PaperlayerView exact revision on arXiv

Source-linked AI summary

Next Day Wildfire Spread: A Machine Learning Data Set to Predict Wildfire Spreading from Remote-Sensing Data

Fantine Huot, R. Lily Hu, Nita Goyal, Tharun Sankar, Matthias Ihme, Yi-Fan Chen

arXiv:2112.02447v2cs.CV

TL;DR

Day-ahead wildfire spread prediction matters for land management and emergency preparedness. The paper introduces a large-scale U.S. dataset of remote-sensing variables and fire masks, and compares a convolutional autoencoder with logistic regression and random forest. It presents the dataset as a resource for benchmarking and further prediction research, while noting that its 1 km daily resolution does not provide tactical-scale information.

  • Problem

    Wildfire spread prediction is relevant to land management and emergency response, while existing satellite fire datasets often lack the aligned explanatory variables and temporal 2D information needed to characterize spread.

    Source: §I · p. 1

  • Method

    The authors aggregate nearly a decade of U.S. fire records and aligned environmental variables into a 1 km, 2D dataset, then compare a convolutional autoencoder with logistic regression and random forest.

    Source: §I · p. 1, A. Motivation · p. 2, A. Models for segmentation · p. 4

  • Results

    The dataset contains 18,545 fire events and is used to demonstrate deep-learning wildfire prediction and compare model performance on day-ahead spread.

    Source: §I · p. 1, §VI · p. 6

  • Takeaways & Limitations

    The publicly available dataset and GEE aggregation code are offered for benchmarking, model comparison, and further development of remote-sensing wildfire predictions.

    Source: §I · p. 2, §VI · p. 6

    MODIS’s 1 km spatial and daily temporal resolution limits prediction resolution and does not provide the fine-scale information required for tactical decision-making.

    Source: §VI · p. 6

Abstract

from arXiv · show

Predicting wildfire spread is critical for land management and disaster preparedness. To this end, we present `Next Day Wildfire Spread,' a curated, large-scale, multivariate data set of historical wildfires aggregating nearly a decade of remote-sensing data across the United States. In contrast to existing fire data sets based on Earth observation satellites, our data set combines 2D fire data with multiple explanatory variables (e.g., topography, vegetation, weather, drought index, population density) aligned over 2D regions, providing a feature-rich data set for machine learning. To demonstrate the usefulness of this data set, we implement a neural network that takes advantage of the spatial information of this data to predict wildfire spread. We compare the performance of the neural network with other machine learning models: logistic regression and random forest. This data set can be used as a benchmark for developing wildfire propagation models based on remote sensing data for a lead time of one day.

I. INTRODUCTION

The paper motivates day-ahead fire-spread prediction for land management and emergency response, then introduces a public, feature-rich U.S. wildfire dataset and demonstrates its use with machine-learning models.

I. INTRODUCTION

  • Day-ahead spread predictions are presented as valuable for emergency response, while wildfire likelihood estimates can inform land management and preparedness.The introduction also connects high-likelihood regions with targeted prevention measures.Source: §I · p. 1
  • The paper presents a public, large-scale dataset combining nearly a decade of U.S. wildfire records with 11 aligned variables over 2D regions at 1 km resolution.It contains 18,545 fire events with snapshots at time t and t + 1 day.Source: §I · p. 1
  • Existing satellite fire datasets generally lack the combination of 1 km resolution, spread-pattern information, and explanatory variables needed for fire prediction.Prior combined datasets were not publicly released or were only partially released.Source: §I · p. 1
  • The authors release the dataset and its GEE aggregation code for benchmarking, model comparison, and further wildfire-prediction research.The code is offered for users to adapt to their use cases.Source: §I · p. 2
  • A convolutional autoencoder uses the dataset’s 2D information to predict next-day spread, with logistic regression and random forest as comparison models.The non-deep-learning models are also compared with the neural network on this task.Source: §I · p. 2

II. DATA SET AGGREGATION

The dataset is assembled in Google Earth Engine by aligning fire records and environmental variables across space and time. Its open aggregation code supports adaptation to other data, regions, and sampling choices.

A. Motivation

  • Each example covers a 64 km × 64 km region at 1 km resolution and pairs fire masks at time t and t + 1 day.The figure presents locations and times in rows and variables in columns; uncertain labels include cloud-covered or unprocessed data.Source: A. Motivation · p. 2, Figure 1 · p. 3

II. DATA SET AGGREGATION

  • GEE combines fire masks with topography, weather, vegetation, drought, energy-release, and population-density variables aligned in location and time.Population density serves as a proxy for anthropogenic activity; vegetation, drought, and weather variables provide information relevant to fuel properties.Source: A. Motivation · p. 2

A. Motivation

  • The authors prioritize regularly updated, publicly available sources to support future predictions, including near-real-time use.Open terms of use are intended to reduce licensing constraints for future prevention and management tools.Source: A. Motivation · pp. 2–3
  • The released data and GEE export code can be resampled or adapted to different regions, scales, time intervals, and additional data sources.The paper notes an example using weekly time sequences to map total burn area.Source: A. Motivation · p. 3

B. Data Sources from Earth Engine

The authors compile the dataset from multiple Earth Engine sources selected for broad geographic and historical coverage and reduced missing data. The released export code lets users change the sources to suit their needs.

B. Data Sources from Earth Engine

  • The selected sources provide extensive geographic and historical coverage while reducing missing data, and users can substitute other sources through the export code.The authors state that the dataset variables come directly from the cited sources rather than being calculated by the authors.Source: B. Data Sources from Earth Engine · p. 3
  • The source set combines daily 1 km MODIS fire masks with 30 m SRTM topography and 4 km GRIDMET weather and drought data.The weather data cover temperature, precipitation, winds, and humidity; drought indices are sampled every five days.Source: B. Data Sources from Earth Engine · p. 3
  • The dataset also uses VIIRS vegetation indices sampled at 0.5 km resolution and GPWv4 population density sampled at 1 km resolution.The cited vegetation product is sampled every eight days; population density is sampled every five years.Source: B. Data Sources from Earth Engine · p. 3

C. Data Aggregation

The authors sample active-fire regions, align heterogeneous data sources, and form training, validation, and test splits. Preprocessing clips and normalizes features and augments examples with randomly cropped regions.

C. Data Aggregation

  • Samples are taken at active-fire locations and times, retaining regions with at least one fire area at time t and treating fires more than 10 km apart as distinct.The input fire mask at time t is paired with the target mask at t + 1 day.Source: C. Data Aggregation · p. 3
  • The contiguous-U.S. data span 2012–2020 and are split by week into training, evaluation, and test sets at an 8:1:1 ratio, with a one-day buffer between sampled weeks.The period was selected based on data availability.Source: C. Data Aggregation · p. 4
  • All inputs are aligned to 1 km; slower-changing variables use the latest timestamp at t, while weather values are averaged over the day at t.Topography and vegetation are downsampled, and weather and drought data use bicubic interpolation.Source: C. Data Aggregation · p. 4
  • Of 18,545 samples, fire area increases from t to t + 1 day in 58% (10,798), decreases in 39% (7,191), and stays the same in the remainder.These are sample-level changes in fire size.Source: C. Data Aggregation · p. 4

III. DATA PREPROCESSING

  • Features other than fire masks are clipped and then normalized separately using training-set means and standard deviations.Clipping addresses extreme values and limits gradient problems; thresholds use physical knowledge or the 0.1% and 99.9% percentiles.Source: §III · p. 4
  • Random 32 km × 32 km crops from the original 64 km × 64 km regions augment examples whose fires are generally centered.The cropping varies fire locations in the machine-learning inputs.Source: §III · p. 4

IV. MACHINE LEARNING APPLICATION

The application frames next-day fire spread as spatial segmentation from prior-day fire masks and remote-sensing features, comparing a convolutional autoencoder with pixel-wise logistic regression and random forest. Training accounts for class imbalance and selects models by AUC PR.

A. Models for segmentation

  • The convolutional autoencoder treats remote-sensing variables as a multi-channel image and the fire mask as a segmentation map for predicting next-day fire.The task is framed as classifying each area as fire or no fire from the previous day's fire location and other features.Source: A. Models for segmentation · p. 4
  • Logistic regression and random forest classify each 1 km × 1 km area using its features and those of its eight neighboring areas, without the autoencoder's 2D modeling.Source: A. Models for segmentation · p. 4

B. Training details and hyperparameter tuning

  • Weighted cross-entropy addresses class imbalance; the deep model uses crop, flip, and rotation augmentation, while grid search selects the model with the best AUC PR.AUC PR is used instead of accuracy because the data contain more no-fire than fire samples.Source: B. Training details and hyperparameter… · pp. 4–5

V. RESULTS

The results compare next-day fire predictions from regions with an ongoing fire against persistence and conventional machine-learning baselines, then examine historical cases and coarser output resolutions. The neural network leads the 1 km² AUC comparison, while larger prediction regions raise reported metrics but reduce localization.

A. Prediction Results

  • 28.4% AUC: the neural network ranks above random forest and logistic regression, and all trained models exceed the 11.5% persistence baseline.The non-deep models' AUCs are within 3% of each other and at least 6% below the neural network's.Source: A. Prediction Results · p. 5
  • 33.6% precision and 43.1% recall: the neural network has nearly logistic regression's precision, while random forest has higher recall but lower precision.Logistic regression has 32.5% precision and 35.3% recall; random forest has 26.3% precision and 46.9% recall.Source: A. Prediction Results · p. 5
  • Predictions roughly match target locations but are smoother, can merge nearby fires, and may carry forward a previous small fire absent from the target.Segmentation errors also include missed small fires and misclassified pixels at fire boundaries and between nearby fire areas.Source: A. Prediction Results · p. 5, Figure 3 · p. 5
  • On six historical fires, AUC was 38.7%, above the test-set AUC, with similar precision and recall; larger fires were recognized, but outlines remained imperfect.The examples include the Bighorn, CZU lightning complex, East Canyon, Evans Canyon, Grizzly Creek, and LNU lightning complex fires.Source: A. Prediction Results · p. 5, Figure 4 · p. 5
  • 66.3% AUC: the best precision/recall trade-off occurs when output areas are eight times larger than input areas, but coarser predictions reduce localization and can miss small fires.The lower-resolution experiment labels an area as fire when any part of it contains fire; its predictions can connect smaller fire areas.Source: A. Prediction Results · pp. 5–6, Figure 5 · p. 6

B. Feature Analysis

Removing the current fire location most reduces AUC, while paired-feature tests identify vegetation and elevation as the strongest companions to the previous fire mask. The study also identifies resolution and MODIS-label constraints and releases modular aggregation code for future extensions.

B. Feature Analysis

  • Removing the current fire location causes the largest AUC decrease; removing other individual features produces similar performance to using all features.The authors note that correlated features, random-cropping variation, and reduced overfitting may help explain similar or slightly improved results after removal.Source: B. Feature Analysis · p. 6
  • With the previous fire mask retained, vegetation and elevation produce the best paired-feature performance; alone, only that mask yields nonzero recall.The previous-fire-mask-only model has 49.6% recall, 29.2% precision, and 22.9% AUC.Source: B. Feature Analysis · p. 6

VI. CONCLUSIONS

  • The dataset combines nearly a decade of remote-sensing data with topography, weather, drought index, vegetation, population density, and historical fire records.The authors present it as an open resource for research and model comparisons.Source: §VI · p. 6
  • MODIS's 1 km spatial and daily temporal resolution does not provide the fine-scale information required for tactical decision-making.GRIDMET weather data also lack the spatial and temporal resolution needed to capture local wind patterns that drive fire spreading.Source: §VI · pp. 6–7
  • MODIS labels miss late-afternoon sampling, do not distinguish wildfires from prescribed fires, and can contain artificial spread patterns despite uncertain-label handling.Fire-suppression efforts also affect the labels and may influence predictive capability; population density accounts for this only to some degree.Source: §VI · p. 7
  • The released modular GEE aggregation code can change data sources, geography, time period, and spatial or temporal sampling.The authors describe extensions to global or longer-term data, recent-fire weighting, simulation data, and other hazards as possible future directions.Source: §VI · p. 7

APPENDIX A MACHINE LEARNING MODEL DETAILS

The appendix describes a convolutional segmentation network and its grid-search tuning, alongside pixel-wise logistic regression and random forest baselines. The non-deep models use neighboring pixels for spatial context and validation-selected settings.

A. Deep Learning Model

  • The segmentation network uses 3 × 3 convolutions, 2 × 2 pooling, dropout 0.1, and a one-filter output layer; training uses learning rate 0.0001 and fire-class weight 3.Most convolutional blocks have 16 filters, while the middle two residual blocks have 32.Source: A. Deep Learning Model · p. 7
  • Grid search selects the deep model with the best validation AUC PR, varying first-block filters, encoder residual-block count, and symmetric decoder filters.The first convolutional block explores 16, 32, or 64 filters; the encoder uses one to four residual blocks, with filters doubling in successive blocks.Source: A. Deep Learning Model · p. 7
  • The deep-model search also varies batch size, learning rate, dropout, and L1 and L2 regularization.The explored batch sizes are 32, 64, 128, and 256; learning rates range from 0.01 to 0.00001.Source: A. Deep Learning Model · p. 7

B. Non-Deep Learning Models

  • Logistic regression and random forest perform pixel-wise segmentation, using each pixel's features and those of its eight neighbors; the image border is excluded.The training set is balanced to approximately 1:1 fire and non-fire examples, while metrics are calculated on the unbalanced test set; unknown labels are ignored.Source: B. Non-Deep Learning Models · p. 7
  • Logistic regression uses default hyperparameters; random-forest depth 15 is selected by validation AUC PR, and the no-fire class weights are 7 and 5, respectively.The class weights are selected by searching values from 1 to 8; the random-forest depth search spans 1 to 20.Source: B. Non-Deep Learning Models · pp. 7–8

C. Random Forest Feature Analysis

The random-forest feature-importance analysis ranks the input pixel's previous fire mask highest, followed by neighboring previous-fire masks and ERC.

C. Random Forest Feature Analysis

  • The input pixel's previous fire mask is the most important random-forest feature, followed by neighboring pixels' previous fire masks.Source: C. Random Forest Feature Analysis · p. 8
  • ERC is the next most important feature after the previous fire masks, while the remaining features have very low importance.Source: C. Random Forest Feature Analysis · p. 8
  • Precipitation for all pixels is the least important feature; feature importance is measured by mean decrease in Gini impurity across tree splits.Features are numbered in row-major order across the 3 × 3 pixel neighborhood, with pixel 5 as the input pixel.Source: C. Random Forest Feature Analysis · p. 8

D. Logistic Regression Feature Analysis

The logistic-regression feature analysis reports the 20 largest-magnitude coefficients and finds the previous fire mask to be its most important feature, as in the other models.

D. Logistic Regression Feature Analysis

  • The 20 largest-magnitude logistic-regression coefficients are reported in Table A-2, with features indexed row-major across the nine pixels in each 3 × 3 example.Pixel 5 is the input pixel; for example, “Min temp. (3)” denotes the top-right neighbor’s minimum temperature.Source: D. Logistic Regression Feature Analysis · p. 8, Table A-2 · p. 11
  • The previous fire mask is the most important feature for logistic regression, similar to the other models.The feature-ablation results are reported in Table A-3.Source: D. Logistic Regression Feature Analysis · p. 8, Table A-3 · p. 11
  • The appendix also includes tables for removing one feature and for retaining the previous fire mask with one other feature.These analyses are titled in Tables IV and V; no outcomes are specified in the supplied passages.Source: Table IV · p. 10, Table V · p. 10

APPENDIX B ADDITIONAL VISUALIZATIONS

Appendix B provides additional examples of fire-spread predictions at the standard and coarser resolutions, including examples captioned as better or worse.

APPENDIX B ADDITIONAL VISUALIZATIONS

  • Figures B-1 and B-2 show fire-spread predictions, while Figures B-3 and B-4 show predictions at coarser resolution.Source: Appendix B · p. 12, Figure B-1 · p. 12, Figure B-2 · p. 12
  • Figure B-1 shows examples captioned as better segmentation predictions, with each row representing a sample.Source: Figure B-1 · p. 12
  • Figure B-2 shows examples captioned as worse segmentation predictions, with each row representing a sample.Source: Figure B-2 · p. 12
  • Figure B-3 shows better coarse-segmentation examples, with each pair of rows representing a sample.Source: Figure B-3 · p. 13
  • Figure B-4 shows worse coarse-segmentation examples, with each pair of rows representing a sample.Source: Figure B-4 · p. 13
Loading 2112.02447v2…