1. Introduction
Satellite Earth observations have been widely used for the study of hazardous phenomena [
1]. Both optical and microwave satellite data have proven to provide essential data about the extents of hazard events, allowing us to infer their spatio-temporal evolution and to support emergency response interventions, as well as to incorporate further prevention plans [
2,
3,
4]. Optical sensors measure passive energy, which usually consists of measuring reflected sunlight, which may be acquired into multispectral (MS) bands. This means that such MS capabilities may be useful to discriminate and characterize different hazardous processes (e.g., [
3,
5,
6,
7,
8]), but at the cost of being limited to daytime and cloud free conditions [
9,
10,
11]. On the other hand, the same processes may be recognized by means of active microwave synthetic aperture radars (SAR) operating on longer wavelengths and acquiring backscattered energy [
12,
13]. Such energy may be filtered according to different polarizations, which may allow features associated with hazard processes to be highlighted [
14]. Furthermore, SAR data may perform independently of sunlight and cloud coverage, which may be crucial for certain events and in certain parts of the globe (e.g., regions with all-year persistent clouds) [
9,
11,
12].
Both optical and SAR data allow multitemporal analysis and change detection procedures to be developed by integrating and comparing images acquired at different times, for which the temporal resolution of the satellite platforms is the most relevant conditioning factor [
15,
16,
17,
18,
19]. This temporal capability of satellite imagery is particularly relevant to analyze the spatial-temporal evolution of hazard processes, as depending on their type they may develop different temporal behaviors [
20,
21].
The access to hazard extent products is often provided by national or international initiatives, or event-response mechanisms [
22]. One of such examples is the Copernicus Emergency Management Service (EMS) [
23], which provides on-demand information for selected emergency situations that arise from natural or man-made disasters anywhere in the world. Despite their recognized efficiency in providing worldwide emergency maps, these services are still dependent on activation protocols, which may delay such processed, or only be initiated for “major” events. This means that, ideally, more customizable and readily available solutions should be useful to local authorities and the public. Such solutions could involve GIS applications to facilitate the combination of multiple Earth observation (EO) datasets, the incorporation of automatic remote sensing (RS) routines and other data processing techniques, whilst also providing expedited and accurate products. Moreover, simplified user interfaces could be more easily accessed by GIS users, without requiring an extended knowledge about the software or RS techniques. Furthermore, if they are distributed as open-source, they also allow experts to acknowledge each processing step to keep developing and adapting beyond the “black-box” tool perspective. Repositories are an extremely valuable resource provided by the GIS community which include scripts, models, and tools for several GIS platforms. In particular, QGIS plugins are a popular way of disseminating free GIS tools, which allows a reduction in the complexity and time necessary to process and produce hazard-related products (e.g., [
24,
25,
26]).
This paper has the objectives of a) presenting the MINDED-FBA tool, an open-source QGIS plugin aimed at determining the extent of flooded and burned areas through unsupervised classifications of MS optical and/or SAR imagery, and b) to demonstrate its application for different case studies. The MINDED-FBA tool comprises the combination, adaptation, and further development of two previous methods by [
9,
27].
The structure of this paper consists of a) presenting the methodological principles of the tool and the main innovations introduced in this paper, b) the presentation of six worldwide flooding and wildfire case studies, and c) the results from the implementation of the MINDED-FBA tool, a discussion of these results, and the main conclusions.
2. Methods
The MINDED-FBA tool which is presented in this paper consists of the adaptation of two previous methods, MINDED [
9] and MINDED-BA [
27], to produce, respectively, flooded and burned area extent maps. These methods have been further developed and integrated as an automatic QGIS plugin tool, with the objective of distributing a new open-source multi-hazard mapping tool, which allows a reduction in complexity and processing times, while facilitating satellite imagery processing by non-RS experts. The theoretical principles of the previously published methods have been preserved, including the multi-index differencing approach, as well as the automatic threshold selection procedure. Nevertheless, besides allowing multispectral data (i.e., Landsat and Sentinel-2 products) to be incorporated, the MINDED-FBA tool also allows the input of SAR data (i.e., Sentinel-1 products). The multispectral and SAR datasets can be used independently or in combination (i.e., data fusion approach) to produce automatic extent maps for flooded or burned areas, with several levels of magnitude of change, as well as to obtain uncertainty maps based on the agreement among the several modeled indices.
The plugin is freely distributed in the QGIS repository (
https://plugins.qgis.org/plugins/minded_fba). The MINDED-FBA tool has been implemented with Python language for the QGIS version 3.22 (the current stable version). The graphical user interface of the plugin is shown in
Figure 1.
Moreover, a detailed workflow of the MINDED-FBA procedures and outputs is represented in the
Appendix A (
Figure A1).
Section 1 of the MINDED-FBA tool interface includes the definition of the study area which will determine both the coordinate reference system and the processing extent of the MINDED-FBA tool procedures. The study area should be previously loaded as a QGIS layer, either vector or raster. For vector data, the MINDED-FBA automatically manages the spatial resolution of all intermediate and final products depending on the type of input data to be selected in
Section 2. When using rasters, the spatial resolution of all intermediate and final products will be limited by the study area layer spatial resolution.
Section 2 “Input data” has been designed to include the folder pathways for the pre-event (
t0) and post-event (
t1) satellite products. The optical multispectral (OMS) module should be selected for using data acquired by passive optical sensors. In order to preserve the simplicity of the procedure and the coherence among the different available products, the tool has been designed to read specific surface reflectance data. These include the Level 2 products from the Collection 2 of the Landsat series, including Landsat 5 TM (LS5), Landsat 7 ETM+ (LS7), Landsat 8 OLI (LS8), and Landsat 9 OLI (LS9) [
28,
29,
30]. Moreover, the multispectral inputs can also include Sentinel-2A (S2A) and Sentinel-2B (S2B) Level 2 products (S2MSI2A) [
31]. For determining the extent of flooded areas, the MINDED-FBA tool performs automatic processing of the same indices described by [
27]: NDVI, NDWI, and MNDWI. For determining the extent of burned areas, the tool processes the indices considered by [
27]: NDVI, NBRs, NBRl, and NBR2. After generating single-index maps for both
t0 and
t1, the MINDED-FBA tool proceeds with the single-index differencing, considering the same temporal order of [
9] and [
27] (i.e.,
t1–t0 for flood-related data and
t0–t1 for burned-related data). The thematic classification of changes is determined by applying the automatic thresholding technique described by [
9] and [
27], which is obtained by analyzing the index-differencing distribution function (
f)
, and more specifically from its first-order (
d1f) and second-order (
d2f) derivative functions. The technique consists of identifying the modal value of
f in order to focus on the index-differencing values towards one of the tails of the distribution. The tail to be analyzed is simultaneously dependent on the index type, the temporal arrangement of the index differencing, and the type of change to be analyzed. Afterwards, the thresholds may be selected from either a change of sign of
d1f, or as local maximums in
d2f [
9]. To maintain the coherence of MINDED-FBA with the previous work, we selected the two threshold values closest to the modal value (independently of them being extracted from either
d1f or
d2f). This means that the first threshold value (
T1) (i.e., closer to the modal value) should correspond to the transition from the no change (Nc) condition to low-magnitude changes (LMc), while the second threshold value (
T2) should correspond to the transition from the condition of LMc to high-magnitude changes (HMc). The thresholds are then used to perform density slicing to produce single-index change detection maps. Finally, MINDED-FBA performs a pixel-by-pixel majority analysis which combines all coeval single-index maps into one overall change detection map related to the OMS module.
Section 2 also includes a SAR module, which is one of the main innovations in comparison to the previous versions of both the MINDED and MINDED-BA [
9,
27]. This module allows the user to select data acquired by both Sentinel-1A and -1B sensors. The tool has been developed for incorporating Sentinel-1 products acquired in Interferometric Wide (IW) swath mode and Ground Range Detected (GRD), which consist of focused SAR data which have been detected, multi-looked, and projected to ground range using an Earth ellipsoid model [
32]. GRD products are available as single or dual polarizations, with reduced speckle and approximately square resolution (at the cost of reduced spatial resolution) and pixel values represent amplitude data (phase information is discarded) [
32,
33]. The application of the MINDED principles to the Sentinel-1 datasets considered several combinations of polarizations (
Table 1), which may be used in a similar way to multispectral indices [
9,
27]. However, considering the previously identified difficulties in processing and finding thresholds for non-normalized indices [
9,
27], we narrowed our research to normalized SAR combinations. Considering that the Normalized Ratio Procedure between Bands (NRPB) [
10] and the Normalized Difference Polarization Index (NDPI) [
34] are essentially additive inverses, we have narrowed the analysis to the following three SAR indices: the Normalized Ratio Procedure between Bands (NRPB) [
10,
35], and the Normalized Differenced Temporal Index (NDTI) for both vertical-vertical (VV) and vertical-horizontal (VH) polarizations. However, since both NDTI combinations represent bi-temporal band differencing, the single-index calculations for
t0 and
t1 are not applicable. For incorporating the above indices within the MINDED-FBA tool, we have verified if these combinations meet the same criteria introduced within the MINDED and MINDED-BA methods: (1) a modal value close to zero corresponding to the No change (Nc) condition; (2) different types of change may be located towards different sides of the frequency distribution of either the index (NDTI) or the index difference (NRPB); (3) higher magnitudes of change (further from the modal value) describing more noticeable changes; (4) change-related thresholds which may be extracted from the analysis of the frequency distribution statistics (through either the first- or second-order derivatives). Depending on the verification of these criteria and their overall performance, we may incorporate one or more SAR indices in the tool, which may be further integrated within the same majority analysis approach defined by the MINDED and MINDED-BA methods.
The OMS and SAR modules may be used separately or combined, with the latter option enabling the automatic fusion of both datasets. The MINDED-FBA tool allows the performance of both spatio-spectral (by combining both multispectral and SAR datasets) and spatio-temporal fusion (depending on the corresponding module image acquisition periods) [
15,
36]. Such fusion is implemented through the pixel-by-pixel majority analysis, which is applied, for every pixel, to the available indices results independently of the module. This means that whenever one of the modules does not provide outputs for a given pixel, as an effect of images extent gaps or masking procedures (see
Section 3 hereunder), only the indices of the other module are used.
Section 3 corresponds to the preprocessing section, including several options for masking features which otherwise could be detected by the model as false-positives. These include the masking of permanent waterbodies, which may significantly improve the performance of the tool for the accurate detection of both flooded and burned areas, when using either the OMS or SAR modules. The masks of permanent water bodies may be based on available thematic data, which may be provided by the user as vector or raster files. Such files should be prepared in such way that those areas to be analyzed by the MINDED-FBA tool are assigned a value of 1, while those areas to be masked (i.e., corresponding to the permanent water bodies) are classified as NoData. The remaining preprocessing options included within the MINDED-FBA tool are exclusive to the OMS module. The “Cloud + Cloud shadows masking” option is available for all the previously mentioned multispectral level 2 data (including Landsat and Sentinel-2 sensors), and it is performed automatically using the respective quality bands provided with these products. The “Highly Reflective Surfaces (HRS) masking” is also available and it is particularly relevant for the detection of burned areas. It translates the procedure described by [
27] and allows for the determination and implementation of sensor-specific masks, according to an
a priori procedure of reflectance thresholding, which has been translated into a sliding bar (from high to low, with high corresponding to larger masking areas). The final preprocessing option corresponds to the “Topographic correction”, which is applied to both t0 and t1 scenes. This option performs simultaneous masking of topographic shadows and surface reflection illumination corrections (using the Cosine method) [
37,
38]. The procedure requires the input of a raster with elevation data, which may correspond to any digital elevation model (e.g., ALOS, SRTM).
Section 4 starts with the selection of the sampling size of the numbers of bins (equally spaced in a base 10 logarithmic scale) to be used for the statistical analysis of the frequency distribution of the index differencing. This procedure, fully described in [
27], accounts for the effects of data binning on the signal-to-noise ratio characterizing the frequency distribution. This in turn influences the quality of the change-related thresholds. The default value of the sampling size (i.e., 15), as considered in previous analyses [
27], may be adjusted by the user, with the software requiring longer processing times for larger values. The final section also allows the user to select the type of events to be determined (i.e., either “Flood areas estimation” or “Burned areas estimation”), which in principle should be mutually exclusive within short time periods. Lastly, the user is required to select the output folder, which will allocate all intermediate and final outputs, including statistics and raster files. The output folder will contain every coeval single-index differencing map, single-index statistics, the coeval single-index thematic maps, the overall change detection map, and the uncertainty maps. The uncertainty maps are also based on the majority analysis among all indices. The considered criteria is based on [
27], with the uncertainty values ranging from lowest to highest uncertainty from 0 to 3: 0—unanimity (identical coeval classifications), 1—absolute majority, 2—relative majority, 3—no majority (tie among different coeval classifications). This analysis is performed pixel-by-pixel, taking into account the available results from either the OMS module, the SAR module, or both.
4. Implementation of the MINDED-FBA Tool
4.1. Input Images Selection
For implementing the MINDED-FBA tool within the chosen study sites, we searched for available scenes to be processed with both the OMS and SAR modules. For this task, we used the USGS EarthExplorer portal [
52] to find Landsat Collection 2 Level 2 products, and the Copernicus Open Access Hub [
53] to search both Sentinel-2 Level 2 products (S2MSI2A) and Sentinel-1 products (GRD). For the selection of scenes, we privileged those with acquisition dates as close as possible to those considered by the ROE datasets. For the OMS module, we focused on obtaining cloud-free images. For the SAR module, we tried to find pre- and post-event images acquired with coherent acquisition modes, i.e., ascending or descending. Moreover, when analyzing mountainous areas with SAR, we tried to match the acquisition mode according to the event areas predominant slope facing orientation (i.e., ascending for south-to-west aspects, and descending for north-to-east aspects).
Table 2 summarizes the selected satellite images used for implementing the MINDED-FBA tool, including both multispectral and SAR images, alongside the ROE dataset post-event date.
For the Mondego floods, we considered a couple of Sentinel-2 scenes to be used with the OMS module (S2A and S2B), and another couple of Sentinel-1 scenes (S1A and S1B) for the SAR module (
Table 2). The post-event S2B scene considered in the OMS module was acquired 3 days after the post-event S1A scene of the SAR module (which was also considered by the EMS for determining the ROE dataset).
Regarding the Richmond floods, given the persistent cloud cover in the region during the period around the event, it was not possible to acquire cloud-free multispectral images. For this reason, the MINDED-FBA tool was implemented with the SAR module alone, using a couple of S1A scenes acquired in ascending mode. In particular, the post-event scene was acquired 7 days after the scene used for determining the ROE dataset.
For the Leiria National Forest fire, we considered a couple of Landsat 8 scenes for the OMS module, with a post-event image being acquired 20 days after the ROE dataset. For the SAR module, we used a couple of S1A images acquired in descending mode, acquired 3 days after the ROE dataset.
For the Monte Pisano fire, we considered the same pre-event and post-event Sentinel-2 scenes used to obtain the ROE. The multispectral data processing was performed in parallel with a couple of Sentinel-1 scenes. Considering that most of the fire took place in the south-facing slopes of this mountainous system, both selected SAR images were obtained in ascending mode.
Given the almost four-month duration of the North Complex fire, for implementing the MINDED-FBA tool we decided to analyze the entire fire season of 2020. To this aim, we considered a couple of LS8 scenes, with the pre-event image being acquired in the previous year (26 October 2019) and the post-event image acquired three days before the fire was officially declared 100% contained (29 November 2020). As for the SAR module, we considered a couple of S1B scenes, acquired in ascending mode, with the pre-event image being acquired in 6 December 2019 and the post-event image in 12 December 2020 (*).
4.2. Preprocessing Approaches and Supporting Data
In order to ensure the best accuracy of results, for this paper we considered the implementation of the MINDED-FBA tool with every available preprocessing option, as described in Chapter 2.
The option for masking permanent water bodies was selected for every case study. For those with the ROE datasets corresponding to Copernicus EMS products, we considered the polygonal hydrography features provided with the vectorial data package. As for the North Complex fire, the water mask was created with the US Census Bureau’s 2016 MAF/TIGER, together with the database NHS major rivers, creeks, lakes, and reservoirs (considering a buffer of 30m around line features).
The remaining preprocessing options integrated in the MINDED-FBA tool are exclusive to the OMS module and were selected whenever that module was used (i.e., every case study, except for the Richmond flood). The “topographic correction“ was implemented using the corresponding tiles from the ALOS World 3D—30m (AW3D30) database (available at [
54]). For the two largest case study areas (i.e., the North Complex fire and the Sindh Province floods) we had to perform patching of several tiles of AW3D30 to include the entire extent. The “Cloud and Cloud Shadow masking” and “Highly Reflective Surfaces (HRS) masking” were also selected when using the OMS module. For the latter, we considered the highest strength of the HRS masking options (i.e., less conservative approach).
As for the SAR module, the current version of the MINDED-FBA tool does not include any further integrated preprocessing options besides the permanent waterbodies masking. Despite it being possible to directly use the Sentinel-1 GRD products, additional preprocessing procedures are strongly recommended. For every case study, we used ESA’s open-source Sentinel Application Platform (SNAP) software (available at [
55]) to perform a series of additional preprocessing steps on the data to be used as inputs in the SAR module. The procedures implemented in SNAP were based on [
9,
33] and included: i) the Orbit File Operator (with Sentinel Precise—Auto Download Orbit State Vectors); ii) the Multilook Operator; iii) Single Product Speckle Filter (Lee Sigma 7x7); and iv) Range-Doppler Terrain Correction (SRTM 3arc-sec Auto Download and Bilinear Interpolation).
4.3. Results
After gathering and preparing all the data required by the MINDED-FBA tool, as well as selecting the preprocessing options, as described in Chapter 4, we ran the tool for each case study. The following sections summarize the main results for each study area, with particular emphasis on the SAR module, which is in one of the main innovations introduced in this paper.
4.3.1. Index Calculation and Differencing
The results of the three SAR indices are illustrated in
Figure 3,
Figure 4,
Figure 5,
Figure 6,
Figure 7 and
Figure 8, in which they are compared with the ROE delimitations of every case study. In
Figure 3,
Figure 4 and
Figure 5, we can verify that flood-related changes are detected by all SAR index differences as negative values, with dNRPB having an apparent reduced signal in comparison to both NDTI polarizations, which have similar results for either the VV or VH polarizations. As for the wildfire case studies, in
Figure 6,
Figure 7 and
Figure 8, we verify that burned-related changes also seem to be detected as negative values for every index. This means that considering the before-mentioned reversed order between pre- and post-event images, in practice they produce inverse responses along the index differencing axis. Nevertheless, in comparison to the flood-related results (
Figure 3,
Figure 4 and
Figure 5), the signal of burned-related changes is less evident, particularly for dNRPB, for which change-related areas are barely visible. Furthermore, for burned areas, the NDTI results seem to be more distinct between both polarizations, with burned-related changes being more evident for VV bands.
4.3.2. Binning and Thresholding
Table 3 and
Table 4 list the optimal bin number values, respectively, for flood and burned case study events, which were determined according to the procedure described by [
27]. With the exception of the Mondego flood case study, we could not observe any significant differences in terms of the optimal bin numbers between the OMS and SAR indices.
The following step consists of determining the thresholds from either the
d1f or
d2f. For the OMS indices, the tool proceeds according to [
9] (for floodings) and [
27] (for burned areas). Regarding the SAR module, since we considered an inverse order of images in the temporal differences of floods and burned areas, for dNRPB, NDPI VV, and NDPI VH the threshold selection analysis only accounts for values smaller than the modal value of frequency distribution function (i.e., focusing only on the left tail of the
d1f and
d2f distributions).
Table 5 and
Table 6 include all thresholds T1 and T2 (if applicable) found for every index and case study.
4.3.3. Single- and Multi-Index Thematic Classifications
After selecting the thresholds, the MINDED-FBA tool performs automatic density slicing for each index, which allows coeval thematic change maps to be obtained for every index. The automatic procedures continue with the application of the majority analysis. Again, for the OMS module, the procedures are implemented according to [
5,
22]. For the SAR module, considering the improved performances of both NDTI polarizations in comparison to dNRPB (as described in
Section 4.3.1), the majority analysis only accounted for NDTI VV and NDTI VH. Nevertheless, we decided to maintain the dNRPB single-index calculations within the tool in order to provide the user with the corresponding outputs. Whenever there was a tie condition between both NDTI indices, they were classified as “Mixed” (with the exception of the LMc + HMc condition, for which they were classified as LMc).
Moreover, whenever both OMS and SAR modules were selected, we also performed a pixel-by-pixel fusion of both datasets. This means that whenever only one of the modules could be determined, the fusion results only include the available module results. In the case of pixels containing results from both modules, we performed the majority analysis among all Multispectral and SAR indices. Hence, the results of the OMS module may include three or five coeval single-index classifications (depending on whether the analysis is performed for either flooded or burned areas), while the SAR module will only contribute two coeval single index classifications. Again, the “Mixed” class is used whenever there is no majority among the available single-index classifications.
Following this, the single- and multi-index thematic classifications were compared to the available dataset of ROE delineation for each case study, using the confusion matrix approach [
56] and the Matthews Correlation Coefficient (MCC) [
57,
58]. The latter is regarded as a more reliable statistical method for unbalanced binary distributions, in which higher scores (i.e., worst MCC value = −1; best MCC value = +1) are only generated when the predictor is capable of correctly classifying both the majority of positive cases and the majority of negative cases [
58]. “Mixed” pixel classifications have been excluded from these comparisons, just like null pixels (e.g., whenever the permanent water body mask is applied).
Table 7 summarizes the results for the Mondego floods event. For the SAR module, we can verify that, in comparison to the ROE dataset, both NDTI polarizations seem to have better overall performances in comparison to the dNRPB, particularly in terms of total change commission errors (Tc CE), observed event omission errors (OE OE), and MCC. These performances seem to be greatly improved by the SAR indices majority analysis. In comparison, the OMS majority analysis results (i.e., NDVI + NDWI + MNDWI) are slightly worse for most parameters. The spatial extent of the OMS data is also smaller, as consequence of the presence of clouds and cloud shadows in both pre- and post-event images. Finally, the fusion results (i.e., majority analysis among both OMS and SAR classifications), seem to achieve the best overall results, including the best OA and MCC. Among all the indices, we verify that most have a tendency to overestimate changes, as the area of the total changes (Tc) is predominantly higher than the delimitation area of the ROE (with the exception of dNRPB, which seems to be the worst performer among all indices).
In
Figure 9, we can observe the change maps generated by the MINDED-FBA tool for the Mondego flood event in comparison to the ROE delineation. When comparing the SAR and OMS maps, we can observe several areas near the Mondego River mouth which have been detected as changes (either LMc or HMc) by the OMS indices only. The upstream areas (to the extreme eastern side) could only be analyzed with the SAR module and were dominated by “Mixed” pixels (which correspond to Nc in ROE). In the fusion map, the classification around the river mouth seems to be similar to the OMS dataset (i.e., mostly as HMc), while the central part of the map includes a few more “Mixed” pixels than either the OMS or SAR modules. Finally, the extreme eastern side consists of the same results from the SAR module, which is the only module contributing to the fusion results.
Regarding the Sindh Province floods, in
Table 8 we can verify that, for the SAR module, the NDTI VV achieved the best individual correlation with the ROE dataset. Nevertheless, the combination of both NDTI polarizations resulted in the best overall correlation with the delineation dataset (including the best OA and MCC). The combined results from the OMS indices show slightly worse correlations with the ROE. Finally, the fusion results have the second-best overall performance for most of the parameters in
Table 8, yet also include the best performance for both HMc accuracy and OE OE.
Figure 10 includes the Sindh Province floods ROE, alongside with the thematic change maps for each module of the MINDED-FBA tool, for the three areas of interest considered by the Copernicus EMS. Among the three MINDED-FBA tool maps, we can verify an overall tendency to underestimate changes in comparison to the ROE. This effect seems to be less evident for the SAR results, although this module displays more “Mixed” pixels than all the other maps. The underestimating of the OMS results seems to be more evident, particularly for the two northern areas (i.e., Jacobabad and Shikarpur). Nevertheless, regardless of the module, all seem to produce false detections of flood-related changes in the northwestern side of the Jacobabad area (i.e., the northern area of interest). In comparison to the remaining datasets, the fusion (SAR + MS) seems to resolve most “Mixed” conditions.
Regarding the Richmond floods, for which we have selected the SAR module only,
Table 9 summarizes the comparison of results with the ROE delineation. Similar to the previous case study, the dNRPB seems to be the worst performer, while the majority analysis with the combination of both NTDI polarizations seems to improve every parameter of the confusion matrix analysis and MCC.
Figure 11 includes the combined NDTI map of the Richmond floods in comparison to the ROE, where we can verify a slight tendency to overestimate changes by the MINDED-FBA tool, particularly to the north side of the study area (mostly as HMc).
Regarding the Monte Pisano fire, in
Table 10 we can observe that the combination of the OMS indices seems to achieve the best performances in comparison to the ROE, being followed by the fusion dataset (which has the second-best statistics for most parameters). In the SAR module results, we verify high rates of omission errors for all indices, and lower commission errors, particularly for the combined NDTI classification.
Figure 12 includes the Monte Pisano change maps alongside the ROE delineation. Here, we can verify the underestimation of burned-related changes in the combined NDTI map when compared to the ROE. Instead, the OMS and fusion maps seem to achieve a much better correlation with the reference dataset.
Regarding the Leiria National Forest fire, in
Table 11 we can once more verify the improved performance of the combined OMS results in comparison to the SAR indices. Once more, the SAR indices have a tendency to underestimate burned-related changes which contribute to their overall worse performance in respect to the ROE. The fusion results have the second-best overall results, including OA and MCC.
In
Figure 13, we confirm the findings of
Table 11 and verify the better correlation of the OMS and fusion maps with the ROE dataset. Furthermore, when compared to the Monte Pisano Fire findings, we have detected a higher proportion of HMc.
Table 12 includes the summarization of the North Complex fire event. The combination of OMS indices resulted in the best correlation in respect to the ROE, followed by the fusion dataset. In comparison to the other fire event case studies, the Tc CEs are higher, suggesting the tendency to overestimate changes in comparison to the reference data. As for the SAR module results, we verify significantly worse correlations with the ROE (including negative MCC), this time with the NDTI polarizations displaying even worse performances than the dNRPB.
In
Figure 14, we can verify the low correlation of the SAR indices in comparison to the reference burned area dataset, with most of the detected changes being located in a range extending from northwest to southeast, away from the ROE data. As for the OMS and fusion maps, the correlation with the ROE is significantly improved, particularly for the North Complex fire area (which is located in the central part of the showcased area). However, the OMS maps also detect a few changes in the same regions resulting from the SAR.
Table 13 summarizes the fusion results for all the case studies. When comparing both types of events, we cannot observe relevant differences between flood and burned events. The Monte Pisano seems to have achieved the best overall statistics (in terms of errors and MCC), while the North Complex case study was generally the worst (except for the LMc and HMc class accuracies). Nevertheless, all events achieved high overall accuracies (i.e., above ca. 90%), and high MCC (over 0.6), indicating good correlations with their respective ROE datasets.
4.3.4. Uncertainty Analysis
Figure 15 represents the uncertainty maps for the fusion datasets for every case study of this paper (apart the
Figure 15c, related to the Richmond flood, which was determined with the SAR module only). These maps show that the higher values of uncertainty (i.e., 2 and 3) tend to occur around and inside the regions of change. In those locations where only the SAR module is available (i.e.,
Figure 15c and parts of
Figure 15a) the scale consists of only two values, i.e., 0 where both NDTI polarizations have identical coeval classifications, and 3 for no majority conditions (different coeval classifications).
5. Discussion
This paper presents the MINDED-FBA tool and demonstrates it as an efficient and effective instrument aimed at extracting both flooded and burned areas by means of both unsupervised classification of RS data and further automatic processing. The tool has been integrated in a freely distributed QGIS plugin, which facilitates running the entire methodological workflow (
Figure 1 and
Figure 1A in
Appendix A). Hence, it facilitates the application of digital change detection procedures even by users with less experience in respect to the tool’s methodological principles, or, more generally, by non-experts in GIS or in RS techniques. The paper highlights the main methodological principles, which are based on the work by [
9,
27], and introduces some significant innovations, such as the incorporation of change detection based on SAR data. Such developments represent a great advancement in comparison to the original work, since they broaden the usable input satellite imagery. Namely, the use of SAR data opens the possibility of studying events even in cases of persistent cloud/smoke coverage, which are particularly relevant when studying flooding and wildfire phenomena.
Despite including a SAR module, the MINDED-FBA tool does not yet incorporate full preprocessing of Sentinel-1 images, which is strongly recommended. Such procedures have been performed with the ESA’s SNAP tool which, despite being also freely available, requires an additional step prior to the MINDED-FBA tool process chain. Nevertheless, the MINDED-FBA tool allows the incorporation of layers for masking permanent water bodies, which would likely result in the detection of false positives by both the SAR and OMS modules.
For the OMS module, the MINDED-FBA tool inherits the same preprocessing procedures presented by [
27], including the “Cloud + cloud shadow masking”, “Topographic correction and topographic shadow masking”, and “Highly Reflectance Surface (HRS) masking”. We verified that the “Topographic correction and topographic shadow masking”, which is particularly relevant for more mountainous areas [
27], required the longest processing times among all the preprocessing options integrated within the MINDED-FBA. The HRS masking procedure is mostly relevant for the accurate detection of burned areas. It consists of applying different levels of masking, which have been derived from the work of [
59], which are then implemented through tasseled cap brightness calculations applied to specific multispectral bands. The reference values used for the HRS masking were previously determined for a study area in Portugal [
59]. The implementation of the MINDED-FBA within further wildfire study areas allowed us to verify that the HRS masking is effective without compromising the detection of burned areas.
Regarding the accuracy of results from the SAR module, we verified that in comparison to the OMS indices, the SAR-derived indices achieved comparable, if not better, overall results, particularly for the flood-related case studies. We found that the theoretical principles behind the MINDED and MINDED-BA methods could be directly applied to analyzing SAR-derived indices. One noticeable result concerns the optimal bin number, which assumes similar values for both the modules, with the only exception being the Mondego flood case study, which shows higher values for the SAR indices. Such response is likely a consequence of the multitemporal image difference being characterized by a strong signal and/or being affected by low statistical noise, hence producing maximum bin ratio values for higher bin number values (see Equation (12) in [
27]).
Considering the disaggregation of the Sentinel-1 as polarized data, the availability of indices for the SAR module is reduced when compared to the full array of indices which may be determined from OMS bands. Furthermore, besides the results shown in
Section 4, we tested the implementation of the SAR module with Sentinel-1 scenes acquired in descending mode for the Monte Pisano and North Complex fires. In those cases, we verified significantly reduced sensitivity for detecting each corresponding burned area in respect to the ascending mode, which should be related with the steep morphology and the predominantly south-facing slopes of the burned areas.
The overall performance of the SAR module, and particularly for both NDTI polarizations, seems to be slightly reduced for the wildfire case studies, resulting in worse correlations when compared to the OMS module. This effect may be related to the lower sensitivity of this index to detect burned-related changes. Indeed, for certain burned-related conditions, the NDTI polarizations may also provide inverted index differencing signals (i.e., positive values instead of negatives). This effect is noticeable in the Leiria National Forest case study, particularly in a scarcely vegetated location which underwent previous burning (14 years before) [
60]. As for the North Complex case study, we also verify the occurrence of presumably burned pixels with inverted signals (e.g., northeast of the Lake Oroville). In this case, the positive SAR index-differencing values also occur in areas which either underwent a previous fire (in 2008; [
51]) or appeared to be clearcutting patches (by image visual interpretation) older than the North Complex fire.
In summary, the results from the considered case studies suggest that the OMS and SAR modules could have different relevancies depending on the type of change. While the OMS module seems to achieve the best performances for determining burned areas, the SAR module seems to be particularly suited to studying flood events. Moreover, the SAR module also enhances the possibilities of selecting recent usable post-event images, an option which is particularly relevant for studying flood events. Indeed, in such cases the spatial and temporal effects caused by water are more dynamic when compared to the short-term effects of fire. Furthermore, floods are often associated with heavy precipitation events, which may correspond to persistent cloud coverage periods, while SAR images may represent the only usable remote sensing data. This was the case of the Richmond floods, for which no cloud-free OMS images were available even several months after the event.
When analyzing floods, the thematic fusion dataset is obtained from a majority analysis with three OMS indices and two SAR indices, while for fires it is based on four OMS indices and two SAR indices. In practice, this means that in the current implementation of the MINDED-FBA, the SAR module has a lower weight in comparison to the OMS module, especially for the case of fires. Nevertheless, the analysis of the five case studies which included the fusion of OMS and SAR modules demonstrates that such combination provided consistently good performances in comparison to the respective ROE delineations. In particular, the fusion dataset provides the second-best OA accuracy and MCC in four of the case studies, as well as the best statistics in the case of the Mondego floods. Moreover, we underline that the tool is open-source, allowing further developments, such as incorporating new indices for either of the modules.
When analyzing the summarized results of
Table 13, we can verify that the best overall accuracies were obtained for the Mondego and Leiria National Forest case studies, while the best MCC was obtained for the Monte Pisano fire. This implies that, for such parameters, there seems to be a correlation with the study area feature processing extent (SAFPE), as the best performances were obtained for the smaller case study areas under analysis, while the worst results seem to correspond to the two largest areas (i.e., the North Complex fire and the Sindh Province floods). These results are to be expected, because for larger areas, the quality of the thresholds extracted from the image differencing statistics could be affected by larger extents of false positives/negatives related to other kinds of changes than either flooding or fire.
The uncertainty results are another important benefit obtained from the majority analysis. It allows us to identify the pixels that result in higher or lower agreement among the considered indices, which we interpret as an indicator of the uncertainty level of classification. As verified in
Figure 15, the higher uncertainty values were found both within and around the larger areas of change. Such an effect is to be expected, and should correspond to different land cover/conditions, both for floods and fires. In the case of floods, certain flooding water conditions (e.g., shallower, turbid, or eutrophicated water), as well as heavily water-saturated soils, or superficially wet natural or artificial surfaces, may be detected as flood-related changes by only a few indices. In the same way, lightly burned areas (e.g., within scarcely vegetated areas or partial burning), post-fire vegetation regrowth, or even certain non-burning conditions (e.g., drought or clearcutting), may also be detected as burned-related changes by only a few indices. We consider that such an approach of uncertainty estimation provides a more reliable representation of both flooding and fire changes, beyond the typical binary maps based on either visual delineation or single-index classifications from either OMS or SAR images.
6. Conclusions
This paper presents the new MINDED-FBA tool, a freely available QGIS plugin for determining flooded and burned areas from satellite remote sensing data. The tool starts by applying an index differencing approach to either optical multispectral and/or SAR imagery. Then, it performs multiple unsupervised classifications based on several automatic statistical procedures. Finally, it implements a majority analysis which allows the user to obtain multi-index thematic change maps, as well as an estimate of their uncertainty.
The MINDED-FBA tool has been applied to six case studies with diverse conditions, geographical locations, extent, and morphology. The outputs have been compared to reference flooded or burned area delineation datasets, mostly obtained from the Copernicus EMS. The results demonstrate the capability of the tool for achieving consistent correlations with such reference products. For this reason, we propose that the MINDED-FBA tool could be used as an unsupervised classification-based near-real-time solution for extracting flooding and burned areas, which could offer essential data to provide better informed emergency response measures.
Despite including here several case studies which encompass a large array of conditions, there is still room for further tests and developments. Hence, future research could include further geographical regions, as well as analyses of different kinds of changes, and the incorporation of further preprocessing steps (particularly for the SAR module) and remote sensing sensors.