Banamali Panigrahi

and 5 more

Machine learning (ML) is increasingly considered the solution to environmental problems where only limited or no physico-chemical process understanding is available. But when there is a need to provide support for high-stake decisions, where the ability to explain possible solutions is key to their acceptability and legitimacy, ML can come short. Here, we develop a method, rooted in formal sensitivity analysis (SA), that can detect the primary controls on the outputs of ML models. Unlike many common methods for explainable artificial intelligence (XAI), this method can account for complex multi-variate distributional properties of the input-output data, commonly observed with environmental systems. We apply this approach to a suite of ML models that are developed to predict various water quality variables in a pilot-scale experimental pit lake. A critical finding is that subtle alterations in the design of an ML model (such as variations in random seed for initialization, functional class, hyperparameters, or data splitting) can lead to entirely different representational interpretations of the dependence of the outputs on explanatory inputs. Further, models based on different ML families (decision trees, connectionists, or kernels) seem to focus on different aspects of the information provided by data, although displaying similar levels of predictive power. Overall, this underscores the importance of employing ensembles of ML models when explanatory power is sought. Not doing so may compromise the ability of the analysis to deliver robust and reliable predictions, especially when generalizing to conditions beyond the training data.
Permafrost plays an important role in the hydrology of arctic/subarctic regions. However, permafrost thaw/degradation has been observed over recent decades in the Northern Hemisphere and is projected to accelerate. Hence, understanding the evolution of permafrost areas is urgently needed. Land surface models (LSMs) are well-suited for predicting permafrost dynamics due to their physical basis and large-scale applicability. However, LSM application is challenging because of the large number of model parameters and the complex memory of state variables. Significant interactions among the underlying processes and the paucity of observations of thermal/hydraulic regimes add further difficulty. This study addresses the challenges of LSM application by evaluating the uncertainty due to meteorological forcing, assessing the sensitivity of simulated permafrost dynamics to LSM parameters, and highlighting issues of parameter identifiability. Modelling experiments are implemented using the MESH-CLASS framework. The VARS sensitivity analysis and traditional threshold-based identifiability analysis are used to assess various aspects of permafrost dynamics for three regions within the Mackenzie River Basin. The study shows that the modeller may face significant trade-offs when choosing a forcing dataset as some datasets enable the representation of some aspects of permafrost dynamics, while being inadequate for others. The results also emphasize the high sensitivity of various aspects of permafrost simulation to parameters controlling surface insulation and soil texture; a detailed list of influential parameters is presented. Identifiability analysis reveals that many of the most influential parameters for permafrost simulation are unidentifiable. These conclusions will hopefully inform future efforts in data collection and model parametrization.

Amin Dezfuli

and 2 more

Fuad Yassin

and 4 more

Hydrologic-Land Surface Models (H-LSMs) have been progressively developed to a stage where they represent the dominant hydrological processes for a variety of hydrological regimes and include a range of water management practices, and are increasingly used to simulate water storages and fluxes of large basins under changing environmental conditions across the globe. However, efforts for comprehensive evaluation of the utility of H-LSMs in large, regulated watersheds have been limited. In this study, we evaluated the capability of a Canadian H-LSM, called MESH, in the highly regulated Saskatchewan River Basin (SaskRB), Canada, under the constraint of significant precipitation uncertainty. A comprehensive analysis of the MESH model performance was carried out in two steps. First, the reliability of multiple precipitation products was evaluated against climate station observations and based on their performance in simulating streamflow across the basin when forcing the MESH model with a default parameterization. Second, a state-of-the-art multi-criteria calibration approach was applied, using various observational information including streamflow, storage and fluxes for calibration and validation. The first analysis shows that the quality of precipitation products had a direct and immediate impact on simulation performance for the basin headwaters but effects were dampened when going downstream. The subsequent analyses show that the MESH model was able to capture observed responses of multiple fluxes and storage across the basin using a global multi-station calibration method. Despite poorer performance in some basins, the global parameterization generally achieved better model performance than a default model parameterization. Validation using storage anomaly and evapotranspiration generally showed strong correlation with observations, but revealed potential deficiencies in the simulation of storage anomaly over open water areas. Keywords: Precipitation Uncertainty, Hydrologic-Land Surface Models, multi-criteria calibration, storage and fluxes validation, Saskatchewan River Basin, Canada

Mohamed Abdelhamed

and 3 more

Permafrost thaw has been observed in recent decades in the Northern Hemisphere and is expected to accelerate with continued global warming. Predicting the future of permafrost requires proper representation of the interrelated surface/subsurface thermal and hydrologic regimes. Land surface models (LSMs) are well suited for such predictions, as they couple heat and water interactions across soil-vegetation-atmosphere interfaces and can be applied over large scales. LSMs, however, are challenged by the long-term thermal and hydraulic memories of permafrost and the paucity of historical records to represent permafrost dynamics under transient climate conditions. In this study, we address the challenge of model initialization by characterizing the impact of initial climate conditions and initial soil frozen and liquid water contents on the simulation length required to reach equilibrium. Further, we quantify how the uncertainty in model initialization propagates to simulated permafrost dynamics. Modelling experiments are conducted with the Modélisation Environmentale Communautaire – Surface and Hydrology (MESH) framework and its embedded Canadian Land Surface Scheme (CLASS). The study area is in the Liard River basin in the Northwest Territories of Canada with sporadic and discontinuous regions. Results show that uncertainty in model initialization controls various attributes of simulated permafrost, especially the active layer thickness, which could change by 0.5-1.5m depending on the initial condition chosen. The least number of spin-up cycles is achieved with near field capacity condition, but the number of cycles varies depending on the spin-up year climate. We advise an extended spin-up of 200-1000 cycles to ensure proper model initialization under different climatic conditions and initial soil moisture contents.

Howard Wheater

and 19 more

Cold regions provide water resources for half the global population yet face rapid change. Their hydrology is dominated by snow, ice and frozen soils, and climate warming is having profound effects. Hydrological models have a key role in predicting changing water resources, but are challenged in cold regions. Ground-based data to quantify meteorological forcing and constrain model parameterization are limited, while hydrological processes are complex, often controlled by phase change energetics. River flows are impacted by poorly quantified human activities. This paper reports scientific developments over the past decade of MESH, the Canadian community hydrological land surface scheme. New cold region process representation includes improved blowing snow transport and sublimation, lateral land-surface flow, prairie pothole storage dynamics, frozen ground infiltration and thermodynamics, and improved glacier modelling. New algorithms to represent water management include multi-stage reservoir operation. Parameterization has been supported by field observations and remotely sensed data; new methods for parameter identification have been used to evaluate model uncertainty and support regionalization. Additionally, MESH has been linked to broader decision-support frameworks, including river ice simulation and hydrological forecasting. The paper also reports various applications to the Saskatchewan and Mackenzie River basins in western Canada (0.4 and 1.8 million km2). These basins arise in glaciated mountain headwaters, are partly underlain by permafrost, and include remote and incompletely understood forested, wetland, agricultural and tundra ecoregions. This imposes extraordinary challenges to prediction, including the need to overcoming biases in forcing data sets, which can have disproportionate effects on the simulated hydrology.
The prediction of future land cover changes is an important step in proper planning and management of watersheds. Various methods exist for this purpose. In this study, land cover changes were investigated in the Hable-Rud River basin in Iran, an arid and semi-arid region, using remote sensing and Geographic Information Systems (GIS). First, a supervised classification technique was applied to Landsat images acquired for 1986, 2000 and 2017 using the maximum likelihood method. Then, using pixel-by-pixel change detection, the land cover changes were predicted for 2017 and 2040 using a Cellular Automata (CA)-Markov model. The descriptive variables used included slope, aspect, elevation, and calculated distances from various land features such as rivers, roads, industrial areas, residential areas, saline land, and land in agricultural production. The predictions for 2017 were validated using the derived map from a Landsat image of 2017 with a resulting standard Kappa index of 0.74. According to the prediction results for 2040, the areas of rangeland and saline land will increase by approximately 6.5% and 2%, respectively, whereas the areas of bare land and agricultural land will decrease by approximately 6% and 2%, respectively. Moreover, the analysis of historical records since 1986 showed that the annual streamflow and precipitation have reduced by almost 44% and 29%, respectively. The reductions, particularly to streamflow, can be attributed largely to agriculture expansion, rapid population growth, and industrial developments. The analysis of the results indicates a need for more effective design, planning, and development of land cover policies for ecosystem protection.