Tea garden identification method and system based on unmanned aerial vehicle image

Abstract

The invention discloses a tea garden identification method and system based on an unmanned aerial vehicle image. The method comprises the steps that an improved UNet model is adopted as a segmentation network, and the improved UNet model comprises a visual Transform module and a cavity convolution module in an encoder: A, the visual Transform module is utilized to segment an input image into patch sequences, and global semantic dependency features are established through a multi-head self-attention mechanism; b, extracting multi-scale local features by using the cavity convolution module and adopting a multi-stage expansion rate combination; and C, the decoder fuses the global semantic dependency features and the multi-scale local features output by the encoder through jump connection to generate a tea garden segmentation result.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing image processing technology, specifically to a method and system for tea garden identification based on UAV imagery. Background Technology

[0002] As a globally important economic crop, the expansion of tea cultivation, while driving regional economic development, has also led to problems such as deforestation and a decline in ecosystem services. Efficiently acquiring information on the spatial distribution of tea gardens is crucial for accurate monitoring, but traditional manual surveys are inefficient. While satellite remote sensing can monitor large areas, its resolution limitations (typically ≥10 meters) make it difficult to identify fragmented tea gardens and clear boundaries, and it is also susceptible to cloud interference, leading to data loss. Especially in complex mountainous tea-growing areas, the phenomena of "same species, different spectra" (tea gardens have similar spectra to trees) and "same spectra, different species" (tea gardens are often confused with shrubs) are prominent, and the accuracy of traditional classification methods is often below 70%.

[0003] The existing technologies mainly include three categories: (1) Traditional machine learning methods (such as SVM and random forest): rely on manually designed spectral indices (such as NDVI) or texture features, and the classification process is easily affected by similar land features. For example, the spectral overlap between tea gardens and arbor forests leads to a classification accuracy of less than 65%, and the misclassification rate in mixed vegetation areas exceeds 30%. (2) Hyperspectral / multispectral methods: improve the ability to distinguish between tea gardens and arbor forests by subdividing the bands (such as the 32-band "Zhuhai-1" satellite), but the data dimension is high and the processing is complex, facing the "curse of dimensionality" problem, and the resolution of free data sources is low (≥8 meters), which is difficult to meet the needs of refined management. (3) Deep learning methods: semantic segmentation models based on CNN (such as DeepLabV3+) significantly improve accuracy through automatic feature extraction (F1 can reach 0.885), but under the resolution of satellite images, the false negative rate of small tea gardens is still over 15%, and the problem of blurred boundary segmentation is prominent.

[0004] Drone remote sensing, with its mobility, flexibility, and high resolution (up to 0.2 meters), has become a new solution for tea garden monitoring. For example, the DJI Phantom 4 Pro, equipped with a multispectral sensor, can complete a 6km² survey in a single day on complex terrain, clearly capturing subtle differences in crop details. However, traditional drone data processing relies on manual feature extraction (such as template matching and image binarization), which is cumbersome and has poor generalization capabilities. Although deep learning models have been introduced (such as the improved CGS-YOLO algorithm with YOLOv8 achieving 97.2% mAP in crop identification), existing models still face two major bottlenecks in tea garden identification: insufficient global semantic modeling: CNN models are limited by local receptive fields, making it difficult to establish long-range spatial dependencies, leading to missegmentation of large-scale tea garden distributions; and loss of detailed features: downsampling operations weaken the expression of small tea gardens and boundary features, with the model's IoU for tea gardens <100m² generally below 0.7.

[0005] Recent studies have attempted to introduce attention mechanisms (such as CBAM and ECANet) to enhance feature selection capabilities, or to use dilated convolutions to expand the receptive field. For example, the hybrid attention module CBAM significantly improves performance in general object detection tasks by using channel-spatial attention-weighted features; DeepLabV3+ combines dilated spatial pyramids (ASPP) to capture multi-scale context, but has high computational complexity (178 GFLOPs) and poor real-time performance; while Vision Transformer (ViT) can model global dependencies, its hard block division leads to smoothing of details and weak edge preservation capabilities.

[0006] Therefore, there is an urgent need for an identification scheme that takes into account long-range semantic understanding, multi-scale detail preservation, and efficient computation to meet the technical requirements of precision agriculture for dynamic monitoring of tea gardens. Summary of the Invention

[0007] To address the technical challenges of high misclassification rates in traditional machine learning methods, low resolution in hyperspectral / multispectral methods (failing to meet the demands of refined management), and blurred boundary segmentation issues in deep learning methods during tea garden identification, this invention proposes a tea garden identification method and system based on UAV imagery. It provides an ultra-high resolution tea garden extraction method based on attention and dilated convolution embedded in a UNet. By introducing attention mechanisms and dilated convolution characteristics into the UNet, it enables effective feature fusion and attention control across feature maps of different scales and resolutions, improving the model's ability to learn complex features and extract tea gardens of varying sizes. This enhances the accuracy and robustness of tea garden semantic segmentation, minimizes information loss during feature extraction, and increases the accuracy of small-scale tea garden detail extraction, improving the clarity of tea garden boundaries in high-resolution images and the overall extraction effect of small-scale tea gardens.

[0008] To achieve this goal, the present invention adopts the following technical solution: A tea garden identification method based on UAV imagery employs an improved UNet model as the segmentation network. The improved UNet model includes a visual Transformer module and a dilated convolution module in the encoder, and includes the following steps: A. The input image is segmented into a patch sequence using the visual Transformer module, and global semantic dependency features are established through a multi-head self-attention mechanism. B. Using the dilated convolution module, multi-level dilation rate combinations are employed to extract multi-scale local features; C. The decoder generates tea garden segmentation results by fusing the global semantic dependency features and multi-scale local features output by the encoder through skip connections.

[0009] In the tea garden identification method based on UAV imagery of the present invention, the step of the visual Transformer module segmenting the input image into a patch sequence and establishing global semantic dependencies through a multi-head self-attention mechanism includes: ResNet34 is used as the front-end feature extraction network to encode the input image into a high-dimensional feature map. Then, each spatial location of the feature map is treated as a token, and its channel dimension is used as an embedding vector to form a token sequence. A learnable one-dimensional positional code is added to this sequence to preserve spatial positional information. Finally, feature transformation is performed through a multi-layer Transformer encoder. Each layer contains a multi-head self-attention sub-layer and a multi-layer perceptron sub-layer. Residual connections and layer normalization are used to model global semantic dependencies.

[0010] In the tea garden identification method based on UAV imagery of the present invention, the dilated convolution module adopts a three-level dilation rate, each level contains parallel convolutional layers, and the output feature maps are spliced ​​and fused through channels.

[0011] In the tea garden identification method based on UAV imagery of the present invention, the encoder of the improved UNet model uses ResNet-34 as the backbone network, and its residual block output feature map is used as the input of the visual Transformer module and the dilated convolution module.

[0012] In the tea garden identification method based on UAV imagery of the present invention, a preprocessing operation of the input image is included before step A. The preprocessing operation includes: Acquire RGB images of the drone, with a spatial resolution ≤ 0.2 meters and a flight altitude ≤ 300 meters; Geometric correction was performed on the RGB images of the UAV, with control point correction errors RMSEx and RMSEy ≤ 0.2 meters; The drone RGB images are cropped to a preset size and then divided into training and validation sets.

[0013] In the tea garden identification method based on UAV imagery of the present invention, image recognition training is performed after obtaining the training set. The image recognition training adopts data augmentation operations, including random rotation and overlapping cropping. Overlapping areas are fused by weighted average to eliminate splicing artifacts.

[0014] In the tea garden identification method based on UAV imagery of this invention, the output of the UNet model is evaluated by the following metrics: Intersection over Union (IoU), F1 score, and Kappa coefficient, where the formula for calculating IoU is: Where TP represents the number of true positive pixels, FP represents the number of false positive pixels, and FN represents the number of false negative pixels.

[0015] The present invention also includes a tea garden identification system based on drone imagery, the tea garden identification system comprising: The drone image acquisition module is equipped with an RGB sensor to acquire drone RGB images with a spatial resolution of ≤0.2 meters; The preprocessing module is used to perform geometric correction on the UAV RGB images, with control point correction errors RMSEx and RMSEy ≤ 0.2 meters; and to crop the UAV RGB images to a preset size and divide them into training and validation sets. The segmentation module employs an improved UNet model as the segmentation network. The improved UNet model includes a visual Transformer module and a dilated convolution module in the encoder. The visual Transformer module segments the input image into a sequence of patches and establishes global semantic dependency features through a multi-head self-attention mechanism. The dilated convolution module extracts multi-scale local features based on a combination of multi-level dilation rates. The decoder fuses the global semantic dependency features and multi-scale local features output by the encoder through skip connections to generate the tea garden segmentation result. The visualization module is used to generate residual analysis plots and mark segmented error regions.

[0016] In the tea garden recognition system based on UAV imagery of the present invention, generating a residual analysis map to mark segmentation error areas includes: calculating the pixel-level difference between the segmentation result and the real label; marking areas with errors exceeding a preset threshold as a first color, and marking missed detection areas as a second color.

[0017] The tea garden identification method and system based on UAV imagery of the present invention have the following technical effects.

[0018] This invention presents a tea garden identification method and system based on UAV imagery, which integrates an improved UNet model combining Visual Transformer (ViT) and dilated convolution. On 0.2-meter resolution UAV imagery, it achieves an Intersection over Union (IoU) of 0.9047 and an F1 score of 0.9498 for tea gardens, representing an improvement of 1.24-3.28 percentage points compared to mainstream models such as DeepLabV3+ and SegFormer. This overcomes the technical bottlenecks of high false negative rates (originally >15%) and blurred boundaries in small tea gardens (<100m²). The ViT module of this invention models long-range spatial dependencies through a self-attention mechanism, solving the problem of missegmentation in complex mountainous tea garden distributions. Multi-level dilated convolution (d=6 / 12 / 18) preserves multi-scale details, improving the clarity of small tea garden boundaries by 35%.

[0019] Furthermore, the tea garden identification method and system based on UAV imagery of this invention optimizes the complex terrain features in UAV remote sensing images of tea gardens by fusing attention mechanisms and dilated convolution techniques. The self-attention module in the model captures long-range dependencies between different regions in the remote sensing imagery, highlighting key targets; while multi-scale dilated convolution significantly expands the receptive field, enabling the model to obtain richer contextual information while maintaining high-resolution features. A dynamic spatial channel fusion attention module is used to extract long-range dependencies, while an extensible decoder is introduced to expand the receptive field, thereby reducing detail loss while preserving global features. The attention mechanism of the tea garden identification method and system based on UAV imagery of this invention enhances the fusion and highlighting of features at different scales, while dilated convolution helps the network efficiently capture multi-scale targets. This structural improvement enables the model to adaptively segment large-leaf tea crowns, fine branches, and complex background boundaries in tea garden scenes, thus achieving a balance between detail restoration and global recognition. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a tea garden identification method based on UAV imagery according to a specific embodiment of the present invention.

[0021] Figure 2 This is a diagram illustrating the overall network architecture of a tea garden identification method based on UAV imagery according to a specific embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the dilated convolution method according to a specific embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings.

[0024] The following detailed exemplary embodiments are disclosed. However, the specific structural and functional details disclosed herein are merely for the purpose of describing exemplary embodiments.

[0025] However, it should be understood that the present invention is not limited to the specific exemplary embodiments disclosed, but covers all modifications, equivalents, and substitutions falling within the scope of this disclosure. Throughout the description of the drawings, the same reference numerals denote the same elements.

[0026] Referring to the accompanying drawings, the structures, proportions, sizes, etc., depicted in the drawings are merely for illustrative purposes to aid those skilled in the art in understanding and reading the content disclosed herein. They are not intended to limit the conditions under which the invention can be implemented and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by the invention, should still fall within the scope of the technical content disclosed herein. Furthermore, the positional limitations used in this specification are merely for clarity of description and are not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.

[0027] It should also be understood that the term “and / or” as used herein includes any and all combinations of one or more of the related listed items. Furthermore, it should be understood that when a component or unit is referred to as “connected” or “coupled” to another component or unit, it may be directly connected or coupled to the other component or unit, or there may be intermediate components or units. In addition, other words used to describe the relationship between components or units should be understood in the same manner (e.g., “between” versus “directly between,” “adjacent” versus “directly adjacent,” etc.).

[0028] Figure 1 This is a flowchart illustrating a tea garden identification method based on UAV imagery according to a specific embodiment of the present invention. As shown in the figure, the specific embodiment of the present invention includes a tea garden identification method based on UAV imagery, employing an improved UNet model as the segmentation network. The improved UNet model includes a visual Transformer module and a dilated convolution module in the encoder. A. The input image is segmented into a patch sequence using the visual Transformer module, and global semantic dependency features are established through a multi-head self-attention mechanism. B. Using the dilated convolution module, multi-level dilation rate combinations are employed to extract multi-scale local features; C. The decoder generates tea garden segmentation results by fusing the global semantic dependency features and multi-scale local features output by the encoder through skip connections.

[0029] UNet is a commonly used image segmentation model. Its core features include a symmetrical U-shaped encoder-decoder structure. The encoder downsamples to extract multi-scale semantic features, while the decoder upsamples to gradually restore spatial resolution. Skip connections are used to directly fuse low-level detail features (such as edges) from the encoder with high-level semantic features (such as target contours) from the decoder, thus mitigating information loss and improving the segmentation accuracy of small targets. This model performs exceptionally well in training with few samples.

[0030] like Figure 2 As shown, the input image for the tea garden identification method based on UAV imagery in this invention is an RGB image from a remote sensing measurement project taken by a UAV. The UAV RGB image is acquired using a rotary-wing aircraft equipped with an RGB sensor. In one example, the RGB sensor is a 1-inch 20-megapixel CMOS sensor with a lens field of view of 84 degrees and a focal length range of f / 2.8 to f / 11. The typical UAV flight altitude is 300m, with a vertical positioning accuracy of 0.1 and a horizontal positioning accuracy of 0.3. The UAV flight is conducted under conditions of wind speed of 2-3 m / s and no precipitation, acquiring multiple images of the area to be detected within a safe flight time.

[0031] To facilitate subsequent processing, data preprocessing is performed first. This includes geometric correction, image stitching, and image cropping of the UAV remote sensing images. First, the control regions in the UAV RGB images are corrected to control points. A terrain-residual regression model is then constructed for accuracy verification. The results show that the pixel errors RMSEx and RMSEy are within the predetermined range, and the maximum residual is also within an acceptable range, proving that there is no systematic bias. The corrected data is then stitched together. During the stitching process, to avoid seam problems, a weighted average method is used to process the images to ensure a natural transition. The mask extraction tool in ArcGIS Pro 3.4 is used to extract the target geographical area from the image, and this extracted area serves as the basis for deep learning extraction of the tea garden.

[0032] Specifically, in the tea garden recognition method based on UAV imagery of this invention, a multi-head self-attention module and a dilated convolution module are added to the UNet encoder. The multi-head self-attention module directly establishes long-range dependencies between any two locations in the image, solving the problem of insufficient global information modeling caused by the local receptive field in existing CNNs. Simultaneously, the embedded dilated convolution module expands the receptive field without reducing the feature map resolution, thereby enhancing the ability to perceive multi-scale contextual information in UAV RGB images.

[0033] The primary objective of dilated convolution methods is to expand the receptive field without reducing spatial resolution through downsampling, making them suitable for multi-scale image segmentation. In a more specific embodiment of the invention, dilated convolutions with dilation rates (d) of 6, 12, and 18 are applied to capture tea garden context features at different scales. Smaller dilation rates (d=6) help preserve local details, such as tea garden boundaries, while intermediate dilation rates (d=12) balance local and global context. In contrast, larger dilation rates (d=18) expand the receptive field to cover a wider range of spatial patterns. By using this progressive design, the network can model both fine-grained structures and large-scale dependencies simultaneously. At the same time, it mitigates the potential meshing effect caused by excessively large dilation rates. This setup follows common practice in semantic segmentation, where d=6, 12, and 18 are frequently used to achieve reliable multi-scale feature representations.

[0034] Multi-head self-attention is a core component of the Transformer encoder. By computing multiple independent attention heads in parallel, it captures information from different subspaces of the input sequence, thereby enhancing the ability to model complex contextual relationships. The multi-head self-attention mechanism divides the input parameters into three matrices—Query, Key, and Value—through a linear transformation, with each matrix corresponding to information from a different subspace. These three matrices are then divided into multiple heads, each independently calculating its attention score and generating a weighted output. Finally, the results from each head are fused through a linear layer.

[0035] Global semantic-dependent features refer to feature representations that enhance the model's discriminative power by integrating the global structure and contextual information of an image. They primarily encompass two core capabilities: (1) The role of global semantic-dependent features: By capturing the overall structure and contextual information of an image, global semantic-dependent features provide high-level semantic guidance for visual understanding, effectively handling occlusion or background interference issues. For example, in semantic segmentation, global features can help the model understand the spatial relationships between objects (such as the relative positions of vehicles and roads), thereby improving segmentation accuracy. (2) Complementarity with local features: Local features focus on fine-grained patterns (such as texture and edges), compensating for the lack of detail in global features. The fusion of both can enhance the model's adaptability to complex scenes: local features handle detailed differences, while global features provide semantic constraints, forming a dual guarantee of "details + structure."

[0036] Specifically, the tea garden identification method based on UAV imagery in this invention uses the encoder's results as input and employs high-dimensional features to proceed with subsequent processes. The later decoder uses a mixture of results from different modules to upsample and finally reconstruct the image.

[0037] Specifically, the image recognition process includes a shrinking path and an expanding path. The shrinking path uses a pre-trained ResNet-34 as a feature extractor, and the feature map output by each residual block of ResNet-34 is used as the encoded feature. Specifically, the convolution and pooling operations of ResNet-34 progressively reduce the spatial resolution and increase the number of feature channels, similar to the convolution and downsampling steps of the original UNet.

[0038] Each step in the expansion path involves upsampling the feature map, then halving the number of feature channels using a 2x2 transposed convolution, concatenating it with the corresponding cropped feature map from the contraction path, and finally applying two 3x3 convolutions and a ReLU activation function. Skip connections introduce high-resolution detail information from the feature maps extracted from each stage of ResNet-34. Since boundary pixels are lost with each convolution, the feature map needs to be cropped to align dimensions.

[0039] In the final layer, 1x1 convolutions are used to map each feature channel to the required number of categories, and the output segmentation map can be seamlessly tiled.

[0040] Figure 2 This diagram illustrates the overall network architecture of a tea garden identification method based on UAV imagery according to a specific embodiment of the present invention. Specifically, in the overall network architecture of this embodiment, the encoder output is passed to subsequent stages where high-dimensional feature representations are further processed. The decoder then reconstructs the image by upsampling the fused output from the heterogeneous modules. The model consists of a shrinking path and an expanding path. A pre-trained ResNet-34 is used as the feature extractor for the shrinking path, employing the feature map generated by each residual block as the encoded feature. The convolution and pooling operations in ResNet-34 progressively reduce the spatial resolution while increasing the channel depth, similar to the convolution and downsampling steps of the original UNet.

[0041] Each step of the expansion path involves upsampling the feature map followed by a 2×2 transposed convolution (i.e., "upconvolution"), halving the number of feature channels. The output of the upconvolution is concatenated with the corresponding cropped feature map from the contraction path, and then processed by two consecutive 3×3 convolutions (with ReLU activation). High-resolution details are reintroduced through skip connections, passing feature maps from each ResNet-34 stage. Since standard convolution operations reduce spatial extent at boundaries, the feature maps are cropped before concatenation to ensure dimensional alignment. In the final layer, a 1×1 convolution is applied to map each feature channel to the desired number of semantic categories. The resulting segmentation map can then be seamlessly tiled.

[0042] In the tea garden identification method based on UAV imagery of the present invention, the step of the visual Transformer module segmenting the input image into a patch sequence and establishing global semantic dependencies through a multi-head self-attention mechanism includes: ResNet34 is used as the front-end feature extraction network to encode the input image into a high-dimensional feature map. Then, each spatial location of the feature map is treated as a token, and its channel dimension is used as an embedding vector to form a token sequence. A learnable one-dimensional positional code is added to this sequence to preserve spatial positional information. Finally, feature transformation is performed through a multi-layer Transformer encoder. Each layer contains a multi-head self-attention sub-layer and a multi-layer perceptron sub-layer. Residual connections and layer normalization are used to model global semantic dependencies.

[0043] Specifically, for the visual Transformer module, the standard Transformer accepts a one-dimensional token embedding sequence as input. To process 2D drone RGB images, the image needs to be... Convert to 2D patch sequence Where (H,W) is the resolution of the original image, C is the number of channels, and (P,P) is the resolution of each image patch. This is the number of patches generated, which also serves as the effective input sequence length for the Transformer module. The Transformer module uses a constant vector size D across all its layers; therefore, in this specific embodiment, a trainable linear projection is mapped to the D dimension, and the output of this projection is embedded as a patch.

[0044] The main goal of dilated convolution is to increase the size of the receptive field in the network, avoid downsampling, and generate a multi-scale frame for segmentation. The basic principle of dilated convolution is as follows: Figure 3 As shown. (a) F1 is generated from F0 via a 1-fold dilated convolution; the receptive field of each element in F1 is 3×3. (b) F2 is generated from F1 via a 2-fold dilated convolution; each element in F2 has a 7×7 receptive field. (c) F3 is generated from F2 via a 4-fold dilated convolution; the receptive field of each element in F3 is 15×15. The number of parameters associated with each layer is the same. The receptive field grows exponentially, while the number of parameters grows linearly.

[0045] In the tea garden identification method based on UAV imagery of the present invention, a preprocessing operation of the input image is included before step A. The preprocessing operation includes: Acquire RGB images of the drone, with a spatial resolution ≤ 0.2 meters and a flight altitude ≤ 300 meters; Geometric correction was performed on the RGB images of the UAV, with control point correction errors RMSEx and RMSEy ≤ 0.2 meters; The drone RGB images are cropped to a preset size and then divided into training and validation sets.

[0046] In the tea garden identification method based on UAV imagery of the present invention, image recognition training is performed after obtaining the training set. The image recognition training adopts data augmentation operations, including random rotation and overlapping cropping. Overlapping areas are fused by weighted average to eliminate splicing artifacts.

[0047] In the tea garden identification method based on UAV imagery of this invention, the output of the UNet model is evaluated by the following metrics: Intersection over Union (IoU), F1 score, and Kappa coefficient, where the formula for calculating IoU is: Where TP represents the number of true positive pixels, FP represents the number of false positive pixels, and FN represents the number of false negative pixels.

[0048] Specifically, the evaluation metrics for the UNet model include Precision, Recall, F1 score, Intersection over Union (IoU), and Kappa coefficient. These metrics are used to evaluate the accuracy of image semantic segmentation. For example, Precision and Recall can be expressed as:

[0049]

[0050]

[0051] Where MCC is the Matthews correlation coefficient, TP represents a true positive (both the prediction and reference are positive), FP represents a false positive (the prediction is positive and the reference is negative), and FN represents a false negative (the prediction is negative and the reference is positive).

[0052] The F1 score, IoU (Intersection over Union), and Kappa coefficient are determined as follows:

[0053]

[0054]

[0055] Where p0 is the number of correctly classified samples divided by the total number of samples, Here, it is assumed that the number of real samples for each type is ai, and the number of predicted samples for each type is bi.

[0056] The specific embodiments of the present invention also include a tea garden identification system based on UAV imagery, the tea garden identification system comprising: The drone image acquisition module is equipped with an RGB sensor to acquire drone RGB images with a spatial resolution of ≤0.2 meters; The preprocessing module is used to perform geometric correction on the UAV RGB images, with control point correction errors RMSEx and RMSEy ≤ 0.2 meters; and to crop the UAV RGB images to a preset size and divide them into training and validation sets. The segmentation module employs an improved UNet model as the segmentation network. The improved UNet model includes a visual Transformer module and a dilated convolution module in the encoder. The visual Transformer module segments the input image into a sequence of patches and establishes global semantic dependency features through a multi-head self-attention mechanism. The dilated convolution module extracts multi-scale local features based on a combination of multi-level dilation rates. The decoder fuses the global semantic dependency features and multi-scale local features output by the encoder through skip connections to generate the tea garden segmentation result. The visualization module is used to generate residual analysis plots and mark segmented error regions.

[0057] In the tea garden recognition system based on UAV imagery of the present invention, generating a residual analysis map to mark segmentation error areas includes: calculating the pixel-level difference between the segmentation result and the real label; marking areas with errors exceeding a preset threshold as a first color, and marking missed detection areas as a second color.

[0058] The tea garden identification method and system based on UAV imagery of the present invention have the following technical effects.

[0059] This invention presents a tea garden identification method and system based on UAV imagery. It integrates an improved UNet model combining Visual Transformer (ViT) and dilated convolution, achieving an Intersection over Union (IoU) of 0.9047 and an F1 score of 0.9498 on 0.2-meter resolution UAV imagery. This represents a 1.24-3.28% improvement over existing models such as DeepLabV3+ and SegFormer, overcoming the technical bottlenecks of high false negative rates (originally >15%) and blurred boundaries in small tea gardens (<100m²). The ViT module of this invention models long-range spatial dependencies through a self-attention mechanism, solving the problem of missegmentation in complex mountainous tea garden distributions. Multi-level dilated convolution with dilation rates (d) of 6, 12, and 18 preserves multi-scale details, improving the clarity of small tea garden boundaries by 35%.

[0060] Furthermore, the tea garden identification method and system based on UAV imagery of this invention optimizes the complex terrain features in UAV remote sensing images of tea gardens by fusing attention mechanisms and dilated convolution techniques. The self-attention module in the model captures long-range dependencies between different regions in the remote sensing imagery, highlighting key targets; while multi-scale dilated convolution significantly expands the receptive field, enabling the model to obtain richer contextual information while maintaining high-resolution features. A dynamic spatial channel fusion attention module is used to extract long-range dependencies, while an extensible decoder is introduced to expand the receptive field, thereby reducing detail loss while preserving global features. The attention mechanism of the tea garden identification method and system based on UAV imagery of this invention enhances the fusion and highlighting of features at different scales, while dilated convolution helps the network efficiently capture multi-scale targets. This structural improvement enables the model to adaptively segment large-leaf tea crowns, fine branches, and complex background boundaries in tea garden scenes, thus achieving a balance between detail restoration and global recognition.

[0061] In a specific embodiment of this invention, a key factor is the combination of ultra-high resolution data with a deep network architecture to achieve fine-grained tea garden classification. First, 0.2-meter UAV imagery is employed, providing spatial detail far exceeding that of Sentinel-2 (10 meters) and Gaofen-2 (1 meter) imagery, which in turn greatly enhances detail. This improved spatial fidelity makes the boundary distinction between tea gardens and heterogeneous adjacent cover (such as forests and wasteland) clearer, alleviating the typical "heterogeneous objects with similar spectral density" and "similarity of similar objects with different spectral density" problems of coarse-resolution satellite data. Furthermore, the specific embodiment of this invention combines dilated convolution and attention mechanisms, thus improving the network's ability to capture multi-scale textures and long-range semantic relationships. As a result of the specific embodiment of this invention, this design effectively distinguishes tea gardens from nearby woodlands or bare areas, even when the textures or edges of these areas appear identical. This advancement surpasses earlier studies that primarily focused on coarse plot extraction, producing a finer classification granularity capable of distinguishing tea garden plots from non-vegetated surfaces, and wasteland from sparsely planted tea gardens. Finally, in terms of fine-scale recognition accuracy, the embodiments of the present invention outperform deep learning techniques and multi-source fusion methods applied to Sentinel-2. These results demonstrate that the embodiments of the present invention provide a promising solution for the accurate mapping and management of tea trees by achieving improved classification granularity and accuracy from high-resolution photographs.

[0062] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as mentioned above, it should be understood that the present invention is not limited to the forms disclosed in this specification and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described in this specification through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A tea garden identification method based on UAV imagery, characterized in that, An improved UNet model is used as the segmentation network. The improved UNet model includes a visual Transformer module and a dilated convolution module in the encoder: A. The input image is segmented into a patch sequence using the visual Transformer module, and global semantic dependency features are established through a multi-head self-attention mechanism. B. Using the dilated convolution module, multi-level dilation rate combinations are employed to extract multi-scale local features; C. The decoder generates tea garden segmentation results by fusing the global semantic dependency features and multi-scale local features output by the encoder through skip connections.

2. The tea garden identification method based on UAV imagery as described in claim 1, characterized in that, The visual Transformer module segments the input image into a sequence of patches, and the steps for establishing global semantic dependencies through a multi-head self-attention mechanism include: ResNet34 is used as the front-end feature extraction network to encode the input image into a high-dimensional feature map. Then, each spatial location of the feature map is treated as a token, and its channel dimension is used as an embedding vector to form a token sequence. A learnable one-dimensional positional code is added to this sequence to preserve spatial positional information. Finally, feature transformation is performed through a multi-layer Transformer encoder. Each layer contains a multi-head self-attention sub-layer and a multi-layer perceptron sub-layer. Residual connections and layer normalization are used to model global semantic dependencies.

3. The tea garden identification method based on UAV imagery as described in claim 1, characterized in that, The dilated convolution module employs a three-level dilation rate, with each level containing parallel convolutional layers, and the output feature maps are spliced ​​and fused through channels.

4. The tea garden identification method based on UAV imagery as described in claim 1, characterized in that, The encoder of the improved UNet model uses ResNet-34 as its backbone network, and its residual block output feature maps serve as inputs to the visual Transformer module and the dilated convolution module.

5. The tea garden identification method based on UAV imagery as described in claim 1, characterized in that, Step A precedes preprocessing of the input image, which includes: Acquire RGB images of the drone, with a spatial resolution ≤ 0.2 meters and a flight altitude ≤ 300 meters; Geometric correction was performed on the RGB images of the UAV, with control point correction errors RMSEx and RMSEy ≤ 0.2 meters; The drone RGB images are cropped to a preset size and then divided into training and validation sets.

6. The tea garden identification method based on UAV imagery according to claim 5, characterized in that, After obtaining the training set, image recognition training is performed. The image recognition training adopts data augmentation operations, including random rotation and overlapping cropping. Overlapping regions are fused by weighted average to eliminate splicing artifacts.

7. The tea garden identification method based on UAV imagery as described in claim 1, characterized in that, The output of the UNet model is evaluated using the following metrics: Intersection over Union (IoU), F1 score, and Kappa coefficient. The formula for calculating IoU is: Where TP represents the number of true positive pixels, FP represents the number of false positive pixels, and FN represents the number of false negative pixels.

8. A tea garden identification system based on UAV imagery, characterized in that, system include: The drone image acquisition module is equipped with an RGB sensor to acquire drone RGB images with a spatial resolution of ≤0.2 meters; The preprocessing module is used to perform geometric correction on the UAV RGB images, with control point correction errors RMSEx and RMSEy ≤ 0.2 meters; and to crop the UAV RGB images to a preset size and divide them into training and validation sets. The segmentation module employs an improved UNet model as the segmentation network. The improved UNet model includes a visual Transformer module and a dilated convolution module in the encoder. The visual Transformer module segments the input image into a sequence of patches and establishes global semantic dependency features through a multi-head self-attention mechanism. The dilated convolution module extracts multi-scale local features based on a combination of multi-level dilation rates. The decoder fuses the global semantic dependency features and multi-scale local features output by the encoder through skip connections to generate the tea garden segmentation result. The visualization module is used to generate residual analysis plots and mark segmented error regions.

9. The tea garden identification system based on UAV imagery as described in claim 8, characterized in that, The process of generating a residual analysis map to mark segmentation error regions includes: calculating the pixel-level difference between the segmentation result and the true label; marking regions with errors exceeding a preset threshold with the first color, and marking missed regions with the second color.

Images