Find the information such as human life, natural resource,agriculture,forestry, biotechnology, biodiversity, wood and non-wood materials.
Blog List
Saturday, 17 December 2016
A MODIS-based burned area assessment for Russian croplands: Mapping requirements and challenges
Published Date October 2016, Vol.184:506–521,doi:10.1016/j.rse.2016.07.022 Open Access, Creative Commons license Author
Joanne V. Hall a,,
Tatiana V. Loboda a
Louis Giglio a
Gregory W. McCarty b
aUniversity of Maryland, Department of Geographical Sciences, 2181 Samuel J. LeFrak Hall, 7251 Preinkert Drive, College Park, MD 20742, USA
bUnited States Department of Agriculture, Beltsville Agricultural Research Service, 10300 Baltimore Ave #302, Beltsville, MD 20705, USA
Received 24 October 2015. Revised 14 June 2016. Accepted 16 July 2016. Available online 4 August 2016.
Highlights
•
A custom MODIS-based burned area product to quantify cropland burning in Russia
•
Coarse resolution (≥ 500 m) satellites inadequate for accurately mapping crop burning
•
Successful mapping requires moderate (10–50 m) resolution, daily overpasses.
•
Future potential with Sentinel-2 and Landsat constellation missions
Abstract Although agricultural burning is banned in Russia, it is still a widespread practice. Accurately monitoring cropland burned area is an important task as these estimates are used in the calculation of cropland burning emissions, which are ultimately utilized in policy making decisions. In this paper we developed an independent estimate of cropland burning in Russia through assessing the capabilities of global burned area products (MCD45A1; Roy et al., 2008 and MCD64A1; Giglio et al., 2009) and our own custom Moderate Resolution Imaging Spectroradiometer (MODIS) based Cropland Regional Area Burned (CRAB) product. An archive of cropland field state samples was generated from very high resolution (< 5 m) imagery allowing us a unique perspective into the challenges of mapping cropland burned area through detailed analysis of the Russian agricultural practices. Our analysis showed all three burned area products were unable to map approximately 95% of burn validation samples, demonstrating that the current coarse resolution (defined here as ≥ 500 m) satellite capabilities are not adequate for accurately mapping burned area associated with agricultural fires. The transient nature of cropland burns and the climatological conditions of the region require multiple subsequent daily overpasses; however, the small spatial scale of the burns requires moderate (defined here as 10–50 m) spatial resolutions. Future opportunities to accurately map cropland burned area may arise with the Sentinel-2 and Landsat constellation missions. Keywords
Burned area
Cropland residue burning
Russian Federation
Remote sensing
MODIS
1 Introduction
Anthropogenic transformations of the terrestrial biosphere have greatly altered the earth's land surface. By 2011, approximately 40% of the ice-free land surface had been modified for agricultural activities (FAOSTAT, 2015a). These managed areas have replaced forests, natural grasslands and savannas and have led to a range of impacts on the planet's ecosystem structures and functions (Ramankutty et al., 2008). The impacts of these activities vary greatly due to the differing management practices applied throughout the various global agricultural regions. The use of fire in agricultural management for removal of excess residue and the control of diseases and pests is a common practice across different regions from developing to developed countries (EPA, US Environmental Protection Agency's Agricultural Centeand McCarty et al., 2012). Although prescribed burning is a useful agricultural management tool, concern is growing in the scientific community regarding the negative consequences on regional climates and air quality due to associated emissions of gases and particulates (Foley et al., 2005).
Impacts from agricultural activities are not limited to areas within these managed ecosystems. Chemical studies of particles deposited on Arctic snow indicate that northern mid-latitude agricultural burning and forest fires are the dominant sources of black carbon (BC) deposition above the Arctic Circle (Hegg et al., 2009, Koch and Hansen, 2005, Pettus, 2009 and Stohl et al., 2006). BC is a product of incomplete combustion of carbonaceous fuels and new estimates place BC as the second most important human emission in regards to climate forcing, behind carbon dioxide (+ 1.82 Wm− 2; Stocker et al., 2013) with a total climate forcing of + 1.1 Wm− 2 with 90% uncertainty bounds of + 0.12 to + 2.1 Wm− 2 (Bond et al., 2013). Unlike the mid- and low latitudes, the Arctic has a unique role in the earth climate system through increased surface reflectance created by the region's high snow and ice albedo. Both natural and human-induced changes to the region's albedo have resulted in the accelerated warming of the Arctic. Specifically the deposition of BC, transported from northern mid-latitudes, is a significant contributor to observed short term warming trends in the Arctic (Bond et al., 2013, Hansen and Nazarenko, 2004, Jacobson et al., 2007, Ramanathan and Carmichael, 2008, Quinn et al., 2008 and Quinn et al., 2011).
The identification of BC emissions from fires as a contributor to Arctic warming has led to a surge of interest in limiting or eliminating prescribed burning, particularly in spring, to reduce the warming effect of BC (Zender, 2007). There are large uncertainties in the current estimates of the sources, source regions, and transport and transformation pathways of BC transported to the Arctic region (Hegg et al., 2009, Shindell et al., 2008 and Quinn et al., 2008) and many of these uncertainties stem from the current inaccurate estimates of cropland burned area. In this paper, we aim to develop an independent estimate of cropland burning in Russia through assessing the capabilities of global burned area products and our own custom Moderate Resolution Imaging Spectroradiometer (MODIS) based burned area product to quantify cropland burning in Russia. Russian agriculture is of global importance, however, relatively little is known about the crop management strategies and in particular the burning practices in rural areas (Pettus, 2009). Russia is the world's fifth-largest wheat exporter and impacts on the wheat harvest can have large implications on global markets (FAOSTAT, 2015b, Grumm, 2011 and Kramer, 2010). Although federal law banned open burning, Russia and Kazakhstan are responsible for the largest portions of BC emissions from biomass burning that reach the Arctic, followed by China, the United States and Canada (Pettus, 2009).
Accurate monitoring of the temporal dynamics of prescribed burning is crucial as studies have found that the timing of the burning plays an important role in determining the magnitude of the BC impact on the Arctic (Stohl, 2006 and Warneke et al., 2010). In particular, early springtime cropland burning has a surprisingly large impact on the Arctic considering cropland fire injection heights are lower than forest fires and they have a much shorter duration of burning. Even with low injection heights, under the right atmospheric conditions, BC particles can be transported long distances. One such method of low-level transport is related to surfaces of constant potential temperature (Stohl et al., 2007). During the spring, Arctic temperatures are higher, thus reducing the temperature contrast between the northern mid-latitude source regions (40°N and above) and the Arctic (Law and Stohl, 2007). This reduction in the temperature contrast leads to the creation of efficient pollution pathways that facilitate low-level transport of BC from northern mid-latitudes to beyond the Arctic Circle; hence it is imperative to understand both the spatial and temporal patterns of cropland residue burning in the northern mid-latitudes.
Currently, accurate monitoring of cropland burning from existing active fire and burned area products is limited. The transient nature and non-contiguous patches of cropland fires requires a tailored mapping algorithm that is designed to capture the subtle nuances of the cropland spectral signature. Many existing burned area algorithms are focused on mapping hotter and larger wildfire events and several have global extents (MCD45A1; Roy et al., 2008 and MCD64A1; Giglio et al., 2009). Even with regional burned area algorithms (Loboda et al., 2007 and McCarty et al., 2008), the inclusion of the MODIS Active Fire Product (MCD14ML; Giglio et al., 2003) as a driver of the mapping algorithm will inhibit the ability to capture the full extent of burning in the cropland regions. For example, McCarty et al. (2008) found that MODIS active fire detections contributed to < 4% of the total burned area estimated in the United States croplands within MODIS tile h10v05, which encompasses a large portion of the Mississippi River Delta and the southern Great Plains. In addition, burned area algorithms which do not address small fires, defined as fires smaller than the spatial resolution of the surface reflectance imagery, potentially underestimate the total burned area particularly in cropland regions where fires are both small and transient (McCarty et al., 2009, Randerson et al., 2012 and Wu et al., 2015). Recent studies have found that utilizing the MODIS Active Fire product helped identify crop residue burning, particularly in small fields which are not able to be detected through traditional difference Normalized Burn Ratio (dNBR) threshold methodologies (McCarty et al., 2009 and Randerson et al., 2012).
In this paper, we focus on the creation of a customized MODIS-based burned area product for the Russian cropland area and the comparison with the existing MCD45A1, MCD64A1 and MCD14ML products to highlight the challenges in mapping cropland burning. Section 2 focuses on the data and methods with particular emphasis on the development of a unique archive of samples extracted from multispectral very high resolution (VHR, defined as < 5 m resolution) imagery (Section 2.1) and the development of a customized MODIS-based burned area product (Section 2.2), including the creation of a multi-sensor clear-surface compositing algorithm (Section 2.3). Section 3 focuses on the accuracy assessment, results and intercomparison between the burned area products, followed by a discussion of the challenges of mapping cropland burned area (Section 4) and concluding with recommendations for future cropland burned area studies.
2 Data and methods
The Cropland Regional Area Burned (CRAB) algorithm was designed to improve quantification of the spatial and temporal variability of burned area from Russian cropland residue burning between 2003 and 2012. The CRAB algorithm maps daily burned area over the Russian cropland at 500 m resolution. The generated outputs include seasonal burned area maps, containing the Julian Date for the mapped burn date for each burn pixel, and confidence layers. The CRAB algorithm is divided into two seasons based on the assumption that croplands can burn more than once a year (e.g. before planting and/or after harvest) thus resulting in two distinct burning seasons: spring (1 March–30 June) and summer (1 July–30 September). These were defined based on the analysis of the frequency distribution of the MODIS Active Fire product (MCD14ML) (Fig. 1) and the general planting and harvesting practices (described below, USDA FAS, 2013). It is further assumed that a given pixel can burn only once during one of the burning seasons (cumulatively no more than twice during a year).
Fig. 1. MODIS Active Fire (MCD14ML; Giglio et al., 2003) counts within the Russian cropland area as defined by the crop mask (see Fig. 4) between 2003 and 2012.
Analysis of the MODIS Active Fire product between March and September 2003 to 2012 revealed an interannual range of approximately 60,000–166,000 active fire counts within the Russian cropland region. The MODIS instrument acquires daily observations of fire activity from two satellites – Terra, launched in late 1999, and Aqua, launched in 2002. Only active fire detections acquired between 2003 and 2012 were included in the analysis to ensure the temporal consistency of the dataset. The variation in the distribution is related to the agricultural practices of the different regions. In particular, a common pattern relates to the spring and winter wheat cycles. Typically the winter wheat campaign begins in mid-August in the Volga District and advances southward. The crops enter dormancy in late October/early November, whereby in an average year, approximately 13% of winter grains are lost over the course of winter due to severe frost damage and other winter weather-related impacts (USDA FAS, 2013). Harvesting of the remaining winter wheat crops begins in late June and concludes in late August. On the other hand, the spring wheat campaign begins in April and the harvest commences in August and is finished by late October. Analysis of the active fire burning distribution, within the Russian cropland region, found that there were generally two distinct burning seasons (Fig. 1), thus the CRAB algorithm was divided into spring (1 March–30 June) and summer (1 July–30 September).
Fig. 2 shows the overall framework of CRAB. Multispectral VHR imagery was acquired from DigitalGlobe's high resolution commercial satellites, QuickBird and World View-2 for the creation of the training and validation samples. The algorithm ingests the daily surface reflectance 500 m Level 2G MODIS imagery acquired from Terra and Aqua (MOD/MYD09GA; Vermote et al., 2011) satellites between 2003 and 2012. A combination of the 1 km quality assessment layers and angle geometry layers were extracted from the Aqua (MYD09GA) and Terra (MOD09GA) datasets and were composited using a multivariate decision making approach to select the best quality observations with intent to minimize false detections (see Section 2.2 for more detail). To ensure the algorithm only processed observations over the Russian cropland region, a crop mask was created using the cropland and cropland/natural vegetation mosaic layers from the 2011 500 m Level 3 International Geosphere-Biosphere Programme (IGBP) classification scheme within the MODIS land cover dataset (MCD12Q1; Friedl et al., 2010). After analyzing the variability in the IGBP cropland and cropland/natural vegetation layers, we are satisfied with utilizing this product in the CRAB algorithm (see Supplementary material for more details). In addition, the MODIS 250 m land-water mask (MOD44W; Carroll et al., 2009) was aggregated to 500 m using a strict aggregation rule to avoid water contamination and applied to the IGBP crop mask. The 500 m grid cell was flagged as water if any of the four 250 m grid cells were identified as water. Since wet soil can often trigger a response in the SWIR bands we opted to reduce any confusion with small water bodies.
Fig. 2. Cropland Regional Area Burned pre-algorithm development and methodological framework.
2.1 Visual assessment of fire practice in Russian croplands from VHR imagery
The high spatial resolution (< 5 m) of the multispectral VHR imagery is a much needed element for the identification and discrimination of various cropland field states, including, active fire verified burn, visually identified burn, plow, residue and bare fields (Fig. 3). A particular challenge with cropland burned area mapping is the difficulty in distinguishing burned from plowed fields (McCarty et al., 2009 and Roy et al., 2005). The VHR imagery allows for more accurate identification of field condition through visual interpretation and from the incorporation of the MODIS Active Fire (MCD14ML) dataset as a means of independent validation for the selection of burned training and validation samples (see Section 2.1.3).
Fig. 3. Very high resolution (< 5 m) field state examples: burned field (a and f), residue field (b), freshly plowed field (c and e) and bare field (d). Fig. 3a shows a burn with a flaming, smoking front which is easily identifiable as a burned field, whereas burned fields without flaming fronts (Fig. 3f) are identified by the natural fingerlike patterns created from a moving flaming front. Typically these burned area examples are located within residue fields. Fig. 3b identifies a residue field which is a brighter field, typically with linear features which correspond to the rows of residue stubble. Plowed fields (Figs. 3c and e) are identified by a dark, smooth texture which tends to start at the field edges. Bare fields (Fig. 3d) are fields which do not contain large amounts of residue or vegetation and do not show any signs of a fresh burn or plow. Images are displayed using the true color composite.
For our archive, dark, smooth textured fields were digitized as plowed. Typically these fields were fairly homogenous and the dark uniform color filled the entire field area (Fig. 3c and e), whereas burns without the flaming front were identified by the rough texture and natural fingerlike patterns created from the moving flaming front (Fig. 3f).
Quantifying the extent and magnitude of cropland burning is a challenging task as there is little consensus amongst the available data. Visual analysis of the VHR imagery reveals a larger proportion of cropland burned area than what is indirectly detected through the MODIS Active Fire product and what has been reported by farmers through telephone surveys undertaken as part of a USDA Black Carbon Initiative project (SovEcon, 2013). According to the survey results, farmers have noticed a significant reduction in agricultural burning, with some respondents claiming agricultural burning has not been observed in the Rostov region for at least the last 10 years, the Stavropol region for 5–7 years, the Krasnodar region for 2–5 years, the Saratov region for 3 years, and the Altai region for 2 years (Fig. S1). However, the survey results also claim that mainly small agricultural enterprises in these regions still burn for a variety of reasons which include the reduction of soil harbored plant diseases in the Rostov region and the removal of straw for the improved crop planting and germination success in areas of the Saratov and Voronezh regions. Further analysis of the active fire product revealed burning in all these areas between 2003 and 2012; however, the magnitude and temporal distribution varied across all regions.
2.1.1 Archive development of field samples under various post-harvest conditions (burn, bare, standing crop residue, plowed) from VHR imagery
We have created an archive of various field state samples within the Russian cropland region from VHR imagery (Fig. 4). The VHR samples were visually identified and digitized by several image analysts (for details see Section 2.1.3). The low repeat frequency of VHR satellites and their small footprint size (e.g., ~ 360 km × 18 km for QuickBird) leads to an opportunistic acquisition strategy and therefore the archive does not cover the full extent of the Russian cropland region.
Fig. 4. Very high resolution (< 5 m) samples (2003 − 2012) acquired from DigitalGlobe's high resolution commercial satellites, QuickBird and World View-2 overlaid over the Russian cropland.
From the analysis of the active fire counts, it is clear there is a substantial amount of burning within Russian cropland. Using the archive of VHR data, we examined these burns in greater detail than can be accomplished through the current MODIS products. This archive allows us a unique perspective into the challenges of mapping cropland burned area through the detailed analysis of the Russian agricultural practices. Table 1 below presents the size distribution of the entire VHR sample archive. Analysis of the size distribution of the cropland burns and field samples found a large proportion of the burn and non-burn samples occupying < 21 ha, which is the approximate equivalent of one 463 m MODIS grid cell.
Table 1. Distribution of burn and non-burn samples digitized from very high resolution (< 5 m) imagery. Values in parentheses are approximate equivalent MODIS 500 m grid cells based on the size of the samples in hectares.
Burn Samples
Field Samples
Minimum, ha (~ grid cell)
0.5 (< 1)
2.4 (< 1)
Maximum, ha (~ grid cell)
411 (20)
656 (31)
Mean, ha (~ grid cell)
57 (3)
88 (4)
Median, ha (~ grid cell)
40 (2)
71 (3)
Standard Deviation, ha (~ grid cell)
56 (3)
68 (3)
Skewness
2.3
1.9
Kurtosis
7.8
5.6
Total Number of Samples
782
6037
Depending on the exact configuration of the field relative to the 500 m MODIS sinusoidal grid, the actual relationship between the MODIS grid cells and the sizes of burned or total field areas can be substantially smaller. Here we are presenting the best case estimates. Furthermore, in general the burned area samples are smaller than the field area as only the burned portion of the field was digitized when creating the burn sample. The distribution of field sizes and burned area both have a positively skewed relationship (Fig. 5) with the majority of burns occupying < 50 ha.
Fig. 5. Burn sample area (above) and field area (below) frequency distributions and cumulative area (× 103 ha) for all VHR polygon samples before the matching process. The vertical dashed line indicates a change in area size categories along the horizontal axis.
2.1.2 Visual examination of burn strategies identified in VHR imagery
Further analysis was carried out on the burn samples to identify any additional potential challenges. Visual examination of burning practices in the available VHR imagery found a variety of burning strategies including, complete field burns, partial field burns and pile burning (Fig. 6).
Fig. 6. Visual burning practices found in available very high resolution (< 5 m) imagery: complete field burn (a), partial field burn (b), partial field burn (c), and pile burn (d).
From visual analysis, the full burns (Fig. 6a) are associated with burns that leave little to no unburned residue in the burn extent that can be seen in the VHR image, whereas the partial burns (Fig. 6b and c) include varying degrees of unburned residue within the burn extent. Depending on a number of factors including the amount of biomass within the field and local burn conditions, these partial burns can cover the entire field extent (Fig. 6c) or can only begin to burn in a small section of the field (Fig. 6b – blue circle). Finally pile burns (Fig. 6d) are associated with setting fire to piles of residue. Analysis of the VHR burn samples revealed that the practice of pile burning has an impact on the burn spectral signature due to the larger percentage of non-burned residue within the burned area extent. Although excluding both pile and partial burn samples would improve the purity of the burn spectral signature, the majority of the burn samples are not full burns and do not give a good representation of the agricultural burn practices within Russia (as seen in the VHR archive). Furthermore, visual analysis of the VHR archive found additional nuances in the agricultural practices including fields which were not burned before plowing (see Fig. S2). These variations in agricultural practices and the hesitancy of farmers to disclose their burning methods leads to further hindrances on the ability to accurately map cropland burned area.
2.1.3 Methodology for the development of the post-harvest field condition archive
To assist the image analysts in the initial sorting of the VHR images, the MODIS Active Fire product was incorporated to highlight any VHR images with a potential active fire within ten days of the image date. Each VHR image was then examined and any burns within approximately a 1 km radius, to account for off nadir viewing angle geolocation inaccuracies, of the Active Fire point were manually digitized and classified as an active fire verified burn. Any burns which were not highlighted by the Active Fire dataset, but were visually identified by the analyst, were classed as a burn and not an active fire verified burn. Visually identifiable burned fields constituted a very small fraction of all fields in the processed VHR imagery. For each processed image, the analysts were instructed to digitize all burns and collect 5 samples per non-burn field state (plow, bare and residue) whether the image contained any burned fields or not. Following the manual digitization of the VHR imagery by an image analyst and subsequent collection of the various field type samples (burn, plow, bare and residue), the samples were georegistered to orthorectified Landsat imagery and then geometrically matched through the rasterization of the VHR polygons to the 500 m Level 2G MODIS imagery (MOD/MYD09GA; Vermote et al., 2011). Approximately 4% of the burn samples (average size ≤ 1 MODIS grid cell) were lost during the matching process as the configuration of the field sample relative to the 500 m MODIS grid cell led to the splitting of the sample between adjacent grid cells and therefore the smaller samples were unable to be rasterized to the coarser MODIS resolution based on the 50% majority aggregation rule. Upon completion of the matching process, the field samples were temporally stratified by typology based on a randomized sampling scheme to separate the samples into training or validation categories based on the dates and field type. The final totals of VHR spring and summer samples are highlighted in Table 2.
Table 2. Very high resolution (< 5 m) spring and summer samples matched to MODIS 500 m grid cells.
Training
Spring
Summer
Burn
480 grid cells
1312 grid cells
Bare
835 grid cells
3171 grid cells
Residue
693 grid cells
3401 grid cells
Plow
662 grid cells
2491 grid cells
Validation
Spring
Summer
Burn
369 grid cells
762 grid cells
Bare
762 grid cells
1872 grid cells
Residue
527 grid cells
1996 grid cells
Plow
579 grid cells
1616 grid cells
2.2 Clear surface analysis
Accurate mapping of cropland burned area requires multiple subsequent daily observations to alleviate one of the largest sources of error being the inability to distinguish burned from plowed fields. This is particularly relevant in the Mollisol regions of southern Russia (McCarty et al., 2009 and Roy et al., 2005). Although MODIS imagery is acquired daily, the time series is not continuous as periods of prolonged clouds, aerosols and snow can limit the number of available clear surface days. A new compositing algorithm has been developed to combine Aqua and Terra data to improve the availability of clear surface views and to avoid forward scattering which results in artificial darkening of the spectral signature. The 1 km MODIS quality assessment (QA) data layers were extracted from the MOD09GA and MYD09GA products for use in the clear surface analysis over the Russian cropland region as defined by the crop mask. Each image contains a corresponding QA layer which holds 16 bit flags that summarize 11 different parameters. For this analysis, only bits related to clouds, aerosols and snow were chosen (Table S1). The QA bits were extracted into daily binary layers, representing the clear surface days for both Aqua and Terra.
2.3 Multi-sensor clear surface composite
Due to the non-Lambertian nature of cropland fields, the bidirectional reflectance characteristics varies considerably, leading to large differences in the observed reflectance (Rahman et al., 1999, Ranson et al., 1985 and Schaaf et al., 2002). Analysis from a sample field's spectral output illustrates the impact on the burn spectral signature (Fig. 7). Reflectance values are expected to decrease in the visible and near infrared bands following a burn due to the deposition of char on the brighter soil surface (Zhang et al., 2003). This decrease is visible in Terra's output following the burn on Julian Date 229 (17 August 2006, 10:04:20 MSK; Fig. 7b), whereas there is a clear increase in reflectance in the spectral output of Aqua (17 August 2006, 12:11:09 MSK; Fig. 7c). Further analysis on the solar and view zenith and azimuthal angles found the discrepancies in reflectance values between the two datasets were related to the differing solar and viewing angle geometries and it is clear that the higher reflectance values correspond to a back scatter scenario with zenith angles far off nadir (Fig. 7c). This same pattern is seen in a number of studies whereby the reflectance values generally decrease from the backward to the forward scattering direction (e.g. Galvao et al., 2009 and Rahman et al., 1999).
Fig. 7. Solar (Ts, As) and sensor (Tv, Av) azimuthal and zenith angles for Terra and Aqua over 16–19 August 2006. The verified active burn occurred on 17 August 2006. The zenith angles are represented by the radius of the black line. Zenith angles scaled between 0° and 50° for visualization. The near infrared reflectance (NIR) values are located in the lower right boxes.
The following method outlining the combination of Aqua and Terra daily observations aims to minimize the impact of bidirectional reflectance distribution function (BRDF) related decreases in surface reflectance with the general aim to lower the commission error of the CRAB product (Fig. 8).
Fig. 8. Methodological framework for the combination of Aqua and Terra clear surface datasets into one BRDF adjusted time series.
In the combination of the two datasets, the solar zenith (SZ), view zenith (VZ), solar azimuth (SA) and view azimuth (VA) angles were extracted from the MOD09GA and MYD09GA products. Following the outline in Fig. 8, the angle geometry algorithm ingested the four angles for the creation of the three variables used in the calculation of the Viewing Indicator Geometry (VIG) index:
Variable 1: Linearly scaled solar zenith angle (SZs) for both Aqua and Terra between 0 (90° off nadir = undesired) and 1 (nadir = desired).
Variable 2: Linearly scaled absolute azimuthal difference (ADs) for both Aqua and Terra between 0 (0° or 180° diff = undesired) and 1 (90° diff = desired). We preferred to avoid forward scattering (artificial darkening) and backward scattering (artificial brightening) and rather considered values between both situations.
Variable 3: Linearly scaled view zenith angle (VZs) for both Aqua and Terra between 0 (90° off nadir = undesired) and 1 (nadir = desired).
The three variables were summed and weighted equally in the calculation of VIG. The difference between VIG Aqua and VIG Terra (dVIG) was evaluated against our chosen threshold. If the difference was less than − 0.15 Aqua was chosen; whereas if the difference was > 0.15 then Terra was chosen. If the value lay between these two thresholds then the smallest azimuthal difference, corresponding to a back scatter scenario, as calculated in variable 2 was chosen. The composition process can be seen in Fig. 7. On 16 August 2006, the VIG values are almost identical, leading to a small dVIG (− 0.03). According to the threshold criteria, the low dVIG requires the lowest ADs value to be chosen. In this instance Terra was slightly closer to 90°, whereas on 17 August 2006 Aqua's geometry was preferred as Terra was closer to a more intense backscatter scenario. In this instance Aqua was preferred as Aqua's VZ angle was also closer to nadir. The combination angle geometry and clear surface layers created a combined multi-sensor clear surface composite which was subsequently utilized in the CRAB algorithm (Fig. 2).
2.4 Burned area map and confidence level development
Upon completion of the clear surface composite, a single daily image of clear surface observations combined from Terra and Aqua is evaluated for burned area detection. The CRAB algorithm uses a decision tree classifier that differentiates between burn and non-burn (residue, plow and bare) classes and assigns a probability to each of the burn terminal nodes. A total of 690 training fields (116 burned and 574 unburned) were utilized in the spring classification tree and 2249 training fields (326 burned and 1923 unburned) were utilized in the summer tree. A set of 12 metrics was created as an input into the decision tree for each burn and unburned training sample split between the two mapping seasons, spring and summer. These inputs were a combination of spectral reflectance bands (ρ); band 1 (620–670 nm), band 2 (841–876 nm), band 3 (459–479 nm), band 4 (545–565 nm), band 5 (1230–1250 nm), band 6 (1628–1652 nm) and band 7 (2105–2155 nm), and derived spectral indices, including; Normalized Burn Ratio ([ρ2 − ρ7]/[ρ2 + ρ7]; Lopez and Caselles, 1991), Normalized Difference Vegetation Index ([ρ2 − ρ1]/[ρ2 + ρ1]; Rouse et al., 1974), Vegetation Index ([ρ5 − ρ7]/[ρ5 + ρ7]; Giglio et al., 2009) and Normalized Difference Water Index (Gao, 1996) for both band 5 ([ρ2 − ρ5]/[ρ2 + ρ5]) and band 6 ([ρ2 − ρ6]/[ρ2 + ρ6]).
An output decision tree was generated for each of the two seasons. The spring tree has 29 terminal nodes of which seven classify burned fields and 22 classify unburned fields with an overall misclassification rate of 11%. The summer tree has six terminal nodes of which one classifies burned fields and five classify unburned fields with an overall misclassification rate of 14%. The spring tree utilized all input metrics for the classification of the burned fields while the summer tree only utilized ρ1, ρ3, ρ6, NBR and NDWI5 in the classification of the burned fields. The classification rules for each season's tree are reported in Table 3.
Table 3. CRAB burned area classification rules (A–G) for spring and (A2) for summer. A field is classified as burned if it satisfies each condition in any of the classification rules.
The algorithm is designed to monitor the spectral signature and look for a persistent burn signal over a five-day moving window. The CRAB product is generated for each season between 2003 and 2012; however, the algorithm is computed on a daily time step. Each seasonal output has a corresponding confidence layer, which was determined through using a five-day weighting scheme (Conf. = 0.4 day1 + 0.2 day2+ 0.1 day3 + 0.1 day4 + 0.1 day5) over five consecutive days to check for signal stability (here dayi represents the original confidence values assigned from the decision tree nodes for the i'th day within the five-day window). The final output records the date of burn detection, which was assigned on the first day that the spectral thresholds were crossed, within the season and the confidence of burning.
The CRAB product is comprised of two separate decision trees because the biophysical conditions are vastly different between spring and summer. The spring burning conditions are extremely challenging as the excess moisture in the soil from the snow melt lowers the surface reflectance across the full visible – SWIR range with a particularly strong impact on the SWIR range resulting in further diminished spectral separability between burned and unburned fields. In contrast, in the summer soils are substantially drier as compared to the post snow melt period in the spring and fields are typically covered in either crop residue or regrowth of weeds, both detectable in the NIR range. However, even in the summer plowed and wet fields can still be easily confused with burns because of the small difference in reflectance across all bands. Due to these differences, we empirically determined that burned grid cells below a confidence of 60% for the CRAB spring product are not as reliable as in the CRAB summer product and therefore we have applied a mapping confidence threshold of > 60% for the CRAB spring product in the subsequent analysis. Fig. 9 illustrates an example of the CRAB output for summer 2012 for all confidence values. The grey denotes the cropland region. The inset highlights the grid cell based Julian Date of burn detection.
Fig. 9. Summer 2012 CRAB product output. A subset of Russia (red circle, top insert) is shown for display purposes. Colors represent Julian Date of burn. Grey area denotes the Russian crop mask.
Upon creation of the CRAB seasonal burned area maps and corresponding confidence layers, the product's accuracy was assessed using the VHR validation samples and compared to the MCD45A1, MCD64A1 and MCD14ML products. All three burned area products (MCD45A1, MCD64A1 and CRAB) are designed to map burned area and are based on differing principles. MCD45A1 is based on the deviation from the anticipated reflectance change due to burning with applied corrections for BRDF. MCD64A1 is based on the dynamic spectral thresholds generated from a burn sensitive vegetation index coupled with the inclusion of the MODIS active fire observations and a measure of the change in temporal texture following a potential burn. CRAB is a decision tree based algorithm constructed from a suite of spectral bands and indices with observations selected to minimize BRDF effects.
3 Results and discussion
3.1 Accuracy assessment
An accuracy assessment was carried out on CRAB using a ten-day window prior to the date of the validation samples. Due to cloud obscuration of the land surface, a ten-day window is required to enhance the probability of mapping the cropland burns. Our decision to use a window of this duration was based on the transient nature of cropland fires and the potential of plowing or reseeding soon after a burn. The persistence of these low biomass burn scars varies as a function of climate, fuel load and crop management practices. Several studies focusing on grass dominated systems such as cropland, savanna and grasslands have found that the persistence of the burn scars is short lived (days to weeks) and therefore long temporal windows are likely to lead to an increase in burn area omission errors (Korontzi et al., 2008 and Trigg and Flasse, 2000). In addition, for regions with double cropping, such as in the Mississippi River Valley, a field can go from harvest to burning and reseeding in a few days to two weeks (Korontzi et al., 2008). This quick turnover and short burn-scar persistence justifies the use of a comparatively short temporal window. While the optimal duration of this value is debatable, it is evident that a 30 day window is too long as the burn signals are virtually always lost to weathering and/or plowing within this period of time (Chuvieco et al., 2008).
Here the samples were applied to the Julian Date of burn detection as described by the CRAB product. If the pixels associated with the validation samples were flagged as burned within a ten-day window prior to the date of the image, then it was counted in the accuracy assessment. An assessment using the VHR spring and summer validation samples was carried out between CRAB, MCD45A1 and MCD64A1 (Table 4). Additionally MCD14ML was assessed to identify the number of VHR validation samples that were intersected by a 1 km radius buffered active fire point within ten days of the image date.
Table 4. Spring and summer accuracy assessment comparison between the Cropland Regional Area Burned (CRAB) product and the two official MODIS burned area products MCD45A1 and MCD64A1. The spring assessment was carried out using the full range of CRAB confidence values and > 60% confidence. The active fire assessment results (MCD14ML shown in italics) are based on the number (percentage) of intersections by a 1 km radius buffered active fire point within ten days of the image date.
Commission error
Omission error
Residue
Bare
Plow
Burn
Spring
CRAB (all conf.)
9.11
9.97
10.36
86.98
CRAB (> 60%)
0.00
0.00
0.00
99.42
MCD45A1
0.00
0.00
0.86
97.30
MCD64A1
0.00
0.00
0.00
99.19
MCD14ML
0.42
0.38
1.84
91.54
Summer
CRAB (all conf.)
0.45
1.28
4.27
96.86
MCD45A1
0.60
0.91
2.66
92.65
MCD64A1
0.00
0.00
0.00
96.73
MCD14ML
1.20
1.90
1.60
89.70
All four products (MCD45A1, MCD64A1, MCD14ML and CRAB) were unable to map approximately 95% of the burn validation samples during summer (93%, 97%, 90% and 97% omission errors, respectively), while in spring there was an improvement using the CRAB product with full confidence values (97%, 99%, 92% and 87% omission error, respectively); however, applying a confidence of > 60% led to a CRAB omission error of 99%. The reduced commission error in the MCD45A1 and MCD64A1 products as compared to the CRAB product are due in part to the restrictiveness of their algorithm's spectral mapping thresholds. The mapping thresholds for CRAB can be constrained further to reduce commission error; however, this causes an increase in omission error.
Due to the opportunistic acquisition strategy of VHR satellites we were unable to obtain temporal stacks of VHR imagery, thus we were consequently unable to utilize pairs of VHR images in a pre- and post-burn validation assessment. Additionally, the VHR samples were not randomly selected (see Section 2.1.3 for further details). The accuracy assessment should therefore be viewed with these limitations in mind.
3.2 CRAB analysis
The total Russian cropland area as defined by the crop mask totals approximately 215 × 104 km2. Using the full range of confidence values, our estimated annual burned area from CRAB within the Russian cropland ranged from approximately 121 × 104 km2 (2007) to approximately 144 × 104 km2 (2010), which equates to roughly 57–67% of the total cropland area. Applying a mapping confidence value of > 60% on the spring output and using the full range of confidence values on the summer output, the annual burned area estimate decreased to between approximately 3.54 × 104 km2 (2003) and 7.92 × 104 km2 (2010), which equates to roughly 1–4% of the total Russian cropland area. The majority of the burned area occurred in spring (defined 1 March–30 June) as compared to summer (defined 1 July–30 September). Table 5 and Table 6 summarize the seasonal burned area estimates for CRAB (full confidence and > 60% confidence values), MCD45A1 and MCD64A1 (see Section 2.3 for intercomparison between the three BA products).
Table 5. Spring cropland burned area (units: × 104 km2 = Mha) comparison between CRAB, MCD45A1 and MCD64A1.
Spring: Total burned area in Russian croplands
CRAB (all conf.)
CRAB (> 60%)
MCD45A1
MCD64A1
2003
126
1.91
6.87
4.78
2004
133
2.70
2.74
0.98
2005
121
1.36
1.41
0.49
2006
128
3.19
4.90
1.45
2007
118
2.70
3.09
1.05
2008
135
2.31
7.79
6.18
2009
139
2.17
5.74
3.00
2010
138
1.86
2.32
0.64
2011
132
2.20
3.27
1.12
2012
131
1.40
3.66
1.32
Table 6. Summer cropland burned area (units: × 104 km2 = Mha) comparison between CRAB, MCD45A1 and MCD64A1.
Summer: Total burned area in Russian croplands
CRAB (all conf.)
CRAB (> 60%)
MCD45A1
MCD64A1
2003
1.63
0.00075
2.04
0.54
2004
2.34
0.0022
2.12
1.08
2005
3.43
0.0014
3.61
1.64
2006
3.38
0.0041
3.40
1.35
2007
3.92
0.0049
4.47
1.89
2008
3.60
0.0020
5.49
2.12
2009
4.86
0.0012
3.21
1.27
2010
6.06
0.0029
4.32
1.21
2011
3.70
0.0025
2.68
0.62
2012
5.63
0.0015
4.01
0.66
Subdividing Russia into administrative units (oblasts) and Federal Districts (okrugs) revealed further information on the spatial and temporal patterns of burning. Analysis on the average fractional burned area per oblast between 2003 and 2012 (Table S2) revealed the general patterns of burned area produced by CRAB correspond with the known winter wheat and spring wheat oblasts. Additionally temporal analysis of the average fractional burned area between 2003 and 2012 revealed information about the relative contribution of spring and summer burning between the various okrugs (Fig. 10, Table S3). CRAB reported larger contributions of spring burning in the eastern okrugs whereas the southern and western okrugs contained a higher contribution of burning in the summer and early fall. This pattern is consistent with the winter and spring wheat agricultural practices in the region.
Fig. 10. Temporal distribution of the average fractional burned area by month between 2003 and 2012 for the Russian cropland region partitioned by Federal Districts (okrugs).
3.3 Intercomparison between CRAB, MCD45A1, and MCD64A1
When comparing the magnitude of mapped burned area for each of the three burned area products within the Russian cropland region it is clear that MCD64A1 and CRAB follow similar patterns of burned area frequency in the spring, while MCD45A1 contains considerable spikes (Fig. 11). Whereas in the summer, MCD45A1 and CRAB show considerable spikes throughout the season with MCD64A1 fairly muted in comparison (Fig. 12). The similarity in mapped burned area between MCD45A1 and CRAB may be related to external factors such as missed clouds or BRDF effects. For illustration purposes one example is shown for each season but the relative patterns are consistent over all years (2003–2012).
Fig. 11. Daily burned area (km2) left axis and MODIS Active Fire counts right axis during spring 2012 for all Russian cropland MODIS tiles. All confidence values > 60% were included for the CRAB product.
Fig. 12. Daily burned area (km2) left axis and MODIS Active Fire counts right axis during summer 2005 for all Russian cropland MODIS tiles. All confidence values were included for the CRAB product.
When comparing the three burned area products' seasonal totals (Fig. 13), MCD45A1 generally has the higher magnitude of burning in the summer; however, there are a few years where CRAB and MCD45A1 have very similar magnitudes. Interestingly CRAB, and to some extent MCD45A1, saw a large spike of burned area in summer 2010 which was the season of extreme drought in Russia where widespread fires in forests and cropland in European Russia were reported.
For further details log on website :
http://www.sciencedirect.com/science/article/pii/S0034425716302796
No comments:
Post a Comment