Zihan Liu

Tracking the temporal dynamics of urban heat island (UHI) is critical for urban heat adaptation and mitigation strategies. However, whether UHI trends have shifted recently and their underlying drivers remain unknown. Here we investigate the variabilities in surface UHI trends and their associated determinants in 2,104 cities worldwide from 2000 to 2022. Our findings reveal that approximately half of the world's cities have experienced notable shifts in surface UHI trends, predominantly characterized by UHI deceleration. These shifts can be primarily attributed to alterations in vegetation trends during the day and to modifications in surface albedo and local warming trends at night. Our study challenges the conventional linear models commonly employed to estimate surface UHI trends, suggesting potential biases in such estimates. Our findings underscore the need for nuanced policies to curtail UHI growth by considering changes in urban underlying surfaces and background climate, particularly from a nonlinear perspective.

Exploring the relationship between the magnitude and temporal trend of the Biodiversity Intactness Index (BII) is critical to assessing current and future changes in biodiversity intactness. However, investigations into the relationship between BII magnitude and trends worldwide have been limited. Using annual BII time series data from 2000 to 2020, we assess the global spatial patterns of BII magnitude and trend, as well as their relationship. Our findings reveal four key insights: First, the global mean BII magnitude exhibits pronounced latitudinal and climatic heterogeneity, with higher values observed in less human-impacted regions. Second, biodiversity trends display contrasting trajectories between areas of differing baseline intactness—regions with initially low biodiversity (BII < 0.50) show recovery potential (−0.007 ± 0.021 decade⁻¹), while high-biodiversity areas (BII > 0.90) face accelerated declines (0.002 ± 0.012 decade⁻¹). Third, continental and climatic disparities are striking: Europe and temperate zones demonstrate stabilizing trends, whereas tropical and polar regions experience marked deterioration. Fourth, climate variables, particularly precipitation seasonality (BIO15) and mean temperature of the coldest quarter (BIO11), show strong negative correlations with the BII trend, indicating climate-linked declines while exhibiting minimal influence on baseline BII magnitude. This study has the potential to help develop more efficient sustainable practices and behaviors to mitigate biodiversity disparities and achieve sustainable development goals.

The monthly composite hourly land surface temperature (LST) under clear sky at 1‐km resolution (denoted as T_mh) plays a critical role in various fields such as urban and agricultural managements. Existing methods for estimating T_mh fall into four categories: single‐source methods employing spatial downscaling or diurnal temperature cycle (DTC) models and multisource methods employing spatiotemporal fusion or DTC models. Despite this methodological diversity, a comprehensive evaluation of their respective strengths and weaknesses remains lacking, posing a challenge for advancing T_mh estimation. To address this critical gap, we performed a wide‐ranging comparison of eight representative approaches for estimating T_mh, comprising two from each category. Their accuracies were assessed over various timescales and conditions, utilizing in situ LST observations from 77 ground‐based stations worldwide. Our evaluations show that DTC‐based multisource approaches exhibit the highest overall accuracy, outperforming both spatial downscaling‐based single‐source approaches and spatiotemporal fusion‐based multisource approaches. Conversely, the performance of DTC‐based single‐source methods exhibits substantial disparity. This observed pattern of overall accuracy remains valid across months, seasons, land cover types, and climatic zones. Furthermore, our assessments indicate that accuracies are time‐of‐day dependent. Spatial downscaling and spatiotemporal fusion approaches are most effective around 2 hr after sunrise, while DTC‐based approaches show better performance around midday. Our findings suggest that this work holds potential significance for generating fine‐resolution hourly LST data with enhanced accuracy.

Mapping thermal anisotropy across global lands is critical for advancing a wide range of Earth science studies. However, a comprehensive understanding of global thermal anisotropy intensity (TAI) and its governing factors remains missing. We introduce a novel data‐driven methodology to quantify global TAI exclusively using multi‐angle MODIS land surface temperature time series observations. Our analysis reveals distinct seasonal and diurnal TAI patterns, with global mean summertime TAI exceeding 2.9°C. Furthermore, we identify strong associations between TAI and key surface and atmospheric parameters, such as leaf area index and downward shortwave radiation. Our findings advocate for a paradigm shift from model‐based to data‐driven approaches in correcting thermal anisotropy, thereby addressing a critical bottleneck in Earth observation.

Biodiversity has important implications for the sustainable development of cities. Given the paucity of ground-based experiments, the responses of biodiversity to urbanization and its associated controls on a global scale remain largely unexplored. We present a novel conceptual framework for quantifying the direct and indirect impacts of urbanization on biodiversity in 1523 cities worldwide using the global 100 m grid biodiversity intactness index data (2017–2020) as a proxy for biodiversity. The results show a pervasive positive impact of urbanization on biodiversity in global cities, with a global mean direct and indirect impact of 24.85 ± 9.97% and 16.18 ± 10.92%, respectively. The indirect impact is relatively large in highly urbanized cities in the eastern United States, Western Europe, and the Middle East. The indirect impact is predominantly influenced by urbanization intensity, population density, and background climate. The correlation between urbanization intensity and indirect impact is most pronounced across all climate zones, while the other driving variables influencing the indirect effect exhibited considerable variations. Furthermore, our findings indicate that the biodiversity responses to urbanization are influenced by the biodiversity and development conditions of cities. Our findings have important implications for understanding the impact of urbanization on biodiversity and for future sustainable urban biodiversity.

A comprehensive comparison of the trends and drivers of global surface and canopy urban heat islands (termed I_s and I_c trends, respectively) is critical for better designing urban heat mitigation strategies. However, such a global comparison remains largely absent. Using spatially continuous land surface temperatures and surface air temperatures (2003–2020), here we find that the magnitude of the global mean I_s trend (0.19 ± 0.006°C/decade, mean ± SE) for 5,643 cities worldwide is nearly six‐times the corresponding I_c trend (0.03 ± 0.002°C/decade) during the day, while the former (0.06 ± 0.004°C/decade) is double the latter (0.03 ± 0.002°C/decade) at night. Variable importance scores indicate that global daytime I_s trend is slightly more controlled by surface property, while background climate plays a more dominant role in regulating global daytime I_c trend. At night, both global I_s and I_c trends are mainly controlled by background climate.

Large‐scale human activity changes in megacities during Chinese New Year (CNY) are believed to significantly affect urban heat islands (UHIs). However, the urban‐rural gradient in UHI variations responsive to human activity changes in cities remains largely unclear. Using in‐situ surface air temperature obtained from a meteorological network that includes 3000‐plus stations, we show that the mean UHI intensity (UHII) in 31 Chinese capitals is 0.52 ± 0.23 K during the CNY holiday and 0.77 ± 0.29 K in the reference period, indicating a UHII reduction of 0.25 ± 0.20 K during the holiday. The reduced UHII decreased more from city core (0.54 K) to city periphery (0.071 K). We find that these UHII reductions were larger at night than during the day and were larger in northern subtropical and warm temperate climates than in other climates. These UHII reductions were mainly attributable to the decline in anthropogenic heat release.

The weekend effect of the canopy urban heat island (UHI) has been long recognized. However, how the UHI weekend effect (UWE) varies with the hour of day and season of year is still unclear; it remains largely unknown on how the UWE is regulated by various controls. To address these knowledge gaps, here we took Beijing, China as an example and investigated the detailed spatiotemporal UWE patterns and the major regulators with a 3‐year data set of in‐situ surface air temperatures. Our results indicate that the annual ΔI_c (the UHI intensity difference on weekends and weekdays) is stronger at night (−0.13 ± 0.12 K; mean ± 1 STD) than during the day (−0.05 ± 0.10 K); at the seasonal scale, ΔI_c reaches the strongest in winter (−0.14 K) and the weakest in summer (−0.05 K). The ΔI_c is strongly regulated by anthropogenic heat flux (AHF), evidenced by a quasi‐synchronous diurnal pattern between ΔI_c and ΔAHF (i.e., the AHF difference between weekends and weekdays). The nighttime ΔI_c is intensely modulated by urban morphology, with a stronger modulation by the landscape shape index than by the distance of the station from the urban center. Weather conditions also modulate the ΔI_c, with the ΔI_c weakening with the increase of cloud coverage and wind speed level. We consider these findings deepen our understanding of the weekly rhythms of UHI as well as the underlying modulators.

Existing knowledge-grounded dialogue systems typically use finetuned versions of a pretrained language model (LM) and large-scale knowledge bases. These models typically fail to generalize on topics outside of the knowledge base, and require maintaining separate potentially large checkpoints each time finetuning is needed. In this paper, we aim to address these limitations by leveraging the inherent knowledge stored in the pretrained LM as well as its powerful generation ability. We propose a multi-stage prompting approach to generate knowledgeable responses from a single pretrained LM. We first prompt the LM to generate knowledge based on the dialogue context. Then, we further prompt it to generate responses based on the dialogue context and the previously generated knowledge. Results show that our knowledge generator outperforms the state-of-the-art retrieval-based model by 5.8% when combining knowledge relevance and correctness. In addition, our multi-stage prompting outperforms the finetuning-based dialogue model in terms of response knowledgeability and engagement by up to 10% and 5%, respectively. Furthermore, we scale our model up to 530 billion parameters and show that larger LMs improve the generation correctness score by up to 10%, and response relevance, knowledgeability and engagement by up to 10%. Our code is available at: https://github.com/NVIDIA/Megatron-LM.