High Spatial And Temporal Fusion Of Multi Source LSTs Based On Dictionary Learning

Recently,remote sensing images obtained from satellite sensors are unable to meet the criterion of high temporal resolution and high spatial resolution owing to the physic design of sensors.How to acquire accurate and diverse spatial and temporal scales sensing images with extensive area coverage becomes a popular research topic with the widespread application of remote sensing.This paper explores the fusion of land surface temperature(LST)with high temporal and spatial resolution in Beijing and Huailai County of Hebei Province.First,analyzing the disadvantages of previous spatial-temporal fusion method:Spatial and Temporal Adaptive Reflectance Model(STARFM)and the Spatial-temporal Integrated Temporal Fusion(STITFM).Then,a improved high temporal and spatial resolution surface temperature fusion method based on dictionary learning for multi-source remote sensing data is proposed,and the fusion results are evaluated by in situ data and real Landsat 8 LST.The contents and research results are mainly as follows:1.The performance of STITFM and STARFM were tested by conducting two experiments in Huailai County,Hebei Province.For EXP1,Landsat 8,MODIS,and FY-2E/F LSTs were used as the input data for STITFM,while for EXP2,only Landsat 8 and FY-2E/F LSTs were used as the input data for STARFM.The performance of the two methods were evaluated with observed Landsat 8 LSTs respectively,and results showed that prediction error for STITFM(R=0.812,RMSE=1.576 K,MAE=1.156 K)is lower than STARFM(R=0.447,RMSE=14.098 K,MAE=13.786 K).2.A nonlinear spatial-temporal fusion framework for multi-source sensor data based on the dictionary learning was proposed in this study.By sampling the high and low resolution image pairs,a higher level of abstract features were extracted to express the change of different remote sensing data.The accuracy of the fusing LSTs with both high temporal and spatial resolution were evaluated by the in situ LST,the results showed that the performance of proposed method based on dictionary decomposition(R=0.993,RMSE=1.032 K,MAE=0.703 K)has a certain improvement compared with STITFM method(R=0.988R,MSE=1.47 K,MAE=1.235 K).3.LSTs with both high temporal and spatial resolution of Beijing were obtained by utilizing improved spatial-temporal fusion framework,and the diurnal variation characteristics of urban heat island in Beijing are analyzed.The results showed that there are obvious urban heat island effects in Beijing,especially in spring,summer and autumn.And a diurnal variation tendency of surface urban heat island intensity(SUHII)was observed:the SUHII starts to decline after sunrise,and is reduced to the lowest around noon,then continues at a low level,after sunset it gradually begins to rise,and finally continues to appear strong heat island during night.The surface temperature of five typical land cover types in Beijing is consistent with the trend of the daytime and the night in the spring,summer and autumn.From high to low,the order is:construction land>farmland>grassland>water body>coniferous forest,and the night are:construction land>water>farmland>grass>deciduous broad-leaved forest.NDVI and SUHII were found to be negatively correlated.Based on the analysis of the urban heat island image factor,it was found that the normalized difference index(NDVI)and UHI were negatively correlated.The relationship between Normalized Difference Water Index(NDWI)and UHI was changed with the seasons,and in addition to winter,basically are negative correlation.This study contributes to the existing spatial-temporal fusion studies by providing a nonlinear spatial-temporal fusion method for multi-source sensor,which are capable of obtaining both high spatial resolution and frequent coverage from multi source data.Such a capability will be beneficial for obtaining long time series data,monitoring quantitative analysis in fine scale and generating long-time series data in large district.