Search

In the coming decades, a changing climate, the loss of high-quality land, the slowing in the annual yield of cereals, and increasing fertilizer use indicate that better agricultural water management strategies are needed. In this study, we designed FarmCan, a novel, robust remote sensing and machine learning (ML) framework to forecast farms' needed daily crop water quantity or needed irrigation (NI). We used a diverse set of simulated and observed near-real-time (NRT) remote sensing data coupled with a random forest (RF) algorithm and inputs about farm-specific situations to predict the amount and timing of evapotranspiration (ET), potential ET (PET), soil moisture (SM), and root zone soil moisture (RZSM). Our case study of four farms in the Canadian Prairies Ecozone (CPE) shows that 8 d composite precipitation (P) has the highest correlation with changes (Δ) of RZSM and SM. In contrast, 8 d PET and 8 d ET do not offer a strong correlation with 8 d P. Using R2, root mean square error (RMSE), and Kling–Gupta efficiency (KGE) indicators, our algorithm could reasonably calculate daily NI up to 14 d in advance. From 2015 to 2020, the R2 values between predicted and observed 8 d ET and 8 d PET were the highest (80 % and 54 %, respectively). The 8 d NI also had an average R2 of 68%. The KGE of the 8 d ET and 8 d PET in four study farms showed an average of 0.71 and 0.50, respectively, with an average KGE of 0.62. FarmCan can be used in any region of the world to help stakeholders make decisions during prolonged periods of drought or waterlogged conditions, schedule cropping and fertilization, and address local government policy concerns.

Synthetic aperture radar (SAR) has great potential for timely monitoring of flood information as it penetrates the clouds during flood events. Moreover, the proliferation of SAR satellites with high spatial and temporal resolution provides a tremendous opportunity to understand the flood risk and its quick response. However, traditional algorithms to extract flood inundation using SAR often require manual parameter tuning or data annotation, which presents a challenge for the rapid automated mapping of large and complex flooded scenarios. To address this issue, we proposed a segmentation algorithm for automatic flood mapping in near-real-time over vast areas and for all-weather conditions by integrating Sentinel-1 SAR imagery with an unsupervised machine learning approach named Felz-CNN. The algorithm consists of three phases: (i) super-pixel generation; (ii) convolutional neural network-based featurization; (iii) super-pixel aggregation. We evaluated the Felz-CNN algorithm by mapping flood inundation during the Yangtze River flood in 2020, covering a total study area of 1,140,300 km2. When validated on fine-resolution Planet satellite imagery, the algorithm accurately identified flood extent with producer and user accuracy of 93% and 94%, respectively. The results are indicative of the usefulness of our unsupervised approach for the application of flood mapping. Meanwhile, we overlapped the post-disaster inundation map with a 10-m resolution global land cover map (FROM-GLC10) to assess the damages to different land cover types. Of these types, cropland and residential settlements were most severely affected, with inundation areas of 9,430.36 km2 and 1,397.50 km2, respectively, results that are in agreement with statistics from relevant agencies. Compared with traditional supervised classification algorithms that require time-consuming data annotation, our unsupervised algorithm can be deployed directly to high-performance computing platforms such as Google Earth Engine and PIE-Engine to generate a large-spatial map of flood-affected areas within minutes, without time-consuming data downloading and processing. Importantly, this efficiency enables the fast and effective monitoring of flood conditions to aid in disaster governance and mitigation globally.

Soil moisture (SM) spatiotemporal variability critically influences water resources, agriculture, and climate. However, besides site-specific studies, little is known about how SM varies locally (1–100-m scale). Consequently, quantifying the SM variability and its impact on the Earth system remains a long-standing challenge in hydrology. We reveal the striking variability of local-scale SM across the United States using SMAP-HydroBlocks — a novel satellite-based surface SM data set at 30-m resolution. Results show how the complex interplay of SM with landscape characteristics and hydroclimate is primarily driven by local variations in soil properties. This local-scale complexity yields a remarkable and unique multi-scale behavior at each location. However, very little of this complexity persists across spatial scales. Experiments reveal that on average 48% and up to 80% of the SM spatial information is lost at the 1-km resolution, with complete loss expected at the scale of current state-of-the-art SM monitoring and modeling systems (1–25 km resolution).

Downward shortwave radiation Rsd determines the surface energy balance, alters evapotranspiration and hydrological conditions, and feeds back to the regional and global climate. Large-scale Rsd estimates are usually retrieved from satellite-based top-of-atmosphere radiation and cloud parameters. These estimates are subject to biases and temporal inhomogeneity due to errors in atmospheric parameters, algorithms, and sensor changes. We found that three satellite products overestimate Rsd by 8%–10% over Asia for 1984–2006, particularly in high latitudes. We used the model tree ensemble (MTE) machine-learning algorithm and commonly used ensemble averaging methods to integrate ground observations and satellite products. Validations based on test stations and independent networks showed that the MTE approach reduces the median relative biases from 8%–10% to 2%, which is more effective than the ensemble averaging methods. We further evaluated the impacts of uncertainty in radiation forcing on surface energy and water balances using the land surface model Noah-MP. The uncertainty of radiation data affects the prediction of sensible heat the most, and also largely affects latent heat prediction in humid regions. Holding the other variables constant, a 10% positive bias in Rsd can lead to a 20%–60% positive bias in the monthly median sensible heat. The simulated hydrological responses to changing radiation forcing are nonlinear as a result of the interactions among evapotranspiration, snowpack, and soil moisture. Our analysis concludes that reducing uncertainty of radiation data is beneficial for predicting regional energy and water balances, which requires more high-quality ground observations and improved satellite retrieval algorithms.

We present a comprehensive global evaluation of monthly precipitation and temperature forecasts from 16 seasonal forecasting models within the NMME Phase-1 system, using Multi-Source Weighted-Ensemble Precipitation version 2 (MSWEP-V2; precipitation) and Climate Research Unit TS4.01 (CRU-TS4.01; temperature) data as reference. We first assessed the forecast skill for lead times of 1–8 months using Kling–Gupta efficiency (KGE), an objective performance metric combining correlation, bias, and variability. Next, we carried out an empirical orthogonal function (EOF) analysis to compare the spatiotemporal variability structures of the forecasts. We found that, in most cases, precipitation skill was highest during the first lead time (i.e., forecast in the month of initialization) and rapidly dropped thereafter, while temperature skill was much higher overall and better retained at higher lead times, which is indicative of stronger temporal persistence. Based on a comprehensive assessment over 21 regions and four seasons, we found that the skill showed strong regional and seasonal dependencies. Some tropical regions, such as the Amazon and Southeast Asia, showed high skill even at longer lead times for both precipitation and temperature. Rainy seasons were generally associated with high precipitation skill, while during winter, temperature skill was low. Overall, precipitation forecast skill was highest for the NASA, NCEP, CMC, and GFDL models, and for temperature, the NASA, CFSv2, COLA, and CMC models performed the best. The spatiotemporal variability structures were better captured for precipitation than temperature. The simple forecast averaging did not produce noticeably better results, emphasizing the need for more advanced weight-based averaging schemes.

In the coming decades, a changing climate, growing global population, and rising food prices will have significant yet uncertain impacts on both water and food security. The loss of high-quality land, the slowing in annual yield of major cereals, and increasing fertilizer use, all indicate that strategies are needed for monitoring and predicting ongoing and future water deficits on farms for better agricultural water management decisions. Most such activities are based on in-situ measurements which are costly, hard to scale, and ignore the wealth of spatial and temporal information from remotely-sensed data. In this study, we designed FarmCan, a novel and robust climate-informed machine learning (ML) framework to predict crop water demand at the farm scale with up to 14 days lead time. We use a diverse set of simulated and observed near-real-time (NRT) remote sensing data coupled with inputs from farmers, a Random Forest (RF) algorithm, and precipitation (P) prediction from MSWEP to predict the amount and timing of evapotranspiration (ET), potential ET (PET), soil moisture (SM), and root zone soil moisture (RZSM). Our study shows that SM and RZSM are the variables that are more correlated with P, while PET and ET do not show a strong correlation with P, SM, and RZSM. Our case study of 4 farms in the Canadian Prairies Ecozone (CPE) using R2, RMSE, and KGE indicators, shows that our algorithm was able to forecast crop water requirements 14 days in advance reasonably well. We also found that during 2020, RF forecasted ET and PET and needed irrigation (NI) with more accuracy than SM and RZSM, although this might vary based on the soil type, location, year of study, and crop type. Due to the forecasting capability and transferability of the mechanism developed, FarmCan is a promising tool for use in any region of the world to help stakeholders make decisions during prolonged periods of drought or waterlogged conditions, schedule cropping and fertilization, and address local government' policy concerns.

Soil moisture (SM) varies widely in space and time. This variability influences agriculture, land-atmosphere interactions and triggers hazards, such as flooding, landslides, droughts, and wildfires. Yet, current observations are limited to a few regional in situ measurement networks or coarse-scale satellite retrievals (9–36-km resolution). As a result, besides site-specific studies, little is known on how SM varies locally (1–100-m resolution). Consequently, quantifying the impact of this variability remains a critical and long-standing challenge in hydrology. This presentation introduces SMAP-HydroBlocks – a novel 30-m resolution SM dataset (2015–2019) that combines hyper-resolution land surface modeling, satellite, and in-situ observations over the United States. Using this data, we reveal the striking variability of local-scale SM across the United States. By mapping the SM spatial variability and its persistence across spatial scales, we show the complex interplay between the landscape and hydroclimate and how this variability is highly scale-dependent. Results show that up to 80% of SM spatial variability information is lost at the 1-km scale, with further losses expected at the scale of current monitoring systems (5–25-km). This high degree of SM variability has a critical influence on freshwater and land ecosystem dynamics. By mapping its spatial variability locally, we provide a stepping-stone towards understanding SM-dependent hydrological, biogeochemical, and ecological processes at local (and so far unresolved) scales.

Previous estimates of tropical forest carbon loss in the twenty-first century using satellite data typically focus on its magnitude, whereas regional loss trajectories and associated drivers are rarely reported. Here we used different high-resolution satellite datasets to show a doubling of gross tropical forest carbon loss worldwide from 0.97 ± 0.16 PgC yr−1 in 2001–2005 to 1.99 ± 0.13 PgC yr−1 in 2015–2019. This increase in carbon loss from forest conversion is higher than in bookkeeping models forced by land-use statistical data, which show no trend or a slight decline in land-use emissions in the early twenty-first century. Most (82%) of the forest carbon loss is at some stages associated with large-scale commodity or small-scale agriculture activities, particularly in Africa and Southeast Asia. We find that ~70% of former forest lands converted to agriculture in 2001–2019 remained so in 2020, confirming a dominant role of agriculture in long-term pan-tropical carbon reductions on formerly forested landscapes. The acceleration and high rate of forest carbon loss in the twenty-first century suggest that existing strategies to reduce forest loss are not successful; and this failure underscores the importance of monitoring deforestation trends following the new pledges made in Glasgow.

Over the past decades, significant changes in temperature and precipitation have been observed, including changes in the mean and extremes. It is critical to understand the trends in hydroclimatic extremes and how they may change in the future as they pose substantial threats to society through impacts on agricultural production, economic losses, and human casualties. In this study, we analyzed projected changes in the characteristics, including frequency, seasonal timing, and maximum spatial and temporal extent, as well as severity, of extreme temperature and precipitation events, using the severity–area–duration (SAD) method and based on a suite of 37 climate models archived in phase 5 of the Coupled Model Intercomparison Project (CMIP5). Comparison between the CMIP5 model estimated extreme events and an observation-based dataset [Princeton Global Forcing (PGF)] indicates that climate models have moderate success in reproducing historical statistics of extreme events. Results from the twenty-first-century projections suggest that, on top of the rapid warming indicated by a significant increase in mean temperature, there is an overall wetting trend in the Northern Hemisphere with increasing wet extremes and decreasing dry extremes, whereas the Southern Hemisphere will have more intense wet extremes. The timing of extreme precipitation events will change at different spatial scales, with the largest change occurring in southern Asia. The probability of concurrent dry/hot and wet/hot extremes is projected to increase under both RCP4.5 and RCP8.5 scenarios, whereas little change is detected in the probability of concurrent dry/cold events and only a slight decrease of the joint probability of wet/cold extremes is expected in the future.

Hydrological extremes, in the form of droughts and floods, have impacts on a wide range of sectors including water availability, food security, and energy production. Given continuing large impacts of droughts and floods and the expectation for significant regional changes projected in the future, there is an urgent need to provide estimates of past events and their future risk, globally. However, current estimates of hydrological extremes are not robust and accurate enough, due to lack of long-term data records, standardized methods for event identification, geographical inconsistencies, and data uncertainties. To tackle these challenges, this article presents the development of the first Global Drought and Flood Catalogue (GDFC) for 1950–2016 by merging the latest in situ and remote sensing datasets with state-of-the-art land surface and hydrodynamic modeling to provide a continuous and consistent estimate of the terrestrial water cycle and its extremes. This GDFC also includes an unprecedented level of detailed analysis of drought and large-scale flood events using univariate and multivariate risk assessment frameworks, which incorporates regional spatial–temporal characteristics (i.e., duration, spatial extent, severity) and global hazard maps for different return periods. This Catalogue forms a basis for analyzing the changing risk of droughts and floods and can underscore national and international climate change assessments and provide a key reference for climate change studies and climate model evaluations. It also contributes to the growing interests in multivariate and compounding risk analysis.

We introduce a set of global high-resolution (0.05°) precipitation (P) climatologies corrected for bias using streamflow (Q) observations from 9372 stations worldwide. For each station, we inferred the “true” long-termPusing a Budyko curve, which is an empirical equation relating long-termP,Q, and potential evaporation. We subsequently calculated long-term bias correction factors for three state-of-the-artPclimatologies [the “WorldClim version 2” database (WorldClim V2); Climatologies at High Resolution for the Earth’s Land Surface Areas, version 1.2 (CHELSA V1.2 ); and Climate Hazards Group Precipitation Climatology, version 1 (CHPclim V1)], after which we used random-forest regression to produce global gap-free bias correction maps for thePclimatologies. Monthly climatological bias correction factors were calculated by disaggregating the long-term bias correction factors on the basis of gauge catch efficiencies. We found that all three climatologies systematically underestimatePover parts of all major mountain ranges globally, despite the explicit consideration of orography in the production of each climatology. In addition, all climatologies underestimatePat latitudes >60°N, likely because of gauge undercatch. Exceptionally high long-term correction factors (>1.5) were obtained for all threePclimatologies in Alaska, High Mountain Asia, and Chile—regions characterized by marked elevation gradients, sparse gauge networks, and significant snowfall. Using the bias-corrected WorldClim V2, we demonstrated that other widely usedPdatasets (GPCC V2015, GPCP V2.3, and MERRA-2) severely underestimatePover Chile, the Himalayas, and along the Pacific coast of North America. MeanPfor the global land surface based on the bias-corrected WorldClim V2 is 862 mm yr−1(a 9.4% increase over the original WorldClim V2). The annual and monthly bias-correctedPclimatologies have been released as the Precipitation Bias Correction (PBCOR) dataset, which is available online (http://www.gloh2o.org/pbcor/).

Poorly monitored river flows in many regions of the world have been hindering our ability to accurately estimate global water budgets as well as the variability of the global water cycle. In situ gauging sites, as well as a number of satellite-based systems, make observations of river discharge throughout the globe; however, these observations are often sparse due to, for example, the sampling frequencies of sensors or a lack of reporting. Recently, efforts have been made to develop methods to integrate these discrete observations to gain a better understanding of the underlying processes. This paper presents an application of a fixed interval Kalman smoother-based model, called inverse streamflow routing (ISR), to generate spatially and temporally continuous river discharge fields from discrete observations. The method propagates the observed information across all reachable parts of the river network (up/downstream from gauging point) and all reachable times (before/after observation time) using a two-sweep procedure that first propagates information backward in time to the furthest upstream locations (inverse routing) and then propagates it forward in time to the furthest downstream locations (forward routing). The ISR methodology advances prediction of streamflow in ungauged basins by accounting for a physical representation of the river system that is not generally handled explicitly in more-commonly applied statistically based models. The key advantages of this approach are that it (1) maintains all the physical consistencies embodied by a diffusive wave routing model (flow confluence relationships on the river network and the resulting mass balance, wave velocity, and diffusivity), (2) updates the lateral influx (runoff) at the pixel level (furthest upstream) to guarantee exhaustive propagation of observed information, and (3) works both with a first guess of initial river discharge conditions from a routing model (assimilation) and without a first guess (pure interpolation of observations). Two sets of experiments are carried out under idealized conditions and under real-world conditions provided by United States Geological Survey (USGS) observations. Results show that the method can effectively reproduce the spatial and temporal dynamics of river discharge in each of the experiments presented. The performance is driven by the density of the gauge network as well as the quality of the data being assimilated. We find that when assimilating the actual USGS observations, the performance decreases relative to our idealized scenario; however, we are still able to produce an improved discharge product at each validation site. With further testing, as well as global application, ISR may prove to be a useful method for extending our current network of global river discharge observations.