Enhancing Deep Learning Soil Moisture Forecasting Models by Integrating Physics-based Models

Accurate soil moisture(SM)prediction is critical for understanding hydrological processes.Physics-based(PB)models exhibit large uncertainties in SM predictions arising from uncertain parameterizations and insufficient representation of land-surface processes.In addition to PB models,deep learning(DL)models have been widely used in SM predictions recently.However,few pure DL models have notably high success rates due to lacking physical information.Thus,we developed hybrid models to effectively integrate the outputs of PB models into DL models to improve SM predictions.To this end,we first developed a hybrid model based on the attention mechanism to take advantage of PB models at each forecast time scale(attention model).We further built an ensemble model that combined the advantages of different hybrid schemes(ensemble model).We utilized SM forecasts from the Global Forecast System to enhance the convolutional long short-term memory(ConvLSTM)model for 1–16 days of SM predictions.The performances of the proposed hybrid models were investigated and compared with two existing hybrid models.The results showed that the attention model could leverage benefits of PB models and achieved the best predictability of drought events among the different hybrid models.Moreover,the ensemble model performed best among all hybrid models at all forecast time scales and different soil conditions.It is highlighted that the ensemble model outperformed the pure DL model over 79.5%of in situ stations for 16-day predictions.These findings suggest that our proposed hybrid models can adequately exploit the benefits of PB model outputs to aid DL models in making SM predictions.

Evaluating Common Land Model Energy Fluxes Using FLUXNET Data

Given the crucial role of land surface processes in global and regional climates, there is a pressing need to test and verify the performance of land surface models via comparisons to observations. In this study, the eddy covariance measurements from 20 FLUXNET sites spanning more than 100 site-years were utilized to evaluate the performance of the Common Land Model （CoLM） over different vegetation types in various climate zones. A decomposition method was employed to separate both the observed and simulated energy fluxes, i.e., the sensible heat flux, latent heat flux, net radiation, and ground heat flux, at three timescales ranging from stepwise （30 rain） to monthly. A comparison between the simulations and observations indicated that CoLM produced satisfactory simulations of all four energy fluxes, although the different indexes did not exhibit consistent results among the different fluxes, A strong agreement between the simulations and observations was found for the seasonal cycles at the 20 sites, whereas CoLM underestimated the latent heat flux at the sites with distinct dry and wet seasons, which might be associated with its weakness in simulating soil water during the dry season. CoLM cannot explicitly simulate the midday depression of leaf gas exchange, which may explain why CoLM also has a maximum diurnal bias at noon in the summer. Of the eight selected vegetation types analyzed, CoLM performs best for evergreen broadleaf forests and worst for croplands and wetlands.

Assessment of global meteorological,hydrological and agricultural drought under future warming based on CMIP6

Quantifying the changes and propagation of drought is of great importance for regional eco-environmental safety and water-related disaster management under global warming.In this study,phase 6 of the Coupled Model Intercomparison Project was employed to examine future meteorological(Standardized Precipitation Index,SPI,and Standardized Precipitation-Evapotranspiration Index,SPEI),hydrological(Standardized Runoff Index,SRI),and agricultural(Standardized Soil moisture Index,SSI) drought under two warming scenarios(SSP2-4.5 and SSP5-8.5).The results show that,across the globe,different types of drought events generally exhibit a larger spatial extent,longer duration,and greater severity from 1901 to 2100,with SPEI drought experiencing the greatest increases.Although SRI and SSI drought are expected to be more intensifying than SPI drought,the models show higher consistency in projections of SPI changes.Regions with robust drying trends include the southwestern United States,Amazon Basin,Mediterranean,southern Africa,southern Asia,and Australia.It is also found that meteorological drought shows a higher correlation with hydrological drought than with agricultural drought,especially in warm and humid regions.Additionally,the maximum correlation between meteorological and hydrological drought tends to be achieved at a short time scale.These findings have important implications for drought monitoring and policy interventions for water resource management under a changing climate.

Plant drought tolerance trait is the key parameter in improving the modeling of terrestrial transpiration in arid and semi-arid regions

The prediction of precipitation depends on accurate modeling of terrestrial transpiration.In recent decades,the trait-based plant hydraulic stress scheme has been developed in land surface models,in order to better predict the hydraulic constraint on terrestrial transpiration.However,the role that each plant functional trait plays in the modeling of transpiration remains unknown.The importance of different plant functional traits for modeled transpiration needs to be addressed.Here,the Morris sensitivity analysis method was implemented in the Common Land Model with the plant hydraulic stress scheme(CoLM-P_(50)HS).Traits related to drought tolerance(P_(50);),stomata,and photosynthesis were screened as the most critical from all 17 plant traits.Among 12 FLUXNET sites,the importance of P_(50);,measured by normalized sensitivity scores,increased towards lower precipitation,whereas the importance of stomatal traits and photosynthetic traits decreased towards drier climate conditions.P_(50);was more important than stomatal traits and photosynthetic traits in arid or semi-arid sites,which implies that hydraulic safety strategies are more crucial than plant growth strategies when plants frequently experience drought.Large variation in drought tolerance traits further proved the coexistence of multiple plant strategies of hydraulic safety.Ignoring the variation in drought tolerance traits may potentially bias the modeling of transpiration.More measurements of drought tolerance traits are therefore necessary to help better represent the diversity of plant hydraulic functions.

Global and regional soil organic carbon estimates:Magnitudes and uncertainties

Globally,soil is the largest terrestrial carbon(C)reservoir.Robust quantification of soil organic C(SOC)stocks in existing global observation-based estimates avails accurate predictions in carbon-climate feedbacks and future climate trends.We investigated the magnitudes and distributions of global and regional SOC estimates(i.e.,density and stocks)based on five widely used global gridded SOC datasets,a regional permafrost dataset developed in 2021(UM2021),and a global-scale soil profile database(World Soil Information Service)reporting measurements of a series of physical and chemical edaphic attributes.The five global gridded SOC datasets were the Harmonized World Soil Database(HWSD),World Inventory of Soil Emission Potentials at 30 arc-second resolution(WISE30sec),Global Soil Dataset for Earth System Models(GSDE),Global Gridded Soil Information at 250-m resolution(SoilGrids250m),and Global Soil Organic Carbon Map(GSOCmap).Our analyses showed that the magnitude and distribution of SOC varied widely among datasets,with certain datasets showing region-specific robustness.At the global scale,SOC stocks at the top 30 and 100 cm were estimated to be 828(range:577–1171)and 1873(range:1086–2678)Pg C,respectively.The estimates from GSDE,GSOCmap,and WISE30sec were comparable,and those of SoilGrids250m and HWSD were at the upper and lower ends.The spatial SOC distribution varied greatly among datasets,especially in the northern circumpolar and Tibetan Plateau permafrost regions.Regionally,UM2021 and WISE30sec performed well in the northern circumpolar permafrost regions,and GSDE performed well in China.The estimates of SOC by different datasets also showed large variabilities across different soil layers and biomes.The discrepancies were generally smaller for the 0–30 cm soil than the 0–100 cm soil.The datasets demonstrated relatively higher agreement in grasslands,croplands,and shrublands/savannas than in other biomes(e.g.,wetlands).The users should be mindful of the gaps between regions and biomes while choosing the most appropriate SOC dataset for specific uses.Large uncertainties in existing global gridded SOC estimates were generally derived from soil sampling density,different sources,and various mapping methods for soil datasets.We call for future efforts for standardizing soil sampling efforts,cross-dataset comparison,proper validation,and overall global collaboration to improve SOC estimates.

高速列车运行控制系统车载子系统故障预测方法研究

列车运行控制系统是高速列车信息控制系统的"神经中枢",其结构复杂、组件繁多,其中,车载子系统是列车运行控制系统的核心组成部分,是保证行车安全、提高运行效率的关键.目前,车载子系统的故障数据处理方式仍以人工处理实现故障定位为主,尚未深入到系统故障机理层面,无法实现有效的故障预测.本文分析了车载子系统的结构及系统故障处置现状,总结了现存的主要问题,结合车载子系统结构及各模块性能参数,提出了基于贝叶斯网络(Bayesian network, BN)的系统级故障预测模型构建方法.利用实际系统现场运行数据,采用基于贝叶斯网络的方法进行故障预测,分别在20, 200, 2000,20000组数据条件下实施了验证,故障预测准确率分别为5%, 27%, 92%, 96.3%,在2000组数据条件下同时对隐马尔科夫模型(hidden Markov model, HMM)、神经网络(neural network, NN)与本文所提出的方法进行了对比,预测结果验证了贝叶斯网络在系统级故障预测方面的显著优势.

Traffic signal control in mixed traffic environment based on advance decision and reinforcement learning

Reinforcement learning-based traffic signal control systems (RLTSC) can enhance dynamic adaptability, save vehicle travelling timeand promote intersection capacity. However, the existing RLTSC methods do not consider the driver’s response time requirement, sothe systems often face efficiency limitations and implementation difficulties.We propose the advance decision-making reinforcementlearning traffic signal control (AD-RLTSC) algorithm to improve traffic efficiency while ensuring safety in mixed traffic environment.First, the relationship between the intersection perception range and the signal control period is established and the trust region state(TRS) is proposed. Then, the scalable state matrix is dynamically adjusted to decide the future signal light status. The decision will bedisplayed to the human-driven vehicles (HDVs) through the bi-countdown timer mechanism and sent to the nearby connected automatedvehicles (CAVs) using the wireless network rather than be executed immediately. HDVs and CAVs optimize the driving speedbased on the remaining green (or red) time. Besides, the Double Dueling Deep Q-learning Network algorithm is used for reinforcementlearning training;a standardized reward is proposed to enhance the performance of intersection control and prioritized experiencereplay is adopted to improve sample utilization. The experimental results on vehicle micro-behaviour and traffic macro-efficiencyshowed that the proposed AD-RLTSC algorithm can simultaneously improve both traffic efficiency and traffic flow stability.