Straw burning area smoke identification method based on deep learning algorithm

Abstract

The invention discloses a straw incineration area smoke recognition method based on a deep learning algorithm, and relates to the technical field of intelligent recognition and environment monitoring cross, and the scheme uses a programming language python and a deep learning framework pytorch to construct a straw incineration area smoke recognition model. The method comprises the following steps: constructing sample data by using Jilin No.1 high-resolution remote sensing image data and Sentinel-2A multispectral remote sensing data, pre-processing the sample data, dividing the obtained remote sensing image data, training a deep learning model by using a training sample, and calculating the accuracy of the deep learning model according to a loss function convergence condition and a precision improvement condition of the verification sample. Selecting optimal parameters for the whole network to obtain an optimal model; and carrying out precision evaluation on a prediction result by adopting four indexes including the precision rate, the recall rate, the average precision and the F1 score, and analyzing the overall performance of the model. Experiments prove that the method shows excellent performance on multi-source remote sensing satellite images and provides key technical support for environment remote sensing monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of intelligent identification and environmental monitoring, and specifically to a method for identifying smoke in straw burning areas based on a deep learning algorithm. Background Art

[0002] Straw burning is a significant cause of regional air pollution and ecological damage. The combustion process releases suspended particulate matter and toxic gases, which not only aggravates the formation of haze, but also directly threatens human respiratory health and causes public health problems such as chronic obstructive pulmonary disease.

[0003] Traditional straw burning monitoring technology mainly relies on flame detection and ground sensor networks. Although flame detection is intuitive, its signal lags behind the initial stage of combustion, making it difficult to meet early warning needs; sensors based on temperature and gas composition are susceptible to environmental interference and have defects such as low sensitivity and high false alarm rate in open spaces. In recent years, with breakthroughs in remote sensing technology, satellite imagery has gradually become an important means of straw burning monitoring due to its wide-area coverage, multispectral perception and periodic observation capabilities. However, the smoke signals obtained by satellite platforms have unique complexity: smoke exhibits low-contrast characteristics in the visible light band, its morphology is significantly affected by atmospheric turbulence and surface background interference, and it often appears as a small target in high-resolution images. These characteristics pose severe challenges to traditional remote sensing interpretation methods.

[0004] To address these issues, deep learning technology, with its end-to-end feature learning capabilities, has provided new insights for smoke detection. Single-stage object detection models, such as the YOLO series, are widely used in remote sensing due to their high detection speed. However, the sporadic temporal nature of straw burning events leads to a scarcity of labeled samples, further limiting the generalization performance of the models. Previous studies have attempted to alleviate this problem through data augmentation or transfer learning, but these strategies are mostly designed for natural scenes and fail to optimize for the domain differences between smoke and complex surface backgrounds, resulting in significantly reduced recall rates in real-world scenarios.

[0005] In response to the above problems, a smoke recognition method for straw burning areas based on deep learning algorithm is proposed to solve the above problems. Summary of the Invention

[0006] In view of this, the technical problem to be solved by the present invention is to propose a method for identifying smoke in straw burning areas based on a deep learning algorithm. This scheme proposes a deep learning model based on an improved YOLOv12, and realizes the accurate identification of straw burning smoke in Jilin-1 high-resolution satellite images through triple optimization of parameter fine-tuning, transfer learning and attention mechanism; this technical path can not only break through the performance bottleneck of existing algorithms, but also provide a new methodological reference for the deep intersection of remote sensing and artificial intelligence.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying smoke in a straw burning area based on a deep learning algorithm, comprising:

[0008] S1, data acquisition: Use multi-source remote sensing image data to construct sample data, and process the image data, including image cropping, image enhancement, positive and negative sample balance and image expansion, to obtain a sample dataset of smoke images in straw burning areas;

[0009] S2, data preprocessing: divide the remote sensing image dataset collected in S1 into training set, validation set and test set;

[0010] S3, model construction: Use the programming language Python and the deep learning framework PyTorch to build a smoke depth recognition model for straw burning areas;

[0011] S4, model training and prediction: Use training samples to train the deep learning model. Perform hyperparameter optimization before training, select the appropriate gradient descent optimizer, loss function, and iterator. At the same time, based on the convergence of the loss function and the accuracy improvement of the validation samples, select the optimal parameters for the entire network to obtain the optimal model.

[0012] S5, Accuracy evaluation and result analysis: Four indicators including precision, recall, average precision, and F1-score are used to evaluate the accuracy of the prediction results and analyze the overall performance of the model.

[0013] Preferably, the image cropping, image enhancement and image expansion described in S1 are written in the programming language Python, and the image calibration is to annotate Labelimg with the help of the target detection tool; wherein the image cropping, image enhancement and expansion and positive and negative sample balancing specifically include:

[0014] Image cropping: Remote sensing image areas containing straw burning were selected to create dataset samples. Each image was cut into 640 × 640 pixel sample slices using a sliding cut method with a 30% overlap rate.

[0015] Image enhancement and expansion: Data enhancement is used to expand samples, performing horizontal flipping, vertical flipping, and diagonal mirroring on each sample to increase the number of samples to 4-6 times the original number;

[0016] Positive and negative sample balance: select pixels containing smoke and assign them a value of 1, and pixels containing non-smoke and assign them a value of 0. Use the histogram to remove samples in which 0-valued pixels account for more than 80% of the total number of single-sample pixels and samples in which 1-valued pixels account for 100% of the total number of single-sample pixels.

[0017] Preferably, the deep recognition model described in S3 adopts the YOLO v12 network as the basic framework and introduces a simple and efficient area attention module (Area Attention, A2), which reduces the computational complexity of attention while maintaining a large receptive field. By introducing the residual efficient layer aggregation network (R-ELAN), it overcomes the optimization problems caused by the introduction of the attention mechanism, thereby improving the calculation speed.

[0018] Preferably, S4 specifically includes:

[0019] S4.1, before training: Use stochastic gradient descent (SGD) or its variant optimization algorithm to train the smoke recognition model, setting the appropriate learning rate, batch size and number of training rounds;

[0020] S4.2, during training: use the cross entropy loss function to calculate the error between the model's prediction results and the true label, and update the model parameters through the backpropagation algorithm;

[0021] S4.3, Post-training: Use the validation set to evaluate the model during training and adjust the model’s hyperparameters based on the validation results to prevent overfitting.

[0022] Preferably, the hyperparameter settings described in S4.3 include setting the batch size (batch size) to 8, selecting the learning rate (lr) to be 0.001, the confidence to be (confidence) 0.001, the threshold value to be 0.5-0.75, when the training epoch (Epoch) is 100 times, the loss function tends to be stable and the model converges, and the stochastic gradient descent (SGD) optimizer is used to optimize the parameter training process; during model training, the ratio of the training set to the test set is 8:2, and the transfer learning strategy is introduced.

[0023] Preferably, the precision described in S5 is evaluated by the average precision AP (average precision) value for the accuracy of the multi-category detection task; the loss function (Loss) value is a measure of whether the model has converged; the F1 score (F1 Score), the recall rate, and the precision rate are used as verification indicators. The specific formula is as follows:

[0024]

[0025] Where TP is the positive sample predicted as positive by the model, FP is the negative sample predicted as positive by the model, and FN is the positive sample predicted as negative by the model.

[0026] Compared with the existing technology, the present invention provides a method for identifying smoke in straw burning areas based on a deep learning algorithm, which has the following beneficial effects:

[0027] Compared with the existing technology, the present invention makes full use of deep learning and image recognition technology, and uses the advantages of deep learning models such as fast speed, high efficiency and automation to effectively identify smoke in straw burning areas.

[0028] (1) This paper uses the YOLO v12 deep learning network structure and introduces transfer learning and attention mechanism technology to achieve efficient recognition of smoke in remote sensing image data; the YOLO v12 deep learning network proposes a simple and efficient regional attention module, which reduces the computational complexity of attention in the simplest and most direct way while maintaining a large receptive field, thereby improving the computational speed. At the same time, a residual efficient layer aggregation network is introduced to solve the optimization problems brought by attention.

[0029] (2) On this basis, the present invention considers that the migration algorithm can play a better role, and adds an attention mechanism again after the two effective feature layers extracted by the backbone network and upsampling. The backbone feature extraction network contains multiple convolutional layers and residual connections. The present invention adds the first attention on the A2C2f module and the Concat layer; adds the second attention at the Conv and A2C2f modules; the Concat layer is used to fuse feature maps from different sources, and adding the third attention at this position can further highlight the key features. The three-time attention mechanism model can focus on the key information in the image more accurately, thereby improving the prediction accuracy and performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a schematic diagram of the steps of the straw burning area smoke identification method based on the deep learning algorithm of the present invention;

[0031] Figure 2 This is a schematic diagram of the smoke identification process in a straw burning area of ​​the present invention;

[0032] Figure 3 This is a schematic diagram of the network structure of the improved YOLO v12 deep learning model of the present invention;

[0033] Figure 4 This is a display diagram of some sample slices in the sample data set constructed by the present invention;

[0034] Figure 5 Schematic diagram of straw burning smoke image after image enhancement and processing according to the present invention;

[0035] in Figure 5 (a) is the original image; (b) is brightness transformation; (c) is grayscale transformation; (d) is random flipping; (e) is adding Gaussian noise; (f) is random color inversion;

[0036] Figure 6 This is a schematic diagram of the boundary extraction result of the straw burning area according to the present invention;

[0037] Figure 7 It is a schematic diagram of the smoke recognition effect of the present invention in a straw burning area. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] The present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0040] Example 1, please refer to Figures 1 to 7 As shown:

[0041] To solve the problems mentioned in the technical solution, the embodiment of the present application provides a method for identifying smoke in straw burning areas based on a deep learning algorithm. First, the Yolov12 model is constructed using the programming language Python and the deep learning framework PyTorch. This model is an attention-centered target detection architecture. By innovating the attention mechanism and the overall network architecture, it achieves state-of-the-art object detection accuracy while maintaining real-time performance. The model introduces a new self-attention method, which divides the feature map into l equal-sized regions horizontally or vertically, thereby avoiding complex operations and maintaining a large effective receptive field. In addition, based on an improved feature aggregation module of the Efficient Layer Aggregation Network (ELAN), a block-level residual connection with scaling function is introduced to solve the optimization problem.

[0042] Among them, in the construction of the model, a 7×7 separable convolution is added to the attention mechanism to implicitly encode the position information. At the same time, an efficient attention mechanism (FlashAttention) is used to minimize memory access overhead, making the model simpler and faster. In order to further enhance the feature extraction capability of the model, attention is added at the connection between the regional attention and cross-stage feature fusion module (A2C2f) and the connection layer (Concat), and at the connection between the convolution layer (Conv) and the regional attention and cross-stage feature fusion module (A2C2f) and the connection layer (Concat) to further highlight the key features. The three-level attention mechanism model can more accurately focus on the key information in the image, thereby improving the prediction accuracy and performance of the model.

[0043] like Figure 2As shown in the figure, sample data was constructed using multi-source remote sensing image data according to the soil classification process. Sample cropping, image enhancement, positive and negative sample balancing, and image expansion were performed on the image data to obtain a sample dataset of smoke images from straw burning areas. When creating the sample dataset, each image was sliced ​​into 640×640 pixel slices using a sliding cut method with a 30% overlap ratio. This ensures efficient use of computer memory and meets the input size of deep learning models. The model can read in blocks during both training and prediction simulation to prevent computer memory overflow caused by inputting the entire image.

[0044] The present invention divides the sample data set constructed from multi-source remote sensing image data into training samples, test samples and verification samples, and uses the training samples to perform model training on the improved YOLO V12 deep learning network;

[0045] The model training process specifically includes:

[0046] (1) Initialize model parameters; optimize model hyperparameters.

[0047] (2) Data loading strategy: construct data enhancement samples and divide the input training samples and validation samples.

[0048] (3) Forward propagation and feature extraction optimization.

[0049] (4) Calculate the error based on the output probability.

[0050] (5) Back propagation: Update model parameters and minimize the error using gradient descent.

[0051] (6) Traverse all training data sets, use basic difficulty samples in the first 20% of iteration cycles, and then gradually introduce complex scenario data. After each epoch, perform validation set evaluation. When the validation loss decreases significantly for five consecutive times, the early stopping mechanism is triggered.

[0052] (7) Model convergence judgment: set a double termination condition. When the number of training iterations reaches the preset threshold or the fluctuation range of the validation set mAP@0.5 indicator is less than 0.2% for 10 consecutive iterations, the optimal model weights are automatically saved, and the integrated model containing uncertainty estimation is finally output.

[0053] Before training, it is necessary to set appropriate hyperparameters and select appropriate gradient descent optimizers and loss functions. Then, based on the convergence of the loss function and the accuracy improvement of the verification sample, the optimal parameters of the entire network are selected to obtain the optimal model. The optimal model is then used to make predictions on the test sample. The prediction requires the model to be read in blocks, and the prediction results are synthesized and then output. Finally, the four indicators of precision, recall, average precision, and F1-score are used to evaluate the accuracy of the prediction results and analyze the overall performance of the model.

[0054] Example 2: Compared with Example 1, this solution was verified in the field in Jilin Province, and the results are as follows:

[0055] Step 1: data collection;

[0056] This plan obtained Jilin-1 sub-meter high-resolution optical satellite image data and Sentinel-2 high-resolution multispectral imaging satellite data for the same time period by investigating straw burning areas within Jilin Province. The remote sensing image data used as training sets did not require preprocessing such as atmospheric correction and radiation calibration.

[0057] The Jilin-1 remote sensing image contains four bands, namely red band, green band, blue band, and near-infrared band, with a resolution of 0.75m. The shooting time is from March to May and October to November from 2022 to 2024, and the Chinese Geodetic Coordinate System model is CGCS2000.

[0058] To ensure the consistency of spatial resolution across all bands of the Sentinel-2 Level 2 Atmospheric Correction (Sentinel-2L2A) remote sensing image data, all bands were resampled using the nearest neighbor assignment method within the Open Source Remote Sensing Data Platform (SNAP) and output to a storage format supported by the Remote Sensing Image Processing Environment (ENVI). Band synthesis was performed using the Remote Sensing Image Processing Environment (ENVI 5.6) and the image was then cropped to a 640×640 pixel image. Figure 4 As shown, there are some examples of positive samples with smoke and negative samples with pure background in the sample set.

[0059] Step 2: Construct sample data set and preprocess data;

[0060] Remote sensing imagery areas containing straw burning were selected to create the dataset samples. To optimize computer memory usage, each image was sliced ​​into 640×640 pixel slices using a sliding cut method with a 30% overlap. To ensure a balanced distribution of positive and negative samples, pixels containing smoke were assigned a value of 1, while pixels containing no smoke were assigned a value of 0. Histogram analysis was used to remove samples where 0 pixels accounted for more than 80% of the total number of pixels in a single sample, and where 1 pixels accounted for 100% of the total number of pixels in a single sample. Furthermore, to ensure sufficient training samples, data augmentation was used to expand the sample size by performing horizontal and vertical flips and diagonal mirroring on each sample, quadrupling the sample size.

[0061] Step 3: Model construction;

[0062] This paper uses Python 3.8 as the development language and the concise and efficient open source deep learning framework Pytorch 1.10 for experiments. The CUDA version is 11.0. The hardware platform of this study is Intel (R) Xeon (R) CPU E5-2696 v3, a graphics card model is NVIDIA GeForce RTX 4060Ti workstation, the video memory size is 64G, the software used in the research is Windows 11 operating system, and the programming language Python and the deep learning framework Pytorch are used to construct a straw burning area smoke depth recognition model.

[0063] The present invention divides the preprocessed dataset into a training set, a validation set, and a test set. The model is trained using stochastic gradient descent (SGD) or its variants, with appropriate learning rates, batch sizes, and number of training rounds. During training, the error between the model's predictions and the true labels is calculated using a cross-entropy loss function, and the model parameters are updated using a backpropagation algorithm. The model is evaluated during training using the validation set, and the model's hyperparameters are adjusted based on the validation results to prevent overfitting.

[0064] In this study, we set the hyperparameters for training the object detection network to a batch size of 8, a learning rate (lr) of 0.001, a confidence level of 0.001, and a threshold of 0.5. When the loss function stabilized and the model converged after 100 training epochs, we used the SGD optimizer to optimize the parameter training process. During model training, the training set to test set ratio was 8:2, and transfer learning was introduced.

[0065] Step 4: Accuracy evaluation;

[0066] The average precision (AP) value is used to evaluate the accuracy of the multi-category detection task; the loss function (Loss) value is used as a measure of whether the model has converged; the F1 score (F1 Score), recall rate (Recall), and precision (Precision) are used as verification indicators. The formula is as follows:

[0067]

[0068] Where TP is the positive sample predicted as positive by the model, FP is the negative sample predicted as positive by the model, and FN is the positive sample predicted as negative by the model.

[0069] Step 6: Result analysis;

[0070] In order to explore the performance of different smoke recognition models, this paper selects several classic target detection algorithms for comparison, including Faster RCNN, SSD, YOLO v8, and YOLO v12.

[0071] Table 1 shows the performance comparison of different target detection models;

[0072] F1 score Recall / % Accuracy / % Average score / % Faster RCNN 0.93 93.52 94.30 91.87 SSD 0.93 91.98 93.71 91.37 YOLO v8 0.95 94.48 95.30 92.67 YOLO v12 0.96 93.51 95.66 93.01

[0073] The results in Table 1 demonstrate that the attention-centered object detection model (YOLO v12) algorithm demonstrates excellent performance across all four accuracy verification metrics. On the test set, it achieved a recognition accuracy of 93.01% and a recall rate of 97.66%, outperforming all compared methods. For images of straw burning, smoke boundaries obtained using traditional algorithms are relatively fuzzy, while the improved algorithm achieves clearer burn area recognition results than traditional algorithms, fully demonstrating the superiority of the improved algorithm.

[0074] In order to make the advantage of the improved algorithm in this paper clearer, the improved model based on the YOLO v12 algorithm was inferred and Table 2 was obtained.

[0075] Table 2 shows the ablation experiment performance results of the YOLO v12 model.

[0076] F1 score Recall / % Accuracy / % Average score / % YOLOv12 0.96 93.51 95.66 93.01 v12-SiLU 0.86 90.00 83.08 87.43 v12+CBAM 0.88 82.22 94.27 94.73 v12+TL 0.93 88.64 88.64 94.79 v12+CBAM+TL 0.96 93.13 96.16 95.52

[0077] The results show that the recall rate of the original YOLOv12 model is good, and the performance of the model decreases after replacing the Sigmoid Linear Unit (SiLU) activation function. After integrating the Convolutional Block Attention Module (CBAM) attention mechanism, the model performance is further improved, and the recognition accuracy reaches 100%. The introduction of transfer learning is mainly reflected in the detection efficiency of the model, but the training weights after migration still make the model detection accuracy improved to a certain extent. The YOLOv12 model of the fusion of pre-learning and attention mechanism proposed in the present invention performs well in smoke recognition, with an accuracy of 90% and a detection speed increased by 5ms. Obviously, the YOLOv12 algorithm that integrates attention mechanism and transfer learning has a faster inference speed and a lighter model. Therefore, the method proposed in the present invention can perform high-precision detection of smoke generated by straw burning.

[0078] In order to analyze the influence of different spatial resolutions on the recognition results, the present invention preprocesses the input band combination of red, green and blue, and the spatial resolutions of 30m, 0.75m and mixed data sets.

[0079] Table 3 compares the results at different spatial resolutions;

[0080] F1 score Recall / % Accuracy / % Average score 50 / % 0.75m 0.91 83.33 99.38 97.82 0.75～30m 0.87 80.21 95.25 92.13 30m 0.77 64.81 96.33 90.01

[0081] As shown in Table 3, when the red, green, and blue bands with a spatial resolution of 0.75m are input into the model for training, the model's recall rate, precision rate, and recognition accuracy are all the highest, at 83.33%, 99.38%, and 97.82%, respectively. This verifies the common understanding that "the higher the spatial resolution, the higher the remote sensing classification accuracy."

[0082] To further improve the accuracy of the model and explore its recognition performance under different channel combinations, this paper first uses the RGB three-channel as the model input, then continues to add the near-infrared band. On this basis, the index is added as a new channel. The present invention obtains Sentinel-2L2A data from the same period to calculate the modified normalized difference fire index, normalized burn rate, burned area index, improved crop residue cover, and normalized difference tillage index for channel synthesis, as shown in Table 4:

[0083] Table 4 shows the index description and calculation method;

[0084] name Abbreviation Band calculation Normalized Difference Fire Index MNDFI (SWIR2-NIR-5%) / (SWIR2+NIR+5%) Normalized firing rate NBR (NIR-SWIR2) / (NIR+SWIR2) Burned Area Index BAI <![CDATA[1 / ((0.1-RED) 2 +(0.06-NIR) 2 )]]> Improved crop residue mulch MCRC (SWIR1-BLUE) / (SWIR1+BLUE) Normalized Difference Tillage Index NDTI (SWIR1-SWIR2) / (SWIR1+SWIR2)

[0085] Table 5 shows the comparison of recognition results of different band combinations;

[0086] Dataset Number of channels Recall / % Accuracy / % Average score 50 / % RGB 3 64.81 96.33 90.01 RGB_band5 4 79.25 80.43 82.83 RGB_band5_band7 5 90.00 83.08 87.43 RGB_MNDFI 4 89.51 94.77 92.91 RGB_band5_MNDFI 5 93.89 94.92 92.92 RGB_MNDFI_NBR 5 91.98 93.71 91.37 RGB_MNDFI_BAI 5 93.83 84.92 91.14 RGB_MNDFI_BAI_MCRC 6 93.13 92.16 95.52

[0087] The results in Table 5 show that adding channels to the original RGB image does not necessarily enhance recognition. Band-calculated imagery helps improve smoke boundary determination. The channel combination (RGB + MNDFI + BAI + MCRC) is the most effective for identifying smoke in straw burning areas, with a recall rate of 93.13% and an accuracy of 95.52%, demonstrating the best model recognition performance.

[0088] In summary, compared with the existing technology, the present invention makes full use of deep learning and image recognition technology, and uses the advantages of deep learning models such as fast speed, high efficiency and automation to effectively identify smoke in straw burning areas.

[0089] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0090] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying smoke in straw burning areas based on a deep learning algorithm, characterized in that: include: S1, data acquisition: Use multi-source remote sensing image data to construct sample data, and process the image data, including image cropping, image enhancement, positive and negative sample balance and image expansion, to obtain a sample dataset of smoke images in straw burning areas; S2, data preprocessing: divide the remote sensing image dataset collected in S1 into training set, validation set and test set; S3, model construction: Use the programming language Python and the deep learning framework PyTorch to build a smoke depth recognition model for straw burning areas; S4, model training and prediction: Use training samples to train the deep learning model. Perform hyperparameter optimization before training, select the appropriate gradient descent optimizer, loss function, and iterator. At the same time, based on the convergence of the loss function and the accuracy improvement of the validation samples, select the optimal parameters for the entire network to obtain the optimal model. S5, Accuracy evaluation and result analysis: Use four indicators including precision, recall rate, average precision, and F1 score to evaluate the accuracy of the prediction results and analyze the overall performance of the model.

2. The method for identifying smoke in straw burning areas based on a deep learning algorithm according to claim 1 is characterized in that: The image cropping, image enhancement, and image expansion described in S1 are written in Python, and image calibration is done by annotating Labelimg with the help of target detection tools. Image cropping, image enhancement, expansion, and positive and negative sample balancing specifically include: Image cropping: Remote sensing image areas containing straw burning were selected to create dataset samples. Each image was cut into 640 × 640 pixel sample slices using a sliding cut method with a 30% overlap rate. Image enhancement and expansion: Data enhancement is used to expand samples, performing horizontal flipping, vertical flipping, and diagonal mirroring on each sample to increase the number of samples to 4-6 times the original number; Positive and negative sample balance: select pixels containing smoke and assign them a value of 1, and pixels containing non-smoke and assign them a value of 0. Use the histogram to remove samples in which 0-valued pixels account for more than 80% of the total number of single-sample pixels and samples in which 1-valued pixels account for 100% of the total number of single-sample pixels.

3. The method for identifying smoke in straw burning areas based on a deep learning algorithm according to claim 1, characterized in that: The deep recognition model described in S3 uses the attention-centric object detection model (YOLO v12) network as its basic framework and introduces a simple and efficient regional attention module. This module reduces the computational complexity of attention while maintaining a large receptive field. By introducing a residual efficient layer aggregation network, it overcomes the optimization problems caused by the introduction of the attention mechanism, thereby improving the calculation speed.

4. The method for identifying smoke in a straw burning area based on a deep learning algorithm according to claim 1, characterized in that: S4 specifically includes: S4.1, before training: Use stochastic gradient descent or its variant optimization algorithm to train the smoke recognition model, setting the appropriate learning rate, batch size and number of training rounds; S4.2, during training: use the cross entropy loss function to calculate the error between the model's prediction results and the true label, and update the model parameters through the backpropagation algorithm; S4.3, Post-training: Use the validation set to evaluate the model during training and adjust the model’s hyperparameters based on the validation results to prevent overfitting.

5. The method for identifying smoke in a straw burning area based on a deep learning algorithm according to claim 4 is characterized in that: The hyperparameter settings described in S4.3 include setting the batch size to 8, selecting a learning rate of 0.001, a confidence level of 0.001, and a threshold value of 0.5-0.

75. When the training epoch is 100, the loss function tends to be stable and the model converges. The stochastic gradient descent optimizer is used to optimize the parameter training process. During model training, the ratio of the training set to the test set is 8:2, and a transfer learning strategy is introduced.

6. The method for identifying smoke in straw burning areas based on a deep learning algorithm according to claim 1, characterized in that: The accuracy described in S5 is used to evaluate the accuracy of the multi-category detection task using the average precision value; The loss function value is a measure of whether the model has converged; the F1 score, recall rate, and precision rate are used as verification indicators. The specific formula is as follows: Where TP is the positive sample predicted as positive by the model, FP is the negative sample predicted as positive by the model, and FN is the positive sample predicted as negative by the model.

Images