Case 6: Urban thermal environment_Green space planning evaluation

Today, I am going to continue sharing a land surface temperature prediction model for green space planning evaluation based on machine learning regression algorithm.

Usually, after analyzing the relationship between LST and land cover (land use type), it can only be used as a guide for planning, which is difficult to implement. In fact, some studies have established regression models only to analyze whether there is a relationship between multiple variables and LST. If we migrate the face completion cases provided by scikit learn, can we predict the spatial distribution of LST according to the land cover type? Therefore, the distribution of vegetation, construction land and water body (interpreted data) is taken as the explanatory variable, and LST is taken as the target variable to train the model. In the data segmentation, the land surface temperature data on August 10, 2018 were cut, and 324 samples with a sample range of 3000 × 3000m were obtained as the target variables, and they were divided into training data set and test data set, with the division ratio of 0.85. At present, because the explanatory variables are relatively rough, simple surface coverage may not reflect the detailed characteristics of temperature changes. Therefore, the calculation is carried out with the goal of analyzing the upper and lower distribution of mean value, as shown in the figure:

Calculate F1_ The distribution of evaluation scores was observed. Linear expression and ridge performed better, with the median of 0.723 and 0.727, and the mean values were 0.723 and 0.726.

This prediction experiment can be further expanded, and several problems need to be solved. For example, the adjustment of explanatory variables should include the index reflecting the land cover structure. According to the existing research, different vegetation distribution structure will image the change of land surface temperature; and the refinement of temperature prediction distribution level is currently based on the two classification of mean temperature, which can be further improved One step refinement can obtain more accurate distribution prediction.