Case 5: Urban thermal environment_Based on LST (land surface temperature)

Jingyi Chen
3 min readDec 10, 2020


When exploring the application of land surface temperature in the field of science, it is found that most of the studies are inclined to analyze the profile temperature or calculate the average temperature change of multiple buffer zones to explore the relationship between the surface temperature and land use (surface cover), but they are not included in specific areas, which is weak for guiding planning. Therefore, we carried out part of the exploration in this direction, and wrote a lot of code, not only convenient for data processing, but also explored some new methods. In this method, the temperature distribution of the land surface is defined by the method of spatial temperature distribution based on convolution and temperature gradient.

The code is shared on GitHub, hoping to drive explorers in this direction for further research. I am going to share some of models.

1 Defining the spatial distribution structure of cold and hot regions based on convolution temperature gradient

Spatial autocorrelation / local indicators of spatial autocorrelation Lisa is to explore the spatial correlation based on the attribute values of spatial points. If the value of a location variable is high, the variable of the value nearby is also high, which is positive spatial autocorrelation, otherwise, it is negative spatial autocorrelation. There are some ready-made methods for analyzing the cold and hot regions (hot spots) of data, such as hot spot analysis based on ArcGIS (based on GETIS ord GI * Statistics). The specific calculation formula can be obtained from ArcGIS help and optimized hot spot analysis tool.

Meanwhile, GeoDa provides global and local spatial autocorrelation based on Moran’s I Moran index. For example, the calculation results are as follows:

The python library pysal also provides (global / local) Moran index calculations. This method will be adopted in the following experiments. Of course, there is another kind in the analysis of spatial relations, that is, convolution. Convolution is usually used in image processing, such as image sharpening, edge detection, relief, mean / Gaussian blur and so on. Because it can extract the features between spatial data, it is an important content in deep learning. Then convolution is used to find out the spatial characteristics of the data, which has been described in detail in previous experiments. This time, the difference between each grid data and the surrounding grid data is searched, and the cold area and hot area are distinguished based on the position where the change value is zero. That is, the hot area is that the grid variable value is higher than the surrounding grid variable value, and the cold area is opposite to analyze the spatial change of LST. As shown in the figure below:

Based on the above analysis results, it is possible to find an independent cold zone and, of course, an independent hot zone:

Appendix: because machine learning scikit learn provides a lot of clustering methods to discover the relationship between spatial data, we try to explore some algorithms and what kind of results can be brought to the analysis, such as spectrum_ Clustering () clustering method, of course, is not deeply analyzed here, so we can explore the value of its calculation results.