Method for monitoring quercus based on environmental simulation enhancement and attention guidance

Abstract

This invention discloses a monitoring method for the genus *Fagus* based on environmental simulation enhancement and attention guidance, belonging to the field of image processing technology. The method includes the following steps: S1, acquiring RGB images using a high-resolution camera mounted on a UAV and preprocessing them to generate a training set; S2, generating an expanded robust training sample set; S3, constructing an attention context guidance network; S4, training the attention context guidance network using the expanded robust training sample set; S5, inputting the images to be monitored into the trained attention context guidance network to generate several indicators. This invention, through the synergistic innovation of the environmental simulation enhancement framework and the attention context guidance network, has achieved significant technical progress in the monitoring of *Fagus* communities; it demonstrates outstanding advantages in segmentation accuracy, generalization ability, and practicality.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a monitoring method for the genus *Fagus* based on environmental simulation enhancement and attention guidance. Background Technology

[0002] Fagus species are core constituent species of broad-leaved and mixed forests in the northern temperate zone, playing an irreplaceable role in maintaining regional biodiversity, conserving water resources, and regulating the global carbon cycle. However, affected by factors such as climate change, pests and diseases, and human disturbance, natural Fagus forests worldwide generally face severe threats such as regeneration barriers, community structure degradation, and shrinking distribution areas. Against this backdrop, achieving accurate and efficient monitoring of the distribution, health status, and dynamic changes of Fagus communities has become an urgent need for forest ecology research and smart forestry management.

[0003] In recent years, with the widespread adoption of UAV remote sensing technology, acquiring high spatial resolution forest canopy images has become possible, providing a data foundation for detailed species-level identification and monitoring. Deep learning-based artificial intelligence technologies, particularly semantic segmentation models (such as U-Net and DeepLabv3+), have proven to have enormous potential in automatically interpreting remote sensing images and extracting target feature information. These models, by learning features from large amounts of labeled data, can achieve pixel-level classification and are theoretically very suitable for segmenting the canopy of *Fagus sylvatica* from aerial images.

[0004] However, when applying existing technologies directly to monitoring practices in the Fagus genus, two interconnected core bottlenecks are encountered:

[0005] First, at the model algorithm level, existing mainstream segmentation models perform poorly in complex forest scenes. Forest environments are highly heterogeneous and complex, specifically manifested in: 1) Variable target scale: Significant differences in the size and shape of the beech crown; 2) Strong background interference: Complex understory backgrounds (such as shrubs, herbs, and shadows) and frequent crown adhesion and occlusion phenomena in dense stands severely interfere with accurate target identification and fine boundary segmentation; 3) Similar visual features: Beech often grows mixed with other tree species, and the interspecific similarity in morphological and spectral features further increases the difficulty of species discrimination. Models such as U-Net have limitations in capturing global contextual information and long-distance dependencies through standard convolution operations, resulting in insufficient ability to distinguish similar features and a tendency to lose spatial details crucial for boundary segmentation during downsampling. While models such as DeepLabv3+ introduce dilated convolutions to expand the receptive field, their computational complexity is high, and their ability to preserve fine-grained edges remains insufficient. The inherent structural flaws in these models make it difficult for them to meet the requirements of practical applications in terms of generalization ability and segmentation accuracy when dealing with the complexity and diversity of forest scenarios.

[0006] Secondly, at the data level, there is a dual challenge of scarce publicly available data and weak model generalization ability. Firstly, there is a severe lack of publicly available high-resolution aerial semantic segmentation datasets specifically for the *Fagus* genus. This data deficiency significantly hinders the development and validation of specialized models. Secondly, to address insufficient data and improve model robustness, traditional data augmentation methods (such as rotation, flipping, and color dithering) are often limited to simple geometric and color transformations, failing to effectively simulate the complex and variable conditions of real-world environments. For example, in actual operations, the dynamic drone data collection process is affected by gusts and flight attitude adjustments, often resulting in motion blur or defocus in aerial images, leading to decreased image sharpness and loss of detail. Variable mountain weather conditions (such as fog and haze, strong light reflection) can easily cause reduced image contrast or local overexposure, all of which severely weaken the model's robustness. This gap between "simple augmentation" and "complex reality" causes the trained model to exhibit a significant performance decline when encountering complex and undesirable environmental conditions.

[0007] Therefore, there is an urgent need in this field for a comprehensive technical solution that not only proposes a novel segmentation network structure that can better adapt to complex forest scenarios to enhance feature extraction and context awareness capabilities, but also develops a data augmentation framework that can effectively simulate real field environments to significantly improve the generalization and robustness of the model, thereby providing a reliable technical tool for the accurate monitoring and sustainable management of Fagus forest resources.

[0008] Existing technology 1 relates to a semantic image segmentation method based on the U-Net architecture. This technical solution was originally designed for biomedical image segmentation, but due to its simple and effective structure, it has been widely used in image segmentation tasks in fields such as remote sensing and forestry.

[0009] The core of this technical solution is an encoder-decoder network with a symmetrical structure: (1) Encoder path (shrinkage path): This path consists of multiple repeated convolutional layers and pooling layers stacked sequentially. Each stage usually performs two consecutive 3×3 convolution operations (each convolution is followed by a modified linear unit activation function) and a 2×2 max pooling downsampling operation. The encoder extracts the input image features layer by layer through this structure, uses convolution to capture local features (such as edges and textures), and expands the receptive field and compresses the feature map size through pooling, thereby generating multi-scale feature maps including shallow features with rich spatial details and deep features containing high-level semantic information. (2) Decoder path (expansion path): This path is symmetrical to the encoder and consists of multiple upsampling stages: Each stage first performs a 2×2 deconvolution (transposed convolution) operation on the deep features from the previous layer to double the spatial size of the feature map; then the upsampled feature map is spliced ​​with the feature map of the same scale in the encoder path through skip connections; the spliced ​​feature map is then fused and refined by two 3×3 convolutional layers, and finally the decoder gradually restores the deep semantic features extracted by the encoder to the spatial size of the input image, outputting a pixel-level segmentation map. (3) Final output layer: The end of the decoder is usually connected to a 1x1 convolutional layer, whose output channel number is equal to the number of categories (e.g., background and beech), mapping the feature map to the category probability of each pixel, thereby achieving pixel-level classification.

[0010] Despite the widespread success of encoder-decoder architectures like U-Net in semantic segmentation tasks, their structural characteristics still exhibit significant limitations when applied to segmenting *Fagus* species in complex forest scenes. First, this model relies on standard convolutional operations for feature extraction, and its fixed local receptive field struggles to adapt effectively to the varying canopy scales and shapes found in forest scenes. This leads to inconsistent recognition of trees with large canopies in the foreground and small canopies in the background, particularly saplings and trees with irregular canopies, resulting in missed detections and missegments. Second, the lack of a dedicated global context-aware mechanism prevents the establishment of semantic relationships between distant pixels in the image. This results in insufficient discrimination ability when distinguishing between different tree species with similar colors and textures or complex forest backgrounds, easily misclassifying the background as the target in scenes with complex backgrounds and varying lighting. Furthermore, while continuous max-pooling operations in the encoder path gradually expand the receptive field, they inevitably cause a continuous loss of spatial detail information, leading to blurred canopy boundaries in the generated segmentation image and a significant decrease in edge localization accuracy, severely impacting the refinement of the segmentation results. Finally, the network structure of the model is relatively fixed, and its feature extraction method has limited adaptability to complex situations such as changes in lighting conditions, overlapping branches and leaves, and phenological changes caused by seasonal changes in the forest environment. In particular, it performs poorly when dealing with imaging quality degradation problems such as dynamic motion blur, cloud and fog interference, and strong light reflection that occur during drone data collection. Ultimately, this results in a significant lack of generalization performance when the model is deployed in real-world scenarios, making it difficult to meet the actual needs of precise forestry monitoring.

[0011] The second prior art involves traditional data augmentation methods, which are widely used to improve the generalization ability of deep learning models. This technical solution expands the dataset by applying basic geometric transformations and pixel-level perturbations to the training data. Its specific implementation includes the following two core operations: (1) The purpose of geometric transformation is to simulate the basic shape and position changes of the target in space. Common geometric transformations include random flipping and random rotation. Random flipping mirrors the input image in the horizontal or vertical direction. Random rotation rotates the image around its center point by a specific angle. (2) Pixel-level perturbation is used to simulate the changes in illumination and sensor noise during the imaging process. Common perturbations include Gaussian noise, Gaussian blur, and contrast adjustment. Gaussian noise is achieved by adding random noise that conforms to a Gaussian distribution to the image pixels. Gaussian blur smooths the image through Gaussian convolution kernels to simulate defocus or motion blur. Contrast adjustment adjusts the contrast and brightness of the image through linear transformation.

[0012] These enhancement operations are typically applied to the training images before training (in the preprocessing stage) and simultaneously applied to the corresponding labeled masks to ensure spatial consistency.

[0013] While traditional data augmentation methods are effective in improving model generalization, they have fundamental limitations when applied to complex forestry remote sensing scenarios. The geometric and color transformations performed by these methods are essentially simple extensions of "idealized" acquired data; their transformation patterns are regular and limited, failing to effectively simulate the complex and varied interference factors in real-world field environments. Specifically, these methods struggle to reproduce the real-world challenges faced by drones operating in forest areas, such as cloud and fog obstruction, localized overexposure due to sudden changes in light, motion blur caused by flight jitter, and random occlusion caused by intertwined branches and leaves. Therefore, a significant "domain gap" exists between models trained using traditional augmentation methods and real-world application scenarios. This leads to a significant decrease in segmentation accuracy and robustness when faced with images acquired in the field containing these complex interferences, failing to meet the stringent performance requirements for accurate forestry monitoring in actual deployment. Summary of the Invention

[0014] To address the above problems, this invention proposes a monitoring method for the genus *Fagus* based on enhanced environmental simulation and attention guidance.

[0015] The technical solution of this invention is: a monitoring method for the genus *Fagus* based on enhanced environmental simulation and attention guidance, comprising the following steps:

[0016] S1. Use the high-resolution camera on the drone to collect RGB images, perform preprocessing, and generate a training set;

[0017] S2. Simulate environmental interference in the RGB image to generate enhanced samples, and merge them with the RGB image to generate an expanded robust training sample set;

[0018] S3. Construct an attention context-guided network;

[0019] S4. Train the attention context guidance network using the expanded robust training sample set;

[0020] S5. Input the image to be monitored into the trained attention context guidance network to generate several indicators.

[0021] Furthermore, S1 includes the following sub-steps:

[0022] S11. Use a high-resolution camera mounted on a drone to capture RGB images;

[0023] S12. Perform Gaussian filtering for noise reduction, geometric correction, and cropping on the RGB image in sequence to complete the preprocessing.

[0024] S13. Annotate the preprocessed RGB image and generate the corresponding annotation mask;

[0025] S14. Divide the labeled RGB image into segments to generate a training set.

[0026] Furthermore, S2 includes the following sub-steps:

[0027] S21. The RGB image is processed by cloud and fog simulation, local overexposure simulation, dynamic motion blur simulation and random occlusion simulation to generate four types of enhanced samples.

[0028] S22. Combine the four types of augmented samples and the training set to generate an expanded robust training sample set.

[0029] Furthermore, in S21, the cloud and fog simulation specifically involves: randomly sampling on the image plane. A circular region is used to smooth the generated mask using a Gaussian filter, resulting in a continuous transparency map. Combine the original image with the hazy color Perform weighted fusion to generate enhanced results. Its transformation formula is:

[0030] ;

[0031] in, Indicates the spatial location of the original image ,aisle Pixel values; This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. The function constrains pixel values ​​to Within the range.

[0032] Furthermore, in S21, the simulation of local overexposure is specifically performed by: randomly generating a local highlight mask, which is then smoothed using Gaussian to form a gradually changing brightness field. And perform additive enhancement processing, the specific transformation formula is as follows:

[0033] ;

[0034] in, Represents the maximum brightness increment. Indicates parameters, Indicates the spatial location of the original image ,aisle pixel values, This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. The function constrains pixel values ​​to Within the range.

[0035] Furthermore, in S21, the dynamic motion blur simulation specifically involves: randomly generating the blur kernel length and rotation angle, constructing a directional linear convolution kernel, performing convolution to obtain a blurred image, and then weightedly fusing the blurred image with the RGB image to complete the dynamic motion blur simulation, the expression of which is:

[0036] ;

[0037] ;

[0038] in, Indicates the spatial location of the original image ,aisle Pixel values; This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. The motion-blurred image is obtained by convolving it with a rotated linear kernel; is the point spread function, representing the generated linear convolution kernel used to describe the motion blur path during camera movement. The function constrains pixel values ​​to Within the range, The intensity blending coefficient controls the blending ratio between the blurred image and the original image. This represents convolution.

[0039] Furthermore, in S21, the random occlusion simulation specifically involves: randomly selecting a rectangular region on the RGB image, setting the pixels within the rectangular region to constant values, and generating an occlusion image, the expression of which is:

[0040] ;

[0041] in, This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. Indicates the spatial location of the original image ,aisle pixel values, Indicates the spatial location of the original image ,aisle pixel values, This indicates the fill value for the occluded area.

[0042] Furthermore, in the attention context-guided network, the encoder's SPConv module includes proportionally divided representative and redundant channels;

[0043] The representative channel undergoes convolution processing; the redundant channel undergoes convolution processing and global average pooling processing, and then weighted fusion is performed to extract features.

[0044] Furthermore, in the attention context-guided network, the encoder's CGB-Down module performs standard convolution and dilated convolution on the features to obtain joint features. The joint features are then processed by global average pooling and multilayer perceptron to generate global context weights, which are then used to process the joint features.

[0045] Furthermore, in the attention context guidance network, the decoder's Criss-Cross attention module convolves the highest-level semantic features output by the encoder to generate Q vectors, K vectors, and V vectors, and calculates the correlation between Q vectors and K vectors in the same row or column to obtain an attention map, which is then used to generate global context information.

[0046] The beneficial effects of this invention are:

[0047] (1) This invention enhances the quality of samples by using simulation algorithms for four environmental interference factors, namely cloud and fog obstruction, sudden changes in illumination, motion blur and random obstruction, as well as parameter configuration and combination rules for each simulation strategy;

[0048] (2) This invention has achieved significant technological progress in the monitoring of Fagus genus communities through the collaborative innovation of environmental simulation enhancement framework and attention context guidance network; it has shown outstanding advantages in terms of segmentation accuracy, generalization ability and practicality. Attached Figure Description

[0049] Figure 1 This is a flowchart of a monitoring method for the genus *Fagus* based on environmental simulation enhancement and attention guidance. Detailed Implementation

[0050] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, this invention provides a monitoring method for the genus *Fagus* based on enhanced environmental simulation and attention guidance, comprising the following steps:

[0052] S1. Use the high-resolution camera on the drone to collect RGB images, perform preprocessing, and generate a training set;

[0053] S2. Simulate environmental interference in the RGB image to generate enhanced samples, and merge them with the RGB image to generate an expanded robust training sample set;

[0054] S3. Construct an attention context-guided network;

[0055] S4. Train the attention context guidance network using the expanded robust training sample set;

[0056] S5. Input the image to be monitored into the trained attention context guidance network to generate several indicators.

[0057] This invention aims to address four key technical challenges faced by existing technologies in segmenting Fagus genus trees in complex forest scenarios: 1) How to improve the efficiency and adaptability of the model for multi-scale canopy feature extraction to accurately identify Fagus canopies of different sizes; 2) How to enhance the model's ability to utilize global contextual information to improve the accuracy of distinguishing similar tree species in complex forest backgrounds; 3) How to better preserve spatial detail information during feature extraction to improve the segmentation quality of canopy boundaries; and 4) How to improve the model's adaptability to complex environmental conditions in real forest areas through effective data augmentation methods to enhance the model's generalization and robustness in practical applications.

[0058] In this embodiment of the invention, S1 includes the following sub-steps:

[0059] S11. Use a high-resolution camera mounted on a drone to capture RGB images;

[0060] S12. Perform Gaussian filtering for noise reduction, geometric correction, and cropping on the RGB image in sequence to complete the preprocessing.

[0061] S13. Annotate the preprocessed RGB image and generate the corresponding annotation mask;

[0062] S14. Divide the labeled RGB image into segments to generate a training set.

[0063] Data Acquisition: A drone equipped with a high-resolution camera (pixel resolution ≥ 0.1m / pixel) was used to conduct aerial photography in the distribution area of ​​the Fagus genus (covering different seasons, terrains, and lighting conditions) to acquire RGB images and avoid quality defects such as blurring and overexposure.

[0064] Data preprocessing: The acquired images are processed sequentially by "Gaussian filtering for noise reduction → geometric correction (correcting aerial distortion) → uniform cropping (size 512×512 pixels)" to select images that meet the quality requirements;

[0065] Pixel-level annotation: Forestry experts used LabelMe, a professional tool, to annotate the pre-processed images, clearly distinguishing between "beech crown (target type)" and "understory vegetation / shade / other tree species (background type)", and generating corresponding annotation masks;

[0066] Dataset partitioning: The labeled images were divided into training and validation sets in an 8:2 ratio to ensure that the distribution of beech and environmental conditions was consistent across the three datasets and to avoid data bias.

[0067] In this embodiment of the invention, S2 includes the following sub-steps:

[0068] S21. The RGB image is processed by cloud and fog simulation, local overexposure simulation, dynamic motion blur simulation and random occlusion simulation to generate four types of enhanced samples.

[0069] S22. Combine the four types of augmented samples and the training set to generate an expanded robust training sample set.

[0070] In this embodiment of the invention, in S21, performing cloud and fog simulation specifically involves: randomly sampling on the image plane. A circular region is used to smooth the generated mask using a Gaussian filter, resulting in a continuous transparency map. Combine the original image with the hazy color Perform weighted fusion to generate enhanced results. Its transformation formula is:

[0071] ;

[0072] in, Indicates the spatial location of the original image ,aisle Pixel values; This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. The function constrains pixel values ​​to Within the range.

[0073] This represents the peak opacity of the transparency coefficient map obtained through Gaussian smoothing. exist Randomly selected within the range; fog color In the high-brightness white area The internal randomization is used to simulate different lighting conditions.

[0074] When simulating haze or scattered light scenes, the ES-REF algorithm approximates the contrast attenuation and high-frequency detail loss caused by atmospheric scattering by constructing a local transparency map and linearly blending it with a preset haze color. Specifically, it involves randomly sampling on the image plane. ( The generated mask is smoothed using a Gaussian filter across several circular regions to create a continuous transparency map. Then the original image was compared with the hazy color. Weighted fusion is used to generate enhanced results. This method aims to reduce the model's dependence on fine textures and encourage it to focus on more robust canopy contours and structural features.

[0075] In this embodiment of the invention, S21, the simulation of local overexposure specifically involves: randomly generating a local highlight mask, and then smoothing it using Gaussian to form a gradient brightness field. And perform additive enhancement processing, the specific transformation formula is as follows:

[0076] ;

[0077] in, Represents the maximum brightness increment. Indicates parameters, Indicates the spatial location of the original image ,aisle pixel values, This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. The function constrains pixel values ​​to Within the range.

[0078] Represents the maximum brightness increment, in Randomly selected within the range to control the peak amplitude of the additional light intensity; This represents a smoothed brightness field, whose initial opacity is determined by the parameter. The decision is made based on this parameter. Random sampling is performed within a certain range to ensure natural saturation levels. The smoothing kernel size is determined based on the image resolution. Dynamic adjustments are made. The range of exposure intensity coefficient values ​​is as follows: .

[0079] To address brightness distortion caused by localized overexposure or strong light reflection, ES-REF designed a lighting occlusion simulation module based on brightness superposition. This module generates randomized local specular masks, which are then Gaussian smoothed to create a gradient brightness field. And perform additive enhancement processing to simulate the loss of detail caused by overexposure.

[0080] In this embodiment of the invention, S21, the dynamic motion blur simulation specifically involves: randomly generating the blur kernel length and rotation angle, constructing a directional linear convolution kernel, performing convolution to obtain a blurred image, and then weightedly fusing the blurred image with an RGB image to complete the dynamic motion blur simulation, the expression of which is:

[0081] ;

[0082] ;

[0083] in, Indicates the spatial location of the original image ,aisle Pixel values; This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. The motion-blurred image is obtained by convolving it with a rotated linear kernel; is the point spread function, representing the generated linear convolution kernel used to describe the motion blur path during camera movement. The function constrains pixel values ​​to Within the range, The intensity blending coefficient controls the blending ratio between the blurred image and the original image. This represents convolution.

[0084] , .

[0085] In this embodiment of the invention, S21, performing random occlusion simulation specifically involves: randomly selecting a rectangular region on the RGB image, setting the pixels within the rectangular region to constant values, and generating an occlusion image, the expression of which is:

[0086] ;

[0087] in, This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. Indicates the spatial location of the original image ,aisle pixel values, Indicates the spatial location of the original image ,aisle pixel values, This indicates the fill value for the occluded area.

[0088] In this embodiment of the invention, in the attention context guidance network, the encoder's SPConv module includes a representative channel and a redundant channel divided proportionally.

[0089] The representative channel undergoes convolution processing; the redundant channel undergoes convolution processing and global average pooling processing, and then weighted fusion is performed to extract features.

[0090] In this embodiment of the invention, in the attention context guidance network, the encoder's CGB-Down module performs standard convolution and dilated convolution on the features to obtain joint features, performs global average pooling and multilayer perceptron processing on the joint features to generate global context weights, and uses the global context weights to process the joint features.

[0091] In this embodiment of the invention, in the attention context guidance network, the Criss-Cross attention module of the decoder convolves the highest-level semantic features output by the encoder to generate Q vector, K vector and V vector, and calculates the correlation between Q vector and K vector in the same row or column to obtain attention map, and generates global context information through attention map.

[0092] In this embodiment of the invention, an improved encoder-decoder architecture is adopted, integrating three core modules sequentially. The specific construction steps are as follows:

[0093] Unet infrastructure construction: Based on the VGG network, a symmetrical architecture of encoder + decoder is constructed. The encoder is responsible for feature extraction, and the decoder is responsible for feature recovery and segmentation.

[0094] Model optimization:

[0095] Replace standard convolutional layers with SPConv modules: proportionally Divide the input channels into representative channels ( ) and redundant channels ( Representative channels via " Group convolution+ "Point convolution" fusion, redundant channels Convolutional processing is then performed, followed by global average pooling to generate attention weights. , According to the formula Weighted fusion enables multi-scale feature extraction. Indicates redundant channels Features obtained through convolution processing Indicates representative channel via " Group convolution+ Features obtained by "point convolution".

[0096] Replace the max pooling layer with the CGB-Down module: via " Standard convolution (local features) + "Diffuse convolution (surrounding context)" is fused into joint features, and then "global average pooling + MLP" is used to generate global context weights. Channel-level recalibration is performed on the joint features to preserve spatial details.

[0097] Embedding a Criss-Cross attention module: After the highest-level semantic features of the encoder, through... Convolution generates Q, K, and V vectors, and the correlation between Q and K in the same row / column is calculated to obtain an attention map.

[0098] End-to-end training of the ACG-Net model:

[0099] Training parameter configuration: Input the robust training sample set into ACG-Net with batch size=8, use cross-entropy loss as the total loss function, Adam optimizer (initial learning rate 1e-4), and cosine annealing learning rate adjustment strategy.

[0100] Model training: Iterative training is performed for 100 rounds in a GPU environment. After each round of training, the average intersection-over-union ratio and average precision are evaluated using the validation set. The model weights with the best performance on the validation set are saved.

[0101] Model fine-tuning: If the average intersection-union ratio of the validation set is <88%, adjust the hyperparameters of ACG-Net training and the simulation parameters of ES-REF (such as the range of α(x,y) and the value of γ), and repeat steps 2-4 until the model performance meets the target.

[0102] Inference from monitoring Fagus communities:

[0103] Preprocessing of images to be monitored: Acquire new aerial images of the distribution area of ​​Fagus sylvatica, and perform denoising, geometric correction, and cropping (unified to 512×512 pixels).

[0104] Model inference: The preprocessed image is input into the trained ACG-Net model, and the pixel-level segmentation map is output through forward propagation calculation (the target class is the canopy of the beech tree).

[0105] Monitoring results output: Statistics on the area, distribution density, and spatial pattern of the canopy of Fagus sylvatica are compiled, and visual monitoring maps and text reports are generated to support forestry resource management decisions.

[0106] Regarding model robustness, as shown in Table 1, the ES-REF framework effectively improves the model's adaptability to complex field conditions by systematically simulating complex disturbance factors in real forest environments. Experimental data show that this enhancement framework increases the average intersection-union ratio (IU / REF) of U-Net and ACG-Net by 0.71 and 2.18 percentage points, respectively, demonstrating its ability to significantly narrow the gap between training data and real-world application scenarios. In particular, for environmental factors that traditional enhancement methods struggle to effectively simulate, such as cloud cover and sudden changes in illumination, this framework provides a more realistic simulation scheme, significantly improving the model's performance in these challenging scenarios.

[0107] Table 1

[0108]

[0109] Regarding model performance, as shown in Table 2, the ACG-Net network achieved an average intersection-over-union (IoU) ratio of 89.55% and an average pixel accuracy of 94.48% on the test set. Compared to the U-Net baseline model, the IoU ratio was improved by 4.31 percentage points. This performance improvement is mainly attributed to the synergistic effect of the three core modules in the network structure: the SPConv module significantly improves the ability to recognize multi-scale tree canopies by optimizing the feature extraction process; the Criss-Cross attention module effectively enhances the model's ability to distinguish similar tree species by establishing global contextual dependencies; and the context-guided downsampling mechanism significantly improves the segmentation quality of tree canopy boundaries by preserving key spatial details.

[0110] Table 2

[0111]

[0112] Compared with existing mainstream segmentation models, as shown in Table 3, our proposed solution surpasses advanced models such as DeepLabv3+ and SegFormer in segmentation accuracy while maintaining reasonable computational complexity, demonstrating its competitive advantage in forest remote sensing segmentation tasks. Furthermore, the high-resolution aerial segmentation dataset of the *Fagus* genus constructed using this solution provides a valuable data foundation for research in this field, promoting the development of related technologies.

[0113] Table 3

[0114]

[0115] In summary, this technical solution not only provides a reliable technical means for automated mapping of Fagus communities, but its proposed environmental simulation enhancement strategy and network structure innovation also provide valuable references for other forestry remote sensing applications, and has important theoretical research significance and practical application value.

[0116] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A monitoring method for the genus *Fagus* based on enhanced environmental simulation and attention guidance, characterized in that, Includes the following steps: S1. Use the high-resolution camera on the drone to collect RGB images, perform preprocessing, and generate a training set; S2. Simulate environmental interference in the RGB image to generate enhanced samples, and merge them with the RGB image to generate an expanded robust training sample set; S3. Construct an attention context-guided network; S4. Train the attention context guidance network using the expanded robust training sample set; S5. Input the image to be monitored into the trained attention context guidance network to generate several indicators; S2 includes the following sub-steps: S21. The RGB image is processed by cloud and fog simulation, local overexposure simulation, dynamic motion blur simulation and random occlusion simulation to generate four types of enhanced samples. S22. Combine the four types of augmented samples and the training set to generate an expanded robust training sample set; In step S21, the cloud and fog simulation specifically involves: randomly sampling on the image plane. A circular region is used to smooth the generated mask using a Gaussian filter, resulting in a continuous transparency map. Combine the original image with the hazy color Perform weighted fusion to generate enhanced results. Its transformation formula is: ； in, Indicates the spatial location of the original image ,aisle pixel values, This represents the pixel value at the corresponding location after enhancement using a haze / scattered light simulation. The function constrains the pixel value to Within the range.

2. The monitoring method for *Fagus* species based on enhanced environmental simulation and attention guidance according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Use a high-resolution camera mounted on a drone to capture RGB images; S12. Perform Gaussian filtering for noise reduction, geometric correction, and cropping on the RGB image in sequence to complete the preprocessing. S13. Annotate the preprocessed RGB image and generate the corresponding annotation mask; S14. Divide the labeled RGB image into segments to generate a training set.

3. The monitoring method for *Fagus* species based on enhanced environmental simulation and attention guidance according to claim 1, characterized in that, In step S21, the simulation of local overexposure is specifically performed by: randomly generating a local highlight mask, and then smoothing it using Gaussian to form a gradient brightness field. And perform additive enhancement processing, the specific transformation formula is as follows: ； in, Represents the maximum brightness increment. Indicates parameters, Indicates the spatial location of the original image ,aisle pixel values, This represents the pixel value at the corresponding location after localized overexposure simulation enhancement. The function constrains the pixel value to Within the range.

4. The monitoring method for *Fagus* species based on enhanced environmental simulation and attention guidance according to claim 1, characterized in that, In step S21, the dynamic motion blur simulation specifically involves: randomly generating the blur kernel length and rotation angle, constructing a directional linear convolution kernel, performing convolution to obtain a blurred image, and then weightedly fusing the blurred image with the RGB image to complete the dynamic motion blur simulation. The expression for this is: ； ； in, Indicates the spatial location of the original image ,aisle Pixel values; This represents the pixel value at the corresponding location after enhancement by dynamic motion blur simulation. This indicates that the motion-blurred image is obtained by convolving it with a rotated linear kernel. Represents the point spread function. The function constrains the pixel value to Within the range, Indicates the intensity mixing coefficient. This represents convolution.

5. The monitoring method for *Fagus* species based on enhanced environmental simulation and attention guidance according to claim 1, characterized in that, In step S21, the random occlusion simulation specifically involves: randomly selecting a rectangular region on the RGB image, setting the pixels within the rectangular region to constant values, and generating an occlusion image, the expression of which is: ； in, This represents the pixel value at the corresponding location after random occlusion simulation enhancement. Indicates the spatial location of the original image ,aisle pixel values, Indicates the spatial location of the original image ,aisle pixel values, This indicates the fill value for the occluded area.

6. The monitoring method for *Fagus* species based on enhanced environmental simulation and attention guidance according to claim 1, characterized in that, In the attention context guidance network, the encoder's SPConv module includes a representative channel and a redundant channel divided proportionally. The representative channel undergoes convolution processing; the redundant channel undergoes convolution processing and global average pooling processing, and then weighted fusion is performed to extract features.

7. The monitoring method for *Fagus* species based on enhanced environmental simulation and attention guidance according to claim 1, characterized in that, In the attention context guidance network, the encoder's CGB-Down module performs standard convolution and dilated convolution on the features to obtain joint features. The joint features are then processed by global average pooling and multilayer perceptron to generate global context weights. These global context weights are then used to process the joint features.

8. The monitoring method for *Fagus* species based on enhanced environmental simulation and attention guidance according to claim 1, characterized in that, In the attention context guidance network, the decoder's Criss-Cross attention module convolves the highest-level semantic features output by the encoder to generate Q vectors, K vectors, and V vectors, and calculates the correlation between Q vectors and K vectors in the same row or column to obtain an attention map, which is then used to generate global context information.

Images