Forest fire area identification method, electronic device and readable storage medium based on deep learning

Abstract

The present invention proposes a forest burn area identification method based on deep learning, comprising the following steps: S1. obtaining a satellite remote sensing image dataset of a forest, and preprocessing the satellite remote sensing image dataset to obtain a sample dataset; S2. constructing an improved Mask2former model; S3. inputting the sample dataset into the improved Mask2former model for training; S4. judging whether the improved Mask2former model has been trained, and if so, entering step S5; if not, adjusting the model parameters and returning to step S3; S5. inputting the image to be identified into the trained improved Mask2former model to obtain a segmentation result; through the above method, the computational complexity can be reduced and the accuracy of identifying forest burn areas and non-burn areas can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image segmentation, and in particular to a forest burnt area identification method based on deep learning, an electronic device, and a readable storage medium. Background Art

[0002] Forest fires are a major threat to forest ecosystems, with devastating impacts on both ecosystems and human society. Understanding post-disaster forest recovery is crucial for developing strategies to protect species diversity and restore ecosystems, and it also helps prevent secondary damage.

[0003] In existing technologies, machine learning models and deep learning models are often used to identify forest burn areas and then observe the recovery of forests after disasters. However, existing models have high computational complexity and low recognition accuracy, and cannot accurately identify burnt areas.

[0004] Therefore, in order to solve the above technical problems, it is urgent to propose a new technical means. Summary of the Invention

[0005] In view of this, the present invention proposes a method for improving the accuracy of identifying forest burnt areas.

[0006] The present invention provides a method for identifying forest burnt areas based on deep learning, comprising the following steps:

[0007] S1. Obtain a satellite remote sensing image dataset of a forest, and preprocess the satellite remote sensing image dataset to obtain a sample dataset;

[0008] S2. Build an improved Mask2former model;

[0009] The improved Mask2former model includes an EASwin-Transformer network, a pixel decoder and a Transformer decoder, wherein the image is input into the EASwin-Transformer network and the Transformer decoder respectively, the output end of the EASwin-Transformer network is connected to the input end of the pixel decoder, and the output end of the pixel decoder is connected to the input end of the Transformer decoder;

[0010] The EASwin-Transformer network includes a Swin-Transformer network and an EfficientlyAdaptive module; the output end of the Swin-Transformer network is connected to the input end of the Efficiently Adaptive module;

[0011] The Efficiently Adaptive module includes a feature grouping module, four average pooling modules, three 1×1 convolution modules, one 3×3 convolution module, two splicing modules, three activation function layers, a group normalization module and two Sigmoid function layers;

[0012] Among them, the output end of the feature grouping module is connected to the input end of the first average pooling module, the second average pooling module and the 3×3 convolution module; the output ends of the first average pooling module and the second average pooling module are both connected to the input end of the first splicing module, the output end of the first splicing module is connected to the input end of the first Sigmoid function layer, the output end of the first Sigmoid function layer is connected to the input end of the group normalization module, the output end of the group normalization module is connected to the input end of the third average pooling module, and the output end of the third average pooling module is connected to the input end of the first activation function layer; the output features of the first 3×3 convolution module are multiplied element-by-element with the output features of the first activation function layer, and the features after element-by-element multiplication are input to the second Sigmoid function layer; the output end of the first 3×3 convolution module is also connected to the first The input ends of the four average pooling modules and the output end of the fourth average pooling module are connected to the input end of the second activation function layer. The output features of the second activation function layer are multiplied element-by-element by the output features of the group normalization module, and the features obtained by element-by-element multiplication are input into the second Sigmoid function layer. The output end of the second Sigmoid function layer is respectively connected to the input ends of the second 1×1 convolution module, the third 1×1 convolution module and the second splicing module. The output end of the second 1×1 convolution module is connected to the input end of the second splicing module, and the output end of the second splicing module is connected to the third activation function layer. The output features of the third activation function layer are multiplied element-by-element by the output features of the third 1×1 convolution module, and the features after element-by-element multiplication are fused with the features output by the second 1×1 convolution module to obtain the output features of the Efficiently Adaptive module.

[0013] S3. Input the sample dataset into the improved Mask2former model for training;

[0014] S4. Determine whether the improved Mask2former model is trained. If so, proceed to step S5; if not, adjust the model parameters and return to step S3;

[0015] S5. Input the image to be recognized into the trained improved Mask2former model to obtain the segmentation result.

[0016] Furthermore, in step S3, the following loss function Loss is used for training:

[0017]

[0018] Among them, i represents the serial number of the category, n represents the total number of categories, j represents the serial number of the sample, b represents the total number of samples, and y ji represents the true label of the j-th sample belonging to category i, represents the predicted probability that the jth sample belongs to category i.

[0019] Furthermore, in step S1, the preprocessing method at least includes image enhancement, noise addition, flipping and noise removal.

[0020] Further, in step S4, when the training of the improved Mask2former model reaches a preset goal, the training of the improved Mask2former model is completed.

[0021] Accordingly, the present invention further provides an electronic device, comprising:

[0022] A memory and a processor, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the above-mentioned forest fire area identification method based on deep learning is implemented.

[0023] Correspondingly, the present invention also provides a readable storage medium, in which computer instructions are stored. When the computer instructions are executed by a processor, the above-mentioned forest fire area identification method based on deep learning is implemented.

[0024] Beneficial effects of the present invention: By improving the Mask2former model, the present invention can reduce the computational complexity while efficiently learning the high-level features of the object of interest, thereby improving segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0026] Figure 1 It is a flow chart of the present invention.

[0027] Figure 2 Schematic diagram of the improved Mask2former model structure of the present invention.

[0028] Figure 3 This is a schematic diagram of the Efficiently Adaptive module structure of the present invention. DETAILED DESCRIPTION

[0029] The present invention will be further described below with reference to the accompanying drawings:

[0030] The present invention provides a method for identifying forest burnt areas based on deep learning, comprising the following steps:

[0031] S1. Obtain a satellite remote sensing image dataset of a forest, and preprocess the satellite remote sensing image dataset to obtain a sample dataset;

[0032] S2. Build an improved Mask2former model;

[0033] The improved Mask2former model includes an EASwin-Transformer network, a pixel decoder, and a Transformer decoder, wherein an image is input into the EASwin-Transformer network and the Transformer decoder, respectively, the output end of the EASwin-Transformer network is connected to the input end of the pixel decoder, and the output end of the pixel decoder is connected to the input end of the Transformer decoder; the output features of the pixel decoder are multiplied element-by-element with the features of the Transformer decoder and then output;

[0034] The EASwin-Transformer network includes a Swin-Transformer network and an EfficientlyAdaptive module; the output end of the Swin-Transformer network is connected to the input end of the Efficiently Adaptive module;

[0035] The Efficiently Adaptive module includes a feature grouping module, four average pooling modules, three 1×1 convolution modules, one 3×3 convolution module, two splicing modules, three activation function layers, a group normalization module and two Sigmoid function layers;

[0036] Among them, the output end of the feature grouping module is connected to the input end of the first average pooling module, the second average pooling module and the 3×3 convolution module; the output ends of the first average pooling module and the second average pooling module are both connected to the input end of the first splicing module, the output end of the first splicing module is connected to the input end of the first Sigmoid function layer, the output end of the first Sigmoid function layer is connected to the input end of the group normalization module, the output end of the group normalization module is connected to the input end of the third average pooling module, and the output end of the third average pooling module is connected to the input end of the first activation function layer; the output features of the first 3×3 convolution module are multiplied element-by-element with the output features of the first activation function layer, and the features after element-by-element multiplication are input to the second Sigmoid function layer; the output end of the first 3×3 convolution module is also connected to the first The input ends of the four average pooling modules and the output end of the fourth average pooling module are connected to the input end of the second activation function layer. The output features of the second activation function layer are multiplied element-by-element by the output features of the group normalization module, and the features obtained by element-by-element multiplication are input into the second Sigmoid function layer. The output end of the second Sigmoid function layer is respectively connected to the input ends of the second 1×1 convolution module, the third 1×1 convolution module and the second splicing module. The output end of the second 1×1 convolution module is connected to the input end of the second splicing module, and the output end of the second splicing module is connected to the third activation function layer. The output features of the third activation function layer are multiplied element-by-element by the output features of the third 1×1 convolution module, and the features after element-by-element multiplication are fused with the features output by the second 1×1 convolution module to obtain the output features of the Efficiently Adaptive module.

[0037] S3. Input the sample dataset into the improved Mask2former model for training;

[0038] S4. Determine whether the improved Mask2former model is trained. If so, proceed to step S5; if not, adjust the model parameters and return to step S3;

[0039] S5. Input the image to be identified into the trained improved Mask2former model to obtain the segmentation result. Through the above method, the forest burn area can be accurately identified and the recognition accuracy can be improved.

[0040] A forest burnt area is an area in a forest that has been burned by fire.

[0041] In this embodiment, in step S1, a satellite remote sensing image dataset of a forest is obtained, and the satellite remote sensing image dataset is preprocessed to obtain a sample dataset;

[0042] Specifically, this embodiment uses a DJI Mavic 3 drone for image acquisition, with an effective pixel count of 20 million, a sensitivity range of 100 to 6400, an image size of 5280×3956, and a storage format of JPEG. The data acquisition area should meet sufficient natural lighting conditions (>15 lux), and the diffuse reflectivity of the materials in the scene should be >20%;

[0043] The preprocessing method includes at least image enhancement, noise addition, flipping and noise removal; it may also include image cropping to obtain more sample data; among which image cropping, image enhancement, noise addition, flipping and noise removal are all existing technologies, and the specific preprocessing methods are not described here.

[0044] Since satellite remote sensing images usually contain multispectral information, more detailed data can be extracted from the images; in addition, the upper limit of the display range of satellite remote sensing images is extremely high, which helps to ignore local interference and conduct global analysis.

[0045] In this embodiment, in step S2, an improved Mask2former model is constructed;

[0046] The improved Mask2former model includes an EASwin-Transformer network, a pixel decoder and a Transformer decoder, wherein the image is input into the EASwin-Transformer network and the Transformer decoder respectively, the output end of the EASwin-Transformer network is connected to the input end of the pixel decoder, and the output end of the pixel decoder is connected to the input end of the Transformer decoder (Efficiently Adaptive Swin-Transformer network), as shown in FIG. Figure 2 As shown, it includes an EASwin-Transformer network, a pixel decoder Pixel Decoder and a Transformer Decoder (Transformer decoder). The output end of the EASwin-Transformer network is connected to the input end of the pixel decoder Pixel Decoder, and the output end of the pixel decoder Pixel Decoder is connected to the input end of the Transformer decoder; the output features of the pixel decoder are multiplied element by element with the output features of the Transformer decoder and output.

[0047] Mask2former is an efficient and flexible segmentation model. Its multi-scale features greatly improve the accuracy of segmentation and is a segmentation method suitable for specific objects. The structure of the Mask2former model, as well as the pixel decoder and transformer decoder are all existing structures and will not be repeated here.

[0048] The EASwin-Transformer network (Efficiently Adaptive Swin-Transformer network) includes a Swin-Transformer network and an Efficiently Adaptive module; the output end of the Swin-Transformer network is connected to the input end of the Efficiently Adaptive module; the output end of the Efficiently Adaptive module is connected to the input end of the pixel decoder;

[0049] Building on the Transformer's superior attention mechanism, Swin-Transformer utilizes hierarchical feature representations to better preserve detailed information while capturing global context, adapting to image features of various scales. Furthermore, the sliding window mechanism in Swin-Transformer applies local window attention to feature maps to reduce computational complexity and enhance the model's ability to capture long-range dependencies. This improves computational efficiency while also enhancing performance when processing high-resolution images. The Swin-Transformer network structure is the existing one and will not be detailed here.

[0050] During the secondary succession process, forests gradually recovering from wildfires face fierce competition between different species, and dominant species are constantly alternating. At the same time, due to the different degrees of fire damage in different areas, they are in different stages of succession, and the situation is complex and changeable. In order to achieve better segmentation effects in such a complex environment, this application introduces the ContextualTransformer module (CoT module). The CoT module integrates contextual information mining and self-attention mechanism, which can better explore contextual information in complex environments and help distinguish different adjacent objects; in order to further enhance the feature extraction ability of the network in classification tasks, this application introduces the idea of ​​Efficient Multi-Scale Attention (MSA) in the CoT module, and finally forms the Efficiently Adaptive module; MSA reshapes some channels to the batch dimension and divides the channel dimension into multiple sub-features, so that the spatial semantic features are well distributed within each feature group. At the same time, MSA recalibrates the channel weights of each parallel branch by encoding global information, thereby more effectively learning the channel content without reducing the channel dimension, and generating better pixel-level attention in high-level feature maps.

[0051] Specifically, the Efficiently Adaptive module includes a feature grouping module, four average pooling modules, three 1×1 convolution modules, one 3×3 convolution module, two splicing modules, three activation function layers, a group normalization module and two Sigmoid function layers; Figure 3 As shown;

[0052] Among them, the output end of the feature grouping module is connected to the input end of the first average pooling module, the second average pooling module and the 3×3 convolution module; the output ends of the first average pooling module and the second average pooling module are both connected to the input end of the first splicing module, the output end of the first splicing module is connected to the input end of the first Sigmoid function layer, the output end of the first Sigmoid function layer is connected to the input end of the group normalization module, the output end of the group normalization module is connected to the input end of the third average pooling module, and the output end of the third average pooling module is connected to the input end of the first activation function layer; the output features of the first 3×3 convolution module are multiplied element-by-element with the output features of the first activation function layer, and the features after element-by-element multiplication are input to the second Sigmoid function layer; the output end of the first 3×3 convolution module is also connected to the first The input ends of the four average pooling modules and the output end of the fourth average pooling module are connected to the input end of the second activation function layer. The output features of the second activation function layer are multiplied element-by-element by the output features of the group normalization module, and the features obtained by element-by-element multiplication are input into the second Sigmoid function layer. The output end of the second Sigmoid function layer is respectively connected to the input ends of the second 1×1 convolution module, the third 1×1 convolution module and the second splicing module. The output end of the second 1×1 convolution module is connected to the input end of the second splicing module, and the output end of the second splicing module is connected to the third activation function layer. The output features of the third activation function layer are multiplied element-by-element by the output features of the third 1×1 convolution module, and the features after element-by-element multiplication are fused with the features output by the second 1×1 convolution module to obtain the output features of the Efficiently Adaptive module.

[0053] The feature grouping module is used to divide the input feature into g sub-features in the channel dimension. Dividing the feature into multiple sub-features in the channel dimension is an existing technology. The average pooling module, convolution module, splicing module, activation function layer (with an activation function), group normalization module (Group norm) and Sigmoid function layer are all existing technologies and will not be described in detail here.

[0054] The Efficiently Adaptive module divides the input feature map into g sub-features in the channel dimension for learning different semantics. Usually, g << c (channels), and the learned attention weight descriptors are used to enhance the feature representation of the regions of interest in each sub-feature. The Efficiently Adaptive module extracts the attention weight descriptors of the grouped feature maps through three parallel routes. Two of the parallel paths are on the 1×1 branches, and the third parallel path is on the 3×3 branch. In the 3×3 branch, only one 3×3 convolutional kernel is used to capture multi-scale feature representations, reducing the computational requirements and enlarging the feature space. The features output from the convolution are sequentially passed through an average pooling module and an activation function layer. In the 1×1 branches, global average pooling operations are respectively adopted along two spatial directions, and the features after global average pooling are concatenated along the height direction of the image and input into a 1×1 convolutional module. After decomposing the output of the 1×1 convolution into two vectors, the nonlinear Sigmoid function is used to fit the two-dimensional binomial distribution after linear convolution. Cross-channel feature interaction is achieved by multiplying the two channels (3×3 branch and 1×1 branch) within each group. The output of the 1×1 branch is encoded using two-dimensional global average pooling and converted into the corresponding dimensional shape. To improve computational efficiency, the nonlinear function Softmax is adopted at the output of the two-dimensional global average pooling to fit the above linear transformation. By multiplying the output of the above parallel processing element-wise with matrix dot product operations, we obtain the first spatial attention map containing spatial position information. The same method is used on the 3×3 branch to obtain the second spatial attention map, and then the two sets of generated spatial attention weight sets are passed through the Sigmoid function to obtain the pixel-level attention map.

[0055] The pixel-level attention map is respectively input into two 1×1 convolutional modules to obtain the input static context representation K1 and the input value V. Then, K1 is merged with the Query of the input pixel-level attention map, and then through the nonlinear function Softmax, the attention matrix A is obtained. The attention matrix A is multiplied element-wise with the input value V to implement the input dynamic context representation K2. The static context representation K1 and the dynamic context representation K2 are added (fused) to output the extracted features.

[0056] The MAS mechanism is applied in the CoT module. While enhancing the capture of context information, the context information is used to better focus on higher-level and deeper features, enabling the network to efficiently learn the high-level features of the attention object while adapting to tasks in complex environments.

[0057] The use of the improved Mask2former model can not only improve the computational efficiency of the model, but also enhance the performance when processing high-resolution images; it can also enhance the ability to capture contextual information and learn key features, thereby improving segmentation accuracy.

[0058] In this embodiment, in step S3, the sample data set is input into the improved Mask2former model for training;

[0059] The learning rate is set to 0.0001, the batch size is 2, and a total of 250 epochs are trained with the following loss function Loss:

[0060]

[0061] Among them, i represents the serial number of the category, n represents the total number of categories, j represents the serial number of the sample, b represents the total number of samples, and y ji represents the true label of the j-th sample belonging to category i, represents the predicted probability that the jth sample belongs to category i.

[0062] By setting the above loss function, we can effectively measure the difference between the model's predicted distribution and the true distribution. By minimizing the cross entropy loss function, we can make the model's predicted probability distribution as close as possible to the probability distribution of the true label.

[0063] In this embodiment, in step S4, it is determined whether the improved Mask2former model is trained. If so, the process proceeds to step S5; if not, the model parameters are adjusted and the process returns to step S3.

[0064] When the improved Mask2former model training reaches a preset target, the improved Mask2former model training is completed. The preset target can be a preset number of training times or a preset degree of convergence, which is not limited here and can be set according to needs. Through the above method, a more accurate prediction model can be obtained.

[0065] In this embodiment, in step S5, the image to be identified is input into the trained improved Mask2former model, which can realize the identification of forests, burned areas and non-forest areas, such as roads, buildings, ponds and construction sites.

[0066] To evaluate the performance of the improved Mask2former model, four commonly used evaluation metrics were used: mAcc (mean Acc), mIou (mean Iou), mDice (mean Dice), and mPrecision (mean Precision). Their values ​​range from 0 to 1. The closer they are to 1, the better the model's predictions. For TP (the true value belongs to a certain class, but the predicted result after segmentation is also that class), TN (the true value does not belong to a certain class, but the predicted result after segmentation is also that class), FP (the true value does not belong to a certain class, but the predicted result after segmentation is that class), and FN (the true value belongs to a certain class, but the predicted result after segmentation is not that class), Acc, Iou, Dice, and Precision are calculated using the following formulas:

[0067]

[0068]

[0069]

[0070]

[0071] Among them, n TP 、n TN 、n FP and n FN represent the number of TP, TN, FP and FN respectively;

[0072] In order to verify the partial effectiveness and versatility of the improved Mask2former model, this application performs segmentation on a total of 19 categories of objects, including "roads", "sidewalks", "buildings", "high walls", "fences", "poles", "traffic lights", "traffic signs", "plants", "terrain", "sky", "pedestrians", "cyclists", "cars", "trucks", "buses", "trains", "motorcycles", and "bicycles" on the open source image segmentation dataset Cityscapes. The segmentation models include: Deeplabv3+ segmentation model with Resnet101 as the backbone network, Segformer segmentation model with MixViT as the backbone network, Mask2former segmentation model with Resnet50 as the backbone network, Mask2former segmentation model with Swin-Transformer as the backbone network, and Mask2former segmentation model with EAswin-Transformer as the backbone network (i.e., the improved Mask2former model in this application), etc., and comparative experiments are conducted on the above segmentation models using similar datasets and training methods. The experimental results are shown in Table 1:

[0073] Segmentation Model backbone network mAcc mIou mDice mPrecision DeepLabV3+ Resnet101 0.7209 0.6361 0.7260 0.7817 Segformer MixViT 0.7689 0.6806 0.7968 0.8433 Mask2former Resnet50 0.7753 0.6732 0.7910 0.8184 Mask2former Swin-Transformer 0.8191 0.7068 0.8213 0.8303 Mask2former EAswin-Transformer 0.8365 0.7281 0.8352 0.8396

[0074] Table 1 Comparative experimental results of the open source dataset Cityscapes

[0075] The above segmentation model was evaluated using the sample dataset of this application, and the experimental results are shown in Table 2:

[0076] Segmentation Model backbone network mAcc mIou mDice mPrecision DeepLabV3+ Resnet101 0.8299 0.7231 0.8373 0.8455 Segformer MixViT 0.8481 0.7460 0.8528 0.8580 Mask2former Resnet50 0.8545 0.7309 0.8421 0.8310 Mask2former Swin-Transformer 0.8576 0.7577 0.8602 0.8633 Mask2former EAswin-Transformer 0.8623 0.7635 0.8642 0.8668

[0077] Table 2 Comparative experimental results of sample datasets

[0078] As can be seen from Tables 1 and 2, the improved Mask2former model in this application performs best in all key indicators, with a high mean intersection over union (mIou). mIou is particularly important in segmentation tasks. A high mIou indicates both high accuracy of the segmentation results and a low misclassification rate of image parts of different categories. It is a key metric for evaluating segmentation model performance, particularly for assessing the consistency between predictions and observations. These results demonstrate that the model in this application can more effectively extract information from burned areas and forest images, improving prediction accuracy.

[0079] To further verify the effectiveness of the Efficiently Adaptive module (EA module), ablation experiments are conducted on the EA module, CoT module, and MAS module respectively. The ablation experiment results are shown in Table 3:

[0080] Segmentation Model CoT MAS mAcc mIou mDice mPrecision Mask2former × × 0.8576 0.7577 0.8602 0.8633 Mask2former √ × 0.8528 0.7599 0.8618 0.8715 Mask2former × √ 0.8607 0.7609 0.8625 0.8659 Mask2former √ √ 0.8623 0.7635 0.8642 0.8668

[0081] Table 3 Ablation experiment results

[0082] Among them, the Mask2former model that includes both the CoT module and the MAS module is the improved Mask2former model in this application; from the ablation experiment results, it can be seen that in order to improve the ability of contextual information to guide dynamic learning and enhance the ability of visual representation, this application adds the CoT module. The results show that there is indeed an improvement in mIou, but although the average precision (mPrecision) reaches the highest among all the improved methods, the average intersection-over-union (mIou) is the lowest. This shows that the improvement of the CoT module sometimes leads to misjudgment, which may be due to insufficient learning of some key, deep features such as texture. The improvement using the MAS module avoids this problem very well. The MAS module improves the ability of information extraction by reshaping some channels to the batch dimension and grouping the channel dimension into multiple sub-features. At the same time, it is more efficient and can avoid boundary blur. The EA module combines the advantages of both, not only improving the ability of information extraction, but also having a stronger learning ability for long-distance data, and effectively avoiding the problem of boundary blur.

[0083] Accordingly, the present invention further provides an electronic device, comprising:

[0084] A memory and a processor, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the above-mentioned forest fire area identification method based on deep learning is implemented.

[0085] Correspondingly, the present invention also provides a readable storage medium, in which computer instructions are stored. When the computer instructions are executed by a processor, the above-mentioned forest fire area identification method based on deep learning is implemented.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A forest burn area identification method based on deep learning, characterized by: The following steps are involved: S1. Obtain a satellite remote sensing image dataset of a forest, and preprocess the satellite remote sensing image dataset to obtain a sample dataset; S2. Build an improved Mask2former model; The improved Mask2former model includes an EASwin-Transformer network, a pixel decoder and a Transformer decoder, wherein the image is input into the EASwin-Transformer network and the Transformer decoder respectively, the output end of the EASwin-Transformer network is connected to the input end of the pixel decoder, and the output end of the pixel decoder is connected to the input end of the Transformer decoder; The EASwin-Transformer network includes a Swin-Transformer network and an Efficiently Adaptive module; the output end of the Swin-Transformer network is connected to the input end of the Efficiently Adaptive module; The Efficiently Adaptive module includes a feature grouping module, four average pooling modules, three 1×1 convolution modules, one 3×3 convolution module, two splicing modules, three activation function layers, a group normalization module and two Sigmoid function layers; Among them, the output end of the feature grouping module is connected to the input end of the first average pooling module, the second average pooling module and the 3×3 convolution module; the output ends of the first average pooling module and the second average pooling module are both connected to the input end of the first splicing module, the output end of the first splicing module is connected to the input end of the first Sigmoid function layer, the output end of the first Sigmoid function layer is connected to the input end of the group normalization module, the output end of the group normalization module is connected to the input end of the third average pooling module, and the output end of the third average pooling module is connected to the input end of the first activation function layer; the output features of the first 3×3 convolution module are multiplied element-by-element with the output features of the first activation function layer, and the features after element-by-element multiplication are input to the second Sigmoid function layer; the output end of the first 3×3 convolution module is also connected to the first The input ends of the four average pooling modules and the output end of the fourth average pooling module are connected to the input end of the second activation function layer. The output features of the second activation function layer are multiplied element-by-element by the output features of the group normalization module, and the features obtained by element-by-element multiplication are input into the second Sigmoid function layer. The output end of the second Sigmoid function layer is respectively connected to the input ends of the second 1×1 convolution module, the third 1×1 convolution module and the second splicing module. The output end of the second 1×1 convolution module is connected to the input end of the second splicing module, and the output end of the second splicing module is connected to the third activation function layer. The output features of the third activation function layer are multiplied element-by-element by the output features of the third 1×1 convolution module, and the features after element-by-element multiplication are fused with the features output by the second 1×1 convolution module to obtain the output features of the Efficiently Adaptive module. S3. Input the sample dataset into the improved Mask2former model for training; S4. Determine whether the improved Mask2former model is trained. If so, proceed to step S5; if not, adjust the model parameters and return to step S3; S5. Input the image to be recognized into the trained improved Mask2former model to obtain the segmentation result.

2. The forest burn area identification method based on deep learning according to claim 1 is characterized by: In step S3, the following loss function Loss is used for training: Among them, i represents the serial number of the category, n represents the total number of categories, j represents the serial number of the sample, b represents the total number of samples, and y ji represents the true label of the j-th sample belonging to category i, represents the predicted probability that the jth sample belongs to category i.

3. The forest burn area identification method based on deep learning according to claim 1 is characterized by: In step S1, the preprocessing method at least includes image enhancement, noise addition, flipping and noise removal.

4. The method for identifying forest burned areas based on deep learning according to claim 1, characterized in that: In step S4, when the improved Mask2former model training reaches a preset goal, the improved Mask2former model training is completed.

5. An electronic device, characterized in that: include: A memory and a processor, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the method for identifying forest burned areas based on deep learning according to any one of claims 1 to 4 is implemented.

6. A readable storage medium, characterized in that: The readable storage medium stores computer instructions, which, when executed by a processor, implement the forest burn area identification method based on deep learning as described in any one of claims 1 to 4.

Images