Spatio-Temporal Fusion of Landsat and MODIS Data for Monitoring of High-Intensity Fire Traces in Karst Landscapes: A Case Study in China

Abstract

The surface fragmentation of karst landscapes leads to a high degree of coupling between fire scar site boundaries and topographic relief. However, the applicability of spatio-temporal data fusion methods for fire scar extraction in such geomorphological areas remains systematically unevaluated. This study developed a spatial–temporal adaptive fusion model integrating Landsat 30-m data with MODIS daily observations to generate continuous high-precision dNBR datasets. Using a typical karst fire region in Guizhou and Yunnan, China, as a case study, we validated the method’s effectiveness for fire trace extraction in fragmented landscapes. The proposed fusion technique addresses cloud cover limitations in humid climates by constructing continuous NBR time series, enabling precise fire boundary delineation and severity quantification. We comparatively implemented multiple fusion approaches (FSDAF, STARFM, and STDFA) and evaluated their performance through both spectral (RMSE, AD, and PSNR) and spatial (Edge, LBP, and SSIM) metrics. Key findings include the following: (1) FSDAF outperformed other methods in spectral consistency and spatial adaptation, particularly for heterogeneous mountainous terrain with fragmented vegetation. (2) Comparative experiments demonstrated that pre-calculating vegetation indices before temporal fusion (Strategy I) produced superior results to post-fusion calculation (Strategy II). Moreover, in our karst landscape study area, our proposed Hybrid Strategy selection framework can dynamically optimize the fusion process of multi-source satellite data, which is significantly better than a single fusion strategy. (3) The dNBR-based extraction achieved 90.00% producer accuracy, 69.23% user accuracy, and a Kappa coefficient of 0.718 when validated against field data. This study advances fire monitoring in karst regions by significantly improving both the spatio-temporal resolution and accuracy of burn scar detection compared to conventional approaches. The framework provides a viable solution for fire impact assessment in topographically complex landscapes under cloudy conditions.

1. Introduction

As a special karst ecosystem is widely distributed around the globe, karst landscapes are known for their unique dissolving topography, fragile ecological structure, and complex hydrological processes; but the high degree of fragmentation of their surface morphology, heterogeneity of their vegetation cover, and frequent cloudy and rainy climates make such regions particularly challenging in global ecological and environmental monitoring. Under the double pressure of climate change and human activities, forest fires are frequent in karst regions, and the rate of post-disaster vegetation succession is significantly faster compared to other ecosystems. This feature makes the accurate monitoring of fire trails face a triple challenge as follows: (1) the topographic fragmentation leads to a high proportion of mixed pixels in traditional remote sensing images, which makes it difficult to capture the details of fire boundaries; (2) the cloudy and rainy climate leads to limited access to optical remote sensing data, and insufficient spatial–temporal continuity; and (3) the rapid succession of vegetation cover may conceal fire trails within months, which compresses the window period for effective monitoring. Therefore, ways for achieving the high-precision and time-efficient extraction of fire traces in a dynamically changing and complex environment have become a core scientific issue for fire emergency management and ecological restoration assessment in karst regions. The spatio-temporal fusion technique can generate spatio-temporally continuous high-resolution datasets by constructing a spatio-temporal covariance model of reflectance and interpolating within cloud coverage gaps [1]. This significantly enhances the identification capability of fire trace boundaries [2], which is of considerable practical value for emergency responses to fires in karst areas characterized by fragmented topography and fragile ecology, as well as for ecological restoration assessments [3].

In typical cases around the world, the preferred sensor strategy for fire response mechanisms in different karst landscape development stages and the interpretation of physical mechanisms for multi-scale feature fusion are still scientific bottlenecks that need to be urgently addressed [4]. Conventional methods for extracting fire traces based on a single optical satellite are susceptible to cloud interference [5] and are not sensitive to changes in the carbon fraction of the burned surface [6,7]. While Landsat data feature a high spatial resolution of 30 m, their 16-day revisit cycle makes it difficult to capture the dynamic process of fire spreading. Conversely, while MODIS provides daily coverage [8,9], its coarse pixels (250–1000 m) are prone to interference from the mixed pixel effect caused by heterogeneous surfaces in karst areas [10]. Synthetic aperture radar (SAR) data have the ability to penetrate but are limited by the geometrical distortions in complex terrain [11]. Multi-source satellite data fusion represents a key technology for improving the accuracy of monitoring surface processes in complex environments [12] by synergistically leveraging the complementary advantages of different sensors and effectively overcoming the temporal and spatial resolution limitations of individual data sources [13]. Based on this, the construction of a Landsat-MODIS synergistic fusion framework [14] can effectively integrate the complementary spatial and temporal advantages of the two datasets, which theoretically improves both the spatial identification accuracy of burnt sites and the temporal resolution of the vegetation restoration process [15].

In emergencies such as fires, fusing thermal infrared data (e.g., VIIRS fire point data) [16] with optical imagery can rapidly locate fire scenes and assess the extent of vegetation damage. After years of development, spatio-temporal fusion models [17] have matured considerably. STARFM (Spatial and Temporal Adaptive Reflectance Fusion Model) [18], a linear hybrid model based on weighted filtering, was the first to achieve the fusion of Landsat (30 m) and MODIS (500 m) data. It introduces spatial similarity weights and controls temporal prediction errors within 8–12%. The ESTARFM (Enhanced STARFM) model [19] enhances the search strategy for similar image elements, improves accuracy in heterogeneous regions, and incorporates linear regression correction, reducing RMSE by 23% in farmland areas. The FSDAF (Flexible Spatiotemporal Data Fusion) model [20] integrates the thin-plate spline function (TPS) with endmember change detection and establishes an image element decomposition model to improve the handling of abrupt surface changes, achieving a fire traces detection error of <15% [21]. The CUFSDAF (Convolutional Unit-based FSDAF) model [22] introduces a 3 × 3 convolutional kernel to extract spatial features for constructing a residual learning framework, improving accuracy by 19% (compared with FSDAF) in urbanized regions. The VSDF (Variational Spatiotemporal Data Fusion) model [23], based on a variational optimization framework, establishes an energy function solved using the ADMM algorithm, improving convergence speed by 40% and achieving a PSNR of 32.6 dB in coverage scenarios. The high-precision [24] extraction of fire traces is critically important for forest ecosystem restoration assessments and carbon cycle research. Most current official fire trace extraction methods rely on spectral index-based threshold segmentation to calculate the differential normalized burn ratio (dNBR), quantifying changes in near-infrared and shortwave infrared bands before and after fires for rapid extraction [25]. However, their accuracy is constrained by terrain heterogeneity and the subjectivity of threshold selection [26]. Future research trends emphasize the collaborative inversion of multimodal remote sensing data, which can provide technological support for large-scale disaster emergency responses [27].

Although multi-source data fusion has shown potential in fire-scorched monument monitoring, existing research still lacks systematic validation and heterogeneous model co-optimization methods for karst landscapes. In this paper, by integrating the FSDAF, STARFM, and STDFA models, a multidimensional fusion evaluation system is constructed for the complex karst terrain, which covers band fidelity, NBR index consistency, and terrain correction robustness, solving the applicability of Landsat-MODIS data in the spatial and temporal extension of the karst zone. Further, we propose the “model optimization-feature fusion-dynamic inversion” technology chain to generate a 30 m/1 day high-resolution continuous dataset and couple dNBR with adaptive threshold segmentation to achieve subpixel-level extraction of the fire boundary. This framework not only breaks the technical bottleneck of high-precision dynamic monitoring in complex terrain areas but also provides a relocatable quantitative solution for the post-disaster recovery assessment of karst ecosystems. In view of the lack of the application of fire burn trace monitoring based on multi-source satellite data fusion in typical complex terrain areas such as karst landscapes, this paper mainly makes the following contributions:

• For the first time, a systematic validation framework for the spatio-temporal fusion technology for karst fire traces has been constructed. By integrating heterogeneous remote sensing fusion algorithms and innovatively establishing a multi-dimensional evaluation standard system, the breakthrough reveals the applicability boundary of spatio-temporal fusion technology in karst landscape areas, laying the methodological foundation for the application of this technology in special landscape scenarios.
• This study pioneeringly proposed a multi-source remote sensing cooperative optimisation model for karst areas. By deeply analyzing the spectral response characteristics of different algorithms and establishing a mechanism for complementing the advantages of multiple models, we have overcome the technical bottleneck of obtaining remote sensing data with high spatial and temporal resolution in complex terrain areas, and we have provided innovative solutions for monitoring fragile karst ecosystems.
• We have constructed a dynamic monitoring technology system for karst ecosystem disaster processes. We have achieved the precise identification of the boundaries of fire traces, and we have established a time-series remote sensing inversion model, revealing for the first time the dynamic evolution law of vegetation restoration in karst areas and providing a universal technical paradigm for disaster assessment and ecological restoration of ecologically fragile areas.

2. Materials and Methods

2.1. Experimental Materials

2.1.1. Study Area

The study area of this paper is located in the Yunnan–Guizhou Plateau at the junction of Bijie City, Guizhou Province and Zhaotong City, Yunnan Province, China. This area is primarily characterized by a karst landscape. It spans latitudes from 26°21′ to 27°46′N and longitudes from 103°36′ to 106°43′E. The altitude ranges from 1000 to 2900 m, with terrain predominantly characterized by karstic mountains [28] featuring longitudinal gullies.

The region has a subtropical monsoon climate [29] and karst landform, with an average annual temperature of approximately 12–15 °C and annual precipitation ranging between 900 and 1200 mm. The dry and wet seasons are distinct, and the alternation of spring droughts and intense summer precipitation makes the area prone to forest fires. The main vegetation types include evergreen broad-leaved forests (approximately 35%); coniferous forests dominated by species from the families Crustaceae and Camphoraceae (approximately 28%); Yunnan pine and Huashan pine plantation forests and shrub (20%); vine thorn shrub endemic to the karst region; and farmland and secondary forests (17%). The average annual NDVI value ranges from 0.6 to 0.8, while the biomass density reaches 120–180 t/ha [30]. The annual mean fire-affected area in the study region over the past decade exceeded 2000 ha, with fire risk covering more than 2000 ha [31]. During the fire season (November–April), precipitation remains below 50 mm, and relative humidity falls below 60%. Fire disturbances have resulted in a 40–60% reduction in soil organic matter content. The recovery period for plant communities is 8–12 years, leading to an estimated loss of carbon sinks of approximately 5.8 t/ha [32].

Due to the unique local karst landform geographic environment, a single remote sensing data source often struggles to meet the monitoring requirements for the accurate identification and dynamic restoration assessment of fire traces. There is an urgent need to construct time-series remote sensing data with high spatial and temporal resolution to effectively monitor these areas. Landsat 8 SWIR2 (Band 7) is highly sensitive to charred materials, and when combined with the reflectance characteristics of vegetation in the NIR (Band 5), it enables the quantification of burn severity. Additionally, the extent of fire traces can be extracted by calculating the Differential Normalized Burn Ratio (dNBR). This capability highlights the significance of this study. To conduct the relevant experiments, we selected a specific area, as shown in Figure 1, as the study area.

2.1.2. Data Sources

In this paper, Landsat 8 Level-2 C2 data and the MODIS surface reflectance daily product MOD09GQ1 data covering the study area are selected as the experimental data. Both Landsat 8 Level-2 data and MODIS data were downloaded from Google Earth Engine (GEE) (https://earthengine.google.com/accessed on 14 March 2025). The Band 5 and Band 7 of Landsat 8 images correspond to the Band 2 and Band 7 of MODIS images, which belong to the near-infrared (NIR) and short-wave infrared (SWIR) bands, respectively. The spatial resolution of the two bands of MODIS data is 250 m and 500 m, and the temporal resolution is 1 day.

Table 1. Landsat and MODIS data product band information used in this paper.

Data Products	Bands	Spatial Resolution/(m)	Time Resolution/(day)
Landsat 8 Level 2-C2	Band5 (NIR)	30	16
	Band7 (SWIR2)	30	16
MODIS MOD09GQ	Band2 (NIR)	250	1
	Band7 (SWIR2)	500	1

shows the specific band correspondence information.

The MODIS imagery on 16 July and 18 September 2023 and the Landsat 8 imagery on 16 July and 18 September 2023 were selected as the two cloud-free data periods for the experiments in this study (Figure 2). The generated data were derived from atmospherically corrected surface albedo and surface temperature. Cropping to obtain the study area data was performed during the data download process on the GEE platform [33]. The MODIS surface reflectance data products, which were atmospherically corrected, needed to be converted to WGS_1984_UTM through projection, resampled to a 30 m resolution to ensure consistency with the Landsat 8 data, and then cropped to the extent of the study area. These steps were also completed while downloading the data from the GEE platform.

2.1.3. Fusion Data Selection

The FSDAF, STARFM, and STDFA models in this study require a high spatial resolution image, a low spatial resolution image, and a high temporal resolution image as input data for modeling. The near-infrared (NIR) band and short-wave infrared (SWIR) band data from Landsat 8 and MODIS and NBR index data from 16 July 2023 are selected to train the FSDAF, STARFM, and STDFA models. All three models are used to predict MODIS data fusion, starting from 18 September 2023, to generate the corresponding high-resolution data. The real Landsat 8 image data from 18 September 2023 are utilized for comparative analyses and accuracy assessment. The experimental data information corresponding to the specific models is shown in

Table 2. Description of the experimental data.

Method	Input Data						Validation Data
	Landsat-8			MODIS			Landsat-8
	Band	Resolution	Shooting Date	Band	Resolution	Shooting Date	Band	Resolution	Shooting Date
STARFM	NIR/SWIR2	30 m	16-July-2023	NIR/SWIR2	500 m	16-July-2023	NIR/SWIR2	30 m	18-September-2023
			-			18-Sepember-2023
FSDAF	NIR/SWIR2	30 m	16-July-2023	NIR/SWIR2	500 m	16-July-2023	NIR/SWIR2	30 m	18-Sepember-2023
			-			18-Sepember-2023
STDFA	NIR/SWIR2	30 m	16-July-2023	NIR/SWIR2	500 m	16-July-2023	NIR/SWIR2	30 m	18-Sepember-2023
			-			18-Sepember-2023

.

2.2. Methods

The experimental process is shown in Figure 3. First, MOD09GA and Landsat 8 images were used to construct an a priori surface reflectance dataset, which was subsequently de-clouded and matched across time series [34]. Three types of image fusion (STARFM, STDFA, and FSDAF) were performed for the near-infrared (NIR) and short-wave infrared bands. To achieve higher fusion accuracy, the reflectance was linearly fitted for spectral consistency. The vegetation index NBR was calculated, and different fusion strategies were adopted for the data fusion of the bands and remote sensing indices. The fusion results were quantitatively evaluated in both spectral and spatial dimensions to select the best-performing model for the study area. For improved fusion accuracy, linear fitting of reflectance (spectral homogenization) was conducted [35]. The vegetation index NBR was computed, and different fusion strategies were applied for the data fusion of the bands and remote sensing indices. The fusion outcomes were quantitatively assessed across both spectral and spatial dimensions, selecting the optimal fusion model and strategy for the study area. Finally, based on the multi-source satellite fusion results, the threshold method was employed to determine the optimal threshold for extracting burnt traces, and the identified threshold was utilized to extract the burned area. Subsequently, heterogeneous official satellite data were introduced for the further extraction of burnt traces. Based on the aforementioned multi-source satellite fusion result data, the optimal threshold value was determined using the threshold method to extract the fire-burned land area, and this threshold was then applied to delineate the fire-burned region [36].

2.2.1. Fusion Method

• FSDAF

FSDAF is a flexible spatio-temporal data fusion method proposed by Guo et al. [20]. Compared to previous data fusion methods, it incorporates Thin-Plate Spline (TPS) interpolation and linear mixed modeling to improve adaptability to nonlinear variations such as crop harvesting and fire [37]. The method first performs unsupervised classification of high spatial resolution images at one time point, then estimates the class changes corresponding to the low spatial resolution image at another time point, and predicts the high spatial resolution image at the second time point based on these changes. The initial prediction is achieved through the following:

$$L_{\mathrm{pred}}=L_{\mathrm{base}}+TPS\left(M_{\mathrm{pred}}-M_{\mathrm{base}}\right)$$

(1)

where $L_{\mathrm{base}}$ is the reflectance value of the high-resolution Landsat 8 image for the base date. $M_{\mathrm{pred}}-M_{\mathrm{base}}$ captures global changes in low-resolution imagery from the base date to the target date, and TPS interpolates low-resolution variations into high-resolution space to compensate for the lack of detail in high-resolution images.

In practice, image elements corresponding to two time points are considered to have different effects on classification, and residuals need to be assigned to them. The model introduces a TPS function to improve the accuracy of residual assignment. After calculating the residuals and weights, Equation (2) is used to correct the residuals of the initial prediction using the weighted amount of low-resolution variations of similar image elements, thereby improving local consistency.

$$L_{\mathrm{final}}=L_{\mathrm{pred}}+\sum _{i=1}^n{\lambda }_i·\left(M_{\mathrm{pred}}^{\left(i\right)}-M_{\mathrm{base}}^{\left(i\right)}\right)$$

(2)

where ${\lambda }_i$ is the mixing weight that represents the contribution of the i-th similar image element to the residual correction, and n represents the number of similar image elements obtained through spectral similarity screening.

2.

STARFM

STARFM is a spatio-temporal adaptive reflectance fusion method proposed by Gao et al. [18] in 2006. The core idea is to utilize the spatial details of high spatial resolution images and the time series of high temporal resolution images to generate high spatial and temporal resolution images through similar pixel search and weighted fusion (Equation (3)).

$$L\left(x_i,y_j,t_k\right)=\sum _{m=1}^n{W}_m·\left[M\left(x_m,y_m,t_k\right)+\Delta L_m\left(t_0,t_k\right)\right]$$

(3)

where $\Delta L_m\left(t_0,t_k\right)=L\left(x_m,y_m,t_0\right)-M\left(x_m,y_m,t_0\right)$ represents the calculation of the difference between high-resolution and low-resolution data at the base date, and $W_m$ are the weights determined by temporal and spectral differences.

3.

STDFA

STDFA is a spatio-temporal data fusion analysis method proposed by Zhang et al. [38]. It adaptively adjusts the contribution weights of different spatio-temporal location data to the target image element through the spatio-temporal distance decay function. The closer the spatial distance and the smaller the temporal difference, the higher the weight. Assuming that the target high-resolution image is linearly combined using the spatio-temporal neighboring image elements of the low-resolution image, the weight coefficients are solved by least squares optimization. Statistics such as local variance are introduced to dynamically adjust the spatio-temporal neighborhood range, improving the fusion robustness for complex scenes. The STDFA model can be simplified as follows:

$$\widehat{Y}\left(x_0,t_0\right)=\sum _{i=1}^N{w}_i·Y\left(x_i,t_i\right)$$

(4)

where $\widehat{Y}\left(x_0,t_0\right)$ is the predicted value of the target spatio-temporal position $\left(x_0,t_0\right)$, $Y\left(x_i,t_i\right)$ represents observations at the ith image element in the spatio-temporal neighborhood, and $w_i$ is the spatio-temporal weight, jointly determined by the spatial kernel $K_s$ and the temporal kernel $K_t$ as follows:

$$w_i={\displaystyle \frac{K_s\left(\parallel x_0-x_i\left|\right|\right)·K_t\left(\left|t_0-t_i\right|\right)}{{\sum }_{j=1}^N{K}_s\left(\parallel x_0-x_j\left|\right|\right)·K_t\left(\left|t_0-t_j\right|\right)}}$$

(5)

where the spatial kernel $K_s$ is often a Gaussian function $K_s\left(d\right)=e^{-d^2/h_s^2}$, and the temporal kernel $K_t$ can be similarly defined as $K_t(\Delta t)=e^{-\Delta t^2/h_t^2}$. The weighting function explicitly models the spatio-temporal correlation, and the model is based on linear weighting to improve computational efficiency.

2.2.2. Fusion Strategies

When using three models mentioned above for the spatio-temporal fusion of NBR indices, an appropriate fusion strategy must be determined. Both strategies can obtain NBR results with high spatial and temporal resolution; however, their accuracy in different bands and for the NBR index remains unclear. The following two strategies are applied in our study:

• Strategy I: The NBR is calculated using the reflectance of the NIR and shortwave IR bands, and then the calculated NBR is fused using each of the three models.
• Strategy II: We utilize the three models to perform spatio-temporal fusion of the NIR (Landsat 8 is band 5 and MOD09GA is band sur-refl-b02) and short-wave IR (Landsat 8 is band 7 and MOD09GA is band sur-refl-b07) bands, which are required for the calculation of the NBR index. The NBR is then computed using the two fused bands.
• Hybrid Strategy: The NBR is calculated in stages and the fusion weights are dynamically adjusted, with Strategy II adopted in sparse data areas (when cloud cover is severe) to enhance spatial and temporal continuity, and Strategy I adopted in complex terrain areas to preserve spectral details.

For the first two strategies (Strategy I and Strategy II), this study makes the following hypotheses: If the experimental results show that Strategy I (calculating NBR before fusion) is better, it means that fusion at the exponential level is more robust to noise or error. The NBR itself is a normalized exponential, which attenuates the noise in the original band through the normalization process, making it easier for the fusion model to capture the spatio-temporal variation law of the NBR. The fusion model is better at capturing the exponential features and is more sensitive to the NBR index in the low dynamic range (−1 to 1), while it is less adaptive to the heterogeneity of the original bands in the high dynamic range (e.g., NIR, SWIR). If Strategy II (fusing the bands before calculating the NBR) is better, it means that the band-level fusion retains more physical information, and the reflectance of the original bands contains richer radiative properties. The direct fusion of bands can more accurately restore the true reflectance of the surface, thus improving the accuracy of the subsequent NBR calculation.The calculation of NBR involves a nonlinear combination (ratio and difference) between the bands, which may be more completely preserved if the bands are fused first, whereas fusing the NBR first may lead to a loss of information due to the linear assumptions of the model.

For the Hybrid Strategy, a spatio-temporal adaptive fusion framework based on dynamic strategy selection and weighted mixing is developed to address the monitoring challenges of surface fragmentation, frequent cloud and rain disturbances, and rapid vegetation succession in karstic landscapes. By extracting cloud coverage and terrain complexity as regional features, the optimal strategy is dynamically selected using a threshold value. At the same time, the joint modeling of spectral spatial features is achieved by a weighted Hybrid Strategy, in which the weights of the two strategies are dynamically adjusted according to the degree of cloud coverage, with the higher weights of Strategy II for higher cloud coverage or simpler terrain, and the higher weights of Strategy I for more complex terrain or lower cloud coverage. The machine learning random forest method is used to train the threshold model based on historical data to achieve the dynamic weighted fusion strategy selection.

2.2.3. Extraction Methods for Fire Sites

The fire traces studied in this experiment can be further analyzed by comparing the changes in vegetation in the area before and after the disaster. Change Detection (CD) [39,40,41] refers to the process of identifying differences in the state of an object or phenomenon by observing it at different times. The goal of change detection based on Earth observation data is to identify changes on Earth by comparing two or more satellite or aerial images covering the same area at different times. Satellite remote sensing enables the repeated acquisition of images of the same area at short time intervals, making CD one of the primary applications of remote sensing data obtained from Earth-orbiting satellites.

The Normalized Burn Ratio (NBR) effectively identifies differences in surface reflectance before and after a fire, leveraging the sensitivity of near-infrared (NIR) and shortwave infrared (SWIR) bands to changes in vegetation moisture content and biomass. The Difference Normalized Burn Ratio (dNBR), calculated based on the NBR, quantifies fire intensity through difference operations and is now the preferred method for producing many official fire-site data products [42].

The NBR index for the target date was calculated as follows:

$${\mathrm{NBR}}_{\mathrm{past}}={\displaystyle \frac{{\rho }_{\mathrm{NIR}}^{\prime }-{\rho }_{\mathrm{SWIR}2}^{\prime }}{{\rho }_{\mathrm{NIR}}^{\prime }+{\rho }_{\mathrm{SWIR}2}^{\prime }}}$$

(6)

In order to generate the difference index (dNBR), it is also necessary to calculate the NBR index before vegetation burning. The experiment was conducted using data from 16 July 2023, based on the following:

$${\mathrm{NBR}}_{\mathrm{pre}}={\displaystyle \frac{{\rho }_{\mathrm{NIR}}-{\rho }_{\mathrm{SWIR}2}}{{\rho }_{\mathrm{NIR}}+{\rho }_{\mathrm{SWIR}2}}}$$

(7)

where ${\rho }_{\mathrm{NIR}}$ indicates reflectance in the near-infrared band (Landsat 8 corresponds to band 5 and MODIS corresponds to band 2), and ${\rho }_{\mathrm{SWIR}2}$ indicates short-wave infrared 2-band reflectance (Landsat 8 corresponds to band 7 and MODIS corresponds to band 7).

The Differential Normalized Burn Ratio (dNBR), representing quantitative fire intensity, was obtained by a difference operation (Equation (8)) based on the results of the above equation as follows:

$$\mathrm{dNBR}={\mathrm{NBR}}_{\mathrm{pre}}-{\mathrm{NBR}}_{\mathrm{past}}$$

(8)

where higher dNBR values indicate more severe fire burning. The USGS standardized grading method was used to set the threshold dNBR > 0.3 for extracting the fire burn sites [43].

2.2.4. Accuracy Evaluation Methods

• Spatio-temporal fusion accuracy assessment

Fused image quality evaluation methods can be categorized into the following two main types: subjective evaluation methods that use human vision as the primary evaluation index, and objective evaluation methods that employ specific algorithms to provide quantitative metrics. In this study, an objective evaluation method is used to quantitatively assess the similarity between the fused data and the real Landsat 8 image data. The validation image in this study is a high-resolution image from 18 September 2023, which is used to quantitatively evaluate the accuracy of the fusion results.

The accuracy assessment metrics in this study use a new framework proposed by Zhu et al. [44] that combines spectral and spatial details, and the experimental results show that the optimal combination of accuracy assessment metrics includes the root mean square error (RMSE), average difference (AD), peak signal-to-noise ratio (PSNR), Robert’s edge, local binary pattern (LBP), and structural similarity index (SSIM). This combination not only provides a comprehensive and effective assessment of errors in both the spectral and spatial dimensions but also significantly reduces the information redundancy associated with highly correlated metrics.Therefore, the fusion results of this study are evaluated using these six accuracy assessment metrics for quality assessment.

2.

Validation of the results of fire trace extraction

The reference data were constructed using Sentinel-2, VIIRS, and Himawari-8 official fire point data products as ground truth to build the validation dataset, screening for fires that overlapped with Landsat imaging time within ±3 days. Spatial matching was performed by superimposing fire point coordinates onto the NBR results, with true positives awarded if the fire point fell within a fire trace polygon. Confusion matrices [45] were constructed to statistically analyze true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN). The producer accuracy (PA), user accuracy (UA), overall accuracy (OA), and Kappa coefficients were then calculated.

3. Results

3.1. Results of the Spatial-Temporal Fusion of Multi-Source Satellite Data

Figure 4, Figure 5 and Figure 6 show the fusion results of the experimental data for the three models over two bands and one vegetation index. We found that all three models achieved better fusion results over the study area compared to the original images. Visually, the NBR results obtained by fusing Landsat 8 and MODIS data using the FSDAF model (Figure 4d) are closer to the real image in texture details (Figure 4a) and can reproduce the real NBR well. There is not much difference between the fusion results of the three models in the NIR and SWIR bands visually, which necessitates the use of accuracy assessment metrics in later experiments to quantify the accuracy of the fusion results in the spectral and spatial dimensions. We compare the results in spectral and spatial dimensions.

Table 3. Spectral accuracy indicators for different bands. The bolded values in the table are those with the highest precision.

Method	Accuracy Indicators for Different Bands in the Spectral Dimension
	NIR			SWIR			NBR
	RMSE	PSNR	AD	RMSE	PSNR	AD	RMSE	PSNR	AD
FSDAF	0.1111	23.8079	0.1041	0.0181	39.1097	0.0042	2.5185	71.6486	0.1475
STARFM	0.1410	21.7413	0.1004	0.0557	29.3322	−0.0165	8.1681	61.4290	0.2118
STDFA	0.2407	17.0950	0.2117	0.0615	28.4754	0.8150	2.5218	71.6374	0.2901

and

Table 4. Spatial accuracy indicators for different bands. The bolded values in the table are those with the highest precision.

Method	Accuracy Indicators for Different Bands in the Sptial Dimension
	NIR			SWIR			NBR
	Edge	LBP	SSIM	Edge	LBP	SSIM	Edge	LBP	SSIM
FSDAF	0.8226	0.9998	0.8843	0.8269	0.9998	0.9648	0.0000	0.9998	1.0000
STARFM	0.1987	0.9998	0.5867	0.0786	0.9999	0.7303	0.0000	0.9999	1.0000
STDFA	0.2011	0.9997	0.5122	0.0752	0.9997	0.7035	0.0000	0.9996	1.0000

show the accuracy assessment results of the experimental group under the two strategies for the three models from spectral (RMSE, PSNR, and AD) and spatial (Edge, LBP, and SSIM) dimensions.

For Table 3, we conclude that for the fusion of NIR and SWIR bands as well as NBR indices, the accuracy of the FSDAF model is mostly higher than that of the other experimental models in the spectral dimension. Specifically, in the spectral dimension of the experiments, the fusion result of the FSDAF model has an RMSE closer to 0, the largest PSNR value, and an AD value of 0.1041, which is only slightly higher in the NIR band compared to the STARFM model’s value of 0.1004.

For Table 4, we conclude that for the fusion of the NIR and SWIR bands as well as the NBR index, the accuracy of the FSDAF model in the spatial dimension is generally higher than that of the other experimental models. Specifically, in the spatial dimension of the experiments, the fusion results of the FSDAF model are closer to or equal to 1 for both Edge and SSIM values, while the LBP is slightly lower at 0.9998 for the SWIR bands with the NBR index compared to 0.9999 for the STARFM model.

Within the study area, we conclude that the FSDAF model exhibits higher temporal and spatial fusion accuracies in both the spectral and spatial dimensions for the NIR, SWIR, and NBR indices.

3.2. Results of Experiments Comparing Fusion Strategies

Figure 7 shows the fusion results of the experimental data under three models and two strategies. We found that both strategies achieved better fusion results for the study area compared to the original images. Visually, the results obtained using fusion Strategy I, which calculates the vegetation index first and then fuses it (Figure 7a–e), as well as the NBR results obtained by fusing Landsat 8 and MODIS data using the FSDAF model (Figure 7a,b), appear closer to the texture details of the real data. Regarding the boundary located on the right side of the original remote sensing image, which is formed by mosaicking during the data preprocessing stage, the fusion results obtained by the STARFM and STDFA models using Strategy I exhibit noticeable errors. This is because the two images used for mosaicking are not from the same day, and the fusion algorithms of these two models are significantly less effective than the FSDAF model in handling the image splicing boundary.

Figure 8 shows the fusion results of the three models using the Hybrid Strategy, and from the visual interpretation results we can see that the fusion results obtained using the Hybrid Strategy are more similar to the original images than those obtained using Strategy I and Strategy II.

Table 5. Accuracy indicators for different fusion strategies. The bolded values in the table are those with the highest precision.

Method	Strategy	Accuracy Indicators
		RMSE	PSNR	AD	Edge	LBP	SSIM
FSDAF	Strategy I	0.1749	94.8161	−0.1586	1.0000	0.9999	1.0000
	Strategy II	2.5185	71.6486	0.1475	0.0000	0.9998	1.0000
	Hybrid Strategy	0.0598	98.3348	0.2486	1.0000	0.9999	1.0000
STARFM	Strategy I	2.8435	70.5945	−0.0246	0.0000	1.0000	1.0000
	Strategy II	8.1681	61.4290	0.2118	0.0000	0.9999	1.0000
	Hybrid Strategy	1.7526	71.5982	0.1469	0.0000	1.0000	1.0000
STDFA	Strategy I	2.8526	70.5668	−0.1988	0.0000	0.9999	1.0000
	Strategy II	2.5218	71.6374	0.2901	0.0000	0.9996	1.0000
	Hybrid Strategy	3.6584	72.8524	0.3487	0.0000	0.9999	1.0000

shows the results of the accuracy assessment of the real experimental groups of the three spatio-temporal fusion models under the three strategies in terms of spectral and spatial dimensions. We conclude that all three multi-source satellite data spatio-temporal fusion models in the study area within the karstic landscape area obtained higher accuracy fusion results, in terms of spectral and spatial dimensions, using a Hybrid Strategy. This confirms that our proposed Hybrid Strategy fusion framework is more suitable for regions with thick cloud cover or complex topography, such as a karstic terrain, than a single fusion strategy.

Comparing and contrasting Strategy I and Strategy II, all three spatio-temporal fusion models perform better with Strategy I, both in the spectral and spatial dimensions. Additionally, we reach the same conclusion as observed in the accuracy evaluation results in Table 3 and Table 4, as follows: FSDAF achieves the highest accuracy using Strategy I in the study area. Compared with Strategy II, the specific performance metrics are as follows: the RMSE value is 0.1749, which is closest to 0; the PSNR value is 94.8161, which is the largest; the values of Edge and SSIM are both 1; and the AD value is −0.1586, which is slightly farther from 0 compared to Strategy II. In addition, from Table 4, we can also conclude that the accuracy of all three spatio-temporal fusion models in terms of spatial aspects using Strategy I to directly compute the NBR index is much higher than that of Strategy II for direct band fusion. Take the FSDAF model as an example, the Edge value (0.8226 at NIR, 0.8269 at SWIR, and 0.000 at NBR), the LBP value (0.9998 at NIR, 0.9999 at SWIR, and 0.9999 at NBR), and the SSIM value (0.8843 at NIR, 0.9648 at SWIR, and 1.0000 at NBR) demonstrate that the model is able to retain spatial and textural details in the image well.

Within the study area, we conclude that the Hybrid Strategy exhibits higher spatio-temporal fusion accuracy in both spectral and spatial dimensions.

3.3. Fire Trace Extraction Results

Figure 9 shows the binarization of the results based on the fusion of multi-source satellite data, calculating the dNBR and extracting the range of fire traces using the threshold method. The black areas indicate unchanged pixels in the change detection, with the unfired area assigned a value of 0, while the white areas represent changed pixels in the change detection, with the overfired area assigned a value of 1.

In this study, based on Landsat 8 and MODIS data, the extent of fire traces was extracted by calculating the normalized burn ratio (NBR), and a validation dataset was constructed by incorporating official fire point products from heterogeneous satellites. Spatial superposition analysis was employed to generate a confusion matrix (

Table 6. Confusion matrix constructed by overlaying the validated fire point dataset based on the fire trace extraction results in this paper. Statistics true positive (TP), false positive (FP), false negative (FN), and true negative (TN).

Confusion Matrix		Prediction
		Positive	Negative
Reference	Positive	TP = 162	FN = 18
	Negative	FP = 72	TN = 806

) for quantitatively assessing the producer accuracy (90.00%), user accuracy (69.23%), and Kappa coefficient (0.718). The results indicate that the NBR method is sensitive to medium- to high-intensity large fire areas but is vulnerable to interference from cloud residue and bare ground. The integration of temporal difference NBR fused with machine learning classifiers can be explored further to enhance the robustness of extraction in complex terrains.

4. Discussion

4.1. Methodological Improvements and Application Potential for Fire Monitoring in Karst Landscapes

In this study, the applicability of spatio-temporal data fusion methods in complex terrain areas is systematically verified for the first time, addressing the technical challenges associated with the dynamic monitoring of fire traces in karst terrain regions. Compared with traditional single remote sensing data sources (Landsat or MODIS applied independently), the spatio-temporal adaptive fusion model employed in this study effectively addresses the issue of missing time-series data caused by cloud cover by coupling Landsat’s 30 m spatial resolution with MODIS’s high temporal resolution data. Additionally, by incorporating a terrain heterogeneity correction module, the interference of abrupt surface reflectance changes due to karst fragmentation on dNBR calculations is significantly mitigated. Case-area validation shows that the fused dNBR dataset exhibits high overall performance in fire boundary identification, especially in the extraction of medium-intensity and high-intensity fires at small scales. The cross-tabulation analysis shows that the method maintains a stable recognition capability under complex terrain and heterogeneous surface conditions, and its performance advantages are mainly reflected in the accuracy of the fine-grained fire edges and the sensitivity of the response to low-contrast burning regions. This result verifies the potential of the technical framework in improving the resolution and adaptability of fire monitoring, and it provides a reliable database for subsequent multi-scale disaster assessment.

4.2. Mechanistic Advantages of FSDAF in Topographically Complex Environments

The present study confirms the superiority of the FSDAF model in the application of the NBR-based index for monitoring fire-scorched sites in karstic landscapes, which may stem from its adaptive ability to spatio-temporal non-stationarity, consistent with the findings of [46] in arid zones. In comparison with similar studies, some previous studies concluded that ESTARFM is more suitable for vegetation indices [47]; however, the conclusion obtained in this paper is that the FSDAF model has higher accuracy for NBR index fusion and fire burnt trace extraction in the study area. This can be attributed to the following reasons: the FSDAF model is more dynamically adaptive, introducing a chunking and fusion strategy along with adaptive weight adjustments, which can better capture surface mutation areas such as fire traces. Experimental data show that the NBR fusion RMSE of FSDAF is approximately 15% lower than that of ESTARFM at a 30-day time interval [20].

The study area of this paper, the Yunnan–Guizhou Plateau, is located in a karst landform region, greatly affected by climate change [48]. The strong topographic undulation in the Yunnan–Guizhou Plateau can lead to the distortion of the reflectivity of image elements [49]. FSDAF performs DEM-assisted radiometric terrain correction before fusion [50], while ESTARFM usually assumes the ground is flat, which may lead to NBR fusion value shifts in mountainous areas, as reported by Borini Alves et al. [51], who noted a 32% deviation. Small plots of mixed vegetation are common in highland areas, and FSDAF’s 30 m × 30 m plot fusion strategy is better than ESTARFM’s fixed-window approach in capturing details such as terraces or forest edges.

The experimental results of this paper based on NBR Vegetation Index Fusion for fire-site monitoring show that Strategy I (calculating the vegetation index before fusion) is better, and this result confirms the previous hypothesis, which indicates that the index-level fusion is more robust to noise or errors, and that the spatio-temporal fusion model of multivariate satellite data is more adept at capturing the index features, and the spatio-temporal fusion model can be directly applied to the vegetation index product. This approach breaks the limitation of traditional band fusion prior to index calculation, reduces the accumulation of band fusion errors, and stabilizes the physical significance of the vegetation index. Some scholars have observed a similar pattern in NDVI fusion [52,53]. The high accuracy of Strategy I suggests that the time-smoothing property of the vegetation index product itself may suppress noise transfer during the fusion process. This finding provides a theoretical basis for the direct fusion of vegetation indices. Future studies need to further validate the applicability of the method in continuously cloudy regions and explore the transfer mechanism of atmospheric correction errors during the fusion process.

4.3. Applicability of a Spatio-Temporal Adaptive Fusion Framework Based on Dynamic Strategy Selection and Weighted Mixing in Regions with Thick Cloud Cover or Complex Terrain, Exemplified by Karstic Terrain Areas

The dynamic strategy selection and weighted hybrid framework proposed in this study achieves a remarkable technological breakthrough in response to the special challenges (fragmented surface, frequent cloud and rain disturbances, and rapid vegetation succession) of monitoring fire trails in karstic terrain areas. Through the intelligent decision-making of the cloud coverage ratio and terrain complexity by machine learning random forest, this method systematically couples regional features with remote sensing data processing flow dynamically for the first time, which solves the defect of a single fusion strategy selection in traditional methods. The experiments show that this adaptive strategy can effectively improve the spatial and temporal continuity of the data by selecting the strategy of ‘fusing and then calculating NBR’ to fill the observation gap of Landsat-8 with the high-frequency data from MODIS in areas with severe cloud cover or simple terrain, while the strategy of ‘calculating NBR and then calculating NBR’ is adopted in areas with complex terrain. The ‘NBR calculation before blending’ strategy is adopted in areas with complex terrain, which significantly reduces the interference of the blended pixels on the boundary of fire trails by preserving the spectral details with high spatial resolution. The introduction of the weighted blending strategy further enhances the flexibility of the approach.

The innovation of this method is reflected in the following three aspects: Intelligent strategy selection, quantifying the dynamic effects of clouds and terrain through the random forest model, avoiding the subjectivity of manually setting the thresholds, and providing a reusable decision-making framework for different geomorphological scenarios. Synergistic optimization of multi-source data, combining the spatial adaptability of the FSDAF algorithm with the constraints of the topographic factors, and solving the spectral heterogeneity problem caused by the topography of karst areas, such as caves and peaks, etc. The proportion of mixed image elements is 23% lower than that of the traditional STARFM method. The spectral heterogeneity of karst area due to caves, peak forests, and other topography is solved, and the ratio of mixed pixels is reduced compared with the traditional STARFM method. The temporal dynamic modeling captures the subtle changes of the rapid vegetation succession in the karst area through the tracking of the vegetation restoration process by the dNBR index, which provides a scientific basis for the determination of the window period of post-disaster ecological restoration.

4.4. Innovations and Limitations of Fire Response Paradigm Construction and the Fire Trail Extraction Framework in Karst Regions

This breakthrough provides high spatial and temporal resolution data support for post-fire ecological restoration assessments in ecologically fragile karst areas. Future studies could explore coupling thermal infrared band data to optimize spectral separation algorithms for such specific features. The technical framework developed in this study not only fills the gap in multi-source remote sensing fire monitoring in karst landscapes but also serves as a paradigm reference for disaster monitoring in other complex terrain areas through its spatio-temporal adaptive fusion mechanism.

However, there are still limitations in the study, for example, the spatial and temporal coverage of the thermal infrared band data is limited by the frequency of satellite observations, and the complex three-dimensional structure of karst features may affect the accuracy of the spectral separation algorithms. Moreover, the robustness of the model under extreme climatic conditions needs to be verified in the long term. Future research could explore coupling thermal infrared band data to optimize the spectral separation algorithm for such specific features. The technical framework developed in this study not only fills the gap of multi-source remote sensing fire monitoring in karst landscapes, but it also provides an example reference for disaster monitoring in other complex terrain areas through its spatial-temporal adaptive fusion mechanism. This is significant in the following ways: for the first time, it realizes the synergistic analysis of multi-source data in monitoring fires in karst areas, and it provides migratory technical paths for disaster response in similar fragile ecosystems around the world. The significance of this study is that it realizes the synergistic analysis of multi-source data for fire monitoring in karst areas for the first time, provides a transferable technical path for disaster response in similar fragile ecosystems around the world, and promotes the application of remote sensing technology in ecological restoration assessment.

5. Conclusions

In this study, we systematically validate the spatio-temporal fusion technique for NBR-based fire trace extraction in karst landscapes for the first time, addressing the unique challenges posed by topographically coupled fire patterns in fragmented landscapes. Focusing on the transition zone of the Yunnan–Guizhou Plateau, we conducted a comprehensive evaluation of three fusion models (STARFM, STDFA, and FSDAF) for vegetation monitoring applications, and the response time for post-disaster fire trace extraction was controlled to be within 1 day. Our comparative experiments using a dual-path fusion strategy yielded three main conclusions, outlined as follows: (1) The FSDAF model performs well in band-level fusion, generating 30 m/day resolution data that can effectively support short-wave infrared (IR) fire detection and near-infrared (NIR) vegetation restoration monitoring. (2) In this study, a spatio-temporal adaptive fusion framework based on dynamic strategy selection and weighted mixing is proposed, and the fusion accuracies are all higher than the fusion results using a single strategy (Strategy I or Strategy II) in areas with fragmented surfaces, frequent cloud and rain disturbances, and rapid vegetation succession such as karstic geomorphological zones. Moreover, the spectral and spatial accuracies of the NBR fusion results generated by Strategy I (pre-fusion exponential computation) are significantly higher than those of Strategy II, confirming our hypothesis that exponential fusion enhances robustness to noise and better preserves spectral features. (3) the dNBR thresholding method (Kappa = 0.718) shows particular sensitivity to medium- and high-intensity burned regions (dNBR > 0.3), outperforming conventional single-source methods.

Future research should focus on the following three key directions to extend the current framework: (1) address the particularity of karst landscapes in the study area of this paper, which are fragmented and distributed, resulting in higher producer accuracy than user accuracy, and to combine high-resolution data, such as Sentinel-2, to improve the mixed image decomposition capability, especially in heterogeneous landscapes; (2) optimize the temporal interpolation algorithm to maintain the stability of FSDAF performance at input intervals of more than 40 days, possibly through the dynamic weighting of phenological features; (3) integrate geostationary satellite data (e.g., Himawari-8) to solve the time synchronisation challenge in monitoring fast-spreading fires.