Accurate spatial population distribution information, especially for metropolises, is of significant value and is fundamental to many application areas such as public health, urban development planning and disaster assessment management. Random forest is the most widely used model in population spatialization studies. However, a reliable model for accurately mapping the spatial distribution of metropolitan populations is still lacking due to the inherent limitations of the random forest model and the complexity of the population spatialization problem. In this study, we integrate gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM) and support vector regression (SVR) through ensemble learning algorithm stacking to construct a novel population spatialization model we name GXLS-Stacking. We integrate socioeconomic data that enhance the characterization of the population’s spatial distribution (e.g., point-of-interest data, building outline data with height, artificial impervious surface data, etc.) and natural environmental data with a combination of census data to train the model to generate a high-precision gridded population density map with a 100 m spatial resolution for Beijing in 2020. Finally, the generated gridded population density map is validated at the pixel level using the highest resolution validation data (i.e., community household registration data) in the current study. The results show that the GXLS-Stacking model can predict the population’s spatial distribution with high precision (R² = 0.8004, MAE = 34.67 persons/hectare, RMSE = 54.92 persons/hectare), and its overall performance is not only better than the four individual models but also better than the random forest model. Compared to the natural environmental features, a city’s socioeconomic features are more capable in characterizing the spatial distribution of the population and the intensity of human activities. In addition, the gridded population density map obtained by the GXLS-Stacking model can provide highly accurate information on the population’s spatial distribution and can be used to analyze the spatial patterns of metropolitan population density. Moreover, the GXLS-Stacking model has the ability to be generalized to metropolises with comprehensive and high-quality data, whether in China or in other countries. Furthermore, for small and medium-sized cities, our modeling process can still provide an effective reference for their population spatialization methods.

Ephedra species are beneficial for environmental protection in desert and grassland ecosystems. They have high ecological, medicinal, and economic value. To strengthen the protection of the sustainable development of Ephedra, we used occurrence records of Ephedra sinica Stapf., Ephedra intermedia Schrenk et C.A. Mey., and Ephedra equisetina Bge., combined with climate, soil, and topographic factors to simulate the suitable habitat of three Ephedra based on ensemble models on the Biomod2 platform. The results of the models were tested using AUC, TSS, and kappa coefficients. The results demonstrated that the ensemble model was able to accurately predict the potential distributions of E. sinica, E. intermedia, and E. equisetina. Eastern and central Inner Mongolia, middle and eastern Gansu, and northeastern Xinjiang were the optimum regions for the growth of E. sinica, E. intermedia, and E. equisetina, respectively. Additionally, several key environmental factors had a significant influence on the suitable habitats of the three Ephedra. The key factors affecting the distribution of E. sinica, E. intermedia, and E. equisetina were annual average precipitation, altitude, and vapor pressure, respectively. In conclusion, the results showed that the suitable ranges of the three Ephedra were mainly in Northwest China and that topography and climate were the primary influencing factors.

Lake water temperature (LWT) is an important indicator of physical processes within a lake, but traditional process‐based and data‐driven models are limited in their ability to estimate long‐term changes in LWT because of simplified physical laws, insufficient onsite measurements and high computational demands. To overcome these limitations, this study proposes a hybrid multi‐parameter scientific knowledge‐guided neural network (MP‐KgNN) for solving 1‐D lake temperature governing equation trained using both simulations of the WRF‐Lake model and onsite LWT measurements based on a novel training framework called physics‐informed deep learning (PIDL) framework and simulates the thermodynamics in a large deep reservoir located in eastern China from 1960 to 2021. The results revealed that the MP‐KgNN can estimate the dynamic changes in LWT with satisfactory accuracy (mean absolute error [MAE] = 1.14 K, root mean square error [RMSE] = 1.49 K). Moreover, it outperformed the pre‐trained MP‐KgNN trained with only the WRF‐Lake model (MAE = 2.43 K, RMSE = 2.77 K), which indicates its successful prediction of the thermal structure of the lake. The prediction derived by MP‐KgNN showed an increasing trend (0.04 K decade⁻¹) of LWT in the Lake Qiandaohu. Specifically, the LWT was experienced to increase at a rate of 0.10 K decade⁻¹ near the lake surface. These changes resulted in an extension and deepening of lake thermal stratification, as indicated by a 0.58 m increase in metalimnion thickness and a 20.46 kJ increase in Schmidt stability. The proposed MP‐KgNN is expected to become a powerful tool for estimating long‐term variations in the thermodynamics of lake ecosystems.