Point Pattern Analysis of Wildfires in Turkey between years 2000 - 2021
Zehra Usta - Rahila Mammadova Huseynova
2024-01-07
fire_df <- read.csv("2000-2021 SINGLE TURKEY FIRE_M-C61_214067.csv")
fire.sf <- st_as_sf(fire_df, coords = c("longitude", "latitude"), crs = 4326)Introduction
Significance of wildfires and their ecological impact increasing with the effect of global warming. This project embarks on a focused exploration of the spatial patterns exhibited by wildfires in Turkey. The Point Pattern Analysis undertaken here aims to unravel the underlying geographic dynamics of wildfire occurrences, shedding light on the clustering, dispersion, and potential hotspot regions within the Turkish landscape. By employing advanced spatial statistical techniques, we seek to discern patterns that may inform the understanding of the historical wildfire data. In this project, we’re studying how wildfires are spread out across Turkey. By looking at patterns in where wildfires happen, our goal is to find helpful information for making smart decisions, managing wildfires better, and understanding how the environment plays a role in causing wildfires in Turkey.
This dataset contains information essential for studying the latest forest fires in Turkey, focusing on specific time intervals. The data originates from recent forest fire occurrences in Turkey, and it has been authorized for use by NASA Portal. The information is gathered from satellites, specifically three separate ones:
MODIS C6.1
SUOMI VIIRS C2
J1 VIIRS C1
General Attributes:
Latitude: The center of the nominal 375 m fire pixel.
Longitude: The center of the nominal 375 m fire pixel.
Bright_ti4 (Brightness temperature I-4): The temperature of the fire pixel measured in Kelvin using VIIRS I-4 channel brightness.
Scan (Along Scan pixel size): The pixel size produced by the algorithm, approximately 375 m at nadir.
Track (Along Track pixel size): Similar to Scan, the pixel size produced by the algorithm, approximately 375 m at nadir.
Acq_Date (Acquisition Date): The date of VIIRS acquisition.
Acq_Time (Acquisition Time): The time of acquisition/overpass of the satellite in Coordinated Universal Time (UTC).
Satellite: The satellite used for data acquisition, such as Suomi National Polar-orbiting Partnership (Suomi NPP).
Confidence: A value indicating the quality of individual hotspot/fire pixels, categorized as low, nominal, or high. Confidence is influenced by factors like sun glint, temperature anomalies, and geographic areas prone to false positives.
Version: Identifies the data collection version and processing source, distinguishing between Near Real-Time (NRT) and Standard Processing.
Additional Attributes:
Bright_ti5 (Brightness temperature I-5): The temperature of the fire pixel measured in Kelvin using the I-5 channel brightness.
FRP (Fire Radiative Power): Represents the pixel-integrated fire radiative power in megawatts (MW). The algorithm is optimized for small fires, and FRP retrieval is viable for small and/or low-intensity fires using a hybrid approach.
Type (Inferred hot spot type):
0 = presumed vegetation fire
1 = active volcano
2 = other static land source
3 = offshore detection (includes all detections over water)
DayNight (Day or Night):
D = Daytime fire
N = Nighttime fire
Locations of wildfires in Turkey between 2000 - 2021
head(fire_df)## latitude longitude brightness scan track acq_date acq_time satellite
## 1 36.8527 35.3594 306.0 2.1 1.4 2000-11-01 815 Terra
## 2 40.5441 35.1786 307.8 1.1 1.0 2000-11-02 857 Terra
## 3 40.1473 35.0444 305.3 1.1 1.0 2000-11-02 857 Terra
## 4 40.5421 35.1909 310.0 1.1 1.0 2000-11-02 857 Terra
## 5 41.2569 31.4172 307.9 1.1 1.0 2000-11-02 857 Terra
## 6 36.7291 36.2161 306.7 1.3 1.1 2000-11-02 858 Terra
## instrument confidence version bright_t31 frp daynight type
## 1 MODIS 60 6.03 294.4 13.2 D 0
## 2 MODIS 66 6.03 296.0 5.8 D 0
## 3 MODIS 46 6.03 293.5 4.3 D 0
## 4 MODIS 69 6.03 296.1 7.3 D 0
## 5 MODIS 0 6.03 290.4 8.0 D 2
## 6 MODIS 24 6.03 294.4 7.9 D 2summary(fire_df)## latitude longitude brightness scan
## Min. :35.82 Min. :25.72 Min. :300.0 Min. :1.000
## 1st Qu.:37.11 1st Qu.:33.04 1st Qu.:311.3 1st Qu.:1.000
## Median :37.71 Median :35.82 Median :322.1 Median :1.200
## Mean :38.10 Mean :35.81 Mean :323.1 Mean :1.519
## 3rd Qu.:38.93 3rd Qu.:39.76 3rd Qu.:331.3 3rd Qu.:1.700
## Max. :42.03 Max. :44.80 Max. :503.5 Max. :4.820
##
## track acq_date acq_time satellite
## Min. :1.000 Length:211309 Min. : 0 Length:211309
## 1st Qu.:1.000 Class :character 1st Qu.: 845 Class :character
## Median :1.100 Mode :character Median :1024 Mode :character
## Mean :1.183 Mean :1128
## 3rd Qu.:1.300 3rd Qu.:1110
## Max. :2.000 Max. :2359
##
## instrument confidence version bright_t31
## Length:211309 Min. : 0.00 Length:211309 Min. :265.2
## Class :character 1st Qu.: 55.00 Class :character 1st Qu.:294.1
## Mode :character Median : 68.00 Mode :character Median :302.8
## Mean : 66.45 Mean :302.5
## 3rd Qu.: 79.00 3rd Qu.:310.0
## Max. :100.00 Max. :400.1
##
## frp daynight type
## Min. : 0.00 Length:211309 Min. :0.000
## 1st Qu.: 8.90 Class :character 1st Qu.:0.000
## Median : 15.30 Mode :character Median :0.000
## Mean : 30.34 Mean :0.139
## 3rd Qu.: 30.70 3rd Qu.:0.000
## Max. :4816.68 Max. :3.000
## NA's :5391ggplot() +
geom_sf(data = TUR$geometry) +
geom_sf(data = fire.sf$geometry, size = 0.5) +
theme_minimal()
When we analysed fire frequency by months and years, we observed that fire frequency increases in august and september which are the hottest months in the year. Also we noticed that between 2008 and 2012 much more fire occurred compared to other years.
df <- fire_df %>%
mutate(acq_date = as.POSIXct(acq_date))
df <- fire_df %>%
mutate(
day = day(acq_date),
month = month(acq_date),
year = year(acq_date)
)
pivot_table <- df %>%
group_by(year, month) %>%
summarise(Fire_Frequency = n()) %>%
spread(key = year, value = Fire_Frequency, fill = 0)## `summarise()` has grouped output by 'year'. You can override using the
## `.groups` argument.pivot_long <- pivot_table %>%
gather(key = "Year", value = "Fire_Frequency", -month)
pivot_long$month <- factor(month.name[pivot_long$month], levels = month.name)
ggplot(pivot_long, aes(x = Year, y = month, fill = Fire_Frequency, label = Fire_Frequency)) +
geom_tile(color = "white") +
geom_text(color = "black", size = 3, show.legend = FALSE) +
scale_fill_viridis_c() +
labs(
x = "Year",
y = "Month",
fill = "Fire Frequency",
title = "Turkey Wildfires: Fire Frequency by Month and Year"
) +
theme_minimal() +
theme(
axis.text.x = element_text(angle = 45, hjust = 1)
)
Literature Review
In the analysis of spatial patterns of wildfires in Turkey from 2000 to 2020, our methodological approach draws inspiration from Rolf Turner’s pivotal article, “Point patterns of forest fire locations,” published in the journal Environmental and Ecological Statistics in 2009. Turner’s work rigorously applied spatial point pattern analysis to a comprehensive dataset of wildfires in New Brunswick, demonstrating the utility of recent advancements in point process modelling for such environmental phenomena. His exploration into the interactions between wildfire events, offers a robust framework for assessing spatial autocorrelation and identifying underlying patterns (Turner 2009).
Our study benefits from Turner’s methodical approach to model fitting, data cleaning, and exploratory analysis, particularly in addressing the challenges inherent in spatial data that exhibit complex interdependencies. Turner’s emphasis on integrating covariate information for more realistic modelling parallels our inclusion of environmental and geographic variables, enhancing the depth and accuracy of our spatial analysis.
By referencing Turner’s article, we align our research with established methodologies while also expanding upon them by adapting these techniques to the unique context of Turkish wildfire data. This not only grounds our work in credible scholarly practices but also ensures that our analysis remains at the forefront of current ecological and statistical research (Turner 2009).
Kernel Density Estimation
Kernel density estimation (KDE) is a statistical method used in analysing wildfires in Turkey. KDE helps create spatial intensity maps that highlight areas with higher wildfire incident densities. By smoothing and interpolating the point data of fire occurrences, KDE provides a continuous surface that accentuates regions with elevated wildfire activity. KDE aids in identifying wildfire hotspots by detecting clusters of incidents. Hotspots represent areas with a higher likelihood of wildfire occurrences. This information is crucial for prioritising resource allocation, implementing preventive measures, and planning emergency responses.
KDE contributes to risk assessment by quantifying the probability of wildfires in different geographical areas. Understanding the spatial distribution of fire risk allows stakeholders to implement targeted mitigation strategies in high-risk zones. Outputs help decision-makers allocate firefighting and prevention resources efficiently. Areas with higher kernel density estimates may require increased monitoring, patrols, or early warning systems to minimise the impact of potential wildfires.
Firstly, we showed the kernel density estimation of wildfires for the whole dataset which contains fire observations between years 2000-2020. As expected, we obtained hotspots in the Southeastern Anatolia Region and Central Anatolia Region. The Southeastern Anatolia Region has a semi-arid climate, characterized by extremely hot summers and mild winters. Similarly, the Central Anatolia Region features hot and dry summers and cold winters. Both regions, particularly during summers, face an elevated risk of wildfires.
coords <- st_coordinates(fire.sf$geometry)
fire <- data.frame(lon = coords[, "X"], Lat = coords[, "Y"])
p <- ggplot() +
ggtitle("Kernel Density Estimation of Wildfires in Turkey between 2000 - 2021") +
geom_sf(data = TUR$geometry, fill = "white", color = "black") +
stat_density_2d(
data = fire,
aes(x = lon, y = Lat, fill = after_stat(level)),
geom = "polygon", linewidth = 0.2,
) +
scale_fill_viridis_c() +
scale_alpha(range = c(0, 0.2)) +
theme_minimal()
print(p)
data <- fire.sf %>%
mutate(date = as.Date(acq_date))
years2000_2005 <- data %>%
filter(between(date, as.Date("2000-01-01"), as.Date("2005-12-31")))
years2005_2010 <- data %>%
filter(between(date, as.Date("2005-01-01"), as.Date("2010-12-31")))
years2010_2015 <- data %>%
filter(between(date, as.Date("2010-01-01"), as.Date("2015-12-31")))
years2015_2020 <- data %>%
filter(between(date, as.Date("2015-01-01"), as.Date("2020-12-31")))Over the years, a notable change has been witnessed in the locations of wildfire hotspots. In the early 2000s, heightened wildfire activities were concentrated in the Central Anatolia Region. However, between 2015 and 2020, there was a notable shift, with the focal points of wildfires moving towards the Southeastern Anatolia Region. This shift holds potential significance for policymakers to consider in their decision-making processes.
grid.arrange(p_years2000_2005, p_years2005_2010, p_years2010_2015, p_years2015_2020, nrow = 2)
fire_0 <- fire.sf %>% filter(type == 0)
fire_2 <- fire.sf %>% filter(type == 2)coords <- st_coordinates(fire_0$geometry)
fire <- data.frame(lon = coords[, "X"], Lat = coords[, "Y"])
type_0 <- ggplot() +
ggtitle("Kernel Density Estimation of Fire Type 0") +
geom_sf(data = TUR$geometry, fill = "white", color = "black") +
stat_density_2d(
data = fire,
aes(x = lon, y = Lat, fill = after_stat(level)),
geom = "polygon", linewidth = 0.2,
) +
scale_fill_viridis_c() +
scale_alpha(range = c(0, 0.2)) +
theme_minimal() +
guides(fill = "none")
coords <- st_coordinates(fire_2$geometry)
fire <- data.frame(lon = coords[, "X"], Lat = coords[, "Y"])
type_2 <- ggplot() +
ggtitle("Kernel Density Estimation of Fire Type 2") +
geom_sf(data = TUR$geometry, fill = "white", color = "black") +
stat_density_2d(
data = fire,
aes(x = lon, y = Lat, fill = after_stat(level)),
geom = "polygon", linewidth = 0.2,
) +
scale_fill_viridis_c() +
scale_alpha(range = c(0, 0.2)) +
theme_minimal() +
guides(fill = "none")In our analysis, we encountered a scenario where our dataset exclusively comprises fire instances categorized as types 0 and 2, disregarding the other two fire types. This led us to conduct a kernel density estimation specifically for fire types 0 and 2.
Upon closer examination, we observed that fire type 2 exhibits a kernel density estimation pattern that closely aligns with the overall dataset. However, the kernel density estimation for fire type 0 reveals a distinct geographical concentration, primarily observed in the Aegean and northwest regions of Turkey.
This geographical insight suggests a localized prevalence of fire type 0 in specific regions, providing valuable information for understanding the spatial distribution and characteristics of different fire types in Turkey.
type_0
type_2
library(tidyverse)
df <- fire_df %>%
mutate(acq_date = as.POSIXct(acq_date))
df <- fire_df %>%
mutate(
day = day(acq_date),
month = month(acq_date),
year = year(acq_date)
)
pivot_table <- df %>%
group_by(year, month) %>%
summarise(Fire_Frequency = n()) %>%
spread(key = year, value = Fire_Frequency, fill = 0)## `summarise()` has grouped output by 'year'. You can override using the
## `.groups` argument.library(tidyr)
pivot_long <- pivot_table %>%
gather(key = "Year", value = "Fire_Frequency", -month)
pivot_long$month <- factor(month.name[pivot_long$month], levels = month.name)joined_data <- st_join(fire.sf, TUR)fire_counts <- joined_data %>%
group_by(NAME_2) %>%
summarise(count = n())merged_data <- merge(as.data.frame(TUR), as.data.frame(fire_counts), by = "NAME_2")merged_data <- st_as_sf(merged_data)Spatial Autocorrelation
The p-value associated with the Moran I test is very low. This p-value provides very strong evidence against the null hypothesis of spatial randomness. The conclusion is that there is a highly significant positive spatial autocorrelation. The Moran I statistic for the dataset is 0.4549143583. The positive value reinforces the presence of a clustering pattern, suggesting that areas with similar values are spatially correlated.
w <- poly2nb(df$geometry.x)
w <- nb2listw(w, zero.policy = TRUE)
variable <- df$count
moran <- moran.test(variable, w, zero.policy = TRUE)
print(moran)##
## Moran I test under randomisation
##
## data: variable
## weights: w n reduced by no-neighbour observations
##
##
## Moran I statistic standard deviate = 22.591, p-value < 2.2e-16
## alternative hypothesis: greater
## sample estimates:
## Moran I statistic Expectation Variance
## 0.4549143583 -0.0011428571 0.0004075449In the context of Moran’s I, Monte Carlo simulation provides a robust method for hypothesis testing, allowing researchers to assess the statistical significance of observed spatial patterns and identify whether there is evidence of spatial autocorrelation beyond what might occur due to random chance.
The small p-value (0.001) provides evidence to reject the null hypothesis, indicating that there is significant positive spatial autocorrelation in the variable. The observed Moran’s I value is relatively strong, as it ranks at the 1000th position among the simulated values. This suggests a notable clustering pattern in the spatial distribution of the variable.
set.seed(19981115)
moran_result <- moran.mc(variable, listw = w, nsim = 999)
print(moran_result)##
## Monte-Carlo simulation of Moran I
##
## data: variable
## weights: w
## number of simulations + 1: 1000
##
## statistic = 0.45491, observed rank = 1000, p-value = 0.001
## alternative hypothesis: greatermoranPlot <- moran.plot(df$count, listw = w, pch = 19)
Conclusion
As we look at the patterns found in this study, we see that understanding these patterns better is really important for dealing with wildfires. Figuring out where wildfires happen a lot and recognizing how they relate to each other gives helpful information for people who make decisions, take care of land, and respond to emergencies.
Even though we mainly looked at past wildfire data in this project, what we learned can help make predictions for the future, decide where to put resources, and plan ways to stop wildfires. These lessons don’t just apply to Turkey, hey give a useful framework for dealing with similar environmental challenges around the world.
To sum it up, our study on where wildfires happen in Turkey is like a first step in finding smarter and more ready ways to handle wildfires in this diverse and ecologically important region.
References
https://www.kaggle.com/datasets/brsdincer/20002021-turkey-fire-points-single-csv-nasa/data
Turner, R. Point patterns of forest fire locations. Environ Ecol Stat 16, 197–223 (2009).