Forest burned area intelligent identification method fusing YOLO and DeepLabv3 +

Abstract

The invention relates to the technical field of deep learning and computer vision, and discloses a YOLO and DeepLabv3 + fused forest burned area intelligent identification method, which comprises the following steps: collecting remote sensing image data of various temporal and spatial resolutions covering a research area, preprocessing the collected data, unifying all the data to the same spatial reference system and pixel size, and identifying the area of the research area. Forming a time-space consistent multi-modal remote sensing data set; on the basis of existing global overfire area products and historical wildfire records, sample labels of overfire areas and non-overfire areas are sketched and generated on the multi-source remote sensing images; a two-stage model cascade architecture is adopted, and an optimal parameter set is searched to minimize an overall loss function; and the trained integrated deep learning model and the multi-source remote sensing image are utilized to realize rapid identification of the forest fire land. According to the method, the accuracy, the robustness and the cross-regional adaptability of forest fire land identification are improved, and the problems of misjudgment and missed judgment caused by data isomerism and environment complexity in an existing method are solved.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and computer vision technology, specifically to an intelligent identification method for forest fire tracks that integrates YOLO and DeepLabv3+. Background Technology

[0002] Wildfires are among the most influential natural disturbances in global terrestrial ecosystems, widely occurring in various ecosystems including forests, grasslands, wetlands, and shrublands (Pausas & Keeley, 2021). Their sudden onset, high intensity, and wide-ranging nature not only directly leads to vegetation destruction, soil structure alteration, and wildlife habitat loss, but also profoundly impacts the health of Earth's ecosystems by altering carbon cycle processes, increasing greenhouse gas emissions, and weakening ecosystem stability (Tedimetal., 2018; Lietal., 2023; Wu et al., 2016; Bai et al., 2020). With the continued intensification of global warming, the frequency of extreme heat and drought events has significantly increased. Coupled with population growth and irrational land use, this has led to a significant increase in the frequency of wildfires in many regions worldwide, with fire intensity and duration also showing an increasing trend (Mansoore et al., 2022; Richardsone et al., 2022; Cunninghametal., 2024). According to the Global Forest Resources Assessment Report and remote sensing data, since 2000, more than 46 million hectares of land have been burned annually worldwide due to wildfires (Luoe et al., 2024). Wildfires have become a significant driver of global ecological and environmental change and a potential source of risk.

[0003] Currently, wildfire identification methods based on remote sensing technology can be mainly categorized into three types: threshold detection methods, time series analysis / change detection techniques, and machine learning and artificial intelligence methods. ① Threshold detection methods: This type of method was the earliest remote sensing identification technique used for wildfire monitoring. It mainly utilizes the brightness temperature thresholds in the mid-infrared and thermal infrared bands to determine the presence of fire points (Ding et al., 2023). Its advantages lie in its simple algorithm, high operating efficiency, and suitability for large-scale preliminary screening; however, it is highly dependent on the threshold setting, relies on prior knowledge, and has a certain degree of subjectivity. In addition, due to significant differences in surface background in different regions, the fire point determination threshold exhibits spatiotemporal instability, and its ability to identify low-temperature, small-scale fire sources is limited, leading to frequent misjudgments and missed detections. ② Time Series Analysis and Change Detection Techniques: Compared to thresholding methods, these methods analyze the dynamic changes in the time series of remote sensing vegetation indices (such as NDVI and NBR), effectively eliminating the influence of seasonality and data noise, and improving the accuracy and universality of fire disturbance identification (Schroeder et al., 2011; Li et al., 2022; Koutsia et al., 2022). For example, Empirical Mode Decomposition (EMD) can separate the trend and seasonal terms from the NDVI time series (Jie et al., 2024), highlighting the fire disturbance signal, and has been successfully applied to forest and grassland fire extraction; multi-time-segment fitting algorithms such as LandTrendr capture abrupt change nodes in the time series trajectory to achieve rapid identification of burned areas and tracking of post-fire recovery processes (DeFerrarie et al., 2024); difference methods based on indices such as NDVI and NBR (such as ΔNDVI and dNBR) directly quantify the degree of change by comparing before and after images, and are suitable for detecting sudden fires (Escuine et al., 2008). This method possesses good versatility and anti-interference capabilities, and has become a research hotspot in remote sensing wildfire identification in recent years. ③ Machine learning and artificial intelligence technologies: In remote sensing image classification and feature recognition, this technology demonstrates strong adaptability and robustness. Supervised classification algorithms such as Random Forest (RF) and Support Vector Machine (SVM) have been widely applied to fire spot detection and burned area identification, exhibiting high classification accuracy and generalization ability (Hultquiste et al., 2014). In recent years, deep learning methods have also been rapidly used in wildfire monitoring research. Architectures such as Convolutional Neural Networks (CNN), Autoencoders (AE), and Graph Neural Networks (GNN) have shown good performance in automatic identification of burned areas, boundary extraction, and hierarchical reconstruction processes (Choie et al., 2024; Bouguettaya et al., 2022). These methods can extract high-order features from remote sensing images, improving the identification accuracy of small-scale fires and complex backgrounds. However, their effectiveness depends on a large number of high-quality training samples and computational resources, and they face problems such as weak model interpretability and weak cross-regional transferability. Summary of the Invention

[0004] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent forest fire area identification method integrating YOLO and DeepLabv3+. This method effectively solves the technical problems of existing fire remote sensing identification methods, such as subjective threshold settings, low sensitivity to low temperatures and small-scale fire sources, insufficient model generalization ability, and inaccurate target boundary extraction in complex backgrounds. It achieves high-precision and high-efficiency identification of large-scale wildfire forest fire areas. Specifically, this invention integrates multi-source remote sensing data and adopts an integrated strategy of "detection-leading segmentation and refinement." First, it uses the YOLO series target detection models to quickly locate suspected fire areas. Then, it uses DeepLabv3+ to extract pixel-level burned patches, thereby improving the accuracy, robustness, and cross-regional adaptability of forest fire area identification. This overcomes the problems of misjudgment and missed detection caused by data heterogeneity and environmental complexity in existing methods.

[0005] (II) Technical Solution The basic idea of ​​this invention is to address the technical challenges in large-scale wildfire remote sensing identification, such as subjective threshold setting, insensitivity to small-scale fire sources, weak model generalization ability, and inaccurate forest fire boundary extraction. It proposes an intelligent forest fire area identification method that integrates YOLO and DeepLabv3+. The core of this method lies in combining the advantages of medium-resolution and high-resolution remote sensing data sources. Through a cascaded deep learning model architecture, it achieves coarse-to-fine identification of forest fire areas and extraction of burned area, effectively improving identification accuracy, robustness, and automation.

[0006] To address the shortcomings of existing technologies that typically employ a single algorithm to process a single data source, making it difficult to balance recognition efficiency and accuracy, and resulting in high false positive and false negative rates in complex geographical contexts, this invention introduces an integrated strategy combining object detection and semantic segmentation. First, an object detection model is used to rapidly scan wide-format remote sensing imagery, locating candidate bounding boxes for suspected fire points or forest fire areas, effectively narrowing down the scope for fine-grained processing. Then, a semantic segmentation model is applied within the candidate bounding box regions for pixel-level classification, accurately delineating the boundaries of forest fire areas. This strategy ensures both computational efficiency for large-scale processing and high-precision mapping of burned areas, and includes the following steps: Step 1: Multi-source remote sensing data collection and processing Remote sensing image data with multiple spatiotemporal resolutions covering the study area were collected, the collected data were preprocessed, and all data were unified to the same spatial reference frame and pixel size to form a spatiotemporally consistent multimodal remote sensing dataset. Step 2: Create a sample dataset Based on existing global burned area products and historical wildfire records, sample labels for forest-burned and non-forest-burned areas are generated on multi-source remote sensing images. Step 3: Construct an ensemble model of YOLO and DeepLabv3+ deep learning. A two-level cascaded model architecture is used to find the optimal parameter set. To minimize the overall loss function The process; Step 4: Train the deep learning ensemble model Using a trained ensemble deep learning model, batch automated processing of multi-source remote sensing images is performed. Step 5: Extraction of forest fire information.

[0007] Preferably, in step 1, the remote sensing image data includes, but is not limited to, MODIS, Landsat series, and Sentinel-2, and the collected data is preprocessed.

[0008] Preferably, step 3 includes: S3.1. Rapidly locate fire points and forest fire areas based on YOLO series models; S3.2 Pixel-level forest fire extraction based on DeepLabv3+ model; S3.3, Model integration and optimization.

[0009] Preferably, the specific implementation method of S3.1 is as follows: The YOLO object detection model was used as the first-level model. The model will preprocess the multimodal remote sensing dataset. Divided into Grid, prediction for each grid There are several bounding boxes. For each bounding box, the network outputs a quintuple. ,in: It is the offset of the bounding box center relative to its grid cells. It is the logarithmic transformation of the width and height of the bounding box relative to the width and height of the prior anchor box. It is the object confidence score, and the actual coordinates of the bounding box. Decode using the following formula: ; ; ; ; in, It is the sigmoid function. These are grid coordinates. The anchor frame dimensions are a priori; Loss function at this stage for: ; in, These are the bounding box coordinate loss, confidence loss, and classification loss, respectively. To balance the weights.

[0010] Preferably, S3.2 is specifically implemented as follows: Each candidate box region output by the YOLO model in S3.1 is cropped from the original high-resolution image and used as input to the second-level segmentation refinement model. This stage adopts the DeepLabv3+ semantic segmentation architecture to achieve pixel-level accurate extraction of the boundaries of forest fire areas. The core operations of a DeepLabv3+ model can be formalized as follows: Let the high-level feature map output by the encoder be... The feature map is fed into the ASPP module, which uses multiple dilated convolutional layers with different dilation rates, a global average pooling layer, and a 1×1 convolutional layer in parallel. The output feature maps of these parallel branches are convolved and upsampled to the same size, and then concatenated to obtain multi-scale fused features. The process is represented as follows: ; in, Indicates the kernel size as Expansion rate The dilated convolution operation, Indicates global average pooling. Indicates an upsampling operation. This indicates a splicing operation along the channel dimension.

[0011] Preferably, in step S3.2, the obtained multi-scale fusion features This is then further fused with low-level features from the encoder network, which are first subjected to channel dimensionality reduction via a 1×1 convolution, and then fused with the upsampled features. The features are then concatenated, and the concatenated features are further refined and restored to the spatial dimensions of the original input candidate regions through a series of 3×3 convolutions and a final upsampling. The final output is a probability map of each pixel belonging to the "overheated" category. ; In this stage, the Dice loss function is preferred as the optimization objective. The Dice loss, based on the Dice coefficient, directly optimizes the overlap of the foreground region, and its definition is as follows: ; in, It is the total number of pixels within a candidate region. It is a pixel The true labels are 1 for "over the top" and 0 for "not over the top". The pixels predicted by the model The probability of falling into the "overdone" category. It is a smoothing term designed to prevent the denominator from being zero and to ensure numerical stability.

[0012] Preferably, S3.3 is specifically implemented as follows: The entire ensemble model can be viewed as a composite function. ,in For input images, To integrate all the parameters to be optimized in the model, its final output is a pixel-level fire probability map of the input image. The training objective of the model is to find the optimal parameter set. This makes it possible to use the training set Overall loss To minimize this overall loss, we have a weighted sum of the detection loss and the segmentation loss: ; ; in, The detection loss of the YOLO model, The segmentation loss for the DeepLabv3+ model. It is a hyperparameter used to balance the contributions of the two losses.

[0013] Preferably, in S3.3, an extrusion excitation module, hereinafter referred to as the SE module, is embedded in the encoder backbone network of DeepLabv3+; The intermediate feature map The SE module operates as follows: First, a "squeezing" operation is performed using global average pooling to compress the global spatial information into a single channel descriptor, generating a channel statistical vector. , its first Each element is calculated as follows: ; in, For the first The global average pooling result of each channel is used for subsequent "stimulation" operations. and These represent the height and width of the feature map, respectively, that is, the dimensions of the feature map in the vertical and horizontal directions. For the first Each channel is located in eigenvalues; The non-linear dependencies between channels are captured using a gating mechanism, and the weight of each channel is output: ; in, and The weights of the fully connected layer, To reduce the ratio, It is the ReLU activation function. The activation function is Sigmoid. Finally, the resulting channel weight vector... Used for processing the original feature map Perform recalibration, also known as "reweighting": ; in, It is a new feature map optimized by the attention mechanism. For the first Attention weights for each channel, The first one after attention reweighting through the channel New feature maps for each channel.

[0014] Preferably, step 4 includes: For each scene of input remote sensing image The trained ensemble model The final output is a probability map with the same size as the input space. Each pixel value This indicates the predicted probability that the location belongs to the "overheated" category; Set a judgment threshold ( The binary segmentation map is generated according to the following rules. : ; in, The pixels were identified as overexposed pixels; Preferably, in step 4, based on the binary segmentation map This allows for precise calculation of the burned area. First, the burned area of ​​each patch is counted. Total number of pixels contained Then, based on the spatial resolution of the input image ,in Calculate the actual physical area of ​​the patch, given the actual surface area represented by a single pixel. : ; For the entire extraction area, the total burned area That is, all within this area The total area of ​​each burned patch: ; Compared with existing technologies, this invention provides a method for intelligent identification of forest fire-affected areas that integrates YOLO and DeepLabv3+, which has the following beneficial effects: 1. This invention innovatively employs an integrated strategy of "detection-leadership segmentation refinement," effectively addressing the issue of insufficient recognition accuracy of single algorithms in complex backgrounds. Traditional thresholding methods have limited ability to identify low-temperature, small-scale fire sources, and threshold setting is highly subjective. This invention uses the YOLO model for rapid region nomination, followed by pixel-level classification within the nominated region using DeepLabv3+. This process combines the efficiency of target detection with the accuracy of semantic segmentation, significantly improving the accuracy of fire zone boundary extraction and the sensitivity of identifying small-scale fires.

[0015] 2. This invention optimizes processing efficiency and computational resource consumption through the collaboration of multi-source remote sensing data and the cascaded architecture of a two-level model. Traditional methods directly perform full-domain fine segmentation on massive high-resolution images, requiring enormous computational resources. This invention utilizes the YOLO model to quickly screen candidate regions, greatly reducing the amount of data that subsequent fine segmentation models need to process. This makes it possible to complete large-scale, long-term forest fire identification with limited computational resources, significantly saving network bandwidth and computation time.

[0016] 3. This invention boasts a high level of automation and strong generalization ability, significantly reducing the method's reliance on prior human knowledge and enhancing its universality and stability. Existing methods often require parameter adjustments for different regions and seasons. This invention, based on a deep learning model, automatically learns spectral, texture, and contextual features from global multi-sample data. It can adapt to different ecosystems, terrains, and climatic conditions, effectively overcoming the spatiotemporal instability of parameters in traditional methods. This provides a fully automated and highly reliable technical solution for producing large-scale wildfire dynamic monitoring products. Attached Figure Description

[0017] Figure 1 This is a technical roadmap for an embodiment of the present invention; Figure 2 This is a flowchart illustrating the global wildfire burned area extraction process based on multimodal remote sensing data using integrated deep learning. Figure 3 This is a diagram illustrating the steps of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-3 A method for intelligent identification of forest fire-affected areas integrating YOLO and DeepLabv3+ includes the following steps: Step 1: Multi-source remote sensing data collection and processing Remote sensing imagery data at various spatiotemporal resolutions covering the study area were collected, including but not limited to MODIS, Landsat series, and Sentinel-2. This multi-source data was preprocessed, and all data were unified to the same spatial reference frame and pixel size to form a spatiotemporally consistent multimodal remote sensing dataset.

[0020] MODIS is used to provide products for thermal anomalies and fire area over a wide range and at high frequency; the Landsat series is used to provide multispectral data with a resolution of 30 meters; and the entinel-2 is used to provide multispectral data with a resolution of 10 meters. Step 2: Create a sample dataset Based on existing global burned area products and historical wildfire records, sample labels for burned and unburned forest areas are manually delineated or semi-automatically generated on multi-source remote sensing imagery. To address differences in data scale, high-resolution imagery is downsampled or medium-resolution imagery is upsampled to construct training sample sets suitable for different model input scales. The sample sets need to cover different ecosystems, seasons, terrains, and other scenarios to ensure the model's generalization ability.

[0021] Step 3: Construct an ensemble model of YOLO and DeepLabv3+ deep learning. This step employs a two-stage cascaded model architecture, and its training process involves finding the optimal parameter set. To minimize the overall loss function The process.

[0022] S3.1 Detection Pilot Phase: Rapid localization of fire points and forest fire areas based on YOLO series models. The YOLO object detection model was used as the first-level model. The model will use preprocessed remote sensing images. Divided into Grid, prediction for each grid There are 10 bounding boxes. For each bounding box, the network outputs a 5-tuple. ,in: It is the offset of the bounding box center relative to its grid cells. It is the logarithmic transformation of the width and height of the bounding box relative to the width and height of the prior anchor box. This is the object confidence score, representing the probability that the bounding box contains the target. The actual coordinates of the bounding box. Decode using the following formula: ; ; ; ; in, It is the sigmoid function. These are grid coordinates. The prior dimensions are the anchor frame dimensions.

[0023] Loss function at this stage for: ; in, These are the bounding box coordinate loss, confidence loss, and classification loss, respectively. To balance the weights.

[0024] S3.2, Segmentation Refinement Stage: Pixel-level forest fire extraction based on the DeepLabv3+ model. Each candidate bounding box region output by the YOLO model in step 3.1 is cropped from the original high-resolution image and used as input to the second-level segmentation refinement model. This stage employs the DeepLabv3+ semantic segmentation architecture to achieve pixel-level accurate extraction of forest fire area boundaries.

[0025] DeepLabv3+ models effectively capture multi-scale contextual information and recover spatial details through their unique encoder-decoder structure and hollow spatial pyramid pooling (ASPP) module. Their core operations can be formalized as follows: Let the high-level feature map output by the encoder be... The feature map is fed into the ASPP module. The ASPP module uses multiple dilated convolutional layers with different dilation rates, a global average pooling layer, and a 1×1 convolutional layer in parallel to capture contextual information under different receptive fields. The output feature maps of these parallel branches are convolved and upsampled to the same size before channel concatenation, a process described as follows: ; in, Indicates the kernel size as Expansion rate The dilated convolution operation, Indicates global average pooling. Indicates an upsampling operation. This indicates a splicing operation along the channel dimension.

[0026] The obtained multi-scale fusion features This is then fused again with low-level features from the encoder network. These low-level features are first reduced in channel dimension through a 1×1 convolution, and then fused with the upsampled features. The features are then concatenated. The concatenated features are then subjected to a series of 3×3 convolutions and a final upsampling to gradually refine and restore the spatial dimensions of the original input candidate regions. The final output is a probability map showing that each pixel belongs to the "overheated" category. .

[0027] The training objective of the model in this stage is to minimize the difference between the predicted probability map and the true label. Since there is often a severe class imbalance between burned and unburned areas, the Dice loss function is preferred as the optimization objective in this stage. The Dice loss, based on the Dice coefficient, directly optimizes the overlap of the foreground region (burned pixels), and its definition is as follows: ; in, It is the total number of pixels within a candidate region. It is a pixel The true labels are 1 for "over the top" and 0 for "not over the top". The pixels predicted by the model The probability of falling into the "overdone" category. It is a smoothing term designed to prevent the denominator from being zero and to ensure numerical stability.

[0028] By minimizing the Dice loss, the model is driven to accurately predict the shape and location of forest fire areas, which is particularly beneficial for improving the identification performance of small forest fire areas or areas with irregular boundaries.

[0029] S3.3 Model Integration and Optimization The goal of this phase is to improve the end-to-end performance of the entire integrated system by using a joint optimization strategy to achieve the global optimum of the parameters of the two-level models (object detection and semantic segmentation).

[0030] The entire ensemble model can be viewed as a composite function. ,in For input images, This integrates all the parameters to be optimized in the model. Its final output is a pixel-level fire probability map of the input image. The training objective of the model is to find the optimal parameter set. This makes it possible to use the training set Overall loss Minimize. The overall loss is a weighted sum of the detection loss and the segmentation loss: ; ; in, The detection loss of the YOLO model, The segmentation loss for the DeepLabv3+ model. It is a hyperparameter used to balance the contributions of the two losses, and is usually set using a validation set before training.

[0031] In model design and training, we introduce an attention mechanism to enhance the model's ability to focus on fire-related features and suppress interference from irrelevant backgrounds. Specifically, a squeeze-and-excitation (SE) module is embedded in the encoder backbone network of DeepLabv3+. This is applied to the intermediate feature maps output by the encoder. The SE module operates as follows: First, a "squeezing" operation is performed using global average pooling to compress the global spatial information into a single channel descriptor, generating a channel statistical vector. , its first Each element is calculated as follows: ; in, For the first The global average pooling result of each channel is used for subsequent "stimulation" operations. and These represent the height and width of the feature map, respectively, that is, the dimensions of the feature map in the vertical and horizontal directions. For the first Each channel is located in eigenvalues.

[0032] Subsequently, an "excitation" operation is performed, using a simple gating mechanism (usually composed of two fully connected layers) to capture the non-linear dependencies between channels and output the weights of each channel: ; in, and The weights of the fully connected layer, To reduce the ratio, It is the ReLU activation function. The activation function is Sigmoid. Finally, the resulting channel weight vector... Used for processing the original feature map Perform recalibration, also known as "reweighting": ; in, It is a new feature map optimized by the attention mechanism. For the first Attention weights for each channel, The first one after attention reweighting through the channel This process creates new feature maps for each channel. This process enables the model to adaptively emphasize channels associated with fire smoke and fire spectral features, thereby improving the discriminative power of features in complex scenarios.

[0033] Step 4: Training the Deep Learning Ensemble Model. This step utilizes the trained "detection-leader-segmentation refinement" ensemble deep learning model to perform automated batch processing of historical archived or near real-time global multi-source remote sensing imagery. This model is defined in the preceding steps. The loss function is optimized and an SE attention mechanism is integrated, which can efficiently and accurately output pixel-level forest fire area identification results.

[0034] For each scene of input remote sensing image The trained ensemble model The final output is a probability map with the same size as the input space. Each pixel value This indicates the predicted probability that the location belongs to the "overheated" category.

[0035] To obtain the final binarized map of forest fire distribution, a decision threshold is set. ( The binary segmentation map is generated according to the following rules. : ; in, The pixels that are contiguous in space are identified as burned pixels. All spatially contiguous burned pixels constitute a single burned patch.

[0036] Based on binary segmentation map This allows for accurate calculation of the burned area. First, the burned area of ​​each patch is statistically analyzed. Total number of pixels contained Then, based on the spatial resolution of the input image... ,in Calculate the actual physical area of ​​the patch, given the actual surface area represented by a single pixel. : ; For the entire extraction area, the total burned area That is, all within this area The total area of ​​each burned patch: ; Through this process, the present invention achieves fully automated generation from raw remote sensing imagery to objective, quantified burned area products. This method overcomes the subjectivity of traditional thresholding methods, provides more accurate boundary information than a single model, and offers reliable key foundational data for large-scale regional wildfire carbon emission assessment, disaster loss statistics, and ecosystem restoration research. Step 5: Extraction of forest fire information.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of forest fire-affected areas integrating YOLO and DeepLabv3+, characterized in that, Includes the following steps: Step 1: Multi-source remote sensing data collection and processing Remote sensing image data with multiple spatiotemporal resolutions covering the study area were collected, the collected data were preprocessed, and all data were unified to the same spatial reference frame and pixel size to form a spatiotemporally consistent multimodal remote sensing dataset. Step 2: Create a sample dataset Based on existing global burned area products and historical wildfire records, sample labels for forest-burned and non-forest-burned areas are generated on multi-source remote sensing images. Step 3: Construct an ensemble model of YOLO and DeepLabv3+ deep learning. A two-level cascaded model architecture is used to find the optimal parameter set. To minimize the overall loss function The process; Step 4: Train the deep learning ensemble model Using a trained ensemble deep learning model, batch automated processing of multi-source remote sensing images is performed. Step 5: Extraction of forest fire information.

2. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 1, characterized in that, In step 1, the remote sensing image data includes, but is not limited to, MODIS, Landsat series, and Sentinel-2, and the collected data is preprocessed.

3. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 1, characterized in that, Step 3 includes: S3.

1. Rapidly locate fire points and forest fire areas based on YOLO series models; S3.2 Pixel-level forest fire extraction based on DeepLabv3+ model; S3.3, Model integration and optimization.

4. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 3, characterized in that, The specific implementation method of S3.1 is as follows: The YOLO object detection model was used as the first-level model. The model will preprocess the multimodal remote sensing dataset. Divided into Grid, prediction for each grid There are several bounding boxes. For each bounding box, the network outputs a quintuple. ,in: It is the offset of the bounding box center relative to its grid cells. It is the logarithmic transformation of the width and height of the bounding box relative to the width and height of the prior anchor box. It is the object confidence score, and the actual coordinates of the bounding box. Decode using the following formula: ； ； ； ； in, It is the sigmoid function. These are grid coordinates. The anchor frame dimensions are a priori; Loss function at this stage for: ； in, These are the bounding box coordinate loss, confidence loss, and classification loss, respectively. To balance the weights.

5. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 3, characterized in that, The specific implementation method of S3.2 is as follows: Each candidate box region output by the YOLO model in S3.1 is cropped from the original high-resolution image and used as input to the second-level segmentation refinement model. This stage adopts the DeepLabv3+ semantic segmentation architecture to achieve pixel-level accurate extraction of the boundaries of forest fire areas. The core operations of a DeepLabv3+ model can be formalized as follows: Let the high-level feature map output by the encoder be... The feature map is fed into the ASPP module, which uses multiple dilated convolutional layers with different dilation rates, a global average pooling layer, and a 1×1 convolutional layer in parallel. The output feature maps of these parallel branches are convolved and upsampled to the same size, and then concatenated to obtain multi-scale fused features. The process is represented as follows: ； in, Indicates the kernel size as Expansion rate The dilated convolution operation, Indicates global average pooling. Indicates an upsampling operation. This indicates a splicing operation along the channel dimension.

6. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 5, characterized in that, In step S3.2, the obtained multi-scale fusion features This is then further fused with low-level features from the encoder network, which are first subjected to channel dimensionality reduction via a 1×1 convolution, and then fused with the upsampled features. The features are then concatenated, and the concatenated features are further refined and restored to the spatial dimensions of the original input candidate regions through a series of 3×3 convolutions and a final upsampling. The final output is a probability map of each pixel belonging to the "overheated" category. ; In this stage, the Dice loss function is preferred as the optimization objective. The Dice loss, based on the Dice coefficient, directly optimizes the overlap of the foreground region, and its definition is as follows: ； in, It is the total number of pixels within a candidate region. It is a pixel The true labels are 1 for "over the top" and 0 for "not over the top". The pixels predicted by the model The probability of it falling into the "overdone" category. It is a smoothing term designed to prevent the denominator from being zero and to ensure numerical stability.

7. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 3, characterized in that, The specific implementation method of S3.3 is as follows: The entire ensemble model can be viewed as a composite function. ,in For input images, To integrate all the parameters to be optimized in the model, its final output is a pixel-level fire probability map of the input image. The training objective of the model is to find the optimal parameter set. This makes it possible to use the training set Overall loss To minimize this overall loss, we have a weighted sum of the detection loss and the segmentation loss: ； ； in, The detection loss of the YOLO model, The segmentation loss for the DeepLabv3+ model. It is a hyperparameter used to balance the contributions of the two losses.

8. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 7, characterized in that, In S3.3, an extrusion excitation module, hereinafter referred to as the SE module, is embedded in the encoder backbone network of DeepLabv3+. The intermediate feature map The SE module operates as follows: First, a "squeezing" operation is performed using global average pooling to compress the global spatial information into a single channel descriptor, generating a channel statistical vector. , its first Each element is calculated as follows: ； in, For the first The global average pooling result of each channel is used for subsequent "stimulation" operations. and These represent the height and width of the feature map, respectively, that is, the dimensions of the feature map in the vertical and horizontal directions. For the first Each channel is located in eigenvalues; The non-linear dependencies between channels are captured using a gating mechanism, and the weight of each channel is output: ； in, and The weights of the fully connected layer, To reduce the ratio, It is the ReLU activation function. The activation function is Sigmoid. Finally, the resulting channel weight vector... Used for processing the original feature map Perform recalibration, also known as "reweighting": ； in, It is a new feature map optimized by the attention mechanism. For the first Attention weights for each channel, The first one after attention reweighting through the channel New feature maps for each channel.

9. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 1, characterized in that, Step 4 includes: For each scene of input remote sensing image The trained ensemble model The final output is a probability map with the same size as the input space. Each pixel value This indicates the predicted probability that the location belongs to the "overheated" category; Set a judgment threshold ( The binary segmentation map is generated according to the following rules. : ； in, The pixels were identified as overexposed.

10. The intelligent identification method for forest fire scars integrating YOLO and DeepLabv3+ according to claim 9, characterized in that, In step 4, based on the binary segmentation map This allows for precise calculation of the burned area. First, the burned area of ​​each patch is counted. Total number of pixels contained Then, based on the spatial resolution of the input image ,in Calculate the actual physical area of ​​the patch, given the actual surface area represented by a single pixel. : ； For the entire extraction area, the total burned area That is, all within this area The total area of ​​each burned patch: 。

Images