Mangrove tree species detection method based on YOLO11 multi-level perception calibration

Abstract

The application discloses a mangrove tree species detection method based on YOLO11 multi-level perception calibration, to solve the problems of mangrove canopy adhesion and small target missed detection under complex background. The method first constructs a UAV aerial dataset through preprocessing, and then improves the YOLO11 model: a perception calibration module (PCS) is introduced at the starting end of the backbone network, a second-order differential operator is used to extract high-frequency edges to enhance the capture of geometric structure; a semantic guidance aggregation module (SGA) is integrated in C3K2, edge or smooth prior attributes are adaptively injected according to the network level, feature pixel-level calibration and noise suppression are realized; a gating interaction module (GIM) is deployed in the neck network and a high-resolution detection head is added, and through the complementary gating mechanism, cross-scale feature dynamic fusion is realized. Experiments show that the application effectively solves the problems of mangrove canopy adhesion and background interference, and also improves the accuracy and robustness of tree species detection, and is suitable for coastal wetland ecological monitoring.

Description

Technical Field

[0001] This invention relates to the field of target detection, and in particular to a method for detecting mangrove tree species based on YOLO11 multi-level perception calibration. Background Technology

[0002] Mangroves, as a unique woody plant community in the tropical and subtropical intertidal zones, play an irreplaceable ecological role in maintaining coastal biodiversity, protecting seawalls, and contributing to the global carbon cycle. Accurate monitoring of mangrove species distribution is a crucial prerequisite for ecological protection, degradation restoration, and carbon sequestration assessment. With the rapid evolution of unmanned aerial vehicle (UAV) remote sensing technology, automated tree species identification using high-resolution UAV imagery combined with deep learning algorithms has become a mainstream research approach in the field of ecological monitoring. This technology provides forestry departments with precise species composition information through real-time analysis of large-scale images acquired by UAVs, demonstrating significant application value in early warning and protection against dynamic changes in the coastal environment.

[0003] When faced with the complex environment of mangrove wetlands, existing general detection algorithms such as YOLO11 still face significant technical bottlenecks. Especially in dense scenes where mangrove communities are highly clustered and canopies overlap, the visual boundaries between individual canopy members are extremely blurred, making it difficult for algorithms to accurately extract discriminative edge features, leading to a significant drop in detection accuracy. Traditional convolutional neural networks, when performing deep feature extraction, often experience severe collapse in the spatial representation of subtle geometric structures in the feature map as the number of downsampling layers increases. Meanwhile, the unique wetland distribution environment of mangroves, such as the specular reflection of tidal water surfaces and the shadow interference from silt mudflats, presents significant challenges to feature fusion. Existing feature pyramid structures, when processing heterogeneous feature flows, typically employ simple channel splicing or element-wise addition. This linear feature aggregation mechanism cannot effectively identify and suppress noise signals from complex backgrounds, resulting in a "semantic dilution" phenomenon during the alignment of high-level semantic features and low-level spatial features. Furthermore, the algorithm exhibits significant representational shortcomings when dealing with the dramatic scale differences between seedlings and mature trees.

[0004] To improve the accuracy of tree species detection and achieve full-scale perceptual alignment in complex mangrove environments, this invention deeply improves the YOLO11 architecture. By introducing a perceptual calibration module (PCS) at the beginning of the backbone network, and using a second-order gradient operator to explicitly capture prior geometric edge properties, the problem of structural information loss in the initial representation stage is solved. Combined with the C3K2-SGA module composed of structure-guided enhancement units (SGAs), the model can continuously calibrate texture representation during deep feature extraction. For the feature transition nodes of the neck network, this invention uses a gated interaction module (GIM) to replace the traditional fusion method, and achieves dynamic alignment of spatial details and semantic information and background suppression through a bidirectional gating mechanism. This invention constructs a multi-level perceptual calibration system, which effectively solves the problems of dense occlusion, background interference, and multi-scale target mismatch without significantly increasing the computational burden, and significantly improves the overall robustness of mangrove tree species detection. Summary of the Invention

[0005] The purpose of this invention is to propose a mangrove tree species detection method based on YOLO11 multi-level perception calibration, in order to solve the problems of low detection accuracy and high false negative rate caused by the complex background of mangrove growth environment, high similarity of features among tree species and uneven distribution of multi-scale targets.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for mangrove tree species detection based on YOLO11 multi-level sensor calibration includes the following steps:

[0008] Step 1: Acquire and preprocess the UAV remote sensing mangrove image dataset;

[0009] Specifically, step 1 includes the following steps:

[0010] Step 1.1: Perform quality checks on the raw images acquired by the UAV, remove blurry and distorted samples, and construct a standardized image set;

[0011] Step 1.2: Label the images to identify mangrove tree species and generate corresponding tag files;

[0012] Step 2: Construct a mangrove detection model based on YOLO11 multi-level perception calibration;

[0013] Step 2 specifically includes the following steps:

[0014] Step 2.1: At the beginning of the backbone network, replace the original shallow convolutional layer with the PCS module to achieve explicit enhancement of edge features;

[0015] Step 2.2: Construct the SGA module and embed it in the feature extraction unit. The module adaptively selects to extract prior attributes using edge operators or smoothing operators according to the depth of the network layers, and injects the prior attributes as guiding signals into the feature map to achieve pixel-level response calibration.

[0016] Step 2.3: Construct a GIM module in the neck network, use a gating mechanism to calculate the saliency weights of high-level semantic features and low-level spatial features, and perform bidirectional weighted fusion based on complementary gating coefficients;

[0017] Step 2.4: Add a high-resolution detection branch to the neck network architecture to enhance the characterization of small and dense tree species targets;

[0018] Step 3: Train the mangrove tree species detection method based on YOLO11 multi-level perception calibration;

[0019] Step 4: Input the remote sensing image of the mangrove forest to be detected into the trained detection model to evaluate the algorithm performance.

[0020] Preferably, step 1 specifically includes the following steps:

[0021] Step 1.1: Perform quality checks on the raw images acquired by the UAV, remove blurry and distorted samples, and construct a standardized image set;

[0022] Step 1.2: Mark the targets on the image and determine the mangrove tree species categories, which include Rhizophora stylosa, Kandelia candel, Excoecaria agallocha, Sonneratia apetala, and Bruguiera gymnorrhiza, and generate the corresponding tag files;

[0023] Preferably, step 2.1 specifically includes the following steps:

[0024] Step 2.1.1: Use 5×5 convolution and Gaussian-Laplacian (LoG) operator to perform feature mapping and edge enhancement on the input image. Generate high-frequency feature map by explicitly capturing the second-order differential information of the image. Then use depthwise separable convolution sequence to perform preliminary spatial downsampling to obtain primary feature mapping with geometric structure prior.

[0025] Step 2.1.2: Gaussian smoothing is applied to the primary feature map for noise reduction, and a dual-path feedback downsampling structure is input. Parallel max pooling and strided convolution branches are used to extract the salient texture and deep semantics of the image, respectively. Finally, channel concatenation and convolution are used to fuse the image and output a P2-scale feature map that balances edge accuracy and semantic information.

[0026] Preferably, step 2.2 specifically includes the following steps:

[0027] Step 2.2.1: The SGA module adaptively extracts input features based on its network depth stage. geometric or semantic prior attributes In the initial stage of the network ( The Scharr operator is used to extract canopy edge features; in the deeper stages of the network ( Spatial smoothing using the Gaussian operator is employed to suppress background noise. The extraction process is as follows:

[0028]

[0029] in, and These are the horizontal and vertical Schaller convolution kernels, respectively. The standard deviation is A two-dimensional Gaussian filter kernel;

[0030] Step 2.2.2: Obtain the prior attributes As a guide signal injected into the original feature In this process, efficient channel attention (ECA) is performed through the SGA module to achieve pixel-level response alignment and spatial interaction, generating calibrated enhanced features. ;

[0031] Step 2.2.3: Perform high-dimensional feature mapping and residual ensemble to enhance features. The input is fed into a multilayer perceptron (MLP) consisting of pointwise convolutions for nonlinear projection, followed by the application of a random depth operator (Drop Path) and batch normalization (BN) layers, and finally connected to the original input via residual connections. Perform weighted summation to obtain the final output features. The process is as follows:

[0032]

[0033] in, It is a feature map obtained by two convolutional operations. To provide a single-layer output for the SGA module, the C3K2-SGA module is constructed by embedding multiple SGA units into the C3K2 architecture.

[0034] Step 2.2.5: Replace layers 1, 3, 5, and 7 of the YOLO network with C3K2-SGA modules in sequence.

[0035] Preferably, step 2.3 specifically includes the following steps:

[0036] Step 2.3.1: The GIM module first performs dimensional alignment and response mapping of heterogeneous features, integrating low-level features of the same scale from the backbone network. With semantic features from deep layers Through respectively The linear mapping layer performs channel adjustment and uses the Sigmoid activation function to generate its respective spatial saliency weight mask, as shown in the following equation:

[0037]

[0038] in, This represents the Sigmoid activation function. and These represent the saliency distribution mappings of the low-level spatial features and the high-level semantic features, respectively, which are used to guide subsequent feature selection.

[0039] Step 2.3.2: Using the generated gating weights as a filtering mechanism, the low-level spatial information and the scale-aligned high-level semantic information are weighted and fused to generate a calibrated feature map. and The interaction logic is shown in the following formula:

[0040]

[0041]

[0042] in, Represents pixel-by-pixel multiplication. This represents bilinear interpolation upsampling. (Through...) Complementary control enables dynamic compensation between heterogeneous features and background noise suppression;

[0043] Step 2.3.3: Enhance the interaction After realigning Perform channel splicing, and through Convolutional layers enable cross-dimensional feature integration and dimensionality reduction, transforming gated and calibrated multi-scale information into a unified output representation of the GIM module with strong discriminative power.

[0044] Step 2.3.4: Deploy the GIM module in the upsampling, stitching and subsequent convolution paths of the YOLOv11 neck network to build a perception calibration link covering four levels: P2, P3, P4 and P5.

[0045] Preferably, step 3 specifically includes the following steps:

[0046] Step 3.1: Load the partitioned dataset, set the iteration number to 300, the batch size to 8, and the optimizer to SGD;

[0047] Step 3.2: Train the mangrove species detection algorithm based on YOLO11 multi-level perception calibration according to the set experimental parameters to obtain the weight file of the trained mangrove species detection model.

[0048] The beneficial effects of this invention are as follows:

[0049] 1. This invention introduces a spatial awareness calibration module into the YOLO11 feature extraction network. Utilizing a spatial weight distribution mechanism, it effectively suppresses complex intertidal water surface reflections and light and shadow noise interference in mangrove habitats, enhancing the separation between canopy edge features and the environmental background. This module, while preserving key spatial information, improves the purity of feature extraction in complex habitats, significantly enhancing the anti-interference capability of the detection task.

[0050] 2. This invention constructs a multi-level perception calibration structure within the YOLO11 neck network. Through feature offset compensation and semantic alignment techniques between levels, it addresses the problem of drastic scale variations in mangrove plants caused by different growth stages (e.g., seedlings versus mature canopies). This structure achieves accurate fusion of deep semantic information and shallow location information without significantly increasing computational overhead, greatly reducing the target false negative rate in densely interwoven forest areas.

[0051] 3. This invention integrates a channel sensing calibration mechanism in front of the YOLO11 detector head, and enhances the model's ability to distinguish fine-grained features of highly similar tree species (such as Rhizophora mangifera and Rhizophora chinensis) by dynamically capturing the unique spectral response and texture features of different mangrove tree species. Attached Figure Description

[0052] The following will provide a detailed description through specific implementation methods and accompanying drawings.

[0053] Figure 1 A schematic diagram of the overall structure of the improved YOLOv11 network;

[0054] Figure 2 This is a schematic diagram of the PCS module.

[0055] Figure 3This is a schematic diagram of the C3K2-SGA module;

[0056] Figure 4 This is a schematic diagram of the GIM module structure;

[0057] Figure 5 This is a flowchart of the mangrove tree species detection algorithm based on YOLO11 multi-level perception calibration proposed in this invention. Detailed Implementation

[0058] To make the embodiments of the present invention and the prior art solutions clearer and easier to understand, the present invention will be described in detail below with reference to the accompanying drawings.

[0059] Reference Figure 5 The steps to obtain the mangrove tree species detection algorithm based on YOLO11 multi-level perception calibration are as follows:

[0060] Step 1: Use a drone equipped with a high-resolution optical sensor to collect images of mangrove-covered areas, perform quality checks on the collected raw images, remove samples with motion blur, overexposure and geometric distortion, and construct a standardized image set;

[0061] Step 1.1: Use image annotation tools to annotate the bounding boxes of mangrove plants in the image to determine the mangrove species categories, including Rhizophora stylosa, Kandelia candel, Excoecaria agallocha, Sonneratia apetala, and Bruguiera gymnorrhiza, and simultaneously generate YOLO format tag files corresponding to the image files;

[0062] Step 1.2: Perform random online data augmentation on the training set, such as Mosaic augmentation, random rotation and scaling, to simulate the visual performance of mangroves under different tidal levels and lighting conditions;

[0063] Step 2: Construct a mangrove detection model based on YOLO11 multi-level perception calibration;

[0064] To address the challenges of mangrove forest detection, the YOLOv11 network was improved in the following three aspects: Figures 1 to 4 As shown:

[0065] Step 2 specifically includes the following steps:

[0066] Step 2.1: At the beginning of the backbone network, replace the original shallow convolutional layer with the PCS module to achieve explicit enhancement of edge features;

[0067] Step 2.2: Construct the SGA module and embed it in the feature extraction unit. The module adaptively selects to extract prior attributes using edge operators or smoothing operators according to the depth of the network layers, and injects the prior attributes as guiding signals into the feature map to achieve pixel-level response calibration.

[0068] Step 2.3: Construct a GIM module in the neck network, use a gating mechanism to calculate the saliency weights of high-level semantic features and low-level spatial features, and perform bidirectional weighted fusion based on complementary gating coefficients;

[0069] Step 2.4: Add a high-resolution detection branch to the neck network architecture to enhance the characterization of small and dense tree species targets;

[0070] Step 2.1 includes the following steps:

[0071] Step 2.1.1: Utilize Ordinary convolution compresses the input image's channels to 32, generating an initial feature map. This feature map is then input into a Gaussian-Laplacian filter, utilizing a predefined... Fixed-parameter convolution kernels perform second-order differentiation operations to extract edge features; Gaussian-Laplacian operator. The kernel parameters follow the formula below:

[0072]

[0073] in, The coordinates are within the convolution kernel. For scale parameters ( Finally, the extracted edge features are added to the initial feature map through residual connections, and then passed through a batch normalization layer and the SiLU activation function in sequence to output the edge-enhanced feature map;

[0074] Step 2.1.2: Utilize Convolutional sequences reduce the spatial resolution of feature maps to half its original value while increasing the number of channels to 64. They also introduce parallel Gaussian kernels with a two-dimensional Gaussian function. The formula is as follows:

[0075]

[0076] in, The coordinates are within the convolution kernel. The scale parameter is set to 0.5. The input features are smoothed and calibrated using convolutional mapping and parallel Gaussian branches, and the calibrated features are fused through residual connections.

[0077] Step 2.1.3: The feature input DRFD module is fused, employing a dual-path decomposition strategy. It simultaneously extracts fine structural features and salient local features from the feature map using convolutional and max-pooling branches, respectively. Multi-dimensional features are integrated through channel concatenation. Convolution integrates information and outputs the final feature maps of the backbone network at layers P2 / 4.

[0078] Step 2.2 includes the following steps:

[0079] Step 2.2.1: Based on the depth level parameters of the SGA module in the backbone network Perform adaptive branch selection, in Construct horizontal convolution kernels and vertical convolution kernel According to the definition of the Scharr operator, the weights are initialized as follows:

[0080]

[0081] use and The input features are grouped and convolved, then the comprehensive edge strength is calculated and superimposed with the original features. The feature projection is performed through bottleneck convolution layers to generate shallow prior attributes.

[0082] Step 2.2.2: In At that time, the SGA module generates a two-dimensional Gaussian weight matrix according to preset parameters and uses it as a fixed depth convolution kernel to filter the feature map to suppress background noise. The generated smooth features are combined with the original input to output deep prior attributes.

[0083] Step 2.2.3: Obtaining Prior Attributes With original features Hybrid modulation, and through Convolutional layers integrate spatial features to generate hybrid features. The calculation process is as follows:

[0084]

[0085] in, This represents pixel-by-pixel multiplication;

[0086] Step 2.2.4: For Efficient channel attention (ECA) computation is performed, using global average pooling (GAP) to compress features into channel descriptors. Then, one-dimensional convolution is used to capture local dependencies across channels, generating channel attention weights and... Enhanced features are obtained by performing weighted calibration. ;

[0087] Step 2.2.5: Enhance Features The input is fed into a multilayer perceptron (MLP) consisting of pointwise convolutions for nonlinear projection, followed by the application of a random depth operator (Drop Path) and batch normalization (BN) layers, and finally connected to the original input via residual connections. Perform weighted summation to obtain the final output features. The process is as follows:

[0088]

[0089] in, It is a feature map obtained by two convolutional operations. To provide a single-layer output for the SGA module, the C3K2-SGA module is constructed by embedding multiple SGA units into the C3K2 architecture.

[0090] Step 2.2.5: Replace layers 1, 3, 5, and 7 of the YOLO network with C3K2-SGA modules in sequence.

[0091] Step 2.3 includes the following steps:

[0092] Step 2.3.1: The GIM module will combine low-level features of the same scale from the backbone network. With semantic features from deep layers Through respectively The linear mapping layer compresses and aligns the input features along the channel dimension, and uses the sigmoid activation function to generate their respective spatial saliency weight masks, as shown in the following equation:

[0093]

[0094] in, This represents the Sigmoid activation function. , representing the saliency distribution mapping of the low-level spatial features and the high-level semantic features, respectively, which are used to guide subsequent feature selection;

[0095] Step 2.3.2: While generating the mask, utilize the other two sets of... Convolution performs a linear transformation on the original aligned features, followed by bidirectional feature interaction based on a complementary gating mechanism, with the gating weights of the current layer... When the current feature is unreliable (i.e., the current feature is not reliable), the system automatically fills the gap by introducing weighted features from another level through bilinear interpolation. This generates a calibrated feature map. and The interaction logic is shown in the following formula:

[0096]

[0097]

[0098] in, Represents pixel-by-pixel multiplication. This represents bilinear interpolation upsampling. (Through...) Complementary control enables dynamic compensation between heterogeneous features and background noise suppression;

[0099] Step 2.3.3: Enhance the interaction After realigning Perform channel splicing, and through Convolutional layers achieve cross-dimensional feature integration and dimensionality reduction, outputting the gated calibration multi-scale information as the final cross-scale calibration features of the GIM module;

[0100] Step 2.3.4: Deploy the GIM module in the upsampling, stitching and subsequent convolution paths of the YOLOv11 neck network to build a perception calibration link covering four levels: P2, P3, P4 and P5.

[0101] Step 3: Train the mangrove tree species detection method based on YOLO11 multi-level perception calibration;

[0102] Step 3.1: Divide the preprocessed mangrove dataset into a training set and a test set in a 7:3 ratio, and input the training set into the improved mangrove species detection algorithm;

[0103] Step 3.2: The input size of the dataset is uniformly adjusted to 640×640, and mosaic data augmentation is performed. The number of iterations is set to 300, the batch size is set to 8, and the optimizer is set to SGD.

[0104] Step 3.3: Execute a training loop in the mangrove detection model based on multi-level perception calibration, calculate the loss function value through forward propagation, and update the network weights using the backpropagation algorithm; monitor the loss changes and accuracy indicators of the validation set in real time during training, and after the model converges, extract and save the model weight file that performs best on the validation set.

[0105] Step 4: Input the remote sensing image of the mangrove forest to be detected into the trained detection model to evaluate the algorithm performance;

[0106] Step 4.1: Load the saved optimal model weight file, switch the model to inference mode, input the mangrove test set into the model for forward computation, output the original prediction tensor containing the target category, confidence level and bounding box coordinates, apply the non-maximum suppression (NMS) algorithm to post-process the original results, filter out low-quality and overlapping redundant prediction boxes according to the preset confidence threshold and intersection-overall reunification (IoU) strategy, and generate the final detection results;

[0107] Step 4.2: Based on the IoU matching principle, compare the final detection results with the ground truth, count the number of true positive (TP), false positive (FP), and false negative (FN) samples, and calculate the core performance indicator, precision. Recall rate The area under the PR curve is used to obtain the average accuracy of a single class. The arithmetic mean of the AP values ​​for all mangrove species was calculated to obtain the mean precision. The formula is as follows:

[0108]

[0109]

[0110] in, The total number of tree species categories is represented by this indicator, which comprehensively quantifies the classification accuracy and positioning robustness of the model in complex mangrove scenarios.

[0111] The mangrove tree species detection method based on YOLO11 multi-level perception calibration provided by this invention fully addresses the problems of canopy adhesion and severe background noise interference in complex mangrove habitats. It enhances the model's ability to perceive geometric edge structures and fine-grained features across scales, effectively solving the semantic misalignment and information loss phenomena in the multi-scale feature fusion process. This enables the model to significantly improve detection accuracy and robustness when dealing with dense small targets and morphologically similar tree species detection tasks, making it particularly suitable for refined management fields such as coastal wetland ecological monitoring, forest resource surveys, and biodiversity conservation.

[0112] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be within the scope of protection of the present invention.

Claims

1. A method for detecting mangrove tree species based on YOLO11 multi-level sensing calibration, characterized in that, Includes the following steps: Step 1: Acquire and preprocess the UAV remote sensing mangrove image dataset; Specifically, step 1 includes the following steps: Step 1.1: Perform quality checks on the raw images acquired by the UAV, remove blurry and distorted samples, and construct a standardized image set; Step 1.2: Label the images to identify mangrove tree species and generate corresponding tag files; Step 2: Construct a mangrove detection model based on YOLO11 multi-level perception calibration; Step 2 specifically includes the following steps: Step 2.1: At the beginning of the backbone network, replace the original shallow convolutional layer with the PCS module to achieve explicit enhancement of edge features; Step 2.2: Construct the SGA module and embed it in the feature extraction unit. The module adaptively selects to extract prior attributes using edge operators or smoothing operators according to the depth of the network layers, and injects the prior attributes as guiding signals into the feature map to achieve pixel-level response calibration. Step 2.3: Construct a GIM module in the neck network, use a gating mechanism to calculate the saliency weights of high-level semantic features and low-level spatial features, and perform bidirectional weighted fusion based on complementary gating coefficients; Step 2.4: Add a high-resolution detection branch to the neck network architecture to enhance the characterization of small and dense tree species targets; Step 3: Train the mangrove tree species detection method based on YOLO11 multi-level perception calibration; Specifically, step 3 includes the following steps: Step 3.1: Load the partitioned dataset and set the algorithm hyperparameters; Step 3.2: Train the constructed mangrove detection model using the target detection dataset, update the model parameters through backpropagation, and obtain the weights of the trained detection model; Step 4: Input the remote sensing image of the mangrove forest to be detected into the trained detection model to evaluate the algorithm performance.

2. The mangrove tree species detection method based on YOLO11 multi-level sensing calibration according to claim 1, characterized in that: Step 2.1 specifically includes the following steps: Step 2.1.1: Use large kernel convolution and Gaussian-Laplacian operator to process the input image in parallel, and extract feature maps and second-order differential edge information respectively; Step 2.1.2: After fusing the feature map with the edge information, input the dual-path feedback downsampling structure. The dual-path feedback downsampling structure includes a parallel max pooling branch and a strided convolution branch, which respectively extract the texture features and semantic features of the image. The outputs of the two branches are then fused through channel concatenation and convolution operations to generate a shallow feature map that balances edge accuracy and semantic information.

3. The mangrove tree species detection method based on YOLO11 multi-level sensing calibration according to claim 1, characterized in that, Step 2.2 specifically includes the following steps: Step 2.2.1: The SGA module adaptively extracts input features based on its network depth stage. geometric or semantic prior attributes In the initial stage of the network, the Scharr operator is used to extract canopy edge features; in the deeper stages of the network, the Gaussian operator is used for spatial smoothing to suppress background noise. The extraction process is as follows: in, and These are the horizontal and vertical Schaller convolution kernels, respectively. The standard deviation is A two-dimensional Gaussian filter kernel; Step 2.2.2: Obtain the prior attributes As a guiding signal, it is nonlinearly projected through a multilayer perceptron and injected into the original input features using residual connections. In this process, calibrated output features are generated.

4. The mangrove tree species detection method based on YOLO11 multi-level sensing calibration according to claim 1, characterized in that, Step 2.3 specifically includes the following steps: Step 2.3.1: The GIM module first performs dimensional alignment and response mapping of heterogeneous features, integrating low-level features of the same scale from the backbone network. With semantic features from deep layers Dimension alignment is performed, and spatial saliency weight masks are generated using a linear mapping layer and an activation function, respectively. and The process is as follows: in, This represents the Sigmoid activation function. and These represent the saliency distribution mappings of the low-level spatial features and the high-level semantic features, respectively, which are used to guide subsequent feature selection. Step 2.3.2: Using the generated weight mask as gating coefficients, cross-weighted fusion of low-level spatial features and high-level semantic features is performed. The interaction logic is shown in the following equation, generating the calibrated feature map. and The interaction logic is shown in the following formula: in, Represents pixel-by-pixel multiplication. This represents bilinear interpolation upsampling. This term is used to achieve complementary control and dynamic compensation between features; Step 2.3.3: Enhance the interactive features and Channel splicing and convolution integration are performed as the output of the GIM module, and the GIM module is deployed in the multi-scale path of the neck network.

Images