Explainable machine learning improves interpretability in the predictive modeling of biological stream conditions in the Chesapeake Bay Watershed, USA

Synergistic effects of environmental factors on benthic diversity: Machine learning analysis

This study examines the water environmental factors of the Cangshan stream and benthic animal communities by using random forest, gradient boosting decision tree, and support vector machine models to analyze the complex response mechanisms of benthic animal diversity and community structure to environmental factors. Feature importance analysis, SHAP values, and 3D response surface analysis are applied to quantitatively assess the non-linear driving effects of environmental factors and their interactions. The findings suggest that total phosphorus and conductivity are central factors influencing benthic animal diversity, with moderate levels fostering community diversity, whereas high levels of total nitrogen and conductivity significantly reduce diversity. Benthic animals exhibit a non-linear response pattern to dissolved oxygen and temperature, with the interaction between dissolved oxygen and temperature highlighting the significant promotion of diversity under low-temperature, high-oxygen conditions, whereas high-temperature, low-oxygen conditions exert evident environmental stress on communities. The results of the multifactor synergistic effect analysis indicate that the moderate synergistic interaction between total phosphorus and conductivity significantly enhances diversity, whereas high total nitrogen levels weaken this positive effect. Model performance comparisons reveal that the RF outperforms the other models in terms of coefficient of determination, mean squared error, and mean absolute error, particularly in capturing complex non-linear relationships and factor interactions. Through machine learning, this study reveals the multidimensional driving mechanisms of environmental factors on benthic animal community characteristics, emphasizing the potential to capture non-linear relationships and multifactor interactions, thereby providing scientific evidence and innovative approaches for stream ecosystem conservation and management.

Using tide for rainfall runoff simulation with feature projection and reversible instance normalization

From LSTMs (Long Short-Term Memory) to Transformers, various networks have been used for runoff forecasting, though many complex structures may be unnecessary. This study introduces RR-TiDE, a simple model based on the Time Series Dense Encoder. RR-TiDE employs fully Multilayer Perceptron architecture for modeling and is specifically designed with understanding of hydrological processes. To manage non-stationarity in hydrological data, RR-TiDE incorporates Reversible Instance Normalization. The model was trained using the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) dataset and evaluated on two tasks: (1) multi-basin runoff simulation; (2) prediction in data-sparse basins. In the first task, RR-TiDE outperformed both Transformer and LSTM-based models across all metrics for 7-day runoff predictions, which indicates that RR-TiDE is highly suitable for rainfall-runoff simulation. In the second task, it achieved a median NSE of 0.82 in 1-day runoff forecasting in 51 watersheds. This suggests that RR-TiDE possesses robust generalization capability, enabling spatial extrapolation. Comparisons were made between models with and without the feature projection layer and RevIN to further understand their individual contributions. Results indicate that the feature projection layer can effectively enhance the performance of RR-TiDE. Although RevIN provided limited overall improvements, it helped stabilize loss fluctuations during training, aiding model convergence.

Enhancing prediction and inference of daily in-stream nutrient and sediment concentrations using an extreme gradient boosting based water quality estimation tool - XGBest

Estimating constituent loads in streams and rivers is a crucial but challenging task due to low-frequency sampling in most watersheds. While predictive modeling can augment sparsely sampled water quality data, it can be challenging due to the complex and multifaceted interactions between several sub-watershed eco-hydrological processes. Traditional water quality prediction models, typically calibrated for individual sites, struggle to fully capture these interactions. This study introduces XGBest, a machine learning-based tool, that integrates hydrological data, land cover, and physical watershed attributes at a regional scale to predict daily concentrations of Total Nitrogen (TN), Total Phosphorus (TP), and Total Suspended Solids (TSS). XGBest leverages 29 environmental variables, including daily and antecedent discharge, temporal features, and landscape characteristics, to comprehensively evaluate water quality dynamics across a large hydrologic region. To explore the robustness of the developed tool, XGBest was validated using observed water quality data in three different hydrologic regions in the eastern United States, encompassing 499 water quality monitoring sites characterized by diverse hydro-climatic conditions and watershed attributes. This study also employed the legacy United States Geological Survey (USGS) tools - LOADEST and WRTDS as benchmarks to evaluate the performance of XGBest in these regions. The results demonstrated that XGBest outperformed LOADEST and WRTDS in predictive accuracy and revealed critical insights into the spatial and temporal variability of nutrient and sediment loads. In addition, SHapley Additive exPlanations (SHAP) values highlighted the importance of integrating static and dynamic watershed attributes, such as land cover, antecedent discharge, and seasonality, in capturing the complex concentration-discharge (C-Q) relationships. This study positions XGBest as a robust and scalable water quality prediction tool that bridges the gap between hydrology and broader environmental management. By combining multiple environmental factors into a unified predictive framework, XGBest enhances our understanding of water quality and supports more effective environmental monitoring and management strategies.

Physical habitat is more than a sediment issue: A multi-dimensional habitat assessment indicates new approaches for river management

Degraded physical habitat is a common stressor affecting river ecosystems and typically addressed in the United States (US) through a regulatory focus on sediment. However, a narrow regulatory focus on sediment may overlook other aspects of physical habitat and the processes for its creation, maintenance, and degradation. In addition, there exist few "ready-to-use" regional assessments of the multiple dimensions of physical habitat to better understand continuous patterns of condition and prioritize management efforts across a large spatial scale. In this study, we use rapid habitat monitoring data to train a machine-learning (i.e., random forest) model to predict twelve physical habitat metrics for nearly 120,000 km of nontidal rivers and streams across the Chesapeake Bay watershed, US. We capture a range of habitat conditions driven by both natural variables and anthropogenic pressures. Covariation among habitat metrics indicated two major dimensions of habitat variation: 1) coarse bed substrate and hydromorphic heterogeneity and 2) bank stability and riparian condition. The model predicted localized changes from 2001 to 2019, and the predicted areas of deterioration roughly balanced improvements across the watershed, indicating little progress towards long-term watershed management goals. To evaluate connections to regulatory and management endpoints, we compared our physical habitat predictions to paired estimates of sediment and flow alteration across the region. Sediment concentrations were greater in reaches with less bank stability and lower riparian quality; however, the relation was weak for coarse bed condition metrics, including embeddedness, which is frequently used for establishing regulatory sediment restrictions. For flow alteration, most habitat metrics had lower scores with altered flow metrics, but metrics of instream habitat heterogeneity and coarse substrate condition were most strongly affected. Increased flashy, high flows negatively affected most metrics, but coarse substrate metrics were also negatively affected by greater low flow severity. This study highlights a potential disconnect between a narrow focus on regulatory sediment targets given the multiple dimensions and responses of physical habitat. A more holistic approach to physical habitat in management interventions - one that considers hydromorphic processes, diversity and variability in microhabitats, and explicit consideration of alterations to both low and high flows - may be warranted. By providing direct estimates of multiple aspects of physical habitat, this model can help support managers in the Chesapeake Bay watershed to better understand the range of habitat conditions, identify high-quality reaches for conservation, and target potential management actions tailored to localized conditions.

A novel coupling interpretable machine learning framework for water quality prediction and environmental effect understanding in different flow discharge regulations of hydro-projects

Machine learning models (MLMs) have been increasingly used to forecast water pollution. However, the "black box" characteristic for understanding mechanism processes still limits the applicability of MLMs for water quality management in hydro-projects under complex and frequently artificial regulation. This study proposes an interpretable machine learning framework for water quality prediction coupled with a hydrodynamic (flow discharge) scenario-based Random Forest (RF) model with multiple model-agnostic techniques and quantifies global, local, and joint interpretations (i.e., partial dependence, individual conditional expectation, and accumulated local effects) of environmental factor implications. The framework was applied and verified to predict the permanganate index (CODMn) under different flow discharge regulation scenarios in the Middle Route of the South-to-North Water Diversion Project of China (MRSNWDPC). A total of 4664 sampling cases data matrices, including water quality, meteorological, and hydrological indicators from eight national stations along the main canal of the MRSNWDPC, were collected from May 2019 to December 2020. The results showed that the RF models were effective in forecasting CODMn in all flow discharge scenarios, with a mean square error, coefficient of determination, and mean absolute error of 0.006-0.026, 0.481-0.792, and 0.069-0.104, respectively, in the testing dataset. A global interpretation indicated that dissolved oxygen, flow discharge, and surface pressure are the three most important variables of CODMn. Local and joint interpretations indicated that the RF-based prediction model provides a basic understanding of the physical mechanisms of environmental systems. The proposed framework can effectively learn the fundamental environmental implications of water quality variations and provide reliable prediction performance, highlighting the importance of model interpretability for trustworthy machine learning applications in water management projects. This study provides scientific references for applying advanced data-driven MLMs to water quality forecasting and a reliable methodological framework for water quality management and similar hydro-projects.

Predictive Modeling Reveals Elevated Conductivity Relative to Background Levels in Freshwater Tributaries within the Chesapeake Bay Watershed, USA

Elevated conductivity (i.e., specific conductance or SC) causes osmotic stress in freshwater aquatic organisms and may increase the toxicity of some contaminants. Indices of benthic macroinvertebrate integrity have declined in urban areas across the Chesapeake Bay watershed (CBW), and more information is needed about whether these declines may be due to elevated conductivity. A predictive SC model for the CBW was developed using monitoring data from the National Water Quality Portal. Predictor variables representing SC sources were compiled for nontidal reaches across the CBW. Random forests modeling was conducted to predict SC at four time periods (1999-2001, 2004-2006, 2009-2011, and 2014-2016), which were then compared to a national data set of background SC to quantify departures from background SC. Carbonate geology, impervious cover, forest cover, and snow depth were the most important variables for predicting SC. Observations and modeled results showed snow depth amplified the effect of impervious cover on SC. Elevated SC was predicted in two-thirds of reaches in the CBW, and these elevated conditions persisted over time in many areas. These results can be used in stressor identification assessments to prioritize future monitoring and to determine where management activities could be implemented to reduce salinization.

Interpretable machine learning for evaluating risk factors of freeway crash severity

Machine learning (ML) models are widely employed for crash severity modelling, yet their interpretability remains underexplored. Interpretation is crucial for comprehending ML results and aiding informed decision-making. This study aims to implement an interpretable ML to visualize the impacts of factors on crash severity using 5 years of freeways data from Iran. Methods including classification and regression trees (CART), K-nearest neighbours (KNNs), random forest (RF), artificial neural network (ANN) and support vector machines (SVM) were applied, with RF demonstrating superior accuracy, recall, F1-score and ROC. The accumulated local effects (ALE) were utilized for interpretation. Findings suggest that light traffic conditions (volume / capacity < 0.5 ) with critical values around 0.05 or 0.38, and higher proportion of large trucks and buses, particularly at 10% and 4%, are associated with severe crashes. Additionally, speeds exceeding 90 km/h, drivers younger than 30 years, rollover crashes, collisions with fixed objects and barriers, nighttime driving and driver fatigue elevate the likelihood of severe crashes.

Using explainable machine learning methods to evaluate vulnerability and restoration potential of ecosystem state transitions

Ecosystem state transitions can be ecologically devastating or be a restoration success. State transitions are common within aquatic systems worldwide, especially considering human-mediated changes to land use and water use. We created a transferable conceptual framework to enable multiscale assessments of state resilience and early warnings of state transitions that can inform strategic restorations and avoid ecosystem collapse. The conceptual framework integrated machine learning predictions with ecosystem state concepts (e.g., state classification, gradients of vulnerability, and recovery potential leading to state transitions) and was devised to investigate possible environmental drivers. As an application of the framework, we generated prediction probabilities of submersed aquatic vegetation (SAV) presence at nearly 10,000 sites in the Upper Mississippi River (United States). Then, we used an interpretability method to explain model predictions to gain insights into possible environmental drivers and thresholds or linear responses of SAV presence and absence. Model accuracy was 89% without spatial bias. Average water depth, suspended solids, substrate, and distance to nearest SAV were the best predictors and likely environmental drivers of SAV habitat suitability. These environmental drivers exhibited nonlinear, threshold-type responses for SAV. All the results are also presented in an online dashboard to explore results at many spatial scales. The habitat suitability model outputs and prediction explanations from many spatial scales (4 m to 400 km of river reach) can inform research and restoration planning.

The nonlinear relationship between air quality and housing prices by machine learning

Using a dataset encompassing 228 cities in China spanning from 2005 to 2019, this study explores the nonlinear relationship between air quality and housing prices and devises a strategy that incorporates the instrumental variable and machine learning to address the endogeneity issue. Both traditional models and machine learning models find air pollution affects housing prices in a diminishing manner. The negative impact of air pollution on housing prices decreases when the degree of air pollution intensifies. Such a characteristic is more pronounced in Eastern China and cities with fewer land resource constraints and larger populations. Mechanism analysis also reveals that air pollution could affect residents' perceived air quality and the industrial structure, further contributing to the nonlinear relationship between air quality and housing prices. The further SHapley Additive exPlanations (SHAP) evaluates the importance of air quality in determining housing prices and finds that air quality's contribution outweighs educational and medical resources. The contribution of air quality also shows a distinct regional disparity and has become increasingly important in recent years. The findings refine the benefit assessment accuracy related to air quality improvement.

Observed and projected functional reorganization of riverine fish assemblages from global change

Climate and land-use/land-cover change ("global change") are restructuring biodiversity, globally. Broadly, environmental conditions are expected to become warmer, potentially drier (particularly in arid regions), and more anthropogenically developed in the future, with spatiotemporally complex effects on ecological communities. We used functional traits to inform Chesapeake Bay Watershed fish responses to future climate and land-use scenarios (2030, 2060, and 2090). We modeled the future habitat suitability of focal species representative of key trait axes (substrate, flow, temperature, reproduction, and trophic) and used functional and phylogenetic metrics to assess variable assemblage responses across physiographic regions and habitat sizes (headwaters through large rivers). Our focal species analysis projected future habitat suitability gains for carnivorous species with preferences for warm water, pool habitats, and fine or vegetated substrates. At the assemblage level, models projected decreasing habitat suitability for cold-water, rheophilic, and lithophilic individuals but increasing suitability for carnivores in the future across all regions. Projected responses of functional and phylogenetic diversity and redundancy differed among regions. Lowland regions were projected to become less functionally and phylogenetically diverse and more redundant while upland regions (and smaller habitat sizes) were projected to become more diverse and less redundant. Next, we assessed how these model-projected assemblage changes 2005-2030 related to observed time-series trends (1999-2016). Halfway through the initial projecting period (2005-2030), we found observed trends broadly followed modeled patterns of increasing proportions of carnivorous and lithophilic individuals in lowland regions but showed opposing patterns for functional and phylogenetic metrics. Leveraging observed and predicted analyses simultaneously helps elucidate the instances and causes of discrepancies between model predictions and ongoing observed changes. Collectively, results highlight the complexity of global change impacts across broad landscapes that likely relate to differences in assemblages' intrinsic sensitivities and external exposure to stressors.