 km spatial resolution AMSR-2 passive microwave bright temperature was downscaled to obtain an all-weather 1 km spatial resolution surface temperature product for use under the MYD11A1 cloud.Surface temperature image element construction.Importance analysis was performed for each eigenfactor:daytime feature’s importance(Importance>5%)7.3GHz V(16.4)>6.9GHz V(15.4)>89GHz V(12.6)>EVI(12.4)>89GHz H(8.0)>23.8GHz V(6.1);nighttime feature’s Importance(Importance>5%)89GHz V(21.4)>6.9GHz V(13.4)>EVI(11.3)>7.3GHz V(10.2)>89GHz H(8.4)>36.5GHz V(6.3).Validation based on MYD11A1 clear sky image elements:daytime and nighttime R2 were0.987 and 0.984,RMSE were 2.82 K and 2.12 K,and MAE were 2.08 K and 1.38 K,respectively;correlation analysis based on soil temperature in the six layers of the site:all p-values were<1.2e-100,which was highly significant level;correlation of ground temperature in the daytime layers(r)5cm(0.98)>10cm(0.97)>15cm(0.97)>20cm(0.966)>40cm(0.95)>100cm(0.89),and the correlation(r)of daytime ground temperature in each layer at night 5cm(0.97)=10cm(0.97)=15cm(0.97)=20cm(0.97)>40cm(0.96)>100cm(0.96)>100cm(0.90),and the correlation r is above 0.8,which is extremely strong.The descending scale accuracy is high and can satisfy the spatial and temporal analysis of regional surface temperature.(4)A theorem about"spatial structure similarity theorem for different scales"was explored,and the"process-oriented land surface temperature downscaling evaluation"algorithm was constructed based on this theorem to evaluate the surface temperature downscaling of multiband passive microwave bright temperature.Based on the assumption of"relational scale invariance",the regularity of images with different spatial resolutions is defined as"structural similarity at different spatial scales",and the connection of images with different spatial resolutions is defined as"structural similarity of remote sensing images at different spatial scales",and the quantitative analysis method of"process-oriented scale extension evaluation"is proposed.Process-oriented downscaling one-factor evaluation:Scale coefficients of variation cSof 06GHz H,06GHz V,07GHz H,07GHz V,10GHz H,10GHz V,18GHz H,18GHz V,23GHz H,23GHz V,36GHz H,36GHz V,89GHz H,89GHz V,EVI at daytime are-1.34、-0.80、-0.66、-0.50、-1.29、-0.78、-0.92、-0.60、-0.64、-0.27、-0.22、-0.17、0.19、-0.45、-0.06,respectively and Scale information entropys Sbare 0.84、0.49、0.86、0.16、0.85、0.44、0.50、0.27、0.27、-0.06、-0.14、-0.01、0.44、-0.05、-0.25,respectively.Scale coefficients of variation ~cSof 06GHz H,06GHz V,07GHz H,07GHz V,10GHz H,10GHz V,18GHz H,18GHz V,23GHz H,23GHz V,36GHz H,36GHz V,89GHz H,89GHz V,EVI at nighttime are-0.51、-0.95、-0.47、-0.70、-0.58、-0.75、-0.85、-0.97、-0.67、-0.41、-0.43、-0.45、-0.06、-0.27、-0.22,respectively and Scale information entropys Sbare 0.15、0.58、0.15、0.33、0.23、0.38、0.46、0.57、0.35、-0.01、-0.15、-0.05、-0.25、-0.06、-0.14,respectively.Process-oriented downscaling multi-factor evaluation:Scale coefficients of variation ~cMand scale information entropy Mbat daytime are-0.5067 and 0.3058,respectively.Scale coefficients of variation ~cMand scale information entropy bMat nighttime are-0.49139 and 0.2411,respectively.Process-oriented downscaling multi-factor evaluation indicated that the nighttime scale loss is lower than the daytime,which is consistent with the MYD11A1 validation of clear sky image elements and the multi-layer ground temperature validation of the site. |
PDF Full Text Request