A Land Cover Recognition-Based Method for Wildfire Early Warning in Transmission Corridor Areas

Abstract

To improve the accuracy of wildfire risk identification in areas adjacent to power transmission corridors, this study proposes a wildfire early warning method that integrates refined land cover segmentation and multimodal feature deep learning. First, an improved bi-branch semantic segmentation network (BuildFormer++) is used to perform refined classification of high-resolution remote sensing images, extracting six types of land cover information, including forest and cultivated land. Second, a multi-dimensional feature set integrating land cover, topography, climate, and human activities is constructed and input into a multimodal wildfire point prediction network for deep feature fusion and probabilistic modeling. Experimental results show that the proposed segmentation network achieves a mean intersection–union ratio (mIoU) of 40.68% in the semantic segmentation task; the early warning model achieves an accuracy of 85.37%, an F1 score of 93.15%, and an ROC-AUC of 85.42% in risk prediction, significantly outperforming comparative methods. The “refined segmentation–feature fusion–risk prediction” framework constructed by this method can provide reliable technical support for the operation and maintenance safety and fire prevention of power transmission corridors.

1. Introduction

As extreme climate events become increasingly frequent, wildfires have emerged as one of the most severe natural disasters threatening the safe operation of power transmission lines [1,2]. Statistics indicate that in many regions across our country, incidents of tripping and shutdowns in transmission and distribution lines are closely linked to mountain fires [3,4]. Consequently, research on early warning systems for wildfires affecting power grid facilities holds significant practical importance.

Existing studies have delineated transmission-line wildfire risk zones using geographic information system (GIS) and remote sensing. In Hubei Province [5,6,7], Ruan Ling et al. built a risk indicator system incorporating land cover, topography, hydrology, and human factors, and produced fire risk zoning via weighted overlay and the analytic hierarchy process [5]. However, these approaches depend on expert-assigned factor weights and struggle to capture the nonlinear relationships between complex land cover patterns and wildfire occurrence [8,9,10].

Internationally, researchers have modeled wildfire risk using probabilistic statistics and machine learning approaches. For example, Nhongo et al. developed a probabilistic model incorporating environmental and socioeconomic factors to estimate wildfire occurrence probabilities [11]; Mohajane et al. combined remote sensing with machine learning to construct a fire hazard susceptibility model [12]; and Moghim and Rahmati applied several machine learning algorithms to delineate fire risk zones [13]. In addition, Díaz-Vázquez et al. integrated remote sensing data with machine learning to model burn probability [14], while Sheriff and Khan used Random Forest (RF) and Classification and Regression Trees (CART) for fire susceptibility mapping [15]. Collectively, these studies suggest that both statistical models and ensemble learning methods hold strong promise for predicting wildfire risk.

In terms of data support, remote sensing satellites provide reliable inputs for wildfire hazard monitoring. Medium- and high-resolution optical imagery—such as Sentinel-2 and Landsat—has been widely used for land cover mapping and burn scar detection [16,17]. For example, Suwanprasit et al. delineated burned areas in Thailand by combining Sentinel-2 data with spectral indices [18], and Taylor et al. verified the spatiotemporal detection capability of Sentinel-2 for grassland fire monitoring [19]. Meanwhile, thermal anomaly and active fire products from MODIS and VIIRS are used globally for historical hotspot detection and situational monitoring, providing long time series data for fire risk modeling [20,21]. In addition, Digital Elevation Models (DEMs) are essential for deriving topographic variables—such as slope and aspect—that influence fire behavior and spread [22,23].

In summary, traditional GIS-based statistical modeling offers strong interpretability and is well-suited for factor overlay analysis; however, its accuracy is often limited in the typically heterogeneous and fragmented landscapes of power transmission corridors. Meanwhile, while traditional wildfire early warning systems based on remote sensing and machine learning perform well in regional-scale risk mapping, they largely rely on pre-classified land use data or medium-resolution imagery, lacking the spatial details required for corridor-scale risk identification. Deep learning-based remote sensing classification excels in automatic feature extraction and achieves high classification accuracy; however, existing deep learning methods mostly focus on single-modal image analysis, lacking a structured mechanism for systematically integrating multi-source features such as topography, meteorology, and human activities. Furthermore, their outputs often fail to fully integrate into the probabilistic risk modeling framework for power grid operation and maintenance. Therefore, a hybrid approach is urgently needed—one that can combine the advantages of both paradigms to achieve refined land cover representation while simultaneously completing probabilistic wildfire risk modeling that deeply integrates multi-source information.

Based on this, this paper takes the area adjacent to the transmission corridor in Xiaogan, Hubei Province, as the study area, and proposes a wildfire early warning method that integrates fine-grained land cover recognition with multi-modal feature learning. The core idea of the research is as follows: First, an improved dual-branch semantic segmentation network (BuildFormer++) is applied to high-resolution remote sensing imagery to accurately classify six land cover types, including forest, cropland, and built-up areas. Second, the obtained land cover information is systematically integrated with topographic, meteorological, and human activity factors to construct a multi-dimensional feature set. Finally, a multi-modal wildfire risk prediction network is employed to model the probability of fire occurrence and produce spatially explicit risk levels. The contributions of this paper are threefold:

(1)

A specialized semantic segmentation network is introduced to achieve refined extraction of land cover features in complex transmission corridor environments;

(2)

A deep learning-based prediction framework is developed to effectively fuse multi-source features and improve wildfire risk identification accuracy;

(3)

The proposed framework provides spatially explicit risk zoning that can be directly integrated into grid operation and maintenance workflows, supporting practical disaster prevention and control.

2. Materials and Methods

2.1. Study Area

This study selects the adjacent areas of the two transmission corridors, Xiaoshi I and Xiaoshi II, in Xiaogan City, Hubei Province, as the research subject in Figure 1. The region lies within the transition zone from the Jianghan Plain to the Dabie Mountains and encompasses diverse landforms, including low-lying plains, hills, and localized mountainous terrain. Forest, cropland, and residential land uses are interwoven, producing a complex fuel mosaic. Coupled with frequent human activities, these conditions heighten susceptibility to forest and grassland fires, posing a substantial risk to the reliable operation of interregional transmission lines. Consequently, the study area is both typical and representative, making it suitable for validating corridor-scale fire risk identification and early warning methods.

2.2. Data Source and Preprocessing

2.2.1. Remote Sensing Image Data

To support refined land cover identification in the power transmission corridor area, this study selected the European Space Agency’s Sentinel-2 multispectral imaging satellite as the primary data source. Images were acquired during the vegetation growing season (May to October) from 2018 to 2023, and six images covering the study area were selected, with a global grid number of T49RBU (Figure 2). The cloud cover rate of each individual image was less than 10%. The images have a maximum spatial resolution of 10 m, enabling refined identification of fragmented and heterogeneous surface cover along the power transmission corridor. In the preprocessing stage, radiometric calibration, atmospheric correction, and orthorectification were performed on the original images sequentially to eliminate sensor and atmospheric interference, ensure spectral and geometric accuracy, and finally mosaic them into a large-format reference image covering the entire study area.

To adapt to the training and inference of the deep learning model, the aforementioned large-format reference image and its corresponding manually labeled ground truth map were regularly segmented into 512 × 512 pixel image sample blocks. An overlapping sliding window strategy (with an overlap of 128 pixels) was used during segmentation to increase the number of samples, enhance data diversity, and mitigate boundary effects during model prediction. Finally, all image label sample pairs were divided into training, validation, and test sets in a 7:2:1 ratio, providing a foundation for the training and evaluation of the BuildFormer++ network.

2.2.2. Multi-Source Geographic Environmental Feature Data

(1)

Climate data

The temperature data used in this study were obtained from Landsat 8 satellite imagery products (Level 2, Collection 2, Tier 1) provided by the U.S. Geological Survey (USGS). Land surface temperature (LST) was retrieved using its ST_B10 thermal infrared band to characterize the spatial heterogeneity of near-surface thermal conditions. Precipitation data were obtained from the Climate Hazards Center Infrared Combined Ground Station Precipitation Data Set (CHIRPS), which integrates satellite infrared information and station observations to generate a long-term, spatially consistent, gridded precipitation product. Evaporation data were obtained from the MOD16A2.061 dataset, which, based on the Penman–Monteith energy balance equation, uses MODIS remote sensing data and meteorological reanalysis data to generate an 8-day synthetic evapotranspiration (ET) product with a resolution of 500 m.

To systematically reveal the climatic background of the study area, we calculated the average surface temperature, average annual precipitation, and average annual evapotranspiration for 15 years from 2009 to 2023, and their spatial distribution pattern is shown in Figure 3. The results show that the climatic elements in the study area exhibit significant spatial differentiation: high surface temperature areas are mainly distributed in urban built-up areas and exposed surfaces; precipitation generally increases from northwest to southeast; evapotranspiration is closely related to vegetation cover and moisture conditions, with high-value areas concentrated around forests and water bodies.

(2)

Human activity data

This study uses multi-source spatial data to quantify the intensity of human activities in the region, serving as a key input for assessing the risk of human-induced wildfires. First, based on OpenStreetMap (OSM) open-source road network data, information on highways, arterial roads, and primary and secondary roads within the study area was extracted. A road proximity spatial layer was generated by calculating the Euclidean distance from each grid cell to the nearest road (Figure 4a). Higher road proximity indicates stronger human accessibility and activity frequency, and a more significant potential fire source risk from transportation, production, and other activities in the corresponding area. Second, the VIIRS monthly composite nighttime light data (Version 1) released by the Colorado School of Mines Earth Observatory was introduced [24]. By calculating the average value of 14 years of data from 2009 to 2023, a stable spatial distribution of nighttime light intensity reflecting human settlements and economic activities was constructed (Figure 4b). Nighttime light intensity can serve as an effective proxy indicator for regional population density, energy consumption, and land development intensity; high-value areas are usually closely related to the high probability of human-induced fire sources. Road proximity and nighttime light intensity together constitute a quantitative expression of human activity factors. They are integrated with land cover, topography and climate data to form a multi-dimensional feature system to support subsequent wildfire risk early warning modeling.

2.2.3. Fire Sample Data

The historical fire sample labels used in this study to train and validate the wildfire risk prediction model were strictly generated based on nine years of historical observation data from 2015 to 2023. Positive samples were sourced from two authoritative data sources: the China Forest Fire Database and the VIIRS (Visible Infrared Imaging Radiometer Suite) active fire point product, ensuring data reliability and spatiotemporal coverage integrity. After spatial registration and data cleaning, a total of 3157 valid historical fire point records were obtained as positive samples, their spatial distribution shown in Figure 5a.

For data quality control, we further applied confidence filtering: only 985 high-quality fire points with confidence scores above 80% were retained for subsequent negative sample generation and model training. This step significantly improved the reliability of the positive samples. To construct a class-balanced and spatially representative training set, negative samples were generated using a constrained random sampling method. In areas within the study area where historical fire points did not appear, 2955 spatial points were randomly generated as negative samples at a ratio of 1:3, their spatial distribution shown in Figure 5b.

A minimum spatial distance constraint was set during the sampling process to ensure that negative sample points maintained a certain distance from the most recent historical fire points, thereby reducing the potential impact of spatial autocorrelation. Ultimately, the positive and negative samples together constituted a structured dataset containing 6112 sample points. Each sample point was associated with multidimensional geographic environmental features of its corresponding location (including land cover composition, topographic features, climate conditions, and intensity of human activities), providing a solid data foundation for subsequent wildfire risk probability modeling.

2.3. Intelligent Remote Sensing Assessment Framework for Wildfire Risk in Transmission Corridors

As a critical infrastructure for energy transmission, the area along the power transmission corridor features complex terrain, diverse surface cover types, and frequent human activities. The occurrence of wildfires not only directly threatens the safe operation of power transmission lines but also may cause large-scale ecological damage and significant economic losses. Therefore, conducting accurate and quantitative assessments of wildfire risks in this region holds important practical significance.

Traditional wildfire risk assessment methods mostly rely on a single semantic segmentation network for surface cover classification. In regions like power transmission corridors, which have small spatial scales and highly heterogeneous surface covers, these methods are prone to misclassification of fine-grained features. Additionally, existing studies mostly adopt simple superposition or traditional statistical models (such as logistic regression) for multi-source feature fusion, making it difficult to fully explore the deep correlations between different modal features and resulting in limited risk prediction accuracy [25,26,27].

To overcome the aforementioned limitations, this study takes multi-source remote sensing and geographical environment data as the core support, and constructs a three-level technical system of “refined surface cover segmentation—multi-dimensional feature fusion—wildfire point risk prediction” (Figure 6) to realize the comprehensive assessment and spatial expression of wildfire risks in mountainous power transmission corridor regions. The technical process is as follows:

Firstly, the remote sensing images of the study area are processed based on an improved dual-branch semantic segmentation network to complete the refined classification of six typical surface cover types, including forest land, cropland, shrubland, water bodies, bare land, and buildings. Meanwhile, the spatial distribution pattern and area proportion of each surface cover type are statistically analyzed, providing high-precision basic data for identifying fire source conditions. Secondly, three core factor categories—terrain, climate, and human activities—are systematically integrated to construct a multi-dimensional feature set. Among them, terrain factors are derived from elevation and slope information extracted from the Digital Elevation Mode); climate factors include land surface temperature retrieved from Landsat satellites, precipitation provided by the CHIRPS dataset, and evapotranspiration obtained from the MOD16 product, which together characterize the regional thermal and moisture conditions; human activity factors are based on road distance extracted from OSM open-source data and VIIRS nighttime light index, used to reflect the intensity of human activities and related ignition risks. Finally, the constructed multi-dimensional feature set is input into a multi-modal wildfire point prediction network. Through processes such as feature encoding, fusion enhancement, and probability output, the quantitative assessment and spatial visualization of regional wildfire risk levels are achieved.

2.3.1. Improved Dual-Branch Semantic Segmentation Network

To address the problems of surface cover misclassification and insufficient fine-grained target segmentation accuracy of traditional semantic segmentation networks in small-scale and highly heterogeneous regions such as power transmission corridors, this study proposes a dual-branch encoder–decoder segmentation architecture based on frequency domain enhancement (Figure 7). Through the collaborative modeling of the frequency domain branch and spatial branch, combined with a specially designed small-target enhancement module, the refinement level and robustness of surface cover classification are significantly improved. The network innovatively adopts a dual-branch parallel processing mechanism: the frequency domain branch focuses on mining the periodic texture structures and edge detail information contained in surface cover targets, while the spatial branch enhances the deep collaboration and interactive optimization of spatial contextual features. The two branches complement each other to form a comprehensive surface cover feature representation, providing accurate surface cover basic data for the subsequent construction of wildfire risk factors.

(1)

Frequency Domain Enhancement Branch (FDEM)

The core goal of the Frequency Domain Enhancement Branch (FDEM) is to extract the periodic textures and edge features of surface covers that are difficult to capture by traditional spatial domain analysis through frequency domain transformation and adaptive enhancement. The specific implementation process is as follows:

Firstly, Fast Fourier Transform ($FFT(·)$) is performed on the input remote sensing image $\mathrm{I}\in {\mathbb{R}}^{\mathrm{H}\times \mathrm{W}\times 3}$ (where H and W represent the height and width of the image, respectively, and 3 denotes the three RGB channels) to map spatial domain information to the frequency domain, realizing the conversion of information dimensions. The transformation formula is expressed as:

$$F(u,v)=FFT(I(x,y))=\sum _{x=0}^{H-1}\sum _{y=0}^{W-1}I(x,y)e^{-j2\pi (ux/H+vy/W)}$$

(1)

where $F(u,v)$ is the transformed frequency domain feature, $(u,v)$ represents the frequency coordinates in the frequency domain, $\mathrm{j}$ is the imaginary unit, and $I(x,y)$ is the pixel value at position $(x,y)$ in the spatial domain. Through this transformation, the periodic texture features of surface covers are converted into specific frequency components in the frequency domain, laying a foundation for subsequent enhancement processing.

To highlight the effective information related to surface cover edges and textures in the frequency domain, an adaptive frequency domain enhancement function is designed:

$$F_{enh}(u,v)=F(u,v)\cdot \mathit{exp}\left({\displaystyle \frac{\Vert (u,v)\Vert -f_0}{\sigma }}\right)$$

(2)

where $f_0$ is a preset low-frequency threshold used to distinguish low-frequency background information from high-frequency detail information; $\sigma$ is an attenuation coefficient used to adjust the enhancement amplitude of high-frequency features; $\Vert (u,v)\Vert$ is the modulus of the frequency coordinate $(u,v)$, used to quantify the frequency level. This function adaptively amplifies high-frequency components through an exponential form while suppressing low-frequency noise, thereby enhancing the representation of edge and texture features of surface covers.

Finally, Inverse Fast Fourier Transform ($IFFT(·)$) is performed on the enhanced frequency domain feature $F_{enh}(u,v)$ to reconstruct the frequency domain information back to the spatial domain, obtaining a spatial domain feature map fused with frequency domain enhancement information and completing the feature extraction process of the frequency domain branch. The expression is:

$$I_{freq}(x,y)=IFFT\left(F_{enh}\right(u,v\left)\right)$$

(3)

The reconstructed feature map $I_{freq}(x,y)$ not only retains the spatial structure information of the original image but also integrates the texture and edge details after frequency domain enhancement, providing high-quality frequency domain feature support for dual-branch feature fusion.

(2)

Cross-Modal Fusion Module (CMFM)

The spatial feature branch adopts MobileNetV3 as the backbone network, fully leveraging its advantages of lightweight design and efficient feature extraction to realize the extraction and fusion of multi-scale spatial contextual features. Through progressive downsampling operations, the backbone network extracts 5 scales of feature maps (from $C_1$ (shallow features) to $C_5$ (deep features)) from the input image. Among them, shallow features retain rich spatial detail information, while deep features contain more abstract semantic category information. To achieve deep collaboration and cross-modal interaction of features at different scales, a Cross-Modal Fusion Module (CMFM) is introduced, which adaptively assigns weights to features of each scale through an attention mechanism to realize optimized feature fusion.

The core of cross-modal fusion is dynamically adjusting the contribution of shallow and deep features through attention weights. The specific fusion formula is:

$$F_{cm}\left(l\right)={\alpha }_l\cdot C_l+(1-{\alpha }_l)\cdot \mathcal{U}\left(C_{l+1}\right)$$

(4)

where $F_{cm}\left(l\right)$ denotes the fused feature of the $l-th$ layer; ${\alpha }_l$ is the attention weight coefficient, which is obtained by performing Global Average Pooling (AvgPool) on the $l-th$ layer feature map $C_l$, inputting the result into a Multi-Layer Perceptron (MLP), and then activating it through a Sigmoid function, i.e.,

$${\alpha }_l=\sigma \left(\mathcal{M}\mathcal{L}\mathcal{P}\right(AvgPool\left(C_l\right)\left)\right)$$

(5)

Its value range is [0, 1], which is used to adaptively balance the importance of the current layer feature and the upper deep layer feature;

$U(\cdot )$ represents the upsampling operation, which upsamples the deep feature of the $\left(l+1\right)-th$ layer to the same size as the $l-th$ layer feature to ensure feature dimension matching. Through this fusion mechanism, the spatial feature branch can effectively integrate multi-scale semantic information and spatial details, improving the ability to distinguish surface cover categories.

(3)

Small-Target Enhancement Module (STEM)

Considering the characteristics of surface covers in the power transmission corridor region, such as fragmentation, small scale, and scattered distribution, traditional segmentation networks have low recognition accuracy for such fine-grained targets. To address this issue, a dedicated Small-Target Enhancement Module (STEM) is designed to achieve accurate segmentation of small-target surface covers through multi-scale feature aggregation and edge enhancement. The core idea of this module is to fully utilize feature information of different scales to strengthen the feature response and edge contours of small-target surface covers. The specific implementation process is as follows:

Firstly, multi-scale feature aggregation is performed. The middle and high-level fused features

$F_{cm}\left(3\right)$ output by the spatial branch, $F_{cm}\left(4\right)$ after one upsampling, and $F_{cm}\left(5\right)$ after two upsamplings are concatenated. Then, 3 × 3 convolution operation is used to unify feature dimensions and fuse information:

$$F_{multi}=C\left(\left[F_{cm}\right(3),U(F_{cm}\left(4\right)),U^2(F_{cm}\left(5\right)\left)\right]\right)$$

(6)

where $\left(\right[\cdot \left]\right)$ denotes the feature concatenation operation, which stacks features of different scales along the channel dimension; $C(\cdot )$ represents the 3 × 3 convolution operation, used to integrate the concatenated multi-scale features, suppress redundant information, and strengthen effective features; $U^2(\cdot )$ denotes the double upsampling operation to ensure that all features participating in aggregation have the same size.

To further enhance the edge contours of small-target surface covers and improve the accuracy of segmentation boundaries, the Laplacian second-order derivative operator is introduced for edge enhancement of the aggregated multi-scale features:

$$F_{edge}=F_{multi}+\lambda \cdot {\mathsf{\nabla }}^2{F}_{multi}$$

(7)

where ${\mathsf{\nabla }}^2$ is the Laplacian operator, used to extract edge information in the feature map; $\lambda =0.15$ is the edge enhancement coefficient, determined through multiple experimental verifications, which is used to balance the contribution of original features and edge-enhanced features and avoid noise amplification caused by excessive enhancement.

Finally, the edge-enhanced feature map $F_{edge}$ is input into a Softmax classifier to obtain the probability distribution of each pixel belonging to different surface cover types, completing the surface cover segmentation process. The probability calculation formula is:

$$P(c|x,y)={\displaystyle \frac{\mathit{exp}(F_{edge}(c,x,y))}{\sum _{k=1}^6\mathit{exp}(F_{edge}(k,x,y))}}$$

(8)

where $P\left(c\right|x,y)$ denotes the probability that the pixel at spatial position $(x,y)$ belongs to the c-th surface cover type; $\mathrm{c}$ is the surface cover type index, ranging from 1 to 6, corresponding to six surface cover types: forest land, cropland, shrubland, water body, building, and bare land, respectively; $F_{edge}(c,x,y)$ denotes the feature response value of the c-th surface cover type at position $(x,y)$ in the feature map. The Softmax function converts feature responses into normalized probability distributions, and the category with the highest probability is finally selected as the surface cover classification result of the pixel.

2.3.2. Multi-Modal Wildfire Point Prediction Network

To achieve accurate prediction of regional wildfire occurrence risks, a wildfire point prediction network combining multi-source feature collaborative modeling and frequency domain enhancement is designed (Figure 8). Through key links such as independent modal encoding, cross-modal attention fusion, frequency domain feature enhancement, and residual refinement, the network fully explores the inherent correlations of features from various modalities and critical information related to wildfire occurrence, and finally outputs the wildfire occurrence probability and risk level.

(1)

Single-Modal Feature Encoding

Due to the significant differences in distribution laws and expression forms among different modal features such as terrain, climate, surface cover, and human activities, the network adopts independent Transformer Blocks to separately encode various types of features to fully model the long-range dependencies of each modal feature and extract unique information of each modality. The features participating in encoding include four categories:

• Surface cover features $X_f\in R^{N\times 6}$ (where $N$ is the number of grid units in the study area, 6 is the number of surface cover types, and each dimension represents the area proportion of the corresponding surface cover type);
• Terrain features $X_t\in R^{N\times 3}$ (including three indicators: elevation, average slope, and slope standard deviation);
• Climate features $X_c\in R^{N\times 3}$ (composed of land surface temperature, precipitation, and evapotranspiration);
• Human activity features $X_h\in R^{N\times 2}$ (road distance and nighttime light index).

The core of the Transformer Block is the self-attention mechanism, which models long-range dependencies by calculating the correlation degree between different positions within the features. The self-attention calculation process is as follows:

$$Attn(Q,K,V)=Softmax\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V$$

(9)

where $Q$ (query matrix), $K$ (key matrix), and $V$ (value matrix) are all obtained by linear transformation of the input modal features; $d_k$ is the dimension of $Q$ and $K$, and dividing by $\sqrt{d_k}$ is used to alleviate the problem of excessively large values caused by dimension growth; the Softmax function is used to normalize the attention weights so that the sum of the weights is 1, ensuring the rationality of feature fusion.

After processing by the self-attention mechanism and feed-forward neural network, each modal feature is fully encoded, and the encoded feature $F_m$ is output. The expression is:

$$F_m=TransformerBlock\left(X_m\right)\hspace{1em}(m\in \{f,t,c,h\left\}\right)$$

(10)

where $\mathrm{m}$ represents different modalities, and $f$, $t$, $c$, $h$ correspond to surface cover, terrain, climate, and human activity features respectively. The encoded feature $F_m$ not only retains the key information of the original modality but also strengthens the internal correlation structure of the features, laying a foundation for subsequent cross-modal fusion.

(2)

Cross-Modal Attention Fusion (CMAF)

To achieve deep interaction and collaborative optimization of multi-source modal features, and improve the consistency and discriminability of key information, a Cross-Modal Attention Fusion (CMAF) module is designed to hierarchically fuse various encoded modal features. The core idea of this module is to adaptively assign fusion weights to each modality by calculating the similarity between different modal features and the reference modal feature, so that modalities contributing more to wildfire prediction obtain higher weights, thereby strengthening effective information and suppressing redundant information.

The mathematical expression of multi-modal feature fusion is:

$$F_{fusion}=\sum _m{\beta }_m\cdot {\mathrm{F}}_{\mathrm{m}}$$

(11)

where $F_{fusion}$ is the integrated feature after fusion; ${\beta }_m$ is the fusion weight of the m-th modal feature, calculated as:

$${\beta }_m={\displaystyle \frac{\mathit{exp}(S(F_m,F_{ref}))}{\sum _k\mathit{exp}(S(F_k,F_{ref}))}}$$

(12)

In the formula, $S(\cdot ,\cdot )$ denotes the cosine similarity calculation function, used to measure the similarity between the current modal feature $F_m$ and the reference modal feature $F_{ref}$.

$F_{ref}$ selects the surface cover feature $F_f$ (since surface cover type is the core foundational factor affecting wildfire occurrence); the denominator is the exponential sum of similarities between all modal features and the reference modal feature, and the weight coefficient ${\beta }_m$ is obtained through Softmax function normalization, ensuring the sum of weights is 1 and realizing adaptive balanced fusion of each modal feature.

(3)

Frequency Domain Enhancement Module

To highlight the periodic features (e.g., climate periodicity caused by seasonal changes, temporal periodicity of human activities) and abrupt features (e.g., extreme temperature changes, sudden shifts in the intensity of human activities) related to fire occurrence, a Frequency Domain Enhancement Module is introduced based on the fused features. This module implements feature mapping to the frequency domain and the extraction of key frequency components using the Fast Fourier Transform (FFT).

First, the Fast Fourier Transform is performed on the fused comprehensive feature, fusion, to map the spatial domain features to the frequency domain, resulting in the frequency domain feature ($F_{\widehat{\omega }}$). The expression is:

$$F_{\widehat{\omega }}=FFT\left(F_{fusion}\right)$$

(13)

The frequency domain feature ($F_{\widehat{\omega }}$) consists of two parts: the amplitude spectrum and the phase spectrum. The amplitude spectrum reflects the intensity of different frequency components, while the phase spectrum reflects the positional information of these components. To enhance the key frequency components relevant to fire occurrence, the amplitude spectrum is subjected to Gaussian window function weighting:

$$M\left(\omega \right)=\left|F_{\widehat{\omega }}\right|\cdot \Gamma \left(\omega \right)$$

(14)

where $\left|F_{\widehat{\omega }}\right|$ is the amplitude spectrum of the frequency domain feature $F_{\widehat{\omega }}$; $\Gamma \left(\omega \right)$ is the Gaussian window function with a center frequency ${\omega }_0$ = 0.3 and a bandwidth $\sigma \omega$ = 0.1, which are determined through multiple experimental optimizations. This function is used to filter the frequency range associated with fire occurrence and suppress irrelevant frequency noise.

Subsequently, the weighted amplitude spectrum is recombined with the original phase spectrum $\angle F_{\widehat{\omega }}$. The frequency domain features are then reconstructed back to the spatial domain using the Inverse Fast Fourier Transform (IFFT), yielding the frequency-domain enhanced feature $F_{freq\_enh}$. The expression is:

$$F_{freq\_enh}=IFFT\left(M\right(\omega )\cdot e^{j\angle F_{\widehat{\omega }}})$$

(15)

His process not only preserves the spatial structural information of the original features but also strengthens the fire-related key features through frequency-domain weighting, thereby improving the discriminative power of the features.

(4)

Residual Refinement and Risk Prediction

To further optimize the quality of frequency-domain enhanced features, suppress noise interference, and ensure stable feature propagation, a Residual Block is introduced for feature refinement. The Residual Block employs a structure of “2-layer convolution + Batch Normalization (BN)”, avoiding the vanishing gradient problem in deep networks through residual connections while enhancing the nonlinear expression capability of features. The feature refinement process is as follows:

$$F_{res}=F_{freq\_enh}+R\left(F_{freq\_enh}\right)$$

(16)

where $R(\cdot )$ denotes the mapping function of the Residual Block, performing convolution and normalization on the frequency-domain enhanced features; residual connections achieve stable feature propagation and gradient backflow by adding the original input features to the processed features, improving the network’s training effect and feature expression capability.

The refined features $F_{res}$ are input to the classification prediction head, processed through two fully connected layers and activation functions, and finally output the fire occurrence probability. The mathematical expression is:

$$P\left(fire\right|x,y)=\sigma (W_2\cdot ReLU(W_1\cdot F_{res})+b)$$

(17)

where $P\left(fire\right|x,y)$ represents the probability of fire occurrence at grid cell $(x,y)$, with a value range of [0, 1]; $W_1$ and $W_2$ are the weight matrices of the two fully connected layers, respectively; $b$ is the bias term; the ReLU function introduces nonlinear transformation to enhance the network’s expression capability; $\sigma$ is the Sigmoid activation function, mapping the output of the fully connected layers to the [0, 1] interval to obtain normalized probability values.

Based on the fire occurrence probability, the Natural Breaks (Jenks) method is used to classify regional fire risk into three levels: Low Risk (p < 0.2), Medium Risk (0.2 ≤ p < 0.5), and High Risk (p ≥ 0.5), achieving quantitative grading and spatial expression of fire risk.

2.4. Evaluation Metrics

To comprehensively evaluate the classification performance of the proposed wildfire risk prediction model, this study adopted a standardized evaluation index system (see

Table 1. Wildfire risk prediction model assessment indicators.

Indicator Category	Name	Formula
Overall Discriminative Ability	ROC-AUC	$AUC={\int }_0^{1}TPR\left(FPR\right)dFPR$
	$\mathrm{Accuracy}(ACC$)	$ACC={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$
Classification decision performance	$\mathrm{Precision}(Pre$)	$Pre={\displaystyle \frac{TP}{TP+FP}}$
	Recall	$Recall={\displaystyle \frac{TP}{TP+FN}}$
	$F1$	$F1=2\times {\displaystyle \frac{Pre\times Recall}{Pre+Recall}}$

). All indices were calculated based on four basic parameters of the classification results: true positives (TP, fires correctly predicted by the model), false positives (FP, fires falsely reported by the model), true negatives (TN, fires not predicted by the model), and false negatives (FN, fires missed by the model).

Overall discriminative power metrics measure the model’s comprehensive potential to distinguish between “fire” and “non-fire” samples, independent of specific classification thresholds. The area under the ROC curve (AUC) is obtained by analyzing the region below the curve formed by the integral true positive rate (TPR) and false positive rate (FPR), reflecting the model’s overall ranking ability across all possible thresholds and serving as a core metric for evaluating the discriminative power of binary classification models. Accuracy directly provides the proportion of correctly predicted samples out of the total sample count. Threshold-based classification decision metrics evaluate the model’s performance at specific decision points after determining the optimal classification threshold. Precision focuses on the reliability of the model’s warnings, that is, the proportion of actual fires among all predicted fires; high precision implies a low false positive rate. Recall focuses on the model’s ability to capture real fires, that is, the proportion of actual fires successfully predicted; high recall implies a low false negative rate. The F1 score, as the harmonic mean of precision and recall, is used to comprehensively evaluate the balance the model achieves between controlling false positives and false negatives.

3. Experimental Analysis

3.1. Semantic Segmentation Results

3.1.1. Comparison Experiment of Different Models

Under the same 512 × 512 input resolution, identical data augmentation and training hyper-parameters,

Table 2. Comparative Experiment on Semantic Segmentation.

Comparative Experiment	Pre	Recall	ACC	mIoU
UNet	0.2934	0.2634	0.8432	0.3214
DaUNet	0.3234	0.3125	0.8763	0.3675
TransuNet	0.3012	0.2873	0.8534	0.3451
BuildFormer	0.3501	0.3357	0.8912	0.3873
MACU-net	0.3572	0.3421	0.8963	0.3954
Our	0.3678	0.3497	0.9091	0.4068

compares BuildFormer++ with five representative networks. U-Net achieves 32.14% mIoU and 26.34% Recall, significantly lower than the others; DA-Unet raises mIoU to 36.75% through dual attention but still below 37%; TransUNet embeds Transformer encoders yet mIoU drops to 34.51%, indicating that global tokens are not directly suitable for landslide edges; BuildFormer attains 38.73% mIoU via a CNN-Transformer hybrid, first exceeding 38%; MACU-net reaches 39.54% mIoU by multi-scale aggregation, becoming the strongest competitor before ours; BuildFormer++ obtains 36.78% Precision, 34.97% Recall, 90.91% Acc and 40.68% mIoU, outperforming MACU-net by 1.06, 0.76, 1.28 and 1.14 percentage points, with the most complete edges and the fewest false alarms in visual comparisons, confirming the superiority of the overall architecture.

To visually demonstrate the segmentation performance of each model, Figure 9 shows a comparison of the semantic segmentation results of different models in a typical transmission corridor region.

3.1.2. Ablation Experiment

Table 3. Ablation Study on Semantic Segmentation Modules.

FDEM	CMFM	STEM	Pre	Recall	ACC	mIoU
✔			0.3461	0.3241	0.8798	0.3708
	✔		0.3125	0.3071	0.8761	0.3691
		✔	0.3045	0.2943	0.8641	0.3641
✔	✔		0.3341	0.3314	0.8801	0.3787
✔		✔	0.3514	0.3453	0.8943	0.3883
	✔	✔	0.3452	0.3351	0.8853	0.3841
✔	✔	✔	0.3678	0.3497	0.9091	0.4068

presents seven ablations on semantic segmentation modules. FDEM alone yields 37.08% mIoU and 34.61% Precision, higher than any single module; individual CMFM and STEM achieve 36.91% and 36.41% mIoU respectively, with stable Recall, showing that channel–spatial complementarity suppresses false positives; among pairwise combinations FDEM + STEM reaches 38.83% mIoU, the best dual-module result; when all three modules are activated mIoU rises to 40.68%, Acc exceeds 90% for the first time, and Precision and Recall simultaneously increase to 36.78% and 34.97%, demonstrating collaborative gain without evident over-fitting.

Ablation experiments demonstrate that the three proposed modules (FDEM, CMFM, and STEM) significantly improve semantic segmentation performance. FDEM exhibits the best single-module performance when used alone, highlighting the advantages of frequency domain information in capturing texture and edge features. CMFM, along with the small target enhancement module and STEM, further optimize the model’s ability to recognize small targets and complex structures through multi-scale fusion and detail enhancement.

In the module combination experiments, the three modules achieved optimal performance (mIoU = 40.68%, accuracy exceeding 90%) when working together, with simultaneous improvements in precision and recall. This indicates a clear complementary and synergistic effect among the modules, effectively enhancing the model’s comprehensive segmentation capability in complex land cover scenarios and laying a reliable foundation for subsequent wildfire risk prediction.

3.2. Wildfire Prediction Results

3.2.1. Analysis of Prediction Performance of Different Models

To comprehensively evaluate the performance of the multimodal wildfire prediction model proposed in this study, we systematically compared it with four classic machine learning methods (random forest, support vector machine, CART, and Naive Bayes) and two mainstream deep learning models (one-dimensional ResNet and one-dimensional visual transformer). The performance metrics are shown in

Table 4. Comparative Experiment on Wildfire Risk Prediction.

Comparative Experiment	ACC	Pre	Recall	F1	ROC
RF	0.782	0.812	0.895	0.851	0.804
SVM	0.765	0.801	0.874	0.835	0.792
CART	0.742	0.776	0.852	0.812	0.765
NB	0.721	0.76	0.835	0.796	0.751
1d resnet	0.8024	0.8395	0.9254	0.8872	0.821
1d vit	0.8152	0.8621	0.9354	0.9154	0.8421
Our	0.8537	0.8972	0.9531	0.9315	0.8542

.

As shown in Table 4, the proposed BuildFormer++ model achieves the best performance across all evaluation metrics. Compared to traditional machine learning methods, its accuracy is improved by more than 5 percentage points; compared to current mainstream deep learning methods, it also shows significant advantages in comprehensive metrics such as F1 score and ROC-AUC. In particular, its excellent performance in recall (95.31%) and F1 score (93.15%) indicates that the model effectively controls the false positive rate while capturing as many real fire points as possible, demonstrating strong practicality and reliability. This further validates the advancement and effectiveness of the deep learning framework integrating multi-source features and a frequency domain-attention dual-branch structure in wildfire risk prediction tasks.

3.2.2. Ablation Experiment

Table 5. Ablation Study on Prediction Model Modules.

CMAF	FFT	ACC	Pre	Recall	F1	ROC
✔		0.8432	0.8735	0.9364	0.9254	0.8456
	✔	0.8324	0.8642	0.9236	0.9164	0.8367
✔	✔	0.8537	0.8972	0.9531	0.9315	0.8542

ablates the prediction-stage modules. CMAF alone gives 84.32% Acc and 92.54% F1; FFT alone gives 83.24% Acc and 91.64% F1; combining both lifts Acc to 85.37% and F1 to 93.15%, with ROC-AUC 85.42%, improving single modules by 2.05, 1.61 and 0.86 percentage points, verifying that parallel frequency–attention branches capture complementary landslide/non-landslide cues.

Table 6. Ablation Study on Data Sources.

Segmentation Prediction Result Data	DEM Data	Anthropogenic Factor	ACC	Pre	Recall	F1	ROC
✔			0.7983	0.8401	0.9198	0.8785	0.8123
	✔		0.7652	0.8045	0.8746	0.8357	0.7928
		✔	0.7827	0.8124	0.8954	0.8543	0.8048
✔	✔		0.8154	0.8567	0.9289	0.9057	0.8245
✔		✔	0.8235	0.8678	0.9354	0.9125	0.8354
	✔	✔	0.8021	0.8435	0.9243	0.8845	0.8165
✔	✔	✔	0.8537	0.8972	0.9531	0.9315	0.8542

reports data-source ablations. Using segmentation-prediction probability alone yields 79.83% Acc; DEM alone 76.52%; human factors alone 78.27%; any two-source combination surpasses single source, among which “segmentation + human factors” reaches 82.35% Acc; fusing all three sources achieves 85.37% Acc, 93.15% F1 and 85.42% ROC-AUC, 3.02 and 1.90 percentage points higher than the best dual-source setting, indicating that multi-source heterogeneous information significantly reduces false alarms and improves landslide sensitivity.

Collectively, the comparative and ablation studies reveal that BuildFormer++ pushes semantic segmentation mIoU to 40.68% via edge–channel–spatial synergy, lifts prediction F1 to 93.15% through frequency–attention dual branches, and attains 85.37% Acc by fusing segmentation probability, DEM and human factors, doubly validating its advancement and robustness for mountainous landslide identification from both model and data perspectives.

3.3. Fire Risk Assessment of Power Transmission Corridor

3.3.1. Fire Risk Threshold Classification and Statistical Analysis

First, based on GIS tools(version 10.8), the fire occurrence probability output by the wildfire prediction model was thresholded to obtain a graded vector surface of fire risk areas (Figure 10). The results show that the numerical range covers 0.124 to 1, divided into three risk levels: low (p < 0.2), medium (0.2 ≤ p < 0.5), and high (p ≥ 0.5).

Further area statistics show that in the entire Xiaogan City area, extremely low-risk areas account for 21.2%, medium-risk areas account for 46.07%, while high-risk areas account for a high 32.73%. This indicates a large-scale distribution of medium- to high-risk fire areas in the study area, requiring close monitoring.

To further verify the rationality of the model’s predictions, this paper performs spatial overlay analysis on 93 fire point samples within a 1 km power transmission corridor and the fire area classification results. Statistical results show that the overlap rate of fire points located in high-risk areas is 62.37%, while the overlap rate of fire points located in medium-risk and above-risk areas is 100%.

The results show that the risk classification predicted by the model not only reasonably reflects the overall risk pattern of the region, but also highly matches the spatial distribution of historical fire points, demonstrating high reliability and practicality. This finding further proves the effectiveness of the logistic regression model in power grid wildfire early warning, and can provide a scientific reference for disaster prevention and mitigation in transmission corridors and line inspection [28,29,30].

3.3.2. Wildfire Risk Analysis of Power Transmission Corridor

Based on the prediction results of the wildfire prediction model, this paper presents a spatial visualization map of wildfire risk within the Xiaogan city area (see Figure 11a). The color gradient in the figure reflects the level of fire risk index, with red areas corresponding to higher risk and green and blue areas representing relatively lower risk levels. From an overall perspective, high-risk areas are mainly concentrated in the southern part of the study area and some hilly areas, exhibiting significant spatial clustering. This distribution characteristic is closely related to the complex terrain, high vegetation cover, and frequent human activities within the region. In contrast, the risk level in the northern and central plain areas is generally lower, indicating that the land cover type and climate environment in this area, to some extent, suppress the possibility of fires. Overall, the spatial distribution results generated by the model effectively reflect the spatial differences in wildfire risk in the Xiaogan area, providing an intuitive reference for power grid operation and maintenance departments to identify key prevention and control areas.

Based on the regional risk distribution, this paper further overlays the prediction results with the spatial orientation of the Xiaogan I and II transmission lines (see Figure 11b). The results showed that some high-risk areas highly overlapped with the transmission line routes, exhibiting a clear characteristic of “risk concentration near corridors.” Specifically, the risk level of the Xiaogan I line significantly increased near sections A and B, while the Xiaogan II line also showed localized high-risk clusters in sections traversing hills and woodlands. These areas typically feature undulating terrain, a mix of woodlands and crops, and frequent human activities such as proximity to roads, thus increasing the potential risk of fire occurrence and spread.

In summary, the spatial distribution results of wildfire risk not only reveal the regional differences in risk levels within the study area, but also further reflect the potential fire hazards along power transmission lines, providing a scientific basis for power grid operation, maintenance, and risk prevention and control.

4. Discussion

The risk mapping results constructed in this study reveal significant spatial differentiation patterns. High-risk areas are clustered in the southern hilly region and show a clear spatial correlation with the transmission corridor. This pattern is the result of the synergistic effect of multiple disaster-causing factors: from a natural perspective, the continuous forest cover and undulating terrain in this area provide a foundation of combustible materials and topographic dynamics for fire spread; from an anthropogenic perspective, the high road density and strong nighttime light intensity indicate frequent human activity, significantly increasing the probability of potential fire sources. Notably, 62.37% of historical fire points fall within the high-risk areas identified by the model. This quantitative evidence not only verifies the reliability of risk zoning but also reveals the dominant role of the “natural–anthropogenic” dual driving mechanism in regional fire occurrence. From an engineering application perspective, the spatial overlap between high-risk areas and transmission corridors provides a clear target area for differentiated power grid prevention and control. Operation and maintenance departments can optimize the allocation of inspection resources based on risk levels to maximize disaster prevention effectiveness.

From an engineering application perspective, the early warning framework proposed in this study aims to complement, rather than replace, the real-time fire monitoring system. The high-resolution land cover segmentation within the framework is an offline or periodically updated component (e.g., quarterly updates), designed to provide stable baseline geographic data. The trained wildfire risk prediction model, as a lightweight probabilistic network, can complete full-area inference within minutes on conventional computing hardware. This design allows the framework to be seamlessly integrated into the existing power grid inspection and risk assessment workflows. Maintenance personnel can proactively identify high-risk corridors before the peak fire season or during regular assessments, thereby driving a shift from a reactive fire response model to a proactive risk prevention and control model. While this “early warning” (identifying areas of increased risk several days or weeks in advance) model differs from real-time fire alarms, it has significant practical engineering value for preventative planning and resource allocation in power grid maintenance.

While the current research has achieved its expected results, it must be acknowledged that it still has limitations in terms of dynamism and universality. The model’s ability to respond dynamically to fire occurrences needs improvement, particularly in how to integrate real-time meteorological data and vegetation phenological changes to enhance early warning timeliness—a key technical challenge that needs to be overcome. Furthermore, while the model has been sufficiently validated in a single geographical region, its ability to adapt to different climate zones and terrain conditions still requires systematic evaluation through multi-regional case studies. Furthermore, the update cycle and spatial resolution limitations of existing data sources still fall short of the real-time monitoring needs of the power grid. Integrating new monitoring data from drones, the Internet of Things, and other sources will be crucial to improving the system’s practicality.

Looking ahead, in-depth exploration in three directions is particularly important: First, developing a dynamic risk early warning system incorporating time-series modeling capabilities, achieving daily updates of risk levels and short-term early warnings by coupling satellite time series observations, weather forecasts, and vegetation dynamic data; second, constructing a transferable cross-regional early warning framework, enabling models to quickly adapt to different geographical environments through transfer learning techniques such as domain adaptation, forming a widely applicable technical solution; and third, promoting the operational integration of the early warning system with the power grid operation and maintenance platform, developing standardized data interfaces and intelligent decision-making modules, ultimately forming a closed-loop management system of “risk identification–assessment–response,” truly realizing the substantial transformation of scientific research into engineering practice. These explorations will not only improve the existing methodological system but also promote a profound shift in fire prevention in transmission corridors from passive response to proactive prevention.

5. Conclusions

This study focuses on the adjacent areas of Xiaoshi I and Xiaoshi II transmission corridors in Xiaogan City, Hubei Province, and proposes a wildfire early warning framework that integrates an improved deep learning-based land cover segmentation network with a multi-modal wildfire point prediction model. By constructing a technical system of “refined land cover segmentation–multi-dimensional feature fusion–wildfire risk prediction”, this framework realizes the quantitative assessment and spatial visualization of wildfire risks in transmission corridor areas, and provides systematic technical support for the safe operation and maintenance of power grids. The main conclusions are summarized as follows:

(1)

Effectiveness of the improved technical framework: The proposed three-level technical system effectively overcomes the limitations of traditional methods such as low fine-grained segmentation accuracy and insufficient multi-source feature fusion. The improved dual-branch semantic segmentation network (BuildFormer++) integrates the Frequency Domain Enhancement Module (FDEM), Cross-Modal Fusion Module (CMFM), and Small-Target Enhancement Module (STEM), which significantly improves the segmentation accuracy of fragmented and small-scale land cover types in transmission corridor areas. The mIoU of land cover segmentation reaches 40.68%, and the overall accuracy exceeds 90%, providing high-precision basic data for subsequent risk assessment. The multi-modal wildfire point prediction network further fuses land cover, terrain, climate, and human activity features, realizing the deep mining of correlations between multi-source heterogeneous data and wildfire occurrence.

(2)

Superior performance of the model system: Comparative experiments show that BuildFormer++ outperforms representative segmentation networks such as U-Net, DA-Unet, and MACU-net in terms of mIoU, Precision, Recall, and Acc. Specifically, its mIoU is 1.14 percentage points higher than that of the second-ranked MACU-net, and the segmentation edges are more complete with fewer false alarms. In terms of wildfire risk prediction, the proposed multi-modal model achieves 85.37% accuracy, 93.15% F1-score, and 85.42% ROC-AUC, which are significantly better than traditional machine learning methods (Random Forest, SVM, etc.) and deep learning baselines (1D ResNet, 1D Vision Transformer). Ablation experiments confirm that each functional module (FDEM, CMFM, STEM, CMAF, FFT) plays a positive role in performance improvement, and the collaborative effect of multi-modules and multi-source data effectively enhances the model’s robustness and predictive ability.

(3)

Rationality of spatial risk zoning and verification: Based on the wildfire occurrence probability output by the model, the study area is divided into three risk levels (Low Risk, Medium Risk, High Risk) using the Natural Breaks (Jenks) method. Spatial mapping results show that high-risk areas are mainly concentrated in the southern hilly area of the study area, which has a high overlap with some sections of Xiaoshi I and Xiaoshi II transmission corridors, showing obvious “corridor-adjacent risk concentration” characteristics. Verification with historical fire point data (2015–2023) shows that 62.37% of historical fire points are distributed in high-risk areas, and all fire points are located in medium-risk and above areas, which fully verifies the rationality and reliability of the risk zoning results.

(4)

Clear guiding significance for engineering applications: The research results have clear practical value for the operation and maintenance of transmission corridors. High-risk sections identified by the model are mostly distributed in areas with undulating terrain, dense woodland, and frequent human activities. This conclusion can guide power grid operation and maintenance departments to carry out targeted patrols and hidden danger rectification, improve the efficiency of disaster prevention and mitigation, and reduce the economic losses caused by wildfires to power grids. Meanwhile, the framework has good scalability and can be popularized and applied to other transmission corridor areas after appropriate parameter adjustment.