1. Introduction
Coal fires, which occur on the surface (primarily in coal waste piles) and in underground coal seams and are caused by spontaneous combustion, natural events (lightning, forest fires, and peat fires), and human activities (mining and domestic fires), cause severe environmental effects. These effects include noxious gas emissions (e.g., SO
2, NO, CO, and CH
4), an increased concentration of heavy metals in the soil (e.g., mercury, zinc, copper, lead, iron, and germanium), and land surface effects associated with fissures, cracks, land subsidence, and collapse [
1,
2,
3,
4,
5,
6]. Coal fires can also be responsible for the total or partial loss of the coal resource. According to the literature [
7], the annual CO
2 emission of the Wuda syncline amounts to 90,000 to 360,000 tons.
The methods involved in the detection of coal fires typically incorporate the identification of changes in the land surface temperature, electrical conductivity, magnetic field, heavy metal concentration, and gas emissions, whereas remote sensing-based coal fire research mainly focuses on coal fire-related thermal anomaly detection [
8]. The first reported study on the detection of spontaneous coal seam combustion and used infrared photography in coal waste pile fires [
9]. Through developments in airborne thermal infrared sensors, coal fires and the depth of burning were detected in Pennsylvania [
10]. Airborne thermal infrared data acquired during the daytime and nighttime have been used to identify high-temperature targets against low-temperature backgrounds [
11]. Then, the use of orbital images (Landsat-5 Thematic Mapper (TM) and NOAA-9 AVHRR) in the isolation of high-temperature areas from cold backgrounds during the night was evaluated [
12]. It has been reported that the Landsat TM bands 4, 5, and 7 performed well when estimating areas affected by fires [
11]. The Landsat TM bands 6 and 7 were then used in the same region [
13]. A dual infrared band algorithm based on the Landsat TM data has been used to delineate areas affected by underground fires [
14]. After the launch of the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor in 1999, ASTER data with five thermal infrared (TIR) bands and a 90 m spatial resolution became a highly respected moderate resolution data source for coal fire research. ASTER has been considered the primary data source for coal fire detection in related studies [
15,
16] and has been used in combination with other data sources as reference data for cross-validation in the literature [
8,
17,
18].
Multiple coal fire anomaly detection methods have been applied using LANDSAT-5 TM, LANDSAT-7 ETM+, and ASTER by many authors [
8]. The density slicing method using a temperature threshold was previously applied [
19,
20,
21]. These thresholds are economical and effective for a regional study site, specific sensor, or certain weather condition. Methods based on sub-pixel analysis have also been used in fire detection. As initially proposed [
22] and recently employed [
2], sub-pixel approaches estimate the temperature of each pixel, considering two pure pixels representing areas with different temperatures. This method is effective to estimate the pixel’s sensitivity to coal fires. However, due to the high variability in a coal fire’s distribution (e.g., spots, lines, and regions), it is difficult to locate uniform proportions to match the fire’s area inside of a pixel. In addition, the sub-pixel algorithm relies on high-resolution images or survey data for sub-pixel modeling, which may not be promptly available. In addition, the contextual or moving window method for thermal anomaly detection has been used by multiple authors [
23,
24,
25]. The moving window method is exquisitely designed and applied to a series of spatial filters, which are square windows in dimensions of continuous odd numbers. A statistical threshold was used to tag the potential coal fire-induced anomaly pixels. Then, the pixels identified as fires were counted, and a cut-off percentage was given to separate the anomaly pixels that were then aggregated to clusters. For fine-tuning, a false alarm removal criterion was applied. Therefore, non-fire patches were excluded by their dimension, standard deviation, and contrast to neighboring background pixels. Through this process, large thermal anomalies, flat temperature areas, and small hot spots (corresponding to water bodies, illuminated slopes, and industrial plants, respectively) were removed. This method is non-interactive and depends on a statistical threshold (the mean value plus the standard deviation) to determine potential fire-related anomaly pixels in the window and a cut-off percentage (70%) to segment the potential fire pixels into the final coal fire pixels from a counting number matrix. The window size depends on the rate of correctly detected pixels. The determination of these thresholds and the window size is statistically based and depends on the known coal fires [
26]. Moreover, a comprehensive or multiple field fusion method has been proposed in the literature [
27,
28,
29], which identifies anomaly pixels by combining the related environmental or geological fields (vegetation coverage, pyro-metamorphic rocks, fumarolic minerals, burn pits, trenches, subsidence, and cracks, along with surface thermal anomalies) and the knowledge of many local experts. This
in situ-based approach has a sound physical basis and considers the direct and indirect factors induced by coal fires but can be costly due to its dependency on field measurements. This approach partly depends on “indigenous” knowledge, which is not accessible for non-local researchers and is not easily reproduced, as mentioned in the literature [
27].
In the Wuda Coalfield, the fire areas with consistently burning fire spots have exhaust gas vents/cracks exposed on the surface that are detectable by remote-based methods and would not be masked by the background temperatures of approximately 300 K. On 27 March 2013 we measured an average temperature of 770 K (497 °C) for the burning fire spots in the field survey, which was conducted to obtain the peak temperatures on the surface. In an experiment of a simulated coal fire [
25], it has been reported that the surface radiant temperatures of the coal fire range from 300 °C to 900 °C (573 K–1173 K).
Based on the coal fire’s thermal level and spatial characteristics, this study proposes a self-adaptive gradient-based thresholding method (SAGBT) for coal fire detection in the absence of basic field/geological data. The thermal spatial characteristics of coal fires can be summed up in two aspects: large-scale homogeneity and poor horizontal thermal conductivity. The coal fire risks were induced by the coal properties and environmental conditions [
30,
31]. In Wuda, the relatively large-scale, homogenous, fine yellow quartz sandstone overlays, and the short and sparse vegetation contribute to the similar thermal conductivity across the region. However, the thermal anomalies induced by coal fire areas do not reach far from the burning centers because they are restricted to the collapsed region, faults, and fissures, which results in poor horizontal thermal conductivity and therefore causes a sharp decrease in the temperature domain on the edge. A given coal fire cannot be far from the burning center, which is supported by the research [
32]. In the field, it has been observed that high temperatures do not extend more than 2–3 m from the observed hot cracks; in addition, the temperature gradients are significant, with temperatures varying by more than 500 °C over a distance of less than 20 m [
32]. These characteristics of coal fires suggest that the attenuation of temperature along the coal fire’s boundaries generates considerable amounts of spots with extremely high gradient values, resulting in uneven gradients in the integrated pixels (ASTER’s 90 m pixel). This SAGBT method is remote sensing-based, primarily depends on the spatial distribution of the most direct coal fire-induced factor and energy release, and uses the basic outer-boundary of the coal-bearing strata to simply exclude false alarms.
This research addresses the following four key issues in coal fire detection: a temperature retrieval method for the TIR images; a gradient calculation algorithm based on supersampled images to guarantee adequate image matching; a skeletonization method to reduce the gradient buffers to single pixel lines; and the determination of thresholds and related convergence analysis for the SAGBT method.
For monitoring the coal fire changes in the Wuda Coalfield, several authors have successfully evaluated the development/shift of coal fire zones, change in area of fires/fire risk regions, and background radiance variations. Comparing two Landsat datasets (TM and ETM+) for 1987 and 2002, Kuenzer
et al. used a maximum likelihood based interactive classification that indicated that the coal-covered surfaces nearly doubled in area in the Wuda Region over a 15-year period [
33]. Tetzlaff (2004) compared two ETM scenes (obtained on 25 September 2001 and 28 September 2002) and two BIRD datasets (16 January 2003 and September 2003), and showed that these datasets are capable of detecting major shifts or activity changes in terms of hot coal fire surface anomalies and background radiance variations in the Wuda area [
17]. Li
et al. (2005) also detected coal fires by a statistical method from 2002 to 2004 based on multi-spectral coal fire demarcation [
34]. Yang
et al. (2005) normalized two Landsat-7 ETM+ images (obtained on 12 August 1999 and 22 September 2002) and extracted thermal anomalies with different surface environmental parameters. The comparison proved that the coal fire area in Wuda enlarged greatly in those years [
35]. Chen
et al. (2007) processed two scenes of Landsat-7 TM and ETM+ data for 1992 and 2002 and separated coal fires as thermally anomalous areas lying in or around coal containing regions with ancillary data and other images [
36]. By three-dimensional mapping and comparison analysis, they discovered that the coal fire changes in the Wuda Coalfield from 1997 to 2002 increased from 16,200 m
2 to 38,610 m
2 [
36]. Kuenzer
et al. (2008) introduced a multi-temporal coal fire mapping technique for all major coal bed fires in the Wuda coal field based on field observations and high-resolution satellite data [
27]. They monitored coal fire developments such as shrinkage and change in area for different coal fire zones and proposed a protection of valuable coal resources in the Wuda syncline [
27]. Kuenzer
et al. (2012) applied the same approach for coal fire change detection from 2000 to 2005 and in 2010 and showed that over the past 10 years a trend can be observed showing underground fires moving eastwards [
29]. Jiang
et al. (2010) monitored coal fires for 1989, 2001, and 2005 in the Wuda Coalfield, and analyzed the spatial distribution, rate of change, and extending direction of coal fires. The results indicated an annual fire area increasing at the rate of 61.3 × 10
3 m
2 (2.48% of total area) [
37]. Kuenzer
et al. (2008) used multi-diurnal MODIS data, especially bands 20, 32, and band ratio, and proved that MODIS has a high potential for the detection of coal fire zones and coal fire hotspots, as well as for regular thermal monitoring activities; however, it is difficult to detect coal fire size changes or intensity changes from a comparison of only two MODIS data sets [
8].
This coal fire change detection generally compared 2–3 scenes of TIR images, and lacked co-analysis with the coal productions. In addition, there is no published research on the use of a remote sensing method to perform a change detection analysis for a continuous time period spanning more than 10 years in the Wuda Coalfield. Our change detection adopted the no-interactive coal fire thresholding algorithm, SAGBT, for estimating the decadal change. On the basis of experiments and validation, this algorithm demonstrated convergence and matching of the observed coal fire areas. Since the thresholds are self-adapted based on the thermal spatial distribution of different images obtained in different seasons, this method provides an opportunity to monitor long-term coal fire changes using the TIR images from ASTER sensor. Our research also explores the possibility of estimating the CO2 emissions due to coal fire propagation using change detection results. A temporal animation performed in Google Earth is used to dynamically visualize these changes. As an application and an estimation of efficiency for the SAGBT, we analyzed 10 years of change in coal fires via a time-series analysis.
7. Conclusions and Vision
We proposed a different coal fire detection method, the Self-Adaptive Gradient-Based Thresholding (SAGBT) method, which uses ASTER thermal infrared data to separate coal fire-induced thermal anomalies.
The analyses of images acquired during different seasons and image pairs acquired on the same date showed that the thermal anomalies present seasonal differences, but the day and night temperature distributions are similar. We observed an attenuation of the coal fire’s temperature on the edges of the coal fire’s boundaries, which resulted in extremely high gradient values along the boundaries. Based on this characteristic, we implemented a gradient-based thresholding method with the following three modules: gradient, temperature, and threshold definition. The image analysis incorporated an extended Sobel filter to process the supersampled TIR images and generate a regional temperature gradient representation. In addition, we used mathematical morphology thinning to skeletonize a potential high gradient buffer image and match the high temperature buffers to estimate the threshold. Better results were achieved using the mean threshold derived from the multiple potential high gradient buffers.
Methods based on ASTER images can effectively detect thermal anomalies. We delineated the anomalies using a TES-MMD method without emissivity and meteorological data. The SAGBT method is remote sensing-based; it primarily depends on the most direct coal fire-induced factor and energy release and uses the basic outer boundaries of the coal-bearing stratum to simply exclude false alarms. The proposed gradient-based thresholding method is a spatially based method to retrieve thermal anomalies from the different images obtained during the nighttime and daytime. This method is non-interactive and programmed by the IDL and primarily relies on the images’ thermal distributions. This method used a limited and basic field/geological dataset; it is a relatively simple and economical method to estimate the intensity of regional coal fires. It also presents an opportunity to detect long-term coal fire changes using the ASTER TIR images from the historical inventory. In addition, this method offers an alternative way to detect unknown coal fires in a certain area without sufficient in situ observations and field surveying data.
Fire maps show that fire areas are located along the coal seam outcrops, particularly the coal seam numbers 2, 4, and 9. The change detection time series plot reveals that, during the initial years, the fire areas started to decrease but soon began to increase moderately from 2003 to 2007. Since 2008, the spread of fires has sharply decreased, which was most likely because of government-sponsored extinguishing efforts. However, the spread of fires sharply increased again after 2010. With the comparison of the historical industrial figures, we found that the fluctuation of coal fire areas is similar to the curve of coal production in the Wuda Coalfield. Meanwhile, the coal fire area also shows a positive correlation with coal productivity in China, supported by the coal production over the last decade. We infer that coal fires are related to human activities, particularly mining. Additionally, the CO2 emissions in the study period (2001–2010) were estimated to be 71.16 × 104 tons.
Using the calibrated SAGBT method and the MMD-TES algorithm, we offer a novel approach to delineate coal fires in northern China and other coal fires that exist in the country using the historical inventory of TIR data without ground measurements and meteorological records. This non-interactive SAGBT method also offers the possibility to create multi-temporal change products using a uniform criterion for the temperature anomalies i.e., coal fires. These products will aid investigation of the resources lost, the environmental impacts, and the greenhouse gas emissions from coal seam combustion in the long term.
However, coal fires are complex from a remote sensing perspective and are difficult to accurately detect. This study focused on a subset of the Wuda syncline, and the geological and land cover/use in this area is relative homogeneous. Thus, if we extend the SAGBT method to other large-scale coal fire areas, then this method must be improved in the following ways: partition by the geological setting, geomorphology and land cover/use, so that the SAGBT method can be used in each small regional area; and study more specific false alarm removal criteria to exclude false alarms (such as water bodies, illuminated slopes, and industrial plants) or map a fire risk area to mask coal fire-related anomalies. In particular, emissivity data are very useful in identifying geological boundaries such as strata boundaries and faults, which could help to divide a large area. When improved, emissivity anomalies could be considered as a data source to determine coal fires or eliminate false fires. We observed that removing solar irradiation can improve the accuracy of coal fire detection for daytime TIR images. In this work, due to the reduced vegetation coverage in the Wuda Coalfield and even in Northwestern China, coal fires are less often confused with wild fires. However, separating coal fires from wild fires is essential in order to extend the SAGBT worldwide. A combination of coal fire risk area, coal fire-induced surface characteristics, and very high temperature areas were expected to be modeled as accepting/rejecting criteria for distinguishing coal fires from wild fires. Further research should also focus on using field survey data collected on the ASTER overpass date and the estimation of the ASTER sensor’s degradation effects on the detection of anomalies.