Changing wildfire regimes in the western US and other fire-prone regions pose considerable risks to human health and ecosystem function. However, our understanding of wildfire behavior is still limited by a lack of data products that systematically quantify fire spread, behavior and impacts. Here we develop a novel object-based system for tracking the progression of individual fires using 375 m Visible Infrared Imaging Radiometer Suite active fire detections. At each half-daily time step, fire pixels are clustered according to their spatial proximity, and are either appended to an existing active fire object or are assigned to a new object. This automatic system allows us to update the attributes of each fire event, delineate the fire perimeter, and identify the active fire front shortly after satellite data acquisition. Using this system, we mapped the history of California fires during 2012–2020. Our approach and data stream may be useful for calibration and evaluation of fire spread models, estimation of near-real-time wildfire emissions, and as means for prescribing initial conditions in fire forecast models.

LargeFires_YYYY.gpkg: Time series of large fire attributes and perimeter in a year LargeFires_2012-2020.gpkg: Time series of attributes and perimeter for all large fires during 2012-2020

Wildfire ignition, intensity, and spread rates are tightly linked with water cycle extremes. The science of wildfire prediction has traditionally encompassed the use of physical and empirical models to quantify the direction and speed of fire spread, plume injection and fire-aerosol impacts on atmospheric composition, predictions of fire season severity on subseasonal-to-seasonal (S2S) time scales, and assessment of the spatial and temporal patterns of fire risk across landscapes. Together with expanding observation networks, machine learning and artificial intelligence (AI) have the potential to revolutionize the application of such models for fire science, saving lives, protecting critical infrastructure, and providing more accurate estimates of wildfire-climate feedbacks.

Wildfires and their emissions reduce air quality in many regions of the world, contributing to thousands of premature deaths each year. Smoke forecasting systems have the potential to improve health outcomes by providing future estimates of surface aerosol concentrations (and health hazards) over a period of several days. In most operational smoke forecasting systems, fire emissions are assumed to remain constant during the duration of the weather forecast and are initialized using satellite observations. Recent work suggests that it may be possible to improve these models by predicting the temporal evolution of emissions. Here, we develop statistical models to predict fire activity one to five days into the future using Moderate Resolution Imaging Spectroradiometer (MODIS) satellite fire counts and weather data from ERA-interim reanalysis. Our predictive framework consists of two-Poisson regression models that separately represent new ignitions and the dynamics of existing fires on a coarse resolution spatial grid. We use ten years of active fire detections in Alaska to develop the model and use a cross-validation approach to evaluate model performance. Our results show that regression methods are significantly more accurate in predicting daily fire activity than persistence-based models (which suffer from an overestimation of fire counts by not accounting for fire extinction), with vapor pressure deficit being particularly effective as a single weather-based predictor in the regression approach.

Abstract Fire emissions of gases and aerosols alter atmospheric composition and have substantial impacts on climate, ecosystem function, and human health. Warming climate and human expansion in fire‐prone landscapes exacerbate fire impacts and call for more effective management tools. Here we developed a global fire forecasting system that predicts monthly emissions using past fire data and climate variables for lead times of 1 to 6 months. Using monthly fire emissions from the Global Fire Emissions Database (GFED) as the prediction target, we fit a statistical time series model, the Autoregressive Integrated Moving Average model with eXogenous variables (ARIMAX), in over 1,300 different fire regions. Optimized parameters were then used to forecast future emissions. The forecast system took into account information about region‐specific seasonality, long‐term trends, recent fire observations, and climate drivers representing both large‐scale climate variability and local fire weather. We cross‐validated the forecast skill of the system with different combinations of predictors and forecast lead times. The reference model, which combined endogenous and exogenous predictors with a 1 month forecast lead time, explained 52% of the variability in the global fire emissions anomaly, considerably exceeding the performance of a reference model that assumed persistent emissions during the forecast period. The system also successfully resolved detailed spatial patterns of fire emissions anomalies in regions with significant fire activity. This study bridges the gap between the efforts of near‐real‐time fire forecasts and seasonal fire outlooks and represents a step toward establishing an operational global fire, smoke, and carbon cycle forecasting system.

Fires in boreal forests of Alaska are changing, threatening human health and ecosystems. Given expected increases in fire activity with climate warming, insight into the controls on fire size from the time of ignition is necessary. Such insight may be increasingly useful for fire management, especially in cases where many ignitions occur in a short time period. Here we investigated the controls and predictability of final fire size at the time of ignition. Using decision trees, we show that ignitions can be classified as leading to small, medium or large fires with 50.4±5.2% accuracy. This was accomplished using two variables: vapour pressure deficit and the fraction of spruce cover near the ignition point. The model predicted that 40% of ignitions would lead to large fires, and those ultimately accounted for 75% of the total burned area. Other machine learning classification algorithms, including random forests and multi-layer perceptrons, were tested but did not outperform the simpler decision tree model. Applying the model to areas with intensive human management resulted in overprediction of large fires, as expected. This type of simple classification system could offer insight into optimal resource allocation, helping to maintain a historical fire regime and protect Alaskan ecosystems.

Advances in both hardware and software are enabling rapid proliferation of in situ plankton imaging methods, requiring more effective machine learning approaches to image classification. Deep Learning methods, such as convolutional neural networks (CNNs), show marked improvement over traditional feature-based supervised machine learning algorithms, but require careful optimization of hyperparameters and adequate training sets. Here, we document some best practices in applying CNNs to zooplankton and marine snow images and note where our results differ from contemporary Deep Learning findings in other domains. We boost the performance of CNN classifiers by incorporating metadata of different types and illustrate how to assimilate metadata beyond simple concatenation. We utilize both geotemporal (e.g., sample depth, location, time of day) and hydrographic (e.g., temperature, salinity, chlorophyll a) metadata and show that either type by itself, or both combined, can substantially reduce error rates. Incorporation of context metadata also boosts performance of the feature-based classifiers we evaluated: Random Forest, Extremely Randomized Trees, Gradient Boosted Classifier, Support Vector Machines, and Multilayer Perceptron. For our assessments, we use an original data set of 350,000 in situ images (roughly 50% marine snow and 50% non-snow sorted into 26 categories) from a novel in situ Zooglider. We document asymptotically increasing performance with more computationally intensive techniques, such as substantially deeper networks and artificially augmented data sets. Our best model achieves 92.3% accuracy with our 27-class data set. We provide guidance for further refinements that are likely to provide additional gains in classifier accuracy.

This paper describes three metrics used to asses the filter diversity learned by convolutional neural networks during supervised classification. As our testbed we use four different data sets, including two subsets of ImageNet and two planktonic data sets collected by scientific instruments. We investigate the correlation between our devised metrics and accuracy, using normalization and regularization to alter filter diversity. We propose that these metrics could be used to improve training CNNs. Three potential applications are determining the best preprocessing method for non-standard data sets, diagnosing training efficacy, and predicting performance in cases where validation data is expensive or impossible to collect.