dominiktraxl/firetracks

The FireTracks Scientific Dataset

Content

• About
• Overview
• Installation
• Getting Orginal Data
• Creating the FireTracks Scientific Dataset
• Loading the FireTracks Scientific Dataset Using Python
• Converting HDF5 files to CSV files (CLI)
• Data Content
  • Active Fire Events Table
  • Active Fire Land Cover Table
  • Spatiotemporal Fire Component Table
  • Spatiotemporal Fire Component Land Cover Table
  • Spatiotemporal Fire Component GeoPackage
  • Spatiotemporal Fire Component (Per Time-Slice) GeoPackage
• Comparison between Active Fire Events Table and MCD14ML
• Comparison between Spatiotemporal Fire Component Table and Global Fire Atlas

About

This is a collection of python scripts that produces the FireTracks Scientific Dataset. The dataset is based on the MODIS Fire Products MOD14A1/ MYD14A1 and the MODIS Land Cover Product MCD12Q1. It entails HDF5-tables/GeoPackages of active fire events, spatiotemporal components of fire events and associated land cover information.

Note: the quality of the FireTracks Scientific Dataset depends entirely on the underlying MOD14A1/MYD14A1/MCD12Q1 datasets, and it is strongly recommended to read the respective user guides before using the data:

• MOD14A1 User Guide
• MYD14A1 User Guide
• MCD12Q1 User Guide

Download: the FireTracks Scientific Dataset, processed for the years 2002-2020, can be downloaded here: https://doi.org/10.5281/zenodo.4461575

Overview

The FireTracks Scientific Dataset is derived from the Aqua and Terra Moderate Resolution Imaging Spectroradiometer (MODIS) Collection 6 (C6) Thermal Anomalies and Fire Data (MOD14A1 and MYD14A1). These satellite products include 1-km gridded fire masks over daily (24-hour) compositing periods on a global scale, from the year 2000 to the present.

Part of the FireTracks Scientific Dataset is the Active Fire Events Table, which contains single pixel fire events in table format extracted from the fire masks of both satellite products. The Active Fire Land Cover Table provides land cover information for each fire event of the active fire events table, derived from the MODIS product MCD12Q1.

The centerpiece of the FireTracks Scientific Dataset is the Spatiotemporal Fire Component Table, providing summarizing characteristics of spatiotemporally tracked fire components. It is derived from the active fire events table by performing a spatiotemporal clustering of the single pixel fire events. A spatiotemporal cluster is identified as the union of nearest neighbors of fire events in the discrete space-time grid prescribed by the resolution of the MODIS fire masks (temporal resolution: 1 day, spatial resolution: 1 km). We consider a 3d-Moore neighborhood for the clustering (26 adjacent cells, 3*3*3 [x*y*time] grid cell box with the fire event in the center). Below, you see an example of a spatiotemporal fire component comprised of 5 single pixel fire events, evolving over 4 consecutive days. At the top of this webpage, you see the temporal evolution of one of the largest spatiotemporal fire components found in the data. It was recorded in the summer of 2018 in California, lasting 48 days and burning a total of 1893 km^2 with an integrated radiative power of 1.614.436 MW.

The active fire events table and the spatiotemporal fire component table are linked to each other via the cp column of the events table (the component membership label) and the cp column of the component table (the component index).

Land cover information for each spatiotemporal fire component is summarized in the Spatiotemporal Fire Component Land Cover Table.

Additionally, the FireTracks Scientific Dataset provides polygon vector data for each spatiotemporal fire component (Spatiotemporal Fire Component GeoPackage), as well as for each time-slice of every spatiotemporal fire component (Spatiotemporal Fire Component (Per Time-Slice) GeoPackage).

For further details of the provided HDF5-tables/GeoPackages, see Data Content.

Installation

No installation is required.

However, to run the python scripts the following packages are required:

• pyhdf
• numpy
• scipy
• pandas
• tables
• deepgraph
• geopandas
• shapely

You can use conda to set up an environment and install all dependencies via

$ conda create -n FT
$ conda activate FT
$ conda install -c conda-forge pyhdf numpy scipy pandas pytables deepgraph geopandas shapely

Getting Original Data

To create the FireTracks Scientific Dataset, you need to first download the original data that it is based on: MOD14A1, MYD14A1 and MCD12Q1.

You can download the data here: https://search.earthdata.nasa.gov/search

Search for the products (MOD14A1, MYD14A1 and MCD12Q1), select the spatial and temporal domain you're interested in, and then download the data.

The data must be stored in their respective folders (MOD14A1, MYD14A1, MCD12Q1) in the same directory as the python scripts. There should be no subdirectories within the data folders.

Note: to associate land cover information with fire events, you need to download the MCD12Q1 files from one year before the actual occurrence of the fire events.

Creating the FireTracks Scientific Dataset

1. Download the python scripts contained in this repository

2. Store the python scripts in the directory that contains the folders with the downloaded original data (MOD14A1, MYD14A1, MCD12Q1)

3. cd into the directory containing the scripts and the downloaded original data folders, then run the scripts in the indicated order:

    $ python 01_create_fire_event_table.py
    ...
    $ python 07_create_component_polygons.py
   

Note: some scripts have positional and/or optional arguments. Use

$ python *script*.py -h

for more information.

Loading the FireTracks Scientific Dataset Using Python

Using Python Pandas/GeoPandas, you can load the HDF5/GeoPackage Data via the following commands:

import os

import pandas as pd
import geopandas as gpd

# active fire events table
v = pd.read_hdf('v.h5')

# active fire land cover table
v_lc = pd.read_hdf('v_LC_Type1.h5')

# spatiotemporal fire component table
cp = pd.read_hdf('cp.h5')

# spatiotemporal fire component land cover table
cp_lc = pd.read_hdf('cp_LC_Type1.h5')

# spatiotemporal fire component shapefile
cp_poly = gpd.read_file('cp_poly.gpkg')

# spatiotemporal fire component shapefile per time-slice
cpt_poly = gpd.read_file('cpt_poly.gpkg')

Note:

• v.h5 and v_*lc*.h5 are sorted and indexed by dtime, allowing for fast queries by time, e.g.:

  v_2019 = pd.read_hdf('v.h5', where='dtime >= "2019-01-01" & dtime < "2020-01-01"')
  v_lc_2019 = pd.read_hdf('v_LC_Type1.h5', where='dtime >= "2019-01-01" & dtime < "2020-01-01"')

• cp.h5 and cp_*lc*.h5 are sorted and indexed by dtime_min, allowing for fast queries by time, e.g.:

  cp_2019 = pd.read_hdf('cp.h5', where='dtime_min >= "2019-01-01" & dtime_min < "2020-01-01"')
  cp_lc_2019 = pd.read_hdf('cp_LC_Type1.h5', where='dtime_min >= "2019-01-01" & dtime_min < "2020-01-01"')

• to load specific rows of cp_poly.gpkg or cpt_poly.gpkg using geopandas, you can pass an int or slice object as argument rows to the gpd.read_file method, e.g.:

  cp_poly_selection = gpd.read_file('cp_poly.gpkg', rows=5)
  cpt_poly_selection = gpd.read_file('cpt_poly.gpkg', rows=slice(10, 20))

Converting HDF5 files to CSV files (CLI)

The h5tocsv.py file contained in this repository allows you to convert the HDF5 files of the FireTracks Scientific Dataset into CSV files using the command line. Via optional arguments, it is possible to subset the original data by time and/or columns before the conversion to CSV.

To run the conversion CLI, the following packages are required:

• pandas
• tables

You can use conda to set up an environment and install all dependencies via

$ conda create -n FTC
$ conda activate FTC
$ conda install -c conda-forge pandas pytables

To show the documentation of the CLI, type

$ python h5tocsv.py -h

This will show the help message of h5tocsv.py and exit.

The following example converts cp.h5 into cp.csv, including only components with ignition dates between 2003-01-01 and 2003-07-01 (excluded). Only the columns cp, maxFRP_sum, duration, area and country are included in the CSV file.

$ python h5tocsv.py cp.h5 cp.csv --from-time 2003-01-01 --to-time 2003-07-01 --columns cp maxFRP_sum duration area country

Data Content

Active Fire Events Table v.h5

The active fire events table is a combination of the MODIS products MOD14A1 and MYD14A1. Fire events from the fire masks of both products are extracted and stored in table format. For fires measured by both satellites, the maximum value of the individual maxFRP values is stored. Furthermore, for each fire, the minimum of the fire pixel classes of all 26 neighboring grid cells (3*3*3 [lat*lon*time] grid cell box with the fire event in the center) is recorded. Possible fire pixel classes are as follows (note that class 3 and 4 are swapped compared to the original fire pixel classes):

Class	Meaning
0	not processed (missing input data)
1	not processed (obsolete; not used since Collection 1)
2	not processed (other reason)
3	cloud (land or water)
4	non-fire water pixel
5	non-fire land pixel
6	unknown (land or water)
7	fire (low confidence, land or water)
8	fire (nominal confidence, land or water)
9	fire (high confidence, land or water)

This minimum value (neigh_int) allows us to see if there are any missing/cloud pixels in the neighborhood of a fire event.

Name	Description	Unit	Valid Range	Data Type
lat	location latitude	degress	[-180, 180]	float64
lon	location longitude	degrees	[-90, 90]	float64
x	x-coordinate on global sinusoidal MODIS grid	-	[0, 36*1200-1]	uint16
y	y-coordinate on global sinusoidal MODIS grid	-	[0, 18*1200-1]	uint16
H	horizontal MODIS tile coordinate	-	[0, 35]	uint8
V	vertical MODIS tile coordinate	-	[0, 17]	uint8
i	row coordinate of the grid cell within MODIS tile (H, V)	-	[0, 1199]	uint16
j	column coordinate of the grid cell within MODIS tile (H, V)	-	[0, 1199]	uint16
dtime	date (YYYY-MM-DD)	-	>= 2002-01-01	datetime64
conf	detection confidence [7: low, 8: nominal, 9: high]	-	[7, 9]	uint8
maxFRP	maximum fire radiative power	MW*10	>= 0	int32
satellite	which satellite detected the fire [MOD, MYD, both]	-	-	string
neigh	string representation of "neigh_int"	-	-	string
t	days since 2002-01-01	days since 2002-01-01	>= 0	uint16
country	country of occurrence	-	-	string
continent	continent of occurrence	-	-	string
neigh_int	minimum of fire pixel classes of neighboring grid cells	-	[0, 9]	uint8
gl	location ID on the global sinusoidal MODIS grid	-	[0, 36*1200*18*1200-1]	uint32
cp	component membership label	-	>= 0	int64

Active Fire Land Cover Table v_*lc*.h5

The active fire land cover table is based on the MODIS product MCD12Q1 and provides land cover information for each fire event of the active fire events table. Since the resolution of MCD12Q1 is twice as high as that of MOD14A1/MYD14A1, each fire event is associated with 4 subpixel land cover type values.

Note: Land cover values are extracted from MCD12Q1 for the year before the fire event.

Name	Description	Unit	Valid Range	Data Type
lc1	land cover type of subpixel 1 (numerical)	-	[0, 255]	uint8
lc2	land cover type of subpixel 2 (numerical)	-	[0, 255]	uint8
lc3	land cover type of subpixel 3 (numerical)	-	[0, 255]	uint8
lc4	land cover type of subpixel 4 (numerical)	-	[0, 255]	uint8
dtime	date (YYYY-MM-DD)	-	>= 2002-01-01	datetime64

Spatiotemporal Fire Component Table cp.h5

The spatiotemporal fire component table provides summarizing characteristics of spatiotemporally tracked fire components.

A fire component is defined as a coherent set of grid cells on fire. Cells are coherent if they can reach each other via nearest neighbor relations considering a 3d-Moore neighborhood (26 adjacent cells, 3*3*3 [lat*lon*time] grid cell box with the fire event in the center).

Name	Description	Unit	Valid Range	Data Type
cp	component index	-	>= 0	int64
n_nodes	number of constituent fire events	-	>= 1	int64
t_min	ignition date (days since 2002-01-01)	days since 2002-01-01	>= 0	uint16
t_max	extinction date (days since 2002-01-01)	days since 2002-01-01	>= 0	uint16
dtime_min	ignition date (YYYY-MM-DD)	-	>= 2002-01-01	datetime64
dtime_max	extinction date (YYYY-MM-DD)	-	>= 2002-01-01	datetime64
lat_mean	mean location latitude	degrees	[-180, 180]	float64
lon_mean	mean location longitude	degrees	[-90, 90]	float64
maxFRP_mean	mean maximum fire radiative power	MW*10	>= 0	float64
maxFRP_sum	sum of maximum fire radiative powers	MW*10	>= 0	float64
neigh_int_min	minimum of "neigh_int" values of constituent fire events	-	[0, 9]	uint8
neigh_min	string representation of "neigh_int_min"	-	-	string
duration	fire duration	days	>= 1	uint16
unique_gls	number of grid locations burnt	-	>= 1	uint32
area	total area burnt	km^2	>= 0.86 (1 MODIS pixel)	float64
expansion	average daily fire expansion	km^2 day^-1	> 0	float64
country	country of occurrence	-	-	string
continent	continent of occurrence	-	-	string

Spatiotemporal Fire Component Land Cover Table cp_*lc*.h5

The spatiotemporal fire component land cover table provides land cover information for each spatiotemporal fire component.

Name	Description	Unit	Valid Range	Data Type
cp	component index	-	>= 0	int64
dlc	dominant land cover type*	-	-	string
lc_X	number of subpixels burnt belonging to land cover X	-	>= 0	int64
plc_X	proportion of subpixels burnt belonging to land cover X	-	[0, 1]	float64
flc_X	number of ignition subpixels belonging to land cover X	-	>= 0	int64
dtime_min	ignition date (YYYY-MM-DD)	-	>= 2002-01-01	datetime64

*a spatiotemporal fire component has a dominant land cover type X, if at least 80% of burnt subpixels belong to land cover X. Otherwise, dlc is set to "None".

Spatiotemporal Fire Component GeoPackage cp_poly.gpkg

The spatiotemporal fire component geopackage provides polygon vector data for each spatiotemporal fire component.

Name	Description	Unit	Valid Range	Data Type
cp	component index	-	>= 0	int64
area	total area burnt	km^2	>= 0.86 (1 MODIS pixel)	float64
perimeter	final perimeter	km	>= 3.71 (1 MODIS pixel)	float64
geometry	(Multi)Polygon vector data of spatiotemporal fire component	-	-	GeometryDtype

Spatiotemporal Fire Component (Per Time-Slice) GeoPackage cpt_poly.gpkg

The spatiotemporal fire component geopackage provides polygon vector data for each time-slice of every spatiotemporal fire component.

Name	Description	Unit	Valid Range	Data Type
cp	component index	-	>= 0	int64
t	days since 2002-01-01	days since 2002-01-01	>= 0	int64
area	total area burnt	km^2	>= 0.86 (1 MODIS pixel)	float64
perimeter	perimeter at given day	km	>= 3.71 (1 MODIS pixel)	float64
geometry	(Multi)Polygon vector data of spatiotemporal fire component	-	-	GeometryDtype

Comparison between Active Fire Events Table and MCD14ML

Apart from serving as the basis for the Spatiotemporal Fire Component Table, the Active Fire Events Table can also be used as an alternative to the MODIS Global Monthly Fire Location Product MCD14ML. MCD14ML is based on the swath products MOD14 and MYD14, rather than the tiled MOD14A1 and MYD14A1 products, and can be downloaded as tables in plain ASCII format.

For each active fire, MCD14ML contains the geographic location, the exact time of measurement, the satellite that made the measurement (Aqua or Terra), the band 21/31 brightness temperatures, the sample number, the fire radiative power, the detection confidence, a day/night algorithm flag, and an inferred hot spot type (e.g. active volcano, offshore or presumed vegetation fire). In comparison, the Active Fire Events Table does not contain the exact time (i.e. hour and minute) of the measurement, since it is based on the daily aggregates MOD14A1 and MYD14A1. The events table also does not contain the brightness temperatures of fire pixels, the sample number, the day/night algorithm flag, or the inferred hot spot type. It does however, in addition to the other columns, contain the neigh_int-column, which provides information as to whether there are any missing/cloud pixels in the neighborhood of a fire event.

To our surprise, we extract a lot more single pixel fire events from the Level 3 MOD14A1 and MYD14A1 products than MCD14ML does from the Level 2 products MOD14 and MYD14. This is surprising, because the MOD14A1 and MYD14A1 products are essentially maximum value composites of the MOD14 and MYD14 products. Our events table should therefore contain strictly less single pixel fire events than MCD14ML. We extract, however, a total of 145.673.360 single pixel fire events for the time period from 2003 to 2016, whereas MCD14ML only contains a total of 67.081.995 fire events over the same period, less than half the amount. When we bin fire events from both tables onto the same regular 0.05° lat/lon grid with a daily resolution, we find a total of 45.242.414 grid cells containing fires. The Active Fire Events Table covers 98.9% of those burning grid cells, whereas MCD14ML only covers 82.5% of grid cells. Unfortunately, we are unaware of the reason for this discrepancy. There is no information in the manuals of the MODIS products that would explain it, and we did not get a reply from the creators of the MODIS products pertaining to this discrepancy.

The figure below depicts the number of single pixel fires by year for both the Active Fire Events Table and MCD14ML, as well the percentage of all fires per year for each respective product. Although the percentage curves of the events table and MCD14ML follow each other relatively closely, the discrepancy between the two datasets is reflected in this figure as well, particularly for the years 2003 and 2004, and the years 2011 and 2012.

Fig. 1 Depicted is the absolute number of single pixel fire events entailed by FireTracks (FT) and MCD14ML per year (black line with circle markers and red line with hexagon markers, respectively). Additionally, the proportion (in percent) of all fire events between 2003 and 2016 of FT and MCD14ML is illustrated as solid black and red lines, respectively.

Overall, we conclude that our events table entails more than twice as many fires as MCD14ML, whilst covering nearly all fires that MCD14ML contains. Additionally, our events table has the advantage that it comes with the Active Fire Land Cover Table, providing land cover information for each fire.

Comparison between Spatiotemporal Fire Component Table and Global Fire Atlas

There are only very few global datasets that provide information on spatiotemporally tracked fire clusters. Two more well known datasets are the Global Fire Atlas and the Global Wildfire Dataset for the Analysis of Fire Regimes and Fire Behaviour. Both of these datasets are derived from the MODIS 500-m Burned Area product MCD64A1. To our knowledge, the FireTracks Scientific Dataset is the only dataset that uses active fires to derive spatiotemporally tracked fire clusters.

Apart from using different input data, the algorithms used to identify spatiotemporal clusters of fires also differ substantially. The FireTracks Scientific Dataset uses a very simple and conservative procedure: clusters are identified as the union of nearest neighbors of active fires in the discrete spacetime grid given by the spatial and temporal resolution of the MOD/MYD14A1 datasets. The Global Fire Atlas and the Global Wildfire Dataset, on the other hand, use much more complex algorithms to detect spatiotemporal clusters.

It is very difficult to assess the quality of any of the aforementioned datasets, since ground based observations are scarce. Additional validation of the fire parameter estimates of the different datasets is still required. Principally, the quality of the datasets depends strongly on the inherent limitations of the fire and land-cover data that serve as input. It is therefore highly recommended to read the respective user guides before using the data. For instance, the smallest identifiable fire size is imposed by the spatial resolution of the input fire data. For the FireTracks Scientific Dataset this is 0.86 km², and for the Global Fire Atlas it is 0.21 km². The Global Fire Atlas is therefore capable of identifying smaller fires. For both datasets, however, the error of the estimated burnt area is expected to grow for smaller fires.

Nevertheless, we can still compare the datasets to each other, and point out some (dis)advantages of the FireTracks Scientific Dataset over the Global Fire Atlas:

• Since the Global Fire Atlas is based on MCD64A1, it cannot account for multiple burns within a month on the same pixel. For instance, fires that last multipe days on the same location can therefore not be captured correctly by the Global Fire Atlas.

• Under relative cloudiness and below overstorey vegetation, the active fires used by the FireTracks Scientific Dataset have an advantage, since their detection is triggered by temperature anomalies that can sometimes be captured even under those circumstances [1]. Changes in surface reflectance, as utilized by MCD64A1, are more easily obscured.

• Although the same land cover product is used in both datasets (MCD12Q1), the Global Fire Atlas uses collection 5.1, whereas the FireTracks Scientific Dataset uses collection 6. This collection includes refinements and new features that cause significant changes in the land-cover classification maps [2].

• The FireTracks Scientific Dataset provides more detailed land-cover information. The Global Fire Atlas provides one dominant land cover type for each fire cluster. The details on how this dominant land cover is assigned are not described in the documentation of the Global Fire Atlas. In Addition to the dominant land cover type, the FireTracks Scientific Dataset quantifies the area of each land use type burnt by a fire cluster. Furthermore, we provide the land use type(s) on which the fire ignited. Also, all land cover schemes provided by MCD12Q1 can be readily downloaded, whereas the Global Fire Atlas is only available with the UMD scheme.

• The FireTracks Scientific Dataset is the only dataset that provides integrated fire radiative power for each fire cluster (FRP). The FRP can be used to quantify the combusted biomass [3] [4] [5].

• Since the clustering approach of the FireTracks Scientific Dataset is maximally conservative (no gaps between neighboring fires are allowed), we expect the distribution of fire sizes to be more skewed towards smaller fires compared to the Global Fire Atlas. We do, however, provide information as to whether a fire is neighbouring any cloud or missing data pixels.

The figure below depicts the absolute frequency distribution of spatiotemporal fire clusters by year for both the FireTracks Scientific Dataset (FT) and the Global Fire Atlas (GFA), as well the percentage of all fire clusters per year for each respective product. The number of fire clusters detected by the FT algorithm is approximately twice as high as that for the GFA (27.8 x 10⁶ versus 13.3 x 10⁶ fire clusters, respectively). Although the percentage curves of FT and GFA follow each other to a certain extent, temporal discrepancies are clearly evident.

Fig. 2 Frequency of spatiotemporal fire components of FT and GFA. The absolute number of spatiotemporal fire components entailed by FT and GFA per year is depicted (black line with circle markers and blue line with diamond markers, respectively). Additionally, the proportion (in percent) of all fire components between 2003 and 2016 of FT and GFA is shown as solid black and blue lines, respectively.

A reason for the difference in the absolute number of fire clusters detected by FT and GFA can be inferred from the figure below. It shows the burnt area frequency distribution of spatiotemporal fire components of FT and GFA. FT is more skewed towards smaller fire sizes. One potential reason for this is our conservative clustering approach, leading to more fragmented fire clusters.

Fig. 3 Burnt area frequency distribution of spatiotemporal fire components of FT and GFA. Depicted is the absolute frequency distribution of the burnt area (in km²) of spatiotemporal fire components for FT (black circles) and GFA (blue diamonds).

Using the powerlaw package, we fit a set of candidate models to the distributions of different spatiotemporal fire cluster characteristics (area, duration, expansion and FRP). These candidates include a log-normal, a stretched exponential, a powerlaw and a truncated powerlaw. In order to determine the optimal parameters for these models with respect to the observed distributions of characteristics, the powerlaw package employs maximum likelihood estimation (MLE) [6]. To compare the likelihood of each candidate, evaluated with the respective MLE-optimal parameters, the powerlaw package uses likelihood-ratio tests [7]. The results are depicted below.

Fig. 4 Probability density functions for different characteristics of spatiotemporal fire components (FT and GFA). (a) Probability density of observed burnt areas (in km²) of spatiotemporal fire components for FT (black circles) and GFA (blue diamonds). A maximum likelihood estimation reveals a powerlaw distribution as the best fit for FT (black solid line), and a stretched exponential distribution as the best distribution for GFA (blue solid line). (b) Probability density of observed durations (in days). For FT, there is no clear winner among the tested candidate distributions. For GFA, a truncated powerlaw distribution is the best model describing the distribution. (c) Probability density of observed expansion speeds (in km² day^-1). For FT, a stretched exponential is the best fit, for GFA a truncated powerlaw. (d) Probability density of observed integrated fire radiative power (in MW). Only FT is shown, since GFA does not contain integrated fire radiative power for spatiotemporal components. A log normal distribution is the clear winner among the tested candidate distributions.

References

[1] Humber, M. L., Boschetti, L., Giglio, L., & Justice, C. O. (2019). Spatial and temporal intercomparison of four global burned area products. International journal of digital earth, 12(4), 460-484.

[2] Sulla-Menashe, D., Gray, J. M., Abercrombie, S. P., & Friedl, M. A. (2019). Hierarchical mapping of annual global land cover 2001 to present: The MODIS Collection 6 Land Cover product. Remote Sensing of Environment, 222, 183-194.

[3] Wooster, M. J., Roberts, G., Perry, G. L. W., & Kaufman, Y. J. (2005). Retrieval of biomass combustion rates and totals from fire radiative power observations: FRP derivation and calibration relationships between biomass consumption and fire radiative energy release. Journal of Geophysical Research: Atmospheres, 110(D24).

[4] Ichoku, C., Giglio, L., Wooster, M. J., & Remer, L. A. (2008). Global characterization of biomass-burning patterns using satellite measurements of fire radiative energy. Remote sensing of Environment, 112(6), 2950-2962.

[5] Kumar, S. S., Roy, D. P., Boschetti, L., & Kremens, R. (2011). Exploiting the power law distribution properties of satellite fire radiative power retrievals: A method to estimate fire radiative energy and biomass burned from sparse satellite observations. Journal of Geophysical Research: Atmospheres, 116(D19).

[6] Clauset, A., Shalizi, C. R., & Newman, M. E. (2009). Power-law distributions in empirical data. SIAM review, 51(4), 661-703.

[7] Neyman, J., & Pearson, E. S. (1933). IX. On the problem of the most efficient tests of statistical hypotheses. Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character, 231(694-706), 289-337.