Tobacco field segmentation method based on multi-modal PIDNet and progressive feature interaction

Abstract

The application discloses a tobacco field segmentation method based on a multi-modal PIDNet and progressive feature interaction, and belongs to the field of agricultural remote sensing intelligent analysis, and comprises the following steps: acquiring multi-modal image data and performing pretreatment; constructing a training data set based on the pretreated multi-modal image data; constructing a double-path progressive learning architecture in the P branch of the PIDNet, obtaining an improved segmentation model, and training the improved segmentation model through the training data set; processing real-time acquired multi-modal image data based on the trained model to obtain a preliminary segmentation mask; and optimizing the preliminary segmentation mask based on NDVI threshold filtering to realize accurate tobacco field segmentation. Through the progressive cross-modal feature guiding mechanism and the deep separable convolution technology, the application realizes efficient fusion and enhancement of RGB and near-infrared features, significantly improves the distinguishing ability of tobacco seedlings and weeds, reduces the model calculation complexity, and meets the demand of unmanned aerial vehicle real-time tobacco field monitoring.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of agricultural remote sensing intelligent analysis, and particularly relates to a tobacco field segmentation method based on multi-modal PIDNet and progressive feature interaction. BACKGROUND

[0002] In the field of precision agriculture and intelligent tobacco planting, high-precision tobacco field monitoring technology is crucial for improving crop yield and quality. With the gradual intelligentization of global agricultural production, the use of precision monitoring technology to assess crop growth conditions, predict crop diseases and pests, and provide support for precision fertilization, irrigation, and disease control has become a key link in improving agricultural production efficiency and sustainable development.

[0003] Traditional RGB image segmentation methods, such as U-Net, DeepLab, etc., have achieved great success in medical image processing and general image segmentation, but still have many shortcomings in agricultural applications. Especially between tobacco seedlings and common weeds, due to their high similarity in color and texture, it is difficult for RGB image segmentation algorithms to accurately distinguish between the two. In addition, factors such as changes in lighting, shadow interference, and changes in plant growth conditions often occur in tobacco fields, which further increase the difficulty of segmentation. Therefore, relying solely on traditional RGB image segmentation methods often results in low segmentation accuracy, and even cannot effectively distinguish between tobacco seedlings and weeds, affecting the accuracy of precision agriculture management and decision-making.

[0004] In recent years, the development of multispectral imaging technology has provided a new solution for agricultural remote sensing. The near-infrared (NIR) band has unique response characteristics to the physiological state of plants (such as chlorophyll content, water status), which can effectively enhance the distinction between healthy crops and weeds, and disease areas. However, existing multi-modal segmentation methods still have obvious shortcomings: on the one hand, simple RGB-NIR channel splicing or post-feature fusion cannot fully exploit the complementarity of cross-modal data; on the other hand, most algorithms have high computational complexity, which cannot meet the real-time requirements of unmanned aerial vehicles or edge computing devices. In addition, general segmentation networks lack optimization design for agricultural scenarios, especially in the feature extraction and post-processing stages, which fail to effectively incorporate agricultural prior knowledge (such as NDVI vegetation index), limiting their practical application effect in tobacco field precision management. SUMMARY

[0005] To solve the above technical problems, the present application proposes a tobacco field segmentation method based on multi-modal PIDNet and progressive feature interaction to solve the problems existing in the prior art.

[0006] To achieve the above purpose, the present application provides a tobacco field segmentation method based on multi-modal PIDNet and progressive feature interaction, comprising:

[0007] Acquire multimodal image data and perform preprocessing; the multimodal image data includes RGB images and near-infrared images.

[0008] A training dataset was constructed based on the preprocessed multimodal image data;

[0009] A dual-path progressive learning architecture is constructed in the P branch of PIDNet to obtain an improved segmentation model;

[0010] The improved segmentation model is trained based on the training dataset. The trained model is then used to process real-time acquired RGB and near-infrared images to obtain a preliminary segmentation mask.

[0011] The initial segmentation mask is optimized based on NDVI threshold filtering to achieve accurate segmentation of tobacco fields.

[0012] Optionally, the preprocessing of the multimodal image data includes: correcting the multimodal image data using a radiometric correction formula, and performing geometric registration of the corrected multimodal image data using SIFT feature point matching and the RANSAC algorithm.

[0013] Optionally, the process of constructing the training dataset includes: labeling the preprocessed multimodal image data with ground truth masks of tobacco field areas, and then expanding the samples through data augmentation techniques to obtain the training dataset, wherein the data augmentation techniques include rotation, scaling, noise addition, and random horizontal flipping operations.

[0014] Optionally, a dual-path progressive learning architecture can be constructed based on inter-modal dynamic guidance mechanisms and intra-modal self-attention reinforcement.

[0015] Optionally, the intermodal dynamic guidance mechanism uses linearly increasing weights. For RGB image features F RGB Near-infrared image features F NIR The gradual fusion is expressed by the following formula:

[0016] ;

[0017] in, For bidirectional attention calculation, This is the fusion factor.

[0018] Optionally, the intramodal self-attention enhancement uses a spatial-channel joint self-attention module to enhance the features of the RGB image branch and the near-infrared image branch. The channel attention module of the spatial-channel joint self-attention module generates channel weights through global average pooling and two fully connected layers. The spatial attention module of the spatial-channel joint self-attention module generates a spatial weight map through convolution operations. The outputs of the channel attention module and the spatial attention module are fused and added to the original features to obtain the enhanced features.

[0019] Optionally, the process of training the improved segmentation model based on the training dataset includes:

[0020] A dataset of farmland images including various crop types is obtained, and the improved segmentation model is pre-trained. The pre-trained segmentation model is then trained and fine-tuned using the training dataset.

[0021] Optionally, the improved segmentation model also includes depthwise separable convolution, which uses the fused features output by the dual-path progressive learning architecture as input to the depthwise separable convolution to extract multimodal features in parallel. The depthwise separable convolution decomposes the standard convolution into depthwise convolution and pointwise convolution.

[0022] Optionally, the output features of the Intensity branch and Dual branch of the improved segmentation model are obtained, and combined with the multimodal features extracted by the depthwise separable convolution, a preliminary segmentation mask is obtained.

[0023] Optionally, the process of optimizing the initial segmentation mask based on NDVI threshold filtering includes:

[0024] After removing the weed areas from the initial segmentation mask using NDVI threshold filtering, a simple linear iterative clustering algorithm is used to perform superpixel clustering on the image, dividing it into multiple superpixel regions with similar features. Morphological erosion is then used to process the tobacco field boundaries of each superpixel region to obtain the final optimized mask.

[0025] Compared with the prior art, the present invention has the following advantages and technical effects:

[0026] This invention proposes a tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction. By constructing a progressive cross-modal feature guidance mechanism, it achieves the gradual fusion and enhancement of RGB texture features and near-infrared spectral features, effectively improving the ability to distinguish tobacco seedlings from weeds. Simultaneously, by employing depthwise separable convolution and multi-branch collaborative optimization strategies, it significantly reduces the model's computational complexity and improves the efficiency of multimodal data fusion, meeting the needs of real-time tobacco field monitoring by UAVs.

[0027] (1) Efficient and progressive modal feature fusion;

[0028] This invention proposes a modal perception guidance mechanism based on dual-path attention, constructing a progressive feature interaction framework between RGB and NIR images. By introducing a cross-modal bidirectional attention module, the guidance relationship between RGB and NIR features is dynamically adjusted, and combined with a depth-increasing fusion factor, modal collaborative optimization is achieved from shallow to deep. This mechanism fully leverages the complementarity of the two modalities in terms of texture details and spectral response, effectively improving the model's ability to identify targets such as tobacco seedlings and weeds, as well as its spatial structure perception capabilities.

[0029] (2) Low computational complexity and real-time performance;

[0030] By employing lightweight designs such as depthwise separable convolution, standard convolution is decomposed into depthwise convolution and pointwise convolution, significantly reducing the number of model parameters and computational cost. This reduces the number of model parameters from 4.2M to 2.5M, and increases the inference speed on embedded devices (Jetson Nano) to 15FPS, meeting the requirements for real-time field segmentation (processing latency <500ms).

[0031] (3) Strong adaptability and generalization ability;

[0032] By combining transfer learning strategies, the model is first pre-trained on a large dataset of farmland images containing various crop types, allowing it to learn common farmland feature patterns. Then, it is fine-tuned on a tobacco field dataset. This training method enables the model to quickly adapt to different tobacco field environments, including varying lighting conditions, topography, and planting density. Whether during periods of drastic lighting changes or in tobacco field areas with complex terrain, the model consistently outputs accurate segmentation results, effectively improving its generalization ability across various real-world scenarios. Attached Figure Description

[0033] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0034] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0035] Figure 2 This is a flowchart illustrating the implementation of depthwise separable convolution in an embodiment of the present invention. Detailed Implementation

[0036] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0037] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0038] Example 1

[0039] like Figure 1 As shown, this embodiment provides a tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction, including:

[0040] S1. Multimodal Image Data Acquisition: Using a DJI Matrice 300 RTK drone equipped with a Micasense RedEdge-MX multispectral camera with synchronous triggering function, RGB and near-infrared (NIR) image data of tobacco fields were acquired. The flight altitude was 120 meters and the flight speed was 4 meters / second, ensuring that the image resolution was not less than 2500×2500 pixels and the NIR band spectral resolution reached 10nm.

[0041] S2. Multimodal image data preprocessing: Radiometric correction and geometric registration are performed on the RGB and near-infrared images of tobacco fields collected by UAVs to construct an RGB-NIR multimodal training dataset;

[0042] Through radiation correction formula The original pixel values ​​of the RGB and near-infrared images of tobacco fields collected by the drone were corrected, among which... Here, D represents the original pixel value, G represents the dark current noise, and G represents the gain coefficient. These are the normalized pixel values ​​after radiometric correction.

[0043] The SIFT feature point matching algorithm is used to extract stable and unique feature points from RGB and near-infrared images. These feature points are key identifiers of the images, representing important structural and textural information. Then, the extracted feature points are filtered and matched using the RANSAC (Random Sample Consensus) algorithm. RANSAC effectively removes mismatched feature points, improving matching accuracy. After filtering and matching, a transformation matrix is ​​calculated, which is used to achieve precise spatial alignment between the two images, ensuring that the spatial location of the same ground feature is completely consistent in both modalities, laying a solid foundation for subsequent feature fusion.

[0044] S3. Training Dataset Construction: Professional annotators used the LabelMe tool to annotate the ground truth mask of the tobacco field area, with the annotation error controlled within 1 pixel. Data augmentation techniques such as rotation (±15), scaling (0.8-1.2 times), and adding noise (Gaussian noise with a standard deviation not exceeding 0.05) were used to expand the sample, and the training set, validation set, and test set were divided in a 7:2:1 ratio.

[0045] S4. Modality-aware feature interaction processing is performed on the preprocessed RGB and near-infrared images. In the P (Parsing) branch of PIDNet, a dual-path progressive learning architecture is constructed to simultaneously achieve inter-modal feature guidance enhancement and single-modal self-attention reinforcement, thereby improving the model's ability to distinguish tobacco seedlings, weeds, and background in complex field environments. In the dual-path progressive learning architecture, after the self-attention module (spatial-channel joint self-attention) of the single-modal path enhances the features of the RGB and NIR branches, the enhanced features are not directly fused through a linear incremental fusion mechanism, but rather participate in the fusion process together with cross-modal interactive features.

[0046] In terms of intermodal guidance, a linearly incremental fusion mechanism is introduced to achieve progressive co-optimization of RGB and NIR features. Its calculation form is as follows:

[0047] ;

[0048] in, For bidirectional attention computation, modeling is performed simultaneously. arrive The relationship between NIR and RGB features. Using fusion factors, and progressively increasing network depth, a guiding relationship is constructed between modalities from shallow to deep. For cross-modal fusion features, Original features of RGB modality These are the original features of the NIR modes.

[0049] Bidirectional attention is calculated as follows:

[0050] ;

[0051] Among them, the input features of RGB and NIR modes are respectively through Convolution generates query vectors ( ), key vector ( ) and value vector ( ), feature dimension is . For RGB modal query vectors, For NIR mode bond vectors, This is the NIR mode value vector. For modal query vectors, For RGB modal bond vectors, This represents the RGB modal value vector. In bidirectional attention computation, the RGB modalities are... and The dot product yields the guided weights, which are then applied to... This enhances NIR features; similarly, NIR modes are enhanced through... and guide The attention enhancement results of the two methods are then weighted and fused to form a joint feature containing bimodal mutual attraction information, which is used for subsequent segmentation processing.

[0052] Within the single-modal path, a joint spatial-channel self-attention module is further introduced for the RGB and NIR branches respectively to model the salient regions and discriminative channel information within the modality. This module consists of a channel attention mechanism and a spatial attention mechanism.

[0053] The channel attention mechanism models the importance of different channels, and its calculation formula is as follows:

[0054] ;

[0055] in, Indicates the feature map Perform global average pooling. and It consists of two fully connected layers. It is the ReLU activation function. This is the Sigmoid function, used to generate channel weights.

[0056] Spatial attention mechanisms focus on the discriminative power of different spatial locations in an image, and their formula is as follows:

[0057] ;

[0058] in, and These represent the average and max pooling operations over the channel dimension, respectively, and are concatenated using... Convolution generates a spatial attention map.

[0059] Finally, channel attention and spatial attention are combined and fused into the original features to obtain the intramodal enhanced feature representation. :

[0060] ;

[0061] in This indicates an element-wise multiplication operation. This feature enhancement strategy effectively improves the responsiveness of single-modal features to key areas in tobacco fields (such as weed clusters and growth margins) and provides stronger discriminative properties for subsequent semantic segmentation.

[0062] S5. Multimodal Feature Extraction: Feature extraction is performed on the fused multimodal features to obtain multimodal features. To fully utilize the advantages of both RGB and NIR modal data, this invention improves PIDNet by expanding its input channels to four channels (RGB+NIR). Simultaneously, depthwise separable convolution technology is employed, which is a highly efficient feature extraction method.

[0063] Depthwise separable convolution decomposes standard convolution into two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs the convolution operation independently on each channel, with the convolution kernel sliding only within a single channel. This approach fully exploits the spatial features within each channel while reducing the number of parameters and computational cost. The depthwise separable convolution process is as follows: Figure 2 As shown.

[0064] Pointwise convolution is used to combine features from different channels. It fuses the features from each channel after depthwise convolution using a 1×1 convolution kernel, enabling the interaction and integration of information between different channels. Through the synergistic effect of depthwise convolution and pointwise convolution, parallel extraction of multimodal features is achieved.

[0065] In the implementation, the spatial convolution kernel size is set to 3×3 with a stride of 1. This parameter setting ensures that the convolution operation covers all regions of the image while fully capturing local spatial details. The depthwise convolution channel expansion ratio is set to 4 to further optimize computational efficiency while maintaining feature extraction effectiveness. The LeakyReLU activation function is used, which effectively alleviates the vanishing gradient problem, enhances the model's non-linear expressive ability, enables the model to better learn complex feature relationships, and improves feature extraction efficiency.

[0066] S6. Multi-branch collaborative optimization: By combining the features of the Intensity branch and the Dual branch, multi-branch collaborative optimization is achieved to generate a more accurate preliminary segmentation mask.

[0067] The Intensity branch primarily handles illumination issues, employing techniques such as Adaptive Histogram Equalization (CLAHE). In actual tobacco field image acquisition, uneven illumination is a common problem, leading to significant brightness differences across different areas of the image and affecting the model's recognition of tobacco field features. CLAHE effectively eliminates the impact of uneven illumination by performing local histogram statistics and contrast enhancement on the image, extracting features with illumination robustness. These features remain relatively stable under different illumination conditions, providing reliable information for the model.

[0068] The Dual branch focuses on capturing detailed information in the image. By calculating the feature differences between adjacent layers, it can keenly capture subtle changes in the image, such as the texture of tobacco leaves and the subtle undulations of tobacco field boundaries. This detailed information is crucial for accurately segmenting tobacco fields, helping the model distinguish between tobacco seedlings and weeds, as well as identify the boundaries of tobacco fields.

[0069] Multimodal features are extracted via depthwise separable convolution and then jointly optimized with features from the Intensity and Dual branches. The Intensity branch handles illumination robustness features, while the Dual branch captures detail features. The three are then concatenated or weighted and fused together via channel concatenation, inputting into the segmentation head (e.g., a 1×1 convolution) to map pixel classification probabilities and generate a preliminary segmentation mask. This multi-branch approach allows for complementary advantages and improves segmentation accuracy.

[0070] To further enhance the model's generalization ability, this invention employs a transfer learning strategy. Pre-training is first performed on a dataset of farmland images containing multiple crop types, with at least 1000 images for each crop type. These crop types encompass different growth forms, colors, and textures. By pre-training on such a diverse dataset, the model can learn general farmland feature patterns, such as general vegetation growth patterns and distinguishing features between farmland and non-farmland features.

[0071] Next, the model was fine-tuned on the tobacco field dataset, specifically optimized for the unique features of tobacco fields. Tobacco fields have specific planting patterns, crop growth cycles, and environmental characteristics; therefore, fine-tuning helps the model better adapt to the tobacco field scenario, improving its performance and generalization ability in tobacco field segmentation tasks. During fine-tuning, the learning rate was set to 1e. -5 Furthermore, the model was optimized using a weighted cross-entropy loss function, enabling it to fit tobacco field data more accurately and thus significantly improving segmentation accuracy.

[0072] S7. Post-processing optimization: NDVI threshold filtering is used to remove falsely detected weed areas. This is based on the average NDVI value of the region. and standard deviation Combined with adjustment coefficient Calculate the threshold:

[0073] ;

[0074] Next, areas below a threshold are considered false positives and removed, thus improving segmentation accuracy. Then, the SLIC superpixel clustering algorithm is used to divide the image into multiple superpixel regions with similar features, simplifying the image structure and reducing the complexity of subsequent processing. The algorithm parameters are optimized: the number of cluster centers N is set to 300-500, dynamically adjusted according to image resolution, with a larger value for high-resolution images to balance detail preservation and feature aggregation; the compactness factor m is set to 15-25 to ensure a balance between geometric compactness and feature similarity of superpixels; the maximum number of iterations is limited to 10 to ensure rapid algorithm convergence.

[0075] Finally, morphological erosion is used to refine the tobacco field boundaries. The erosion operation iterates through the segmentation mask using a structuring element (a 3×3 square), removing boundary pixels that do not match the structuring element. Specifically, for the binarized tobacco field segmentation mask, if non-tobacco field pixels exist within the area covered by the structuring element, the central pixel is identified as boundary noise and deleted. This operation effectively removes boundary burrs, isolated pixels, and small protrusions, smoothing the tobacco field contour and making the segmentation result more closely resemble the actual shape of a tobacco field.

[0076] Through the synergistic effect of the above post-processing steps, a high-precision tobacco field segmentation result with continuous boundaries is finally output.

[0077] The present invention exhibits significant technical effects in several aspects, specifically with the following advantages:

[0078] (1) Efficient and progressive modal feature fusion;

[0079] A modal perception guidance mechanism based on dual-path attention is proposed, constructing a progressive feature interaction framework between RGB and NIR images. By introducing a cross-modal bidirectional attention module, the guidance relationship between RGB and NIR features is dynamically adjusted, and combined with a depth-increasing fusion factor, modal co-optimization is achieved from shallow to deep. This mechanism fully leverages the complementarity of the two modalities in terms of texture details and spectral response, effectively improving the model's ability to identify targets such as tobacco seedlings and weeds, as well as its spatial structure perception capabilities.

[0080] (2) Low computational complexity and real-time performance;

[0081] By employing lightweight designs such as depthwise separable convolution, standard convolution is decomposed into depthwise convolution and pointwise convolution, which significantly reduces the number of model parameters and computational cost.

[0082] (3) Strong adaptability and generalization ability;

[0083] By combining transfer learning strategies, the model is first pre-trained on a large dataset of farmland images containing various crop types, allowing it to learn common farmland feature patterns. Then, it is fine-tuned on a tobacco field dataset. This training method enables the model to quickly adapt to different tobacco field environments, including varying lighting conditions, topography, and planting density. Whether during periods of drastic lighting changes or in tobacco field areas with complex terrain, the model consistently outputs accurate segmentation results, effectively improving its generalization ability across various real-world scenarios.

[0084] This invention focuses on the intersection of smart agriculture and advanced image processing technology. For the first time, it introduces the progressive fusion concept of RGB-NIR multimodal images into the PIDNet framework, constructing a tobacco field image segmentation system with high precision, high efficiency, and strong generalization capabilities. It breaks through the technical constraints of traditional single RGB image segmentation in areas such as feature complementarity mining, multi-source data fusion, and computational efficiency adaptation. With its accurate tobacco field region identification and segmentation results, it lays a solid technological foundation for the smart tobacco planting industry. This method not only significantly improves the precision and intelligence of tobacco planting management, substantially reduces resource consumption, and increases production efficiency, but also provides a new approach for technological iteration and upgrading in the field of smart agriculture through innovative multimodal processing and lightweight design concepts, demonstrating enormous application potential and significant economic and social value in the process of agricultural modernization.

[0085] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction, characterized in that, Includes the following steps: Acquire multimodal image data and perform preprocessing; the multimodal image data includes RGB images and near-infrared images. A training dataset was constructed based on the preprocessed multimodal image data; A dual-path progressive learning architecture is constructed in the P branch of PIDNet to obtain an improved segmentation model; The improved segmentation model is trained based on the training dataset, and the trained model is used to process real-time acquired RGB and near-infrared images to obtain a preliminary segmentation mask. The initial segmentation mask is optimized based on NDVI threshold filtering to achieve precise segmentation of tobacco fields; A dual-path progressive learning architecture is constructed based on inter-modal dynamic guidance mechanism and intra-modal self-attention reinforcement. The intermodal dynamic guidance mechanism uses linearly increasing weights. For RGB image features F RGB Near-infrared image features F NIR The incremental collaborative optimization is performed using the following formula: ； in, For bidirectional attention calculation, The fusion factor increases layer by layer according to network depth; Bidirectional attention is calculated as follows: ； Among them, the input features of RGB and NIR modes are respectively through Convolution generates query vectors ( ), key vector ( ) and value vector ( ), feature dimension is ; For RGB modal query vectors, For NIR mode bond vectors, This is the NIR mode value vector. For modal query vectors, For RGB modal bond vectors, It is an RGB modal value vector.

2. The tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction according to claim 1, characterized in that, The preprocessing of the multimodal image data includes: correcting the multimodal image data using a radiometric correction formula, and performing geometric registration of the corrected multimodal image data using SIFT feature point matching and the RANSAC algorithm.

3. The tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction according to claim 1, characterized in that, The process of constructing the training dataset includes: labeling the tobacco field region ground truth mask on the preprocessed multimodal image data, and then expanding the samples through data augmentation techniques to obtain the training dataset. The data augmentation techniques include rotation, scaling, noise addition, and random horizontal flipping operations.

4. The tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction according to claim 1, characterized in that, The intramodal self-attention enhancement uses a spatial-channel joint self-attention module to enhance the features of the RGB image branch and the near-infrared image branch. The channel attention module of the spatial-channel joint self-attention module generates channel weights through global average pooling and two fully connected layers. The spatial attention module of the spatial-channel joint self-attention module generates a spatial weight map through convolution operations. The outputs of the channel attention module and the spatial attention module are fused and added to the original features to obtain the enhanced features.

5. The tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction according to claim 1, characterized in that, The process of training the improved segmentation model based on the aforementioned training dataset includes: A dataset of farmland images including various crop types is obtained, and the improved segmentation model is pre-trained. The pre-trained segmentation model is then trained and fine-tuned using the training dataset.

6. The tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction according to claim 1, characterized in that, The improved segmentation model also includes depthwise separable convolution, which takes the fused features output by the dual-path progressive learning architecture as input to the depthwise separable convolution and extracts multimodal features in parallel. The depthwise separable convolution decomposes the standard convolution into depthwise convolution and pointwise convolution.

7. The tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction according to claim 6, characterized in that, The output features of the Intensity and Dual branches of the improved segmentation model are obtained, and combined with the multimodal features extracted by the depthwise separable convolution, a preliminary segmentation mask is obtained.

8. The tobacco field segmentation method based on multimodal PIDNet and progressive feature interaction according to claim 1, characterized in that, The process of optimizing the initial segmentation mask based on NDVI threshold filtering includes: After removing the weed areas from the initial segmentation mask using NDVI threshold filtering, a simple linear iterative clustering algorithm is used to perform superpixel clustering on the image, dividing it into multiple superpixel regions with similar features. Morphological erosion is then used to process the tobacco field boundaries of each superpixel region to obtain the final optimized mask.

Images