When using machine learning to model environmental systems, it is often a model’s ability to predict extreme behaviors that yields the highest practical value to policy makers. However, most existing error metrics used to evaluate the performance of environmental machine learning models weigh error equally across test data. Thus, routine performance is prioritized over a model’s ability to robustly quantify extreme behaviors. In this work, we present a new error metric, termed Reflective Error, which quantifies the degree at which our model error is distributed around our extremes, in contrast to existing model evaluation methods that aggregate error over all events. The suitability of our proposed metric is demonstrated on a real-world hydrological modeling problem, where extreme values are of particular concern.

Peat is formed by the accumulation of organic material in water-saturated soils. Drainage of peatlands and peat extraction contribute to carbon emissions and biodiversity loss. Most peat extracted for commercial purposes is used for energy production or as a growing substrate. Many countries aim to reduce peat usage but this requires tools to detect its presence in substrates. We propose a decision support system based on deep learning to detect peat-specific testate amoeba in microscopy images. We identified six taxa that are peat-specific and frequent in European peatlands. The shells of two taxa (Archerella sp. and Amphitrema sp.) were well preserved in commercial substrate and can serve as indicators of peat presence. Images from surface and commercial samples were combined into a training set. A separate test set exclusively from commercial substrates was also defined. Both datasets were annotated and YOLOv8 models were trained to detect the shells. An ensemble of eight models was included in the decision support system. Test set performance (average precision) reached values above 0.8 for Archerella sp. and above 0.7 for Amphitrema sp. The system processes thousands of images within minutes and returns a concise list of crops of the most relevant shells. This allows a human operator to quickly make a final decision regarding peat presence. Our method enables the monitoring of peat presence in commercial substrates. It could be extended by including more species for applications in restoration ecology and paleoecology.

Increasing penetration of variable and intermittent renewable energy resources on the energy grid poses a challenge for reliable and efficient grid operation, necessitating the development of algorithms that are robust to this uncertainty. However, standard algorithms incorporating uncertainty for generation dispatch are computationally intractable when costs are nonconvex, and machine learning-based approaches lack worst-case guarantees on their performance. In this work, we propose a learning-augmented algorithm, RobustML, that exploits the good average-case performance of a machine-learned algorithm for minimizing dispatch and ramping costs of dispatchable generation resources while providing provable worst-case guarantees on cost. We evaluate the algorithm on a realistic model of a combined cycle cogeneration plant, where it exhibits robustness to distribution shift while enabling improved efficiency as renewables penetration increases.

Sustainable agricultural practices have become increasingly important due to growing environmental concerns and the urgent need to mitigate the climate crisis. Digital agriculture, through advanced data analysis frameworks, holds promise for promoting these practices. Pesticides are a common tool in agricultural pest control, which are key in ensuring food security but also significantly contribute to the climate crisis. To combat this, Integrated Pest Management (IPM) stands as a climate-smart alternative. We propose a causal and explainable framework for enhancing digital agriculture, using pest management and its sustainable alternative, IPM, as a key example to highlight the contributions of causality and explainability. Despite its potential, IPM faces low adoption rates due to farmers’ skepticism about its effectiveness. To address this challenge, we introduce an advanced data analysis framework tailored to enhance IPM adoption. Our framework provides (i) robust pest population predictions across diverse environments with invariant and causal learning, (ii) explainable pest presence predictions using transparent models, (iii) actionable advice through counterfactual explanations for in-season IPM interventions, (iv) field-specific treatment effect estimations, and (v) assessments of the effectiveness of our advice using causal inference. By incorporating these features, our study illustrates the potential of causality and explainability concepts to enhance digital agriculture regarding promoting climate-smart and sustainable agricultural practices, focusing on the specific case of pest management. In this case, our framework aims to alleviate skepticism and encourage wider adoption of IPM practices among policymakers, agricultural consultants, and farmers.

Photovoltaic (PV) energy grows rapidly and is crucial for the decarbonization of electric systems. However, centralized registries recording the technical characteristics of rooftop PV systems are often missing, making it difficult to monitor this growth accurately. The lack of monitoring could threaten the integration of PV energy into the grid. To avoid this situation, remote sensing of rooftop PV systems using deep learning has emerged as a promising solution. However, existing techniques are not reliable enough to be used by public authorities or transmission system operators (TSOs) to construct up-to-date statistics on the rooftop PV fleet. The lack of reliability comes from deep learning models being sensitive to distribution shifts. This work comprehensively evaluates distribution shifts’ effects on the classification accuracy of deep learning models trained to detect rooftop PV panels on overhead imagery. We construct a benchmark to isolate the sources of distribution shifts and introduce a novel methodology that leverages explainable artificial intelligence (XAI) and decomposition of the input image and model’s decision regarding scales to understand how distribution shifts affect deep learning models. Finally, based on our analysis, we introduce a data augmentation technique designed to improve the robustness of deep learning classifiers under varying acquisition conditions. Our proposed approach outperforms competing methods and can close the gap with more demanding unsupervised domain adaptation methods. We discuss practical recommendations for mapping PV systems using overhead imagery and deep learning models.

Machine-learning (ML) methods have shown great potential for weather downscaling. These data-driven approaches provide a more efficient alternative for producing high-resolution weather datasets and forecasts compared to physics-based numerical simulations. Neural operators, which learn solution operators for a family of partial differential equations, have shown great success in scientific ML applications involving physics-driven datasets. Neural operators are grid-resolution-invariant and are often evaluated on higher grid resolutions than they are trained on, i.e., zero-shot super-resolution. Given their promising zero-shot super-resolution performance on dynamical systems emulation, we present a critical investigation of their zero-shot weather downscaling capabilities, which is when models are tasked with producing high-resolution outputs using higher upsampling factors than are seen during training. To this end, we create two realistic downscaling experiments with challenging upsampling factors (e.g., 8x and 15x) across data from different simulations: the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) and the Wind Integration National Dataset Toolkit. While neural operator-based downscaling models perform better than interpolation and a simple convolutional baseline, we show the surprising performance of an approach that combines a powerful transformer-based model with parameter-free interpolation at zero-shot weather downscaling. We find that this Swin-Transformer-based approach mostly outperforms models with neural operator layers in terms of average error metrics, whereas an Enhanced Super-Resolution Generative Adversarial Network-based approach is better than most models in terms of capturing the physics of the ground truth data. We suggest their use in future work as strong baselines.

Achieving net-zero carbon emissions by 2050 necessitates the integration of substantial wind power capacity into national power grids. However, the inherent variability and uncertainty of wind energy present significant challenges for grid operators, particularly in maintaining system stability and balance. Accurate short-term forecasting of wind power is therefore essential. This article introduces an innovative framework for regional wind power forecasting over short-term horizons (1–6 h), employing a novel Automated Deep Learning regression framework called WindDragon. Specifically designed to process wind speed maps, WindDragon automatically creates Deep Learning models leveraging Numerical Weather Prediction (NWP) data to deliver state-of-the-art wind power forecasts. We conduct extensive evaluations on data from France for the year 2020, benchmarking WindDragon against a diverse set of baselines, including both deep learning and traditional methods. The results demonstrate that WindDragon achieves substantial improvements in forecast accuracy over the considered baselines, highlighting its potential for enhancing grid reliability in the face of increased wind power integration.

This study shows the impact of black carbon (BC) aerosol atmospheric rivers (AAR) on the Antarctic Sea ice retreat. We detect that a higher number of BC AARs arrived in the Antarctic region due to increased anthropogenic wildfire activities in 2019 in the Amazon compared to 2018. Our analyses suggest that the BC AARs led to a reduction in the sea ice albedo, increased the amount of sunlight absorbed at the surface, and a significant reduction of sea ice over the Weddell, Ross Sea (Ross), and Indian Ocean (IO) regions in 2019. The Weddell region experienced the largest amount of sea ice retreat ($ \sim \mathrm{33,000} $ km2) during the presence of BC AARs as compared to $ \sim \mathrm{13,000} $ km2 during non-BC days. We used a suite of data science techniques, including random forest, elastic net regression, matrix profile, canonical correlations, and causal discovery analyses, to discover the effects and validate them. Random forest, elastic net regression, and causal discovery analyses show that the shortwave upward radiative flux or the reflected sunlight, temperature, and longwave upward energy from the earth are the most important features that affect sea ice extent. Canonical correlation analysis confirms that aerosol optical depth is negatively correlated with albedo, positively correlated with shortwave energy absorbed at the surface, and negatively correlated with Sea Ice Extent. The relationship is stronger in 2019 than in 2018. This study also employs the matrix profile and convolution operation of the Convolution Neural Network (CNN) to detect anomalous events in sea ice loss. These methods show that a higher amount of anomalous melting events were detected over the Weddell and Ross regions.

Ambient air pollution remains a global challenge, with adverse impacts on health and the environment. Addressing air pollution requires reliable data on pollutant concentrations, which form the foundation for interventions aimed at improving air quality. However, in many regions, including the United Kingdom, air pollution monitoring networks are characterized by spatial sparsity, heterogeneous placement, and frequent temporal data gaps, often due to issues such as power outages. We introduce a scalable data-driven supervised machine learning model framework designed to address temporal and spatial data gaps by filling missing measurements within the United Kingdom. The machine learning framework used is LightGBM, a gradient boosting algorithm based on decision trees, for efficient and scalable modeling. This approach provides a comprehensive dataset for England throughout 2018 at a 1 km2 hourly resolution. Leveraging machine learning techniques and real-world data from the sparsely distributed monitoring stations, we generate 355,827 synthetic monitoring stations across the study area. Validation was conducted to assess the model’s performance in forecasting, estimating missing locations, and capturing peak concentrations. The resulting dataset is of particular interest to a diverse range of stakeholders engaged in downstream assessments supported by outdoor air pollution concentration data for nitrogen dioxide (NO2), Ozone (O3), particulate matter with a diameter of 10 μm or less (PM10), particulate matter with a diameter of 2.5 μm or less PM2.5, and sulphur dioxide (SO2), at a higher resolution than was previously possible.

Understanding the complex dynamics of climate patterns under different anthropogenic emissions scenarios is crucial for predicting future environmental conditions and formulating sustainable policies. Using Dynamic Mode Decomposition with control (DMDc), we analyze surface air temperature patterns from climate simulations to elucidate the effects of various climate-forcing agents. This improves upon previous DMD-based methods by including forcing information as a control variable. Our study identifies both common climate patterns, like the North Atlantic Oscillation and El Niño Southern Oscillation, and distinct impacts of aerosol and carbon emissions. We show that these emissions’ effects vary with climate scenarios, particularly under conditions of higher radiative forcing. Our findings confirm DMDc’s utility in climate analysis, highlighting its role in extracting modes of variability from surface air temperature while controlling for emissions contributions and exposing trends in these spatial patterns as forcing scenarios change.

Forests play a crucial role in the Earth’s system processes and provide a suite of social and economic ecosystem services, but are significantly impacted by human activities, leading to a pronounced disruption of the equilibrium within ecosystems. Advancing forest monitoring worldwide offers advantages in mitigating human impacts and enhancing our comprehension of forest composition, alongside the effects of climate change. While statistical modeling has traditionally found applications in forest biology, recent strides in machine learning and computer vision have reached important milestones using remote sensing data, such as tree species identification, tree crown segmentation, and forest biomass assessments. For this, the significance of open-access data remains essential in enhancing such data-driven algorithms and methodologies. Here, we provide a comprehensive and extensive overview of 86 open-access forest datasets across spatial scales, encompassing inventories, ground-based, aerial-based, satellite-based recordings, and country or world maps. These datasets are grouped in OpenForest, a dynamic catalog open to contributions that strives to reference all available open-access forest datasets. Moreover, in the context of these datasets, we aim to inspire research in machine learning applied to forest biology by establishing connections between contemporary topics, perspectives, and challenges inherent in both domains. We hope to encourage collaborations among scientists, fostering the sharing and exploration of diverse datasets through the application of machine learning methods for large-scale forest monitoring. OpenForest is available at the following url: https://github.com/RolnickLab/OpenForest.

Shifting to cycling in urban areas reduces greenhouse gas emissions and improves public health. Access to street-level data on bicycle traffic would assist cities in planning targeted infrastructure improvements to encourage cycling and provide civil society with evidence to advocate for cyclists’ needs. Yet, the data currently available to cities and citizens often only comes from sparsely located counting stations. This paper extrapolates bicycle volume beyond these few locations to estimate street-level bicycle counts for the entire city of Berlin. We predict daily and average annual daily street-level bicycle volumes using machine-learning techniques and various data sources. These include app-based crowdsourced data, infrastructure, bike-sharing, motorized traffic, socioeconomic indicators, weather, holiday data, and centrality measures. Our analysis reveals that crowdsourced cycling flow data from Strava in the area around the point of interest are most important for the prediction. To provide guidance for future data collection, we analyze how including short-term counts at predicted locations enhances model performance. By incorporating just 10 days of sample counts for each predicted location, we are able to almost halve the error and greatly reduce the variability in performance among predicted locations.

Simulated data from the Coupled Model Intercomparison Project Phase 6 (CMIP6) has been very important for climate science research, as they can provide wide spatio-temporal coverage to address data deficiencies in both present and future scenarios. However, these physics-based models require a huge amount of high-performance computing (HPC) resources. As an alternative approach, researchers are exploring if such simulated data can be generated by Generative Machine Learning models. In this work, we develop a model based on Pix2Pix conditional Generative Adversarial Network (cGAN), which can generate high-resolution spatial maps of global sea surface temperature (SST) using comparatively less computing power and time. We have shown that the maps generated by these models have similar statistical characteristics as the CMIP6 model simulations. Notably, we trained and validated our cGAN model on completely distinct time periods across all ensemble members of the EC-Earth3-CC and CMCC-CM2-SR5 CMIP6 models, demonstrating satisfactory results and confirming the generalizability of our proposed model.

The increasing size and severity of wildfires across the western United States have generated dangerous levels of PM2.5 concentrations in recent years. In a changing climate, expanding the use of prescribed fires is widely considered to be the most robust fire mitigation strategy. However, reliably forecasting the potential air quality impact from prescribed fires, which is critical in planning the prescribed fires’ location and time, at hourly to daily time scales remains a challenging problem. In this paper, we introduce a spatio-temporal graph neural network (GNN)-based forecasting model for hourly PM2.5 predictions across California. Utilizing a two-step approach, we use our forecasting model to predict the net and ambient PM2.5 concentrations, which are used to estimate wildfire contributions. Integrating the GNN-based PM2.5 forecasting model with simulations of historically prescribed fires, we propose a novel framework to forecast their air quality impact. This framework determines that March is the optimal month for implementing prescribed fires in California and quantifies the potential air quality trade-offs involved in conducting more prescribed fires outside the peak of the fire season.

Federal and local agencies have identified a need to create building databases to help ensure that critical infrastructure and residential buildings are accounted for in disaster preparedness and to aid the decision-making processes in subsequent recovery efforts. To respond effectively, we need to understand the built environment—where people live, work, and the critical infrastructure they rely on. Yet, a major discrepancy exists in the way data about buildings are collected across the United SStates There is no harmonization in what data are recorded by city, county, or state governments, let alone at the national scale. We demonstrate how existing open-source datasets can be spatially integrated and subsequently used as training for machine learning (ML) models to predict building occupancy type, a major component needed for disaster preparedness and decision -making. Multiple ML algorithms are compared. We address strategies to handle significant class imbalance and introduce Bayesian neural networks to handle prediction uncertainty. The 100-year flood in North Carolina is provided as a practical application in disaster preparedness.

Prediction of dynamic environmental variables in unmonitored sites remains a long-standing challenge for water resources science. The majority of the world’s freshwater resources have inadequate monitoring of critical environmental variables needed for management. Yet, the need to have widespread predictions of hydrological variables such as river flow and water quality has become increasingly urgent due to climate and land use change over the past decades, and their associated impacts on water resources. Modern machine learning methods increasingly outperform their process-based and empirical model counterparts for hydrologic time series prediction with their ability to extract information from large, diverse data sets. We review relevant state-of-the art applications of machine learning for streamflow, water quality, and other water resources prediction and discuss opportunities to improve the use of machine learning with emerging methods for incorporating watershed characteristics and process knowledge into classical, deep learning, and transfer learning methodologies. The analysis here suggests most prior efforts have been focused on deep learning frameworks built on many sites for predictions at daily time scales in the United States, but that comparisons between different classes of machine learning methods are few and inadequate. We identify several open questions for time series predictions in unmonitored sites that include incorporating dynamic inputs and site characteristics, mechanistic understanding and spatial context, and explainable AI techniques in modern machine learning frameworks.