Introduction

In 2019 and 2020 media reported a steep increase in the intentional burning of Iraqi and Syrian farmlands. Some of them were attributed to remnants of the Islamic State (IS) and raised concerns among international organizations such as the United Nations (United Nations (UNITAD) 2020). The attacks on agricultural production were not without precedent, but their intensity was high compared to earlier periods. Based on prior research, press reports, witness accounts from an international observer, and personal communications, we suggest distinguishing between four different phases, during which the IS and its predecessor organizations had distinct approaches to Iraq’s water-food complex (see “Background: the IS and Iraq’s water-food complex”): before its territorial conquests in 2013/2014 when it was still lingering in relative obscurity (Phase 1: “in waiting”), its short period of proto-statehood in 2014–2017 (Phase 2: “in power”), its military withdrawal in 2015/2016–2017 (Phase 3: “in withdrawal”), and the onset of its renewed clandestinity in 2017/2018 (Phase 4: “in hiding”).

Recently, research has progressed understanding of fires and conflict in Iraq and/or Syria, e.g., Eklund et al. (2021), Linke and Ruether (2021), Rasul et al. (2021), Schon et al. (2021), Zubkova et al. (2021). For instance, Rasul et al. (2021) found an increasing trend in fire events in Iraq during 2001–2019 and attributed that to wind and maximum air temperature, without exploring the link between fire and conflict. Schon et al. (2021) studied fire activity in Syria and attributed unusually high fire activity in 2019 not only to conflict but also to clearing and recultivation of land in some areas where conflict has receded. Moreover, the influence of crop seasonality and weather variability on violence patterns in Syria has been investigated by (Linke and Ruether 2021). It has been shown that government and rebel strategies of armed conflict include the increase in violent activity during local crop growing season (wheat and barley) either to capture or destroy agriculture resources.

Therefore, there has been consensus that during 2019, agricultural burned area in Iraq and Syria increased significantly. However, due to the complex political and social background of the region, as well as the conflicting results in the literature as to the reasons behind these fires, it is difficult to confidently identify what is causing this increase or “who” was behind these fires.

Remotely sensed data have proven to be beneficial in estimating agricultural changes within unsafe conflict zones, where data collection is limited (Jaafar and Woertz 2016). Remote sensing can also provide valuable insights on land use and land cover change, including burned area. The focus of application has shifted “from land use planning and management to land-use change impact and driving factors” (Chang et al. 2018; Li et al. 2017). Conflict can be one such driving factor of land-use change, but its effects are not unidirectional (Baumann and Kuemmerle 2016), which poses a challenge in studying land change and attribution to armed conflict. For example, deforestation (Stevens et al. 2011) and harvest failure (Huang et al. 2019) have been examined as an outcome of armed conflict. Land abandonment can be another conflict outcome that may or may not increase with proximity to conflict (Yin et al. 2019) and can be followed by reforestation (Burgess et al. 2015) and increased plot sizes of remaining farmers (Adelaja and George 2019).

Burned area estimation and mapping is another challenging task in war-torn countries. Remote sensing and GIS analysis have been widely implemented in mapping burned areas (Boer et al. 2008; Pereira 2003; Quintano et al. 2011), identifying burn severity (Cansler and McKenzie 2012; De Santis and Chuvieco 2007; Gitas et al. 2009), and fire management (Keane et al. 2001; Xiao-rui et al. 2005). Several methods have been proposed to estimate the damage caused by fires including image differencing, spectral indices, fire indices, and image classification (Cocke et al. 2005; Çömert et al. 2019; Pinto et al. 2020; Ramo and Chuvieco 2017; Veraverbeke et al. 2011).

Existing global fire data products provide valuable information on the estimation of burned area, fire location, and the timing of fires. Global fire products rely on coarse spatial resolution and high temporal resolution sensors such as the moderate resolution imaging spectro-radiometer (MODIS) (Justice et al. 2002) and the visible infrared imaging radiometer suite (VIIRS) (Schroeder et al. 2014). These sensors have shown to significantly underestimate small agricultural fires (Alonso-Canas and Chuvieco 2015; Oliva et al. 2011). Recently, the use of moderate spatial resolution sensors, including Landsat satellite data (30 m) and Sentinel-2 data (10 m), has shown to better estimate small fires (Deshpande et al. 2022; Liu et al. 2021; Wang et al. 2018). The main limitation in detecting agricultural fires using MODIS is the size of the fields. Coarse-resolution satellite imagery provides high temporal resolution time-series, which is useful for statistical analysis. However, it does not accurately detect agricultural burned area. Previous research has shown that MODIS burned area (BA) product detects a small fraction of small-sized burned fields (< 120 ha) (Palomino-Angel and Anaya-Acevedo 2012; Zhu et al. 2017). However, MODIS BA product can capture large uncontrolled fires, which tend to expand and spread more quickly and adversely as compared to regular traditional cropland fires. At least some of the fires of 2019/2020 in Iraq and Syria were very large and therefore captured by MODIS (Rasul et al. 2021).

In this study, we build on previous research that has explored the influence of conflict and social factors on fire activity in conflict-affected regions. We examine the extent of the burned area damage in agricultural fields at different periods of the year in a region that was once controlled by IS and includes both Kurdish territories and non-Kurdish territories. The importance of investigating a region with such characteristics is that several armed groups exist or have existed where the strategic differences among these groups can also be explored. We also examine the effect of using different satellite products on the results of such an analysis and analyze the interlinkages between the changes in the percentage of burned area, the environmental variables, and the four phases of the IS approach to Iraq’s water-food complex described above. Here, we explore the causes and the extent of irregular and uncontrolled fire damage on farmlands in parts of Iraq and Syria, for the years 2019 and 2020 and compare it to the historical record using various sources of satellite imagery. We analyze remote sensing data of different resolutions and ancillary weather and conflict data and the percentage of burned agricultural areas (pBA) in Syria and Iraq during 2001–2020 in an attempt to better understand the relations between fires, conflicts, and weather variability. We use MODIS burned area (BA) product for a 20-year time-series analysis and for more accurate burned area estimation; for years 2019 and 2020, we resort to higher resolution imagery (Landsat 8 and Sentinel-2) and construct a comparative analysis of the impact of resolution on the detection of fires.

Background: the Islamic State and Iraq’s water-food complex

In the wake of the US invasion of 2003, a hybrid political system emerged in Iraq that combined formally democratic elements with a reconstructed authoritarianism with sectarian undertones (Dodge 2013). Sunni grievances over the loss of power played a major role in the civil war that followed and contributed to the emergence of the Islamic State of Iraq and Syria. It expanded in the Northeast of Syria in 2013 and then conquered large swaths of territory in Iraq in 2014. In June 2014, it shortened its name to just Islamic State (IS), underscoring its aspiration of statehood that went beyond the immediate financing needs of its military, police, and security services. These accounted for as much as three-quarters of its budgetary spending (Al-Tamimi 2015), but there was also the provision of public goods such as health care, food provision, and waste disposal that the IS broadcasted in its public relations. It wanted to be seen as a state by domestic and international audiences, whether friends or foes, and spent money to this end.

Main sources of IS income included the cannibalization of Syrian oil fields (Hansen-Lewis and Shapiro 2015; Solomon et al. 2015), the taxation of agricultural value chains (Jaafar and Woertz 2016), extortion, looting, and ransom, as well as foreign donations (Eklund et al. 2017; Eklund and Lange 2018; Hansen-Lewis and Shapiro 2015; House of Commons 2016a, b, c; Jaafar and Woertz 2016; Jones and Solomon 2015; Solomon et al. 2015; Solomon and Jones 2015). Ransom and looting were some of the sources of income that were not recurrent. The oil aspect received somewhat exaggerated attention in the media as well and declined over time because of military action. In comparison, the role of foreign donations and taxation of agricultural value chains received less attention but were of crucial importance (House of Commons 2016c; Jaafar and Woertz 2016).

Around May 2015, the IS had reached its peak territorial expansion. Afterward, it entered a retreat that accelerated in 2017. Mosul and Raqqa fell in 2017. By the end of that year, the IS had lost most of its territory. In March 2019, it was driven from its last physical stronghold, Baghuz on the border between Syria and Iraq. In the same year, 2 years after its 2017 annus horribilis, ominous burnings of fields intensified in Iraq that would repeat themselves in 2020 and were attributed to remnants of the IS.

During its “in waiting” Phase 1, the predecessor organization of Islamic State of Iraq and Syria (ISIS) and then the IS, the Islamic State in Iraq (ISI, est. 2006), had cultivated a support base in the countryside when it was forced to withdraw in 2007 after tribal leaders turned against it and were helped by the Surge of US troops. The manual management of savagery (idarat al-tawahhush) that enjoyed an avid readership within IS circles regarded access to food-producing territories as crucial for the control of conquered areas, advocating consolidation of power in the countryside to take the fight to population centers at a later point in time (McCants 2015; Naji 2004). IS pursued a strategy of instigating a sectarian civil war between Sunnis and Shiites, during which farmlands of opponents were burnt as a strategy of ethnic cleansing and reprisals (Arango 2013), but the tactic was not used to the same extent as it would be later on in 2019–2020, as we show in this article. Due to its strategic priorities in the “in waiting Phase,” there was not an immediate interest in sabotage acts of agricultural production.

This did not change during the “in power” Phase 2. The IS did not deliberately destroy agricultural production facilities within its territory. Quite the contrary with the Syrian Gezirah and Ninewa province in Iraq, it had conquered the food baskets of Syria and Iraq, respectively. Like oil production facilities, it sought to take over farms, silos, dams, and processing factories intact so it could tax them, trade for its own account, distribute spoils of wars to its members and clients, and guarantee food supplies on its territory. Helped by better rainfalls, the production of rainfed winter crops like wheat and barley on IS territory increased compared to the pre-conflict average based on remote sensing data (Jaafar and Woertz 2016). At the same time, the population about halved to not more than four million people due to displaced persons fleeing IS territory as Jaafar and Woertz (2016) argue based on cities’ illumination on satellite imagery, pre-war census data, and reports on refugee flows by international organizations. Hence, the IS presumably had an exportable surplus which it smuggled into the subsidized Iraqi Public Distribution System (PDS) and to Turkey where prices were higher. Only irrigated summer crops that rely on machinery that is vulnerable to military action and pilfering suffered heavily. The IS did, however, use water as a weapon of war by deliberately flooding areas to repel advancing government forces and temporarily withholding it or poisoning it with diesel to harm populations in territories opposing it (von Lossow 2016).

During its “in withdrawal” Phase 3 in 2015/2016–2017, the IS used scorched earth tactics to hold back advancing forces of the Iraqi army and the Kurdish Peshmerga and out of spite against populations or displaced proprietors who it deemed disloyal to its cause. Sabotage acts included the destruction of wells, the chopping down or burning of orchards, and the dismantling of electricity lines as interview-based reports have shown for the region of Sinjar and other parts of North Iraq (Amnesty International, 2018; Schwartzstein 2016). When IS slipped back into renewed clandestinity during its “in hiding” Phase 4 from 2017 onwards, there was not an immediate uptick in sabotage acts, which raises questions about their timing, motivation, and impact. Only a few of the farm attacks were officially claimed by IS and its media channels such as Al-Amaq according to the UN (United Nations (UNITAD) 2020). The attacks on farms co-occurred with other acts of sabotage such as the destruction of surveillance cameras and electricity transmission towers. As possible reasons for the burning of farmland, the UN identifies the added economic burden on Iraqi authorities and targeted farmers, to create obstacles for counter-attacks against the IS, to instigate protests against the government that demand more protection and diverts government forces, and intimidation to facilitate extortion.

Data and methods

Study area

The study area covers North and North-East Iraq, including part of Syria (Fig. 1). Iraq consists of eighteen governorates; here, we focus on five governorates — Ninewa, Kirkuk, Erbil, Diyala, and As-Sulaymaniya — in addition to Al-Hassakeh — the breadbasket of Syria. The study region is generally known for its aridity and semi-aridity (northern parts of Syria and Iraq) (Jaafar and Woertz 2016). About 50–300 mm of precipitation occurs between November and April during the winter season. Agriculture is a key component of the region’s economy with wheat being the major food crop cultivated. Wheat is mainly cultivated in the northern and central areas of Iraq (almost half of the cultivated agriculture is in northern Iraq) and Al-Hassakeh of Syria (Jaafar and Woertz 2016). Other crops cultivated include barley, corn, fodder, vegetables, and cotton. Wheat is planted in December and harvested in the summer (June–July).

Fig. 1
figure 1

Location of study area in Iraq and Syria showing the agricultural burned area in 2019 and selected validations sites

Full size image

Data

A variety of imagery sources is required to estimate the burned area at the field scale. From NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS), we use the burned area product provided in monthly format (MCD64A1, (Giglio et al. 2018)), the yearly land cover product (MCD12Q1 (Friedl et al. 2010)), and the Nadir BRDF-Adjusted Reflectance Daily product (MCD43A4 (Schaaf and Wang 2015)). MODIS burned area product returns global monthly, gridded per-pixel burned area with a spatial resolution of 500 m. Because MODIS provides the longest high-temporal resolution time series available for burned area, which is useful for statistical analysis, we use it as a reference against which our analysis of burned area is compared (2001–2020). For a more precise estimation of burned area at the field scale, we use higher spatial resolution satellite imagery for years 2017–2020. We use satellite imagery from the European Space Agency Copernicus mission (Sentinel-2, 10–20 m resolution and a 5-day revisit time), and National Aeronautics and Space Administration-US Geological Survey (NASA-USGS) mission of Landsat 8 (30 m resolution and16-days revisit time). We estimate burned area of small farms from Sentinel-2 and Landsat 8 imagery in 2017–2020. We perform geospatial analysis using Google Earth Engine (Gorelick et al. 2017), a cloud-based platform used for planetary-scale analysis.

Burned area: estimation and trend

First, we examine the burned area pattern in the region using MODIS burned area product (500 m) from 2001 until 2020. Our aim is not to map agricultural fires over the whole year, but rather to focus on two key periods in wheat’s growth cycle: pre-harvest (May–June) and post-harvest (July–August). Fires do occur naturally in non-agricultural areas, so we use MODIS Land Cover Type Global data to restrict the analysis to arable lands and to identify the croplands changes through time. Croplands, i.e., class 12 (croplands; at least 60% of area is cultivated cropland) from land cover type 1 (Annual International Geosphere-Biosphere Programme -IGBP- classification) were selected. For better detection of wheat fields, we apply an NDVI threshold mask. Using MODIS Surface Reflectance product, the mean NDVI was calculated for March, and NDVI values greater than 0.6 were extracted from the MODIS croplands classification per year (Jaafar and Ahmad 2015; Jaafar and Woertz 2016). Using Landsat 8 and Sentinel-2, we use the normalized burn ratio (NBR) and the difference in NBR (dNBR) to quantify burned area; we define the pre-fire time-period 4–6 weeks before May and the post-fire period as 2–3 weeks after June to detect the May–June burned area. Unlike forest fires where the burn scars are visible in the imagery for months, traditional cropland burn scars tend to disappear quite quickly. Either a farmer has plowed over the burn scar, or they have planted a new crop, or wind and rain can erode the spectral signature. The dNBR classification we perform does not aim to map cropland burned areas that tend to disappear quickly (e.g., when farmers plow the land or plant new crops) but rather to map burned areas that are visible for a longer period of time such as intentional fires that take place during cereal harvest time. These burns are likely caused by IS, and therefore, the burn scars will be visible for longer.

We estimate the agricultural burned area for each year from 2001 until 2020 using MODIS burned area and perform trend detection analysis using the Mann–Kendall test and Sen’s slope. Previous studies on fires and burned area estimation have used Mann–Kendall test for trend analysis (Jiménez-Ruano et al. 2017; Rasul et al. 2021; Rodrigues et al. 2013; San-Miguel-Ayanz et al. 2012; Vadrevu et al. 2019). Mann–Kendall is a non-parametric trend analysis statistical test used to detect trends in a time series of data. It is appropriate to be used in detecting both linear and non-linear trends. The significance of the trend is evaluated using Z-value, where we test the null hypothesis (trend does not exist) against the alternative hypothesis (trend exists). However, Mann–Kendall’s test is non-dimensional and does not quantify the magnitude of the trend. It is only a measure of the correlation of a variable (in our study: burned area) with time; it offers a measure of the significance of an observed trend (Salmi 2002). Thus, we use Sen’s slope to estimate the magnitude (true slope) of the trend being observed. We perform Mann–Kendall test and Sen’s slope using MAKESENS 1.0 software (Salmi 2002).

Second, we resort to a higher spatial resolution (10–30 m) imagery to better highlight the burned area in small fields in years when Landsat 8 and Sentinel-2 imagery were commonly available (2017–2020). For Landsat 8, we apply a cloud masking function on GEE which masks out clouds, cloud shadows, and snow. We use the bitwise operation along with the AND operator to create the masks. We perform the masking operation on the quality assessment band (“pixel_qa”) using the “bitwiseAnd” function on GEE. Similarly, for Sentinel 2 a cloud masking function was also used to mask clouds. We compare burned areas obtained using Landsat 8 and Sentinel-2 (medium spatial resolution: 10–30 m) to those derived from MODIS Burned Area Product. Among the most widely used indices of burn severity mapping is the normalized burn ratio (NBR) (Fassnacht et al. 2021). NBR has been used in many studies to map small fires including agricultural fires (Franco et al. 2020; Loboda et al. 2007; Wang et al. 2018; Zhu et al. 2017). NBR is defined as the normalized difference of near-infrared reflectance (NIR), and short-wave infrared range reflectance (SWIR) (Key and Benson 2006)) calculated using the following equation:

$$NBR= \frac{{\rho }_{NIR }- {\rho }_{SWIR }}{{\rho }_{NIR }+ {\rho }_{SWIR}}$$

NBR makes use of the way fires alter the optical properties of vegetation. Specifically, a drop in NIR and SWIR is observed post-fire due to vegetation elimination and ash formation (Key and Benson 2006). Vegetation elimination is detected by a change in NIR, since healthy vegetation shows high reflectance in NIR unlike burned area; and the ash formation is detected by a change in SWIR reflectance (Key and Benson 2006). Difference normalized burn ratio (dNBR) is derived from NBR by subtracting post-fire NBR from pre-fire NBR.

The higher the dNBR value, the more severe the fire is, and negative dNBR values indicate vegetation regrowth. To interpret the dNBR results and burn severity, the United States Geological Survey (USGS) burn-severity classification table was used (Table S1).

For estimation of burned area in May and June, we calculate dNBR using pre-fire NBR of March and/or April and post-fire NBR of July. Details on the date ranges for each year and month along with the number of images used are provided in Supplementary material (Table S2). After obtaining dNBR values for each governorate over the years, we compute the area of each dNBR class. We visually interpret the dNBR classification results (Figure S1); we find that the “high severity” class corresponds to actual burned area.

Correlating armed conflict and weather variables with burned area

The IS officially claimed responsibility for a series of fire attacks on wheat and barley fields in villages in Iraq and Syria during the harvest season of 2019 (Cordesman and Toukan 2019). We use a dataset compiled from press reports and witness accounts of an international observer of security issues based in Iraq who had to remain anonymous. From this dataset, we identify 75-point locations of farmland fires that were attributed to IS attacks in governorates of interest. The dataset includes the name of the village where IS burned fields with the date of burning. We use satellite imagery to identify the exact coordinates of burned fields by observing pre-fire and post-fire imagery of each reported location. We study the proximity of IS attacks in 2019 to previously detected burned areas. We examine the spatial correlation between IS attacks and the extent of damage in 2019 by mapping point locations of the reported IS attacks over the post-fire burned area. We look at true color and false color pre-fire Sentinel-2 images because false color composites allow us to visualize wavelengths difficult to see by the naked eye. We use the near-infrared, red and green band to distinguish vegetation from its surrounding easily and to examine whether fires/and reported attacks have caused damage to healthy vegetation or harvested agriculture.

To disentangle the effect of increases in agricultural activity (and therefore available biomass for fueling fires), and weather conditions on fire trend, we selected air temperature from the European Centre for Medium‐range Weather Forecasts (ECMWF), and precipitation from the Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) (Dee and Uppala 2009; Funk et al. 2014), as well as our mapping of the agricultural areas in May–June within governorates of interest. We use the daily aggregates from the ECMWF (available on GEE) which has a spatial resolution of 9 km and a temporal resolution of 1-h. As for the CHIRPS, it is 6-h precipitation data provided at a spatial resolution of 0.05°. We estimate agricultural area per governorate for the past 20 years and use it as an agricultural mask over the detected burned area and to study the relation between agricultural area and percent burned area. We derive the relationship between agricultural area, standardized anomalies of precipitation, temperature, and percent burned area. We analyze the interlinkages between the changes in percent burned area, the environmental variables, and the IS phases.

Burned area accuracy assessment

Burned area accuracy assessment was performed for fires in May–June 2019. We use Sentinel-2 post-fire images along with visual interpretation to delineate the burned areas (our reference data) to evaluate the estimated burned area obtained from MODIS and Landsat 8. Validation regions were distributed over the most affected governorates — five validation sites (Fig. 1). The manually delineated boundaries cover burns that may span multiple fields. Here, we test against larger burn polygons that encompass several fields (i.e., a larger burn scar). Therefore, reference polygons were manually drawn over large burn scars which include multiple burned fields that are visually clear using Sentinel-2 post-fire images. A sample of how reference maps were created is shown in Figure S2. We evaluate the accuracy of satellite-derived burned area (from MODIS and Landsat 8) using accuracy metrics including confusion matrix, commission error (false alarm fires), and omission error (underestimation of fires) (Olofsson et al. 2014).

We also perform independent validation using the active fire data product from the visible infrared imaging radiometer suite (VIIRS) instrument (Schroeder et al. 2014) for years 2018 and 2019. Active fire products were filtered over croplands, only active fires with “high” confidence were selected for validation, and we manually filtered the active fires points through visual interpretation (Supplementary S1.2). We report the percent of active fire points that are within 0.5 km from the burned area obtained from Sentinel-2 and Landsat 8.

Results and discussion

Agricultural area vs % burned agricultural area and trend analysis results

We find a strong positive correlation between agricultural areas and percent burned agricultural areas in all governorates (Table 1, correlation in As-Sulaymaniya was positive when using Landsat 8 but not MODIS, Table S4). An interesting finding is that the percent burned agricultural area increases with the increase in agricultural area during May–June (Figure S3), i.e., a greater proportion of agricultural land was burned when the absolute agricultural area was larger. One explanation could be that (1) ISIS would have larger access to these areas and (2) it is easier for the fire to spread; hence it would be more difficult to contain. Biomass increase will create more fuel for fires which will extend to more areas at or before harvest times. During pre-harvest period, cereals become dry and are expected to cause greater fire intensity and spread fire to nearby fields. A significant (p-value < 0.05) increasing trend in percent burned area pre-harvest period for Al-Hassakeh and Diyala in 2001 — 2020 was noted (Fig. 2), contrary to the post-harvest season, where no trend exists (Figure S4).

Table 1 Pearson correlation results using percent burned area from MODIS (r = Pearson correlation coefficient). Significant values in bold
Full size table
Fig. 2
figure 2

Trend analysis results performed on the percent burned agricultural area (for months May and June and period 2001–2020). Only Al-Hassakeh and Diyala governorates show a significant increase in the trend of percent burned area

Full size image

Influence of weather variables on burned agricultural area

Impact of precipitation

Results show a positive relation (significant in Al-Hassakeh, Ninewa, and Kirkuk) between precipitation and agricultural burned area in all governorates except in As-Sulaymaniya (Table 1). In Erbil and Diyala, the relation is significantly positive when using Landsat 8 dNBR–derived burned areas for 2017–2020. The positive relation between precipitation and agricultural area (Table S3) implies a similar relation between percent burned area and precipitation. A higher precipitation means there is more biomass and therefore more fuel for fire especially during pre-harvest period. Our findings are in agreement with Schon et al. (2021) who reported that a wet year in 2019 prompted more wheat plantations which in turn created more fuel for intentional fires. Results of the standardized precipitation anomalies for 2001–2020 show that precipitation was significantly higher than the mean (2001–2020) only during 2019 for all governorates, which was the year of increased fires (Figure S6).

Pearson correlation results show a strong positive correlation between agricultural area and precipitation for all governorates except As-Sulaymaniya (Table S3). For Al-Hassakeh governorate, it is evident from Figure S7 that agricultural area varies with precipitation (mm/yr) for most years except 2018. Although precipitation (mm/yr) increased in 2018, compared to previous year, the agricultural area did not. This decrease has occurred following 2 years (2016 & 2017) of a decrease in precipitation (mm/yr) and agricultural area. Moreover, Amnesty International (2018) reported complaints from farmers that ISIS was destroying agricultural fields, wells, and irrigation systems after its loss of territorial control by end of this period (2017) — which could possibly explain the decrease in agricultural area during 2018.

For Ninewa governorate, a similar pattern is observed except for the years 2015 and 2016. During 2015, the agricultural area in Ninewa increased regardless of a decrease in precipitation (mm/yr). A possible interpretation of this result may be due to ISIS controlling 60% of croplands in Ninewa. It was mentioned by Jaafar and Woertz (2016) that during 2015, wheat and barley production was 48% greater than “pre-conflict mean” in ISIS-controlled Ninewa. Later in 2016, agricultural area decreased although precipitation did not because of conflict — where ISIS started to lose its territorial control in the region.

Kirkuk governorate shows a similar relation between agricultural area and precipitation (mm/yr) over the years. However, during 2018, similar to Al-Hassakeh governorate, agricultural area decreased although precipitation (mm/yr) was higher than the 20-year average precipitation (mm/yr). During 2018, ISIS intentionally set farmlands on fire (Amnesty International 2018) which we believe is one of the main reasons agricultural area decreased in Kirkuk. It is evident that the agricultural area in Kirkuk began to increase during ISIS’s invasion (i.e., 2013).

Air temperature and burned area

We find no correlation between mean air temperature (January to June) and percent burned agricultural area in all governorates except in Diyala, where a significant positive relationship is observed (Table 1). In 2010, Diyala experienced the highest percent BA, while the mean air temperature during 2010 was significantly higher than the mean (2001–2020) (Fig. 3). We also find that, in 2018, the mean temperature (Jan-June) z-score was significantly high in year 2018 for all governorates except Diyala and As-Sulaymaniyah (Fig. 3). However, no increase in percent burned agricultural area is observed.

Fig. 3
figure 3

Standardized temperature anomalies (January to June) against percent agricultural burned area (MODIS) for governorates of interest

Full size image

We find that when using percentages of agricultural areas rather than the burned areas metric, different results and therefore conclusions can be reached. For example, our findings that percent burned area is not correlated with temperature except in Diyala are not in agreement with those of Rasul et al. (2021), who found that drought and high temperature are the major reasons for the increase in agricultural burned area.

We further map the 2019 standardized anomalies in air temperature for the months May–June as compared to the mean air temperature of May–June from 1982 until 2020 (Figure S8). We observe that the significant increase in temperature was only observed in parts of Al-Hassakeh, Ninewa, and Duhok. The location of the agricultural burned area and the areas that have witnessed a significant increase in temperature during May–June 2019 do not overlap.

Comparing burned area from MODIS, Landsat 8 dNBR, and Sentinel-2 dNBR

Both Landsat 8 and Sentinel-2 detected more burned areas during the wheat harvest season in 2019 compared to MODIS. MODIS estimated a larger burned area during the post-harvest season compared to Landsat 8 and Sentinel-2, both of which estimated less than 10 km2 of burned land (less than 0.1% of agricultural fields) (Figure S9, Figure S10). An explanation of MODIS estimating a larger burned area during post-harvest is that some pixels are counted twice, once pre-harvest and another time during the post-harvest season, which was also noted by Vetrita et al. (2021). It is noteworthy that there is less productive and green vegetation post-harvest; as a result, the difference between NBR images is reduced due to the reduced variability in vegetation vigor, unlike pre-harvest burned area which is detected with better contrast due to vegetation being productive pre-fire (Veraverbeke et al. 2010).

Figure 4 shows the intersection between MODIS burned area product, burned area obtained from Landsat 8 dNBR, and Sentinel-2 dNBR. It is evident from Fig. 4 that Landsat 8 burned area and Sentinel 2 burned area agree more compared to MODIS burned area. The percent agreement in the burned area (Figure S11) between Landsat 8 and Sentinel 2 is the highest (32.6%) unlike the agreement between MODIS and Sentinel 2 or Landsat 8, which is less than 20%.

Fig. 4
figure 4

a Difference between MODIS and Landsat 8 burned area, b Difference between MODIS BA and Sentinel-2 BA c Difference between Sentinel-2 BA and Landsat-8 BA, and d the agreement between the three burned area products. The burned area (BA) from Landsat 8 in a is hashed, similarly the BA from Sentinel-2 in b to better show the overlap and spatial intersection in BA obtained from different sources

Full size image

Burned area accuracy assessment of MODIS and Landsat 8 burned areas

The results of the commission and omission errors obtained from the confusion matrices by comparing the reference maps (from Sentinel-2) to MODIS and Landsat 8 are shown in Table 2.

Table 2 Confusion Matrices of MODIS BA and Landsat 8 BA showing commission and omission errors, and overall accuracy
Full size table

When compared to Sentinel-2, MODIS misses 45% of small agricultural burned fields and overestimates 15%, due to pixel size. Similar findings were observed by Hall et al. (2021), where MODIS BA product missed up to 63% of burned cropland. Our results demonstrate that Landsat 8 burned area misses 28% of burned area but has fewer false fire alarms (2%) than MODIS (15%). Moreover, we find that Landsat 8 is 15% more accurate than MODIS when validated using Sentinel-2.

Figure 5 shows the percent of active fires that are within 0.5 km from burned areas obtained from Landsat 8 and Sentinel 2 dNBR. From Fig. 5, in May–June 2018, 79% of active fires were detected by the burned area from Sentinel-2 and 83% were detected by the burned area from Landsat 8. Hence, we infer that only 16–21% of active fires were missed by the dNBR burned area method. For July–August 2018, results show that Sentinel-2 burned area missed > 40% of active fire points and Landsat 8 missed > 90%. This could be due to images being more cloudy or less available during this year/month. We find that the burned areas in May–June 2019 show a higher agreement with the VIIRS active fire points. More than 90% of active fires were within 500 m from burned areas in May–June 2019 (obtained from Landsat 8 and Sentinel 2 dNBR). This means that less than 10% of active fires were missed in May–June 2019.

Fig. 5
figure 5

Percent of VIIRS active fires that are within 500 m from burned area obtained by Landsat 8 and Sentinel 2 dNBR

Full size image

Relationship between Islamic State attacks and burned fields

In Kirkuk, 42% of the burned agricultural area is within a 5-km radius from reported IS attacks during 2019, and 73% of the total burned agricultural area is within a 10-km radius from the IS attacks (Fig. 6). Similar findings are observed in Ninewa governorate and to a lesser extent in Diyala. However, in Erbil and As-Sulaymaniyah, most burned areas (> 80%) are more than 10 km away from reported IS attacks.

Fig. 6
figure 6

Percent of burned agricultural area (km2) during May–June 2019 proximity to reported IS attacks on farmlands in 2019

Full size image

We further examine the correlation between the number of IS-related attacks (in 2019) and the percent burned agricultural area per governorate. Results revealed a significant positive relationship (r = 0.83) between the number of IS-related incidents and the percent burned agricultural area during 2019. We conclude that burning of agricultural fields during May–June 2019, in Ninewa and Kirkuk, is more related to IS attacks as compared to Erbil, As-Sulaymaniyah, and Diyala.

We map the point locations of the reported IS attacks over the post-fire burned area to observe the spatial correlation (Fig. 7c and 7d). It is clearly shown that the IS attacks have caused huge damage to agricultural fields, where the burned area extends to more than 10 km. Figure 7 shows pre-and post-fire false-color imagery using Sentinel-2. The possible reason these fires were extreme and have caused larger burned area is that the fields were burning when the biomass was still on the field. This is in comparison to traditional/regular field burning when the crops have been already harvested and only thing left is the straw stubble. Stubble burning is usually over within a few hours, and it is often followed by plowing or planting by farmers and therefore is likely to result in a reduced extent of burned area damage compared to that shown in Fig. 7.

Fig. 7
figure 7

Agricultural burned area visualization. First column shows pre-fire image false color (based on bands 8, 4, 3). Second column shows during fire image (or post-fire image) false color where smoke is visible. Pre-fire image showing vegetation and post-fire image showing burned agricultural fields

Full size image

Interlinking Islamic State of Iraq and Syria phases and burned area

In Al-Hassakeh of Syria, no significant burnings have occurred during the waiting phase of IS (2006–2012). A small increase (2.5%) was noted in 2013 which we attribute to clashes between several armed groups, including Kurdish forces and IS that was trying to take control of several villages in Al-Hassakeh. We find a decrease in pBA during 2014/2015 when IS took control of Al-Hassakeh) [Phase 2Al -Hassakeh: 2014 & 2015]. In May–June 2019, pBA in Al-Hassakeh increased to 8%, which is twelve times greater than average pBA2001–2020. Similar to 2019, an increase in pBA is observed during May–June 2020 but to a lesser extent (seven times greater than average pBA2001–2020). Apart from outliers in 2010 in Ninewah and Diyala and Kirkuk in 2006, fires increased considerably during the withdrawal phase and then again during the in-hiding phase (see Fig. 8).

Fig. 8
figure 8

Percent burned agricultural area (obtained from MODIS BA product) for years 2006–2020

Full size image

A new finding of our study is that in contrast to other provinces the % burned area in Iraq’s Ninewa province at the end of the power phase and the beginning of the withdrawal phase (2015–2016) exceeded the 2019–2020 percent burned area. Our analysis reveals the extent of the deliberate arson of farmland in Iraq and Northeast Syria in 2019–2020. In comparison, only a few fires of farmland were observed during the in-waiting phase in 2007–2013 in Iraq, and Al Hassakeh, Syria, that became a stronghold of the IS in 2013. In 2014–2017, we found evidence for increased farmland fires in various locations that closely tracked the military withdrawal of IS between 2015/2016 and 2017 when it lost most of its territory.

When IS slipped back into renewed clandestinity from 2017 onwards, there was not an immediate uptick in sabotage acts. The extent of the arson in 2019 and 2020 raises questions about the motivations of the enhanced use of intentional burning of farmland by remnants of the IS. These questions cannot be answered by remote sensing and are beyond the scope of this study. Only detailed field studies could unearth dynamics at play and they are difficult to undertake in the areas in question. Such dynamics could include spite against population segments that IS suspected of collaboration with victorious forces of the Iraqi army, the Shiite Popular Mobilization Forces (al-hashd al-sha’bi), and the Kurdish Peshmerga a new way to subsist as a protection racket after the IS’ state aspirations had evaporated or an expectation that contrary to the “in waiting” period before 2013/2014 the group itself does not expect to return to power and rule territory any time soon.

What can be said at this stage is that the IS has manipulated the ecology in its conflict with its adversaries (Le Billon and Duffy 2018) in a way that marks a departure from its earlier approaches to the water-food complex, adding another layer to the complex interrelations between environmental damage and conflict in Iraq. For the time being, it is important to monitor whether intentional burning of farmland and related scorched earth tactics will be a regular occurrence in Iraq and the Northeast of Syria. This could be done with the remote sensing analysis and methodology suggested in this article. The implications of such fires for socio-economic development, conflict dynamics, and the security situation could then be investigated by future field-based studies.

Conclusion

Our findings stress the importance of selecting the appropriate metric when trying to identify the causes and extents of burned area. We highlight the importance of validating remote sensing results especially those with a coarse spatial resolution (MODIS), and running the same trend analysis using higher-resolution datasets (such as Sentinel-2 and Landsat), as sometimes different results can be attained (for example correlation between agricultural areas and percent burned areas in As-Sulaymaniyah — a Kurdish controlled region — was positive when using Landsat 8 data but not when using MODIS data).

We analyze the burned area using different satellite products to examine the extent of the unusual increase in agricultural burned area in 2019 and 2020 in parts of Syria and Iraq. The spike in agricultural burned area that we examine in this study occurred during May and June during the harvest season of wheat (wheat is harvested end of June), which has significant implications on food security and stability in both countries. Knowing that the number of violent conflicts during May and June of 2019 has not increased or changed as compared to previous years (Figure S12), we believe that ISIS has lost control and power in 2019/2020; hence, ISIS has shifted to a different technique: destroying agriculture because they do not have enough power to control the agriculture areas and wheat/barley harvest. In an aim to improve our understanding of the possible causes of this increase, our study highlights the complex relationship between climatic variables, conflict, agricultural activity, and burned area. Unless we are able to quantify the area of irregularly shaped burned agricultural fields, it remains difficult to estimate the extent of agricultural damage caused by ISIS or other armed groups. More research is needed to overcome the limitations of detecting small field fires and classify them based on their shape, especially in inaccessible conflict-affected areas.