Evaluation of SADFAT model performance in daily forecast of Land Surface Temperature in the city of Tehran



Abstract

The high spatial and temporal limitations of TIR images for use in urban climatology have been identified as a current scientific challenge. Therefore, the use of Data Fusion Algorithms in Remote Sensing has been considered. In the old methods, two bands of one sensor were used for Data Fusion. In these methods, a panchromatic band was used to increase spatial accuracy, so only spatial resolution was increased. To solve this problem, the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) was used to integrate the images of two Landsat and Modis gauges to increase the spatial and temporal resolution of the reflection. but, this algorithm is designed for pixels and unmixing areas that are the same in Modis and Landsat pixels. The use of this model was not suitable for urban areas with a different of landuse. Therefore, the Enhanced STARFM model (ESTARFM) was developed. The ESTARFM model was improved in 2014 to predict thermal radiation and LST, taking into account the annual temperature cycle and the unevenness of the earth's surface, and the SADFAT model was introduced.

In this study, the performance of SADFAT model in the use of OLI spatial resolution and MODIS temporal resolution in LST forecast in urban areas was examined. The metropolis of Tehran has different surface covers and multiple microclimates. So if the algorithm works successfully, This model can be used in other cities to improve urban heat island studies. The inputs for the algorithm are thermal radiance of Modis and Landsat images, the red and near infrared band of Landsat for daily production of LST in 2017 in the city of Tehran. The algorithm uses two pairs of Modis and Landsat images at the same time and sets of Modis images at the time of prediction and then calculate the conversion coefficient for relating the thermal radiance change of a mixed pixel at the coarse resolution to that of a fine resolution. In this way, LST is generated in areas with a variety of landuse.

All the estimated pixels were compared to the base image pixels in that range to evaluate the results of the model. The comparison results for the autumn days with the average correlation coefficient of 0.86 and RMSE equal to 0.122, showed that the model has the highest accuracy in this season and in other seasons with the average correlation coefficient of 0.76 and RMSE about 0.4, has provided good accuracy.

Visual interpretation of the results of SADFAT showed that this model is able to accurately predict the LST of the land cover in different surface coatings and even in areas where one or more urban land uses are mixed in one MODIS pixel.

However, the borders are well separated and the features are not combined. Although the boundaries are clearly defined, in some land uses, the predicted LST is somewhat higher than the observational image.

Landsat and Modis satellites pass through an area with a small time difference, so they are suitable for combining with each other. But in predicting reflectance with the SADFAT algorithm, there are systematic and variable errors that we need to be aware of in order to increase the output accuracy. One of the systematic and unavoidable errors is the instability of the Terra and Aqua satellites passing through at any point, ie at each satellite pass, the location of the study area in Swath and the size of the pixel changes. Due to the distance of the study area from the vertical center of measurement on the ground (Nadir), the amount of this error varies on different days and should be checked for each day. The preventable error is the sudden change in one or more images used (16 days of the same pass time interval for Landsat) is high for estimating surface reflectance with spatial and temporal resolution. These changes may be due to human factors such as air pollution or natural factors. Natural factors such as clouds and dust storms are the main sources of error in using the SADFAT model because they are sudden and temporary and cover a wide area. The occurrence of these two factors has a great impact on reflectance. Therefore, a sudden change in these factors, in one or more images, causes a large error in the calculations.

The study also found minor spatial errors in the prediction, so that even on days when the results were better, points were observed where the values ​​in the predicted LST images did not match exactly with the OLI sensor. The reason for this may be due to changes in vegetation. Although there are some systematic and variable errors in the images and the implementation of the algorithm The results of this study showed that the performance of this model is reliable for predicting the daily LST with a spatial resolution of 30 meters in Tehran.

This method is able to support urban planning activities related to climate change in cities, so it is recommended that its performance be examined separately for different land cover in the city and the efficiency of this algorithm be evaluated with other sensors such as Copernicus Sentinels.



Key words: Spatial and Temporal Data Fusion, SADFAT, Heat island, LST, Urban climatology

.