Hill tobacco field identification method, equipment and medium

Abstract

The invention relates to the field of remote sensing intelligent monitoring, in particular to a hilly tobacco field identification method and device and a medium, the process of the method is as follows: constructing a tobacco field sample data set, the tobacco field sample data set comprising a plurality of preprocessed tobacco field samples; constructing a tobacco field identification network based on direction self-adaption and multi-scale perception; training the tobacco field identification network by adopting a tobacco field sample data set to obtain a tobacco field identification model; and processing the tobacco field image by using the tobacco field identification model to complete tobacco field identification. According to the method, local fine-grained information, scale semantic expression and global context information are utilized, and the identification ability and boundary expression ability of the model to the tobacco field area in the complex terrain background are improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing intelligent monitoring, and in particular to a hilly tobacco field identification method, equipment and medium. Background Art

[0002] Due to the complex terrain and scattered plantings in hilly areas, traditional monitoring methods relying on manual inspections and ground measurements are inefficient and costly. Remote sensing technology has demonstrated significant advantages in vegetation identification and crop monitoring and is gradually being introduced into tobacco field extraction. However, traditional methods rely heavily on spectral features and hand-designed rules, making feature extraction difficult and segmentation accuracy insufficient for tobacco fields of varying scales and complex shapes. In recent years, with the advancement of high-resolution remote sensing data and deep learning technologies, researchers have combined remote sensing imagery with semantic segmentation networks, achieving initial automation of tobacco field extraction. However, in complex areas such as hilly and mountainous terrain, mainstream models such as FCN and U-Net still suffer from significant limitations in robustness and generalization. Firstly, the complex shapes and varying sizes of tobacco fields make it difficult to accurately capture their boundaries and internal structure during feature extraction. Secondly, the high elevation differences and background vegetation in hilly areas can lead to feature confusion, missed detections, and false detections, making precise tobacco field identification difficult.

[0003] In summary, there is an urgent need for a hilly tobacco field identification method, equipment and medium to solve existing problems. Summary of the Invention

[0004] To address the above issues, the present invention proposes a method, device, and medium for identifying hilly tobacco fields that combines direction perception with multi-scale feature fusion. The method comprehensively utilizes local fine-grained information, scale semantic expression, and global context information to enhance the model's ability to identify and express boundaries of tobacco fields in complex terrain. The specific technical solution is as follows:

[0005] A method for identifying hilly tobacco fields, the process is as follows:

[0006] S1: Construct a tobacco field sample dataset, which includes several preprocessed tobacco field samples;

[0007] S2: Construct a tobacco field recognition network based on direction adaptation and multi-scale perception;

[0008] S3: using a tobacco field sample data set to train the tobacco field recognition network to obtain a tobacco field recognition model;

[0009] S4: Use the tobacco field recognition model to process the tobacco field image and complete tobacco field recognition.

[0010] Optionally, in S1, the preprocessing process of tobacco field samples is as follows:

[0011] Acquire ground images, annotate the ground images, generate initial tobacco field samples, enhance the initial tobacco field samples through rotation, flipping and hue conversion, and obtain expanded tobacco field samples.

[0012] Optionally, in S2, the tobacco field recognition network includes a Swin Transformer feature extraction module, a direction-adaptive attention mechanism, a multi-scale feature enhancement module, and a classification head;

[0013] The Swin Transformer feature extraction module is used to model global context information;

[0014] The direction-adaptive attention mechanism is used to capture low-frequency information and high-frequency information in different directions in local details;

[0015] The multi-scale enhancement module is used to extract key features of different scales;

[0016] The classification head is used to map features to class predictions.

[0017] Optionally, in S2, the feature extraction module is a Swin Transformer feature extraction module, which is used to extract first feature maps of different levels.

[0018] Optionally, in S2, the direction-adaptive attention mechanism is specifically as follows:

[0019] Receive the first feature map and perform max pooling and average pooling in the horizontal dimension using the horizontal pooling kernel;

[0020] Max pooling and average pooling are performed using vertical pooling kernels in the vertical dimension;

[0021] The convolutional layer is used to align features and the nonlinear activation function is used to adjust the attention weight to obtain the second feature map.

[0022] Optionally, in S2, the multi-scale enhancement module includes a feature pyramid network and a SE module:

[0023] Input the second feature map into the feature pyramid network to fuse high-level semantic features with low-level features;

[0024] The SE module is used to recalibrate the channels of features at different scales, and global information embedding, adaptive correction and reweighting are completed in sequence to obtain the third feature map.

[0025] Optionally, in S3, the tobacco field recognition network is trained as follows:

[0026] Calculate the total loss of the tobacco field recognition network, backpropagate the total loss, update the parameters of the tobacco field recognition network until the parameters converge, save the trained parameters, and obtain the tobacco field recognition model.

[0027] Optionally, in S3, the training loss includes binary cross entropy loss and semantic segmentation loss, expressed as follows:

[0028] L=αL bce +βL dice ;

[0029]

[0030] Among them, L represents the total loss of tobacco field recognition network, L bce represents the binary cross entropy loss, L dice Represents semantic segmentation loss, α and β represent the weight ratio of binary cross entropy loss and semantic segmentation loss respectively, N represents the set of image pixels, a represents the ath element, y represents the true label, Represents the prediction result, X represents the pixel category of the predicted image, and Y represents the pixel label of the real image.

[0031] Additionally, the present invention also includes a computer device comprising a memory and a processor;

[0032] The memory is used to store a computer program that can be executed on the processor;

[0033] The processor is configured to implement the steps of the above-mentioned hilly tobacco field identification method when executing the computer program.

[0034] In addition, the present invention also includes a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the hilly tobacco field identification method as described above are implemented.

[0035] The application of the technical solution of the present invention has the following beneficial effects:

[0036] (1) The present invention provides a method for identifying hilly tobacco fields. The directional adaptive attention mechanism in the method of the present invention constructs maximum pooling and average pooling operations in different directions, and combines the low-frequency and high-frequency features obtained by the attention mechanism to be enhanced, and finally fuses them with the original image features, thereby maximally retaining important details, reducing information loss, and significantly improving the model's ability to represent complex terrain scenes.

[0037] (2) The multi-scale feature enhancement module in the method of the present invention effectively improves the diversity and discriminability of feature expression by fusing high-level semantic information with low-level detail features, enabling the model to more accurately identify tobacco fields of different scales and significantly improving the detection ability of tobacco field targets. In addition, the feature extractor in the method of the present invention can effectively model the global and local contextual dependencies in the image, enhancing the model's ability to express complex terrain and heterogeneous tobacco fields, thereby improving overall recognition accuracy and robustness.

[0038] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 It is a flowchart of the steps of the hilly tobacco field identification method in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, this embodiment provides a method for identifying hilly tobacco fields, and the process is as follows:

[0043] S1: Construct a tobacco field sample dataset, which includes several preprocessed tobacco field samples;

[0044] S2: Construct a tobacco field recognition network based on direction adaptation and multi-scale perception;

[0045] S3: using a tobacco field sample data set to train the tobacco field recognition network to obtain a tobacco field recognition model;

[0046] S4: Use the tobacco field recognition model to process the tobacco field image and complete tobacco field recognition.

[0047] Optionally, in S1, the preprocessing process of tobacco field samples is as follows:

[0048] After orthorectification and image fusion, the acquired ground imagery was cropped to create a remote sensing image of the study area. Plot-level detailed manual annotation was performed on the ArcGIS platform to construct an initial tobacco field sample set. To further improve sample quality and annotation accuracy, drone imagery with a spatial resolution of 0.36 meters was collected from a representative area covering approximately 20% of the study area. Cross-validation was performed on 165 field sample points with GPS location information to generate high-quality semantic segmentation labels. Furthermore, enhancement operations such as rotation, flipping, and hue shifting were performed on the initial tobacco field samples to construct a more diverse tobacco field sample dataset, enhancing the model's robustness and generalization capabilities.

[0049] Optionally, in S2, the tobacco field recognition network includes a Swin Transformer feature extraction module, a direction-adaptive attention mechanism, a multi-scale feature enhancement module, and a classification head;

[0050] The feature extraction module is used to model global context information;

[0051] The direction-adaptive attention mechanism is used to obtain low-frequency information and high-frequency information in different directions in local details;

[0052] The multi-scale enhancement module is used to extract key features of different scales;

[0053] The classification head is used to map features to class predictions.

[0054] Optionally, in S2, the feature extraction module is a Swin Transformer feature extraction module, which is used to extract first feature maps of different levels.

[0055] Optionally, in S2, the direction-adaptive attention mechanism is specifically as follows:

[0056] Receive the first feature map and perform maximum pooling and average pooling in the horizontal dimension using the horizontal pooling kernel (1, W);

[0057] Max pooling and average pooling are performed in the vertical dimension using the vertical pooling kernel (H, 1);

[0058] The convolutional layer (1×1 convolution) is used to align the features, and the attention weight is adjusted in combination with the nonlinear activation function to obtain the second feature map.

[0059] Optionally, in S2, the multi-scale enhancement module includes a feature pyramid network and a SE module:

[0060] The second feature map is input into the feature pyramid network, and high-level semantic features are combined with low-level features through upsampling to obtain tobacco field features at different scales.

[0061] The SE module is used to recalibrate the channels of features at different scales, and global information embedding, adaptive correction and reweighting are completed in sequence to obtain the third feature map.

[0062] Global information embedding: By compressing features in the spatial dimension, the feature map of size W×H×C is compressed into a feature vector Z of size 1×1×C. The process is expressed as:

[0063]

[0064] Among them, u c (ij) represents the eigenvalue of the cth channel at position (i, j), z c represents the global response of channel c.

[0065] Adaptive correction: The information in the compression operation is further processed through a two-layer fully connected gate mechanism. The expression is as follows:

[0066] s = σ(W2·ReLU(W1·z));

[0067] The weight s has a dimension of 1×1×C and is used to characterize the weights of C feature maps. z is the feature vector after global average pooling, W1 is the weight matrix of the first fully connected layer, which compresses the number of channels from C to C / r, and W2 is the weight matrix of the second fully connected layer, which restores the number of channels from C / r to C. σ represents the sigmoid activation function.

[0068] Reweighting: The weight s is regarded as the importance index of each feature, and is applied to each channel feature through channel-by-channel multiplication weighting to recalibrate the multi-scale features of the tobacco field and obtain the final output The expression is as follows:

[0069]

[0070] Among them, s c Represents the weight factor of the c-th channel.

[0071] Optionally, in S3, the tobacco field recognition network is trained as follows:

[0072] Calculate the total loss of the tobacco field recognition network, backpropagate the total loss, update the parameters of the tobacco field recognition network until the parameters converge, save the trained parameters, and obtain the tobacco field recognition model.

[0073] Optionally, in S3, the training loss includes binary cross entropy loss and semantic segmentation loss, expressed as follows:

[0074] L=αL bce +βL dice ;

[0075]

[0076] Among them, L represents the total loss of tobacco field recognition network, L bce represents the binary cross entropy loss, L dice Represents semantic segmentation loss, α and β represent the weight ratio of binary cross entropy loss and semantic segmentation loss respectively, N represents the set of image pixels, a represents the ath element, y represents the true label, Represents the prediction result, X represents the pixel category of the predicted image, and Y represents the pixel label of the real image.

[0077] In order to verify the technical effect of the method of this embodiment, this embodiment quantitatively compares the method of this embodiment with four advanced deep learning models, FCN, U-Net, DeepLabV3+, and Swin Transformer, on the tobacco field identification task in a certain place. At the same time, an ablation experiment is designed to evaluate the contribution of each module. The experimental results are shown in Tables 1 and 2, respectively. Table 1 is the quantitative evaluation results, Table 2 is the results of the ablation experiment, mIoU is the intersection over union ratio, OA is the overall accuracy, and F1-score is the F1 score.

[0078] Table 1 Quantitative evaluation results of tobacco field recognition task

[0079]

[0080] As can be seen in Table 1, while the Swin Transformer model is capable of capturing long-range contextual information, its limited ability to refine local details and boundaries makes it perform poorly when extracting small-scale tobacco fields against complex backgrounds. The Swin Transformer's mIoU, OA, and F1-score values ​​are all the lowest in Table 1, at 65.49%, 83.85%, and 76.44%, respectively. The FCN model, with mIoU of 72.25%, OA of 76.92%, and F1-score of 83.89%, performs poorly primarily because the FCN lacks fine-grained contextual information and struggles to recover boundary details, resulting in insufficient performance when dealing with scenes with complex backgrounds and diverse scales, such as tobacco fields. Although the U-Net model combines low-level details with high-level semantic information through skip connections, its ability to capture multi-scale objects in remote sensing imagery remains limited. This is particularly true for small-scale objects in tobacco field extraction, resulting in mIoU (mean Intersection of Two-Dimensional Objects) scores of 68.52%, OA (Open Area Recognition) scores of 87.51%, and an F1-score of 81.32%. The DeepLabV3+ model, which uses dilated convolutions to expand the receptive field, performs well in some scenarios. However, when faced with tobacco fields with complex backgrounds and blurred object boundaries, dilated convolutions fail to capture local information, affecting segmentation accuracy. This results in mIoU scores of 69.75%, OA (Open Area Recognition) scores of 88.21%, and an F1-score of 82.18%. The DAMP network achieved the highest mIoU, OA, and F1-score scores for tobacco field prediction, at 77.25, 94.47, and 87.17, respectively. These results show that the DAMP network is significantly better than the current mainstream semantic segmentation methods when dealing with data with complex backgrounds, fragmented plots, and diverse plot shapes and scales in semantic segmentation tasks, which is well reflected in the remote sensing image dataset of a certain place in the present invention. DAMP introduces a multi-scale feature fusion module, which enhances the model's ability to capture tobacco fields of different scales by fusing features of different scales, especially when dealing with tobacco fields with complex backgrounds, it can accurately segment targets of different scales. In addition, DAMP's contextual information enhancement module further improves the ability to capture details of tobacco fields, making it perform better than other semantic segmentation models in remote sensing image scenes with complex backgrounds and varied scales. From the experimental results and the above analysis, it can be seen that the direction-adaptive attention mechanism in the method of this embodiment has an excellent effect on the accurate representation of plots of various shapes in hilly areas.

[0081] Table 2 Results of ablation experiments

[0082] method mIoU (%) OA (%) F1-score (%) (a) Backbone 65.49 83.85 76.44 (b) Backbone+DAAM 73.11 95.24 78.58 (c) Method of this Example (DAMP) 76.68 96.68 81.76

[0083] As can be seen from Table 2, after adding DAAM, the overall accuracy of the model is compared with the basic model (Backbone), the mIoU is improved by 7.62%, the OA coefficient is increased by 11.39%, and the F1-score is improved by 2.14%. This result shows that the direction-adaptive attention mechanism has a significant enhancement effect in capturing the low-frequency and high-frequency features of different directions of local details of hilly and mountainous tobacco fields. As can be seen from Table 2, after adding the multi-scale feature enhancement module, the mIoU is improved by 3.57%, the OA value is improved by 1.44%, and the F1-score is improved by 3.18% compared with the baseline model plus DAMP. This result shows that the multi-scale feature enhancement module has significant effectiveness in capturing multi-scale key features and effectively addresses the challenges brought by the scale variation of hilly and mountainous plots. As can be seen from Table 2, the direction-adaptive attention mechanism and the multi-scale feature enhancement module in the method of this embodiment can both improve the effect of the model. The results of the ablation experiment verify the superiority of the method of this embodiment in the remote sensing intelligent extraction of hilly and mountainous tobacco fields.

[0084] In addition, this embodiment also includes a computer device, including a memory and a processor;

[0085] The memory is used to store a computer program that can be executed on the processor;

[0086] The processor is configured to implement the steps of the above-mentioned hilly tobacco field identification method when executing the computer program.

[0087] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in the computer device.

[0088] The computer device may be a mobile phone, desktop computer, laptop, PDA, cloud server, or other computing device. The computer device may include, but is not limited to, a processor and memory. For example, the computer device may also include input and output devices, network access devices, buses, etc.

[0089] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the computer device, connecting various parts of the entire computer device using various interfaces and lines.

[0090] The memory can be used to store the computer program and / or module, and the processor implements the computer program by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created based on the use of the mobile phone (such as audio data, a phone book, etc.). In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0091] Wherein, if the module / unit integrated in the computer device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0092] In addition, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the above-mentioned hilly tobacco field identification method are implemented.

[0093] The present embodiment provides a method for identifying hilly tobacco fields, including a feature extraction module, a direction-adaptive attention mechanism, and a multi-scale feature enhancement module, wherein the direction-adaptive attention mechanism constructs multi-directional maximum pooling and average pooling operations, and combines the low-frequency and high-frequency features obtained by the attention mechanism to be enhanced, and finally fuses them with the original image features, thereby retaining important details to the maximum extent, reducing information loss, and significantly improving the model's representation ability for complex terrain scenes. The multi-scale feature enhancement module in the method of this embodiment effectively improves the diversity and discriminability of feature expression by fusing high-level semantic information with low-level detail features, enabling the model to more accurately identify tobacco field areas of different scales and greatly improving the detection ability of tobacco field targets. In addition, the feature extraction module in the method of this embodiment can efficiently model the global and local contextual dependencies in the image, enhance the model's expression ability for complex terrain and heterogeneous tobacco field areas, and thus improve the overall recognition accuracy and robustness.

[0094] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0095] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for identifying a hilly tobacco field, characterized in that: The process is as follows: S1: Construct a tobacco field sample dataset, which includes several preprocessed tobacco field samples; S2: Construct a tobacco field recognition network based on direction adaptation and multi-scale perception; S3: using a tobacco field sample data set to train the tobacco field recognition network to obtain a tobacco field recognition model; S4: Use the tobacco field recognition model to process the tobacco field image and complete tobacco field recognition.

2. The hilly tobacco field identification method according to claim 1, characterized in that: In S1, the preprocessing process of tobacco field samples is as follows: Acquire ground images, annotate the ground images, generate initial tobacco field samples, enhance the initial tobacco field samples through rotation, flipping and hue conversion, and obtain expanded tobacco field samples.

3. The hilly tobacco field identification method according to claim 2, characterized in that: In S2, the tobacco field recognition network includes a feature extraction module, a direction-adaptive attention mechanism, a multi-scale feature enhancement module, and a classification head; The feature extraction module is used to model global context information; The direction-adaptive attention mechanism is used to capture multi-directional low-frequency information and high-frequency information in local details; The multi-scale enhancement module is used to extract key features of different scales; The classification head is used to map features to class predictions.

4. The hilly tobacco field identification method according to claim 3, characterized in that: In S2, the feature extraction module is a Swin Transformer feature extraction module, which is used to extract first feature maps of different levels.

5. The hilly tobacco field identification method according to claim 4, characterized in that: In S2, the direction-adaptive attention mechanism is as follows: Receive the first feature map and perform max pooling and average pooling in the horizontal dimension using the horizontal pooling kernel; Max pooling and average pooling are performed using vertical pooling kernels in the vertical dimension; The convolutional layer is used to align features and the nonlinear activation function is used to adjust the attention weight to obtain the second feature map.

6. The hilly tobacco field identification method according to claim 5, characterized in that: In S2, the multi-scale enhancement module includes a feature pyramid network and a SE module: Input the second feature map into the feature pyramid network to fuse high-level semantic features with low-level features; The SE module is used to recalibrate the channels of features at different scales, and global information embedding, adaptive correction and reweighting are completed in sequence to obtain the third feature map.

7. The hilly tobacco field identification method according to claim 1, characterized in that: In S3, the tobacco field recognition network is trained as follows: Calculate the total loss of the tobacco field recognition network, backpropagate the total loss, update the parameters of the tobacco field recognition network until the parameters converge, save the trained parameters, and obtain the tobacco field recognition model.

8. The hilly tobacco field identification method according to claim 7, characterized in that: In S3, the training loss includes binary cross entropy loss and semantic segmentation loss, which are expressed as follows: L=αL bce +βL dice ; Among them, L represents the total loss of tobacco field recognition network, L bce represents the binary cross entropy loss, L dice Represents semantic segmentation loss, α and β represent the weight ratio of binary cross entropy loss and semantic segmentation loss respectively, N represents the set of image pixels, a represents the ath element, y represents the true label, Represents the prediction result, X represents the pixel category of the predicted image, and Y represents the pixel label of the real image.

9. A computer device, characterized in that: including memory and processor; The memory is used to store a computer program that can be executed on the processor; The processor is configured to implement the steps of the hilly tobacco field identification method according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the hilly tobacco field identification method according to any one of claims 1 to 8.

Images