Terrain correction-based multi-modal remote sensing tea garden segmentation and area estimation method and system and medium

Abstract

The application discloses a terrain correction-based multi-modal remote sensing tea garden segmentation and area estimation method and system and a medium. The method comprises the following steps: constructing a multi-modal time series dataset, which comprises high-resolution remote sensing image data, medium-resolution time series remote sensing image data and digital elevation model data; inputting the dataset into a terrain correction-based multi-modal tea garden segmentation model, extracting time series phenology aggregation features through a lightweight time series attention encoder, introducing a learnable C correction coefficient through a terrain attention correction module to correct terrain by wave band and fuse the core spatial features, fusing the features through a multi-modal feature fusion module, outputting a tea garden area segmentation result through a multi-scale segmentation module; and correcting terrain projection distortion based on the segmentation result and slope information to obtain a corrected total tea garden area. The application realizes accurate image recognition and area estimation of tea gardens under complex terrain.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of remote sensing image processing, deep learning, and agricultural monitoring information technology, specifically to a multimodal remote sensing method, system, and medium for tea garden segmentation and area estimation based on terrain correction. Background Technology

[0002] Tea gardens, as an important economic crop in my country, are widely distributed and concentrated in complex and fragmented terrain areas such as mountains and hills, making rapid and accurate identification and area estimation of large-scale tea gardens extremely difficult. With the development of remote sensing technology, remote sensing data has evolved into a comprehensive observation base encompassing multiple sources, modalities, and time series, and remote sensing technology is gradually becoming the main technical means for large-scale tea garden identification. Against the backdrop of the current trend of artificial intelligence development, using machine learning or deep learning for large-scale, rapid, and precise crop segmentation has become a research hotspot. Current technical solutions can be mainly divided into two categories: single-modal and multi-modal segmentation. Single-modal methods often rely on a single remote sensing data source (such as high-resolution imagery, medium-resolution multispectral imagery, etc.) to segment tea garden planting areas. The paper "Research on Accurate Identification of Tea Gardens Based on Satellite Remote Sensing Imagery" (authors: Li Dongxue et al., published in the journal *Shandong Agricultural Sciences*, 2025, No. 4) utilizes a high-resolution image-annotated tea garden segmentation dataset to construct a high-precision tea garden segmentation model, focusing on the differences in texture and color features between tea gardens and other land features, thereby achieving the segmentation of tea garden distribution and boundaries. The invention patent with publication date of December 23, 2022 and patent publication number CN115512218A proposes a single-mode method using medium-resolution multispectral images. It mainly uses the spectral band information of the multispectral images to construct spectral characteristic indices such as tea garden phenology index, highlighting the physiological characteristics and phenological differences of tea garden vegetation and ground cover, thereby achieving the segmentation of tea garden distribution.

[0003] Multimodal remote sensing fusion methods, by combining the advantages of different data sources, improve the accuracy of tea garden segmentation to a certain extent compared to methods using single-modal data sources. Patent CN117011702B, published on December 9, 2025, proposes a method for automatically identifying tea gardens based on phenological indices and topographic features from medium-resolution multispectral images of Sentinel-2 and Landsat, constructing a decision tree model and using classification thresholds. Patent CN113221806B, published on February 1, 2022, proposes a method for automatically identifying tea gardens based on a cloud platform that fuses multi-source satellite imagery and tea tree phenology. It captures key phenological periods of tea gardens by fusing Landsat 7 / 8 and Sentinel-2 data, and constructs tea garden growth time-series feature curves by dividing the growth cycle, achieving automated and precise identification of tea gardens.

[0004] Accurate segmentation of tea gardens using remote sensing technology relies on stable and reliable data support. Currently, the segmentation scheme based on single-modal remote sensing images is as follows: The distribution areas of tea gardens in China are fragmented, with large differences in shape and mixed vegetation. In addition, the tea garden planting areas are affected by climate factors such as clouds, rain and fog for a long time. The current technical solution is difficult to solve the problem of accurately segmenting the distribution and boundaries of tea gardens.

[0005] Single-modal segmentation schemes have obvious limitations: although high-resolution images can capture the fine texture features of tea gardens, tea gardens are perennial woody evergreen plants, and their spectral information features are easily confused with evergreen vegetation such as woodlands and shrublands; although multispectral images have rich spectral information and can reflect the physiological characteristics of tea garden vegetation, their spatial resolution is low, making it difficult to accurately capture the details of tea garden boundaries, and their utilization of temporal information is also low.

[0006] In complex and fragmented terrain scenarios, multimodal remote sensing data solutions suffer from significant problems: the large topographic relief in mountainous and hilly areas leads to projection distortion and spectral aberration in remote sensing images, while steep slopes result in noticeable pixel mixing. Existing multimodal fusion methods mostly fuse data from medium-resolution or coarse-resolution sources, failing to consider the influence of terrain factors during the fusion process. This results in segmentation results with blurred boundaries, misclassification, and omissions, especially in terraced fields, steep slopes, and shady slopes, where the segmentation accuracy is insufficient for practical applications.

[0007] Furthermore, existing tea plantation area estimation methods mostly rely on direct pixel statistics based on segmentation results, without correcting for image scale deviations caused by terrain undulations. They also lack differentiated area calculation strategies for different terrain types such as terraces and steep slopes, leading to significant errors in area estimation. The feature fusion methods in existing models are also mostly simple stitching or fixed-weight fusion, failing to adapt to the feature differences of different terrains and data sources, further limiting the accuracy of segmentation and area estimation.

[0008] To address the technical challenges of low accuracy in tea garden segmentation and difficulty in area estimation under complex and fragmented terrain conditions, there is an urgent need for a method that can effectively eliminate the influence of terrain, achieve adaptive fusion of multimodal features, and balance segmentation accuracy with area estimation accuracy. Summary of the Invention

[0009] The purpose of this invention is to provide a multimodal remote sensing tea garden segmentation and area estimation method, system, and medium based on terrain correction. By introducing a terrain correction module to eliminate the influence of terrain undulation on the projection distortion of remote sensing images, and designing an adaptive multimodal feature fusion module, combined with a terrain-weighted loss training strategy, the invention achieves accurate segmentation and accurate area estimation of tea gardens under complex terrain, effectively alleviating the technical problems of low segmentation accuracy and large area estimation error of existing methods in complex terrain scenarios.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multimodal remote sensing method for tea garden segmentation and area estimation based on terrain correction includes the following steps: A multimodal time-series dataset is constructed, which includes high-resolution remote sensing image data, medium-resolution time-series remote sensing image data, and digital elevation model data for the same target area. The multimodal time-series dataset is input into a pre-built multimodal tea garden segmentation model based on terrain correction to obtain tea garden area segmentation results; Based on the tea garden area segmentation results and combined with the slope information in the digital elevation model data, the tea garden area is corrected for topographic projection distortion according to the following formula to obtain the corrected total tea garden area:

[0011] In the formula: This represents the revised total area of ​​the tea plantation. The horizontal projection area of ​​a single pixel in high-resolution remote sensing image data; pixel location of the tea garden area The corresponding radian slope value; This is the set of pixels in the segmentation result that are all identified as tea garden areas.

[0012] Furthermore, the terrain-corrected multimodal tea garden segmentation model processes the input data in the following manner: The core spatial features of high-resolution remote sensing image data are extracted using a high-resolution image spatial feature extraction module. The medium-resolution image multi-temporal phenological feature extraction module uses a lightweight temporal attention encoder to perform temporal modeling on the medium-resolution temporal remote sensing image data. The lightweight temporal attention encoder uses a multi-head attention mechanism to weight and aggregate features from different temporal windows, and outputs temporal phenological aggregated features. The terrain attention correction module calculates the slope and aspect based on the digital elevation model data, and calculates the cosine of the solar incidence angle by combining the solar zenith angle and solar azimuth angle; a learnable C correction coefficient is introduced, and an optimized C correction algorithm is used to perform band-by-band terrain correction on the high-resolution remote sensing image data; the extracted terrain features after correction are fused with the core spatial features by attention to obtain high-resolution core features. The high-resolution core features, the temporal phenological aggregate features, and the spectral information enhancement features are spliced ​​together through the multimodal feature fusion module, and the weights of each channel are dynamically adjusted through the channel attention gating mechanism to obtain the final fused features. The multi-scale segmentation module decodes the final fused features and the multi-scale features output by the high-resolution image spatial feature extraction module to output the segmentation result of the tea garden area.

[0013] Furthermore, the lightweight temporal attention encoder is specifically processed according to the following steps: For each frame of the medium-resolution time-series remote sensing image data, spectral features are initially extracted and then attention weighted by the lightweight three-dimensional attention module SimAM to obtain a time-series spectral feature sequence. The temporal spectral feature sequence is position-encoded using sine and cosine functions to obtain temporal features after injecting position information; The multi-head attention module maps the time-series features after the injection of location information into query vectors, key vectors, and value vectors, and calculates the time-series features after multi-head attention weighting. The temporal features weighted by multi-head attention are aggregated into temporal aggregated features using a multilayer perceptron, and then adjusted to the same spatial scale as the core spatial features through bilinear interpolation to obtain the temporal phenological aggregated features.

[0014] Furthermore, the terrain attention correction module specifically processes the following steps: Obtain the solar zenith angle and solar azimuth angle from high-resolution remote sensing image data, and calculate the cosine of the solar incidence angle using the following formula, combined with the slope and aspect. :

[0015] In the formula: The solar zenith angle; This is the solar azimuth angle; Slope in radians; The slope is in radians; band-by-band adaptive C-correction is performed according to the following formula:

[0016] In the formula: For the first Image data corrected for each band; For the first Raw image data for each band; The learnable C-correction coefficient is initialized to 0.1. For band identification.

[0017] Furthermore, in the terrain attention correction module, the step of performing attention fusion between the corrected extracted terrain features and the core spatial features includes: The core spatial features and the terrain features are concatenated in the channel dimension, and the attention weights of the terrain features are calculated by adaptive average pooling, 1×1 convolution and S-shaped growth curve activation function. The terrain features are multiplied element-wise with their attention weights to obtain weighted terrain features; After concatenating the weighted terrain features with the core spatial features, the high-resolution core features are output through 3×3 convolution, batch normalization, and linear rectified activation function.

[0018] Furthermore, the multi-scale segmentation module specifically includes: constructing a feature pyramid based on the spatial features at multiple scales output by the high-resolution image spatial feature extraction module and the final fused features; processing the feature levels in the feature pyramid through spatial decoding branches and edge decoding branches respectively, and splicing and fusing the outputs of the two branches in the channel dimension to obtain decoded features; upsampling the decoded features to the scale of the original high-resolution remote sensing image, and processing them through a segmentation head containing 3×3 dilated convolutions to output binary segmentation results for the tea garden area and the non-tea garden area.

[0019] In specific implementation, the segmentation head includes the following components connected in sequence: (1) 3×3 dilated convolution with a dilation rate of 2, keeping the number of channels constant, is used to expand the receptive field and extract multi-scale contextual features; (2) The batch normalization layer (BN) is followed by the ReLU activation function, which is used to accelerate convergence and stabilize the feature distribution and nonlinear feature refinement.

[0020] (3) 1×1 convolution is used to map the number of channels to two channels, which is used to distinguish between tea gardens and non-tea gardens; Alternatively, the segmentation head includes: a cascaded 3×3 dilated convolution, a batch normalization layer, a ReLU activation function, and a 1×1 convolution; after processing by the segmentation head, the probability distribution of each pixel belonging to a tea garden or not is output through the Softmax activation function.

[0021] Furthermore, the step of constructing the multimodal time-series dataset further includes: acquiring high-resolution remote sensing image data of the target area during key phenological periods of tea, wherein the high-resolution remote sensing image data is a 4-channel image containing red band, green band, blue band and near-infrared band; acquiring medium-resolution remote sensing images of the target area at multiple time points throughout the entire growth cycle of tea, and constructing medium-resolution time-series remote sensing image data by dividing the time series window and filtering by the maximum value of the normalized difference vegetation index; acquiring digital elevation model data of the target area, and calculating slope and aspect information based on the digital elevation model data.

[0022] Furthermore, the step of screening the maximum value of the normalized difference vegetation index includes: Let the first... Each timing window contains The medium-resolution remote sensing observations correspond to the normalized difference vegetation index (NDVI) image set as follows: Best observation after screening Determined according to the following formula:

[0023] In the formula: For time-series window index; For the first Number of observations within a time window; For the first Normalized difference vegetation index values ​​of each observation. ; This is a function to find the maximum value. For the first The best observation selected from each time series window; in time series locations without observations, if there are valid observations within the time series window, the average of the valid observations within the time series window is used to fill the gap; if there are no valid observations within the time series window, the default value is 0.

[0024] Furthermore, the multimodal tea garden segmentation model based on terrain correction also includes a spectral information feature enhancement and extraction module, which is used to calculate the normalized difference vegetation index, enhanced vegetation index, normalized red edge index and modified soil-regulated vegetation index based on the medium-resolution time-series remote sensing image data, and extract spectral information enhancement features.

[0025] Furthermore, the spatial resolution of the high-resolution remote sensing image data is better than 1 meter; the medium-resolution temporal remote sensing image data comes from the Sentinel-2 satellite; and the digital elevation model data is the ALOS 12.5-meter digital elevation model.

[0026] Furthermore, the step of correcting the topographic projection distortion of the tea garden area further includes: extracting the brightness spatial gradient using the grayscale distribution of the high-resolution remote sensing image data, and constructing a local topography restoration function; using the brightness spatial gradient as an adjustment variable for micro-topographic undulations, performing sub-pixel-level spatial downscaling correction on the initial slope field provided by the digital elevation model data, and generating a sub-pixel-level slope value matrix that corresponds one-to-one with the pixel coordinates of the high-resolution remote sensing image data; based on the sub-pixel-level slope value matrix, performing slope normal projection compensation on the horizontal projection area of ​​each pixel point within the tea garden area, and obtaining the corrected total tea garden area through differential accumulation.

[0027] Furthermore, the step of generating the sub-pixel level slope value matrix includes: The high-resolution remote sensing image data is converted to grayscale, and the pixel-level brightness spatial gradient is calculated. The brightness spatial gradient is standardized to eliminate differences in light intensity; A sensitivity adjustment coefficient is introduced to linearly map the standardized brightness spatial gradient to the correction amount of the initial slope value. By superimposing the correction amount, the local topological smoothing or refinement adjustment of the initial slope value is achieved.

[0028] On the other hand, the present invention also discloses a multimodal remote sensing tea garden segmentation and area estimation system based on terrain correction, comprising: The dataset construction unit is used to construct a multimodal time-series dataset, which includes high-resolution remote sensing image data, medium-resolution time-series remote sensing image data, and digital elevation model data for the same target area. The model segmentation unit is used to input the multimodal time-series dataset into a pre-constructed terrain-corrected multimodal tea garden segmentation model to obtain tea garden area segmentation results; wherein, the terrain-corrected multimodal tea garden segmentation model processes the input data in the following manner: The core spatial features of high-resolution remote sensing image data are extracted using a high-resolution image spatial feature extraction module. The medium-resolution image multi-temporal phenological feature extraction module uses a lightweight temporal attention encoder to perform temporal modeling on the medium-resolution temporal remote sensing image data. The lightweight temporal attention encoder uses a multi-head attention mechanism to weight and aggregate features from different temporal windows, and outputs temporal phenological aggregated features. The terrain attention correction module calculates the slope and aspect based on the digital elevation model data, and calculates the cosine of the solar incidence angle by combining the solar zenith angle and solar azimuth angle; a learnable C correction coefficient is introduced, and an optimized C correction algorithm is used to perform band-by-band terrain correction on the high-resolution remote sensing image data; the extracted terrain features after correction are fused with the core spatial features by attention to obtain high-resolution core features. The high-resolution core features, the temporal phenological aggregate features, and the spectral information enhancement features are spliced ​​together through the multimodal feature fusion module, and the weights of each channel are dynamically adjusted through the channel attention gating mechanism to obtain the final fused features. The multi-scale segmentation module decodes the final fused features and the multi-scale features output by the high-resolution image spatial feature extraction module to output the tea garden area segmentation result. The area correction unit, based on the tea garden area segmentation result and combined with the slope information in the digital elevation model data, corrects the terrain projection distortion of the tea garden area according to the following formula to obtain the corrected total tea garden area:

[0029] In the formula: This represents the revised total area of ​​the tea plantation. The horizontal projection area of ​​a single pixel in high-resolution remote sensing image data; pixel location of the tea garden area The corresponding radian slope value; This is the set of pixels in the segmentation result that are all identified as tea garden areas.

[0030] In addition, the present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction as described above.

[0031] Compared with the prior art, the present invention has the following beneficial effects: This invention addresses the issues of spectral and projection distortion in remote sensing images of tea gardens under complex terrain. It designs and optimizes a C-correction terrain attention correction method, introduces learnable C-correction coefficients to adaptively adapt to tea garden segmentation scenarios, and achieves terrain correction of high-resolution images at the feature extraction layer. Simultaneously, the terrain data is encoded, and the terrain features are fused with high-resolution spatial features through an attention mechanism to enhance the feature representation of complex terrain areas (terraces, steep slopes) and alleviate the problem of poor recognition accuracy caused by terrain undulations.

[0032] This invention combines the phenological growth patterns of tea gardens, uses a time-series windowing method to encode the location of time-series data, and integrates high-resolution and medium-resolution time-series images with scale alignment technology to construct a multimodal time-series dataset adapted to tea garden segmentation, comprehensively capturing key phenological information of tea gardens.

[0033] This invention introduces a lightweight temporal attention encoder (L-ATE), which combines a multi-head attention mechanism to quantify the importance of different temporal windows. It avoids temporal confusion by using sine and cosine position encoding, and adaptively aggregates key phenological features of tea gardens, thus solving the problems of low efficiency in existing temporal feature extraction and inaccurate capture of key phenological information.

[0034] This invention integrates three modalities: high-resolution spatial features after terrain fusion, medium-resolution temporal phenological features, and spectral enhancement features. It designs an adaptive attention fusion mechanism, calculates the feature weights of each modality through attention gating, and dynamically adjusts the weight allocation according to the feature importance of different terrain regions and different phenological periods. This highlights effective features and suppresses ineffective features, achieving efficient complementarity of multimodal features and improving the distinguishability of tea gardens from similar vegetation.

[0035] This invention employs a multi-branch decoding and fusion structure based on feature pyramids to achieve accurate multi-scale segmentation of tea gardens. By combining the image pixel scale after terrain correction with the segmentation results, a slope correction formula is introduced to eliminate the influence of projection distortion, thereby completing accurate statistics of tea garden area. This solves the problems of large area estimation errors and inability to adapt to complex terrain scenarios in existing methods. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0037] Figure 1 This is the lightweight temporal attention encoder (L-TAE) network structure described in this invention.

[0038] Figure 2 This is a schematic diagram of the multimodal tea garden segmentation model based on terrain correction of the present invention. Detailed Implementation

[0039] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0040] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] Example 1: This embodiment discloses a multimodal remote sensing method for tea garden segmentation and area estimation based on terrain correction. The specific technical solution is as follows: (1) Construction of multimodal time series dataset Data Acquisition Considering the differences in phenological characteristics between tea gardens and ordinary crops during their growth cycle, as well as their surrounding climate, this invention selects high-resolution multispectral images of the key phenological period of tea bud emergence and medium-resolution multispectral remote sensing images of multiple phenological nodes throughout the entire tea growth cycle to capture spectral, textural, and temporal phenological information during key phenological periods in tea gardens. Specifically, the key phenological period for tea bud emergence is April, and the entire tea growth cycle is from February to November. The high-resolution multispectral remote sensing images used are domestically produced 0.5-meter high-resolution images with four channels (red band R, green band G, blue band B, and near-infrared band NIR). The medium-resolution multispectral remote sensing images used are Sentinel-2 multispectral remote sensing images, whose rapid re-observation capability provides rich temporal information for tea garden growth monitoring. Topographic correction data uses the ALOS 12.5-meter digital elevation model.

[0042] Construction of multimodal time series datasets Although tea is a perennial evergreen woody plant, its growth status can be captured through long-term, intensive remote sensing observations. According to the phenological cycle of tea, tea leaves generally begin to sprout in late February, reaching their peak growth period around April, until the first round of harvesting and pruning. They then sprout again around August, reaching peak growth again, followed by harvesting and pruning before overwintering in November. Considering the influence of topographical differences, this invention selects high-resolution remote sensing images from April or August, as well as medium-resolution multi-temporal remote sensing images from February to November, and combines them with the ALOS digital elevation model to construct a multimodal temporal dataset for tea garden segmentation. For the medium-resolution multi-temporal images, multiple time-series remote sensing images from February to November are selected, and cloud removal processing is performed using the built-in cloud masking of the images. Subsequently, using a temporal alignment method, the period from February to November was evenly divided into nine temporal windows. Medium-resolution multispectral images from February to November were mapped to their corresponding temporal window positions using temporal window matching. Then, the best observation within each window was selected based on the maximum value of the normalized difference vegetation index (NDVI) to construct the temporal dataset. The NDVI calculation method is as follows:

[0043] In the formula: The near-infrared reflectance of the multispectral image; This represents the red band reflectance of the multispectral image.

[0044] Let the first Each timing window contains The corresponding set of normalized difference vegetation index (NDVI) imagery from medium-resolution remote sensing observations is as follows: The best observation after screening the maximum value of the normalized difference vegetation index. for In time series regions with no observations, if valid observations exist within that time series window, the mean of the valid observations within that window is used for filling; if there are no valid observations within that time series window, linear interpolation is used to fill the gap using valid observations from the nearest time series window. The final constructed medium-resolution time series dataset... The organization method is as follows:

[0045] In the formula: The time series length of the medium-resolution remote sensing image; The number of channels for Sentinel-2 imagery; Image height; Image width; It is the field of real numbers.

[0046] The final constructed multimodal time series dataset as follows:

[0047] In the formula: For high-resolution remote sensing image data, For high-resolution image channels, For high-resolution image height, For high-resolution image width; This is a medium-resolution time series dataset; This is ALOS digital elevation model data.

[0048] (2) Construction of a multimodal tea garden segmentation model based on terrain correction Integrating texture information from high-resolution imagery with temporal and spectral information enhancement from medium-resolution imagery is an effective method for distinguishing tea gardens from other vegetation. This invention designs a terrain-corrected multimodal tea garden segmentation model using a pre-constructed multimodal temporal dataset. The model mainly includes: a high-resolution imagery spatial feature extraction module, a medium-resolution imagery multi-temporal phenological feature extraction module, a spectral information feature enhancement extraction module, a terrain attention correction module, a multimodal feature fusion module, and a multi-scale segmentation module.

[0049] 2.1 High-resolution image spatial feature extraction module The high-resolution image spatial feature extraction module uses ResNet50 as the backbone network and supports input data of high-resolution 3-channel (red band R, green band G, blue band B) and 4-channel (red band R, green band G, blue band B, near-infrared band NIR) image data with a resolution better than 1 meter. This module is primarily responsible for extracting fine spatial texture and color features from high-resolution remote sensing images of tea gardens, adapting to the detailed capture of tea garden boundaries under complex terrain conditions. To improve the model's feature extraction efficiency and generalization ability, the backbone network is initialized using ImageNet pre-trained weights, and the gradient vanishing problem in deep networks is mitigated through residual structures. The residual block calculation formula is as follows:

[0050] In the formula: Output features for residual blocks; Indicates the weight parameters The residual mapping function consists of a 1×1 convolutional layer, a 3×3 convolutional layer, a batch normalization layer, and a linear rectified activation function. The input features are for the residual block.

[0051] Multi-scale spatial features at four different stages were extracted using the ResNet50 backbone network, mainly including: fine-grained scale texture features. Mesoscale texture features Coarse-scale texture features Deep semantic features .in, The core spatial features are used for fusion with other multimodal features, including the spatial and textural features of the tea garden in high-resolution images, to enhance the model's ability to segment the tea garden at a spatial scale. Furthermore, an adaptive channel adaptation layer is added to the input of the ResNet50 backbone network. If the input data is 4-channel image data, the pre-trained 3-channel convolutional kernel weights are averaged and expanded to 4 channels, thus preserving the feature extraction capability of the pre-trained weights and avoiding feature extraction errors caused by channel mismatch.

[0052] 2.2 Multi-temporal phenological feature extraction module for medium-resolution images The medium-resolution image multi-temporal phenological feature extraction module uses a multimodal temporal dataset. Medium resolution time series data As input, a lightweight temporal attention encoder is used as the main structure to extract the temporal phenological features of tea gardens from medium-resolution images. This captures the spectral variation patterns of tea gardens at different growth stages, and utilizes temporal differences to distinguish tea gardens from other similar vegetation, thereby improving the feature representation of tea garden temporal phenology. The different growth stages include budding stage, vigorous growth stage, harvesting stage, and overwintering stage. The specific implementation process of this module is as follows: (1) Spectral feature extraction: For time series data In the temporal frames, spectral features are initially extracted using consecutive 3×3 convolutions to capture the spectral details of each frame. Simultaneously, a lightweight 3D attention module, SimAM, is introduced to weight the extracted single-frame spectral features, thereby highlighting spectral features related to the tea garden and suppressing interference from soil, water, and building backgrounds. The weighted fusion then yields a temporal spectral feature sequence. .in For batch size, For timing length, This represents the number of time-series spectral characteristic channels. and These represent the height and width of the medium-resolution image, respectively. The lightweight 3D attention module SimAM is calculated as follows:

[0053]

[0054] In the formula: After being weighted by the lightweight 3D attention module SimAM, the th Channel, First line, number The column's characteristic pixel values; The first feature map of the original data Channel, First line, number The column's characteristic pixel values; This refers to the channel index of the feature map. ; , For the row and column indices of the feature map; This represents the average value of all neurons within a single channel. ; The variance of all neurons within a single channel; The activation function for the S-shaped growth curve; The positive constant of the local minimum value takes the value of . This is used to avoid dividing the denominator by zero.

[0055] (2) Temporal Attention Aggregation: The temporal attention aggregation module processes temporal spectral feature sequences. Aggregates into temporal aggregate features along the time dimension. This method adaptively captures key phenological information from tea gardens, avoiding the loss of crucial phenological information due to simple stacking of time-series data. The key process is as follows: ① Location encoding: For time series datasets Position encoding is performed using sine and cosine functions to provide a time-series dataset. Timing position information is written to avoid loss of timing sequence information. Its mathematical expression is as follows:

[0056]

[0057]

[0058] In the formula: For time-series location index; Indexed by channel dimension; The feature dimension for location encoding is the number of temporal spectral feature channels. ; This is the position encoding matrix; The temporal features after injecting location information.

[0059] ② Multi-head attention computation: Utilizing a multi-head attention module, the temporal features after injecting location information are processed. The query vector matrix is ​​mapped through three independent linear transformation layers. Key vector matrix Value vector matrix The temporal attention weights are then calculated using scaled dot product attention. Subsequently, the generated query vector matrix... Key vector matrix Value vector matrix The query vector matrix is ​​split based on the number of heads and an attention score is calculated to measure its performance. and key vector matrix The temporal feature similarity between the features is calculated, and the attention-weighted features are calculated. The calculation formula is as follows:

[0060]

[0061]

[0062]

[0063] In the formula: , , These are the learnable weight matrices used to generate query vectors, key vectors, and value vectors, respectively. To query the vector matrix, The key vector matrix, It is a value vector matrix; Let be the dimension of the key vector. , The number of heads receiving multi-head attention; It is a normalized exponential function; , , The first The query vector matrix, key vector matrix, and value vector matrix of each attention head; Indicates the first Temporal features after attention head weighting; This indicates a splicing operation along the channel dimension; The output projection matrix for multi-head fusion; The temporal features are weighted by multi-head attention.

[0064] ③ Temporal aggregation: Utilizing a multilayer perceptron to weight the temporal features after multi-head attention. Perform temporal feature aggregation and refine it into temporal aggregated features. This method integrates temporal variation information throughout the entire growth cycle of a tea garden, effectively distinguishing tea gardens from other vegetation types. Subsequently, bilinear interpolation is used to aggregate the temporal features. Adjust to and core spatial features At the same spatial scale, temporal phenological aggregation characteristics are obtained. This is used for subsequent multimodal feature fusion. This represents the number of time-series aggregated feature channels. and Core spatial features Height and width.

[0065] 2.3 Spectral Information Feature Enhancement and Extraction Module The spectral information feature enhancement and extraction module is based on a multimodal time series dataset. Medium resolution time series datasets By combining the spectral characteristics of tea gardens, multiple vegetation indices sensitive to tea gardens are calculated. Through multi-index fusion, the spectral distinguishability of tea gardens from other land features and similar vegetation is enhanced, compensating for the insufficient distinguishability of single spectral features and providing reliable spectral feature support for multimodal feature fusion. The specific implementation process is as follows: (1) Calculation of sensitive vegetation index in tea garden: Combining medium-resolution time series dataset Based on the band characteristics, the Normalized Difference Vegetation Index, Enhanced Vegetation Index, Normalized Red Edge Index, and Modified Soil-Regulated Vegetation Index were selected as the index dataset for enhancing spectral information features. These indices are highly sensitive to the vegetation cover and growth status of tea gardens and can effectively suppress interference from soil background and shading background. Their calculation formulas are as follows:

[0066]

[0067]

[0068]

[0069] In the formula: Indicates near-infrared reflectivity; Indicates the reflectivity in the red light band; Reflectivity in the blue light band; The red-edge band reflectivity corresponds to the time-series datasets. The corresponding band of the image. During the calculation, a minimum value is added to the denominator. To avoid division by zero errors.

[0070] (2) Spectral Enhancement Information Feature Extraction: The multiple vegetation indices calculated above are stacked to obtain an index dataset. Furthermore, deep feature extraction was performed using two layers of 3×3 convolution to enhance the exponentially enhanced spectral information features that differentiated the tea garden from other vegetation. Subsequently, bilinear interpolation was used to adjust the spectral enhancement features to match the core spatial features. At the same spatial scale, spectral information enhancement features are obtained. This is used for subsequent multimodal feature fusion. To enhance the number of feature channels for spectral information.

[0071] 2.4 Terrain Attention Correction Module The terrain attention correction module is primarily used to alleviate the difficulties in extracting tea garden features and the low accuracy of tea garden identification caused by image distortion and mixed pixel effects in complex terrain environments. This module performs terrain correction on high-resolution remote sensing imagery based on an optimized C-correction algorithm, eliminating spectral and projection distortions caused by terrain undulations. Simultaneously, it extracts the encoded terrain features and fuses them with high-resolution spatial features through an attention mechanism, adaptively increasing the feature weights of complex terrain areas such as terraced fields and steep slopes, thereby improving the model's identification accuracy. Its specific implementation process is as follows: (1) Data preprocessing: The digital elevation model data of the target area is obtained as input. Based on the digital elevation model, the slope and aspect topographic information are calculated. Within a 3×3 grid window, the slope is calculated as follows:

[0072]

[0073]

[0074] In the formula: Slope is expressed in radians; It is the arctangent function; and These represent the rate of change of elevation per meter of horizontal distance on the Earth's surface in the east-west and north-south directions, respectively. to This represents the elevation values ​​of each cell within a 3x3 window, with the subscripts ordered from left to right and top to bottom. This represents the pixel spatial resolution of the digital elevation model, in meters.

[0075] The slope aspect is calculated as follows:

[0076]

[0077] In the formula: Slope direction in radians; Slope direction in angle system; It is the arctangent function in the four quadrants; Pi is the mathematical constant of a circle.

[0078] (2) Adaptive C-correction: High-resolution images are corrected band by band by introducing learnable C-correction coefficients. It adapts to 4-channel high-resolution imagery and can be flexibly adjusted according to image type. Compared to the traditional C-correction algorithm, the learnable coefficients can adaptively adapt to the spectral characteristics of tea garden scenes, improving terrain correction effects. (Solar incidence angle cosine) The calculation method is as follows:

[0079] In the formula: The solar zenith angle is read from the metadata file of the high-resolution remote sensing image. The solar azimuth angle is read from the metadata file of the high-resolution remote sensing image. Slope in radians; The slope is measured in radians.

[0080] The calculation method for adaptive C-correction is as follows:

[0081] In the formula: Indicates the first Image data corrected for each band; Indicates the first Raw image data for each band; This represents the cosine value of the solar zenith angle under flat terrain. The C correction coefficient is a learnable coefficient, initialized to 0.1, and updated through backpropagation during model training. Band identifiers for high-resolution images.

[0082] (3) Topographic feature coding: By encoding the slope Slope aspect Elevation values ​​are channel-stitched and encoded, and then batch normalization is used to normalize the terrain data. Subsequently, terrain features are extracted using consecutive 3×3 convolutions and the lightweight 3D attention module SimAM. This enhances the ability to distinguish topographic features, highlighting the characteristics of tea gardens in complex terrain areas such as terraced fields and steep slopes. This represents the number of terrain feature channels.

[0083] (4) Terrain Feature Attention Fusion: The extracted terrain features are then fused together. High-resolution spatial features Attention fusion is performed, and terrain weights are adaptively assigned based on terrain complexity, with higher weights given to areas with steep slopes, thereby enhancing the feature representation capabilities of complex terrain regions. The terrain feature attention fusion process is as follows: High-resolution spatial features and terrain features The splicing feature is obtained by splicing along the channel dimension. Subsequently, the terrain feature attention weights are calculated using adaptive average pooling, 1×1 convolution, and a sigmoid growth curve activation function. Then, the terrain features are multiplied by the attention weights to obtain the weighted terrain features. And then with high-resolution spatial features After stitching, the high-resolution core features are output through 3×3 convolution. :

[0084]

[0085]

[0086] In the formula: The activation function for the S-shaped growth curve; This is a 1×1 convolution operation; For adaptive average pooling operation; This represents element-wise multiplication; It is a linear rectification activation function; For batch normalization; This is a 3×3 convolution operation; This indicates a splicing operation along the channel dimension.

[0087] 2.5 Multimodal Feature Fusion Module The multimodal feature fusion module integrates high-resolution core features obtained after terrain fusion. Medium-resolution temporal phenological aggregation features Spectral information enhancement features The three modalities are used, and an adaptive attention fusion mechanism is employed to dynamically adjust the channel-level weights of each modal feature, highlighting effective features and suppressing ineffective features, thereby improving the model's ability to segment tea gardens with complex terrain. The specific process is as follows: (1) Feature concatenation: The three modal features are concatenated into a fused feature along the channel dimension. ,in The calculation formula is as follows:

[0088] Subsequently, the concatenated features are fused and reduced in dimensionality using 3×3 convolution, batch normalization, and linear rectified activation function. A dropout method is then used to prevent overfitting and enhance the correlation of features, resulting in intermediate fused features. Its mathematical expression is:

[0089] In the formula: This indicates the discard regularization operation; It is a linear rectification activation function; For batch normalization; This is a 3×3 convolution operation.

[0090] (2) Adaptive Attention Weighting: The weights of each channel in the intermediate fusion features are calculated using a channel attention gating mechanism. This dynamically adjusts the channel-level feature responses, highlighting effective feature channels relevant to tea garden identification and suppressing ineffective feature channels. Its mathematical expression is as follows:

[0091]

[0092] In the formula: This is the channel attention weight vector; The activation function for the S-shaped growth curve; For adaptive average pooling operation; This is a 1×1 convolution operation; It is a linear rectification activation function; This indicates element-wise multiplication.

[0093] (3) Feature Refinement: Finally, the attention-weighted features are refined using 3×3 convolutions to remove redundant information, enhance the discriminative ability of the fused features, and output the final fused features. Its mathematical expression is:

[0094] In the formula: The core features after multimodal fusion integrate four major types of features: spatial features, temporal features, spectral features, and terrain features, which are used for subsequent multi-scale segmentation.

[0095] 2.6 Multi-scale segmentation module The multi-scale segmentation module employs a feature pyramid network combined with a multi-branch decoding fusion structure to achieve accurate segmentation of tea garden areas, balancing segmentation precision and boundary accuracy. It also provides high-quality segmentation results for subsequent accurate tea garden area estimation, addressing the issues of blurred tea garden boundaries and low segmentation accuracy in complex terrain. The specific implementation process is as follows: (1) Feature pyramid construction: Based on high-resolution branch features at four scales , , , and multimodal fusion features A feature pyramid is constructed to achieve multi-scale feature complementarity, and a 1×1 convolution is used to unify the number of channels, finally resulting in five corresponding feature levels. , , , , .

[0096] (2) Multi-branch decoding fusion: Two sub-branches are designed: a spatial decoding branch and an edge decoding branch. The spatial integrity and boundary accuracy are optimized respectively. The final decoded features are obtained after fusion. The spatial decoding branch and the edge decoding branch are implemented as follows: ① Spatial decoding branch: For feature levels to By progressively upsampling and fusing features through a process of "upsampling + feature stitching + convolutional refinement," the image scale is gradually restored, and spatial segmentation features are output. .

[0097] ② Edge Decoding Branch: Utilizes a multi-scale pooling strategy to fuse multimodal features with a uniform number of channels. Extracting edge features This separates the main body from edge features, thereby optimizing boundary accuracy.

[0098] ③ Multi-branch fusion: By splicing along the channel dimension, spatial segmentation features are fused. and edge features The fused features used for decoding are obtained. .

[0099] (3) Segmentation Output: Decoded features are segmented using 4x bilinear interpolation. The image is upsampled to the original high-resolution image scale, consistent with the scale of the topographically corrected high-resolution image. Then, a segmentation head is sequentially passed through 3×3 dilated convolution, 1×1 convolution, batch normalization, and a linear rectified activation function to output binary segmentation results for the tea garden area and the non-tea garden area. The mathematical expression of the segmentation head is as follows:

[0100]

[0101] In the formula: This represents a 3×3 dilated convolution with a dilation rate of 0.5%. ; This represents a 1×1 convolution operation; It is a linear rectification activation function; For batch normalization, the segmentation results are finally transformed into a probability map using the Softmax activation function, outputting the final segmentation results for tea garden areas / non-tea garden areas.

[0102] (4) Tea garden area estimation: Using the binary map of the tea garden segmentation results obtained after topographic correction. Count the number of pixels in the tea garden area The total area of ​​the tea garden was calculated using the actual pixel size. The high-resolution image has a spatial resolution of 0.5 meters and a single pixel area of Square meters. Simultaneously, by incorporating slope information from the topographic data, the area of ​​steep slope regions was corrected to obtain the corrected total tea garden area. To further improve the accuracy of tea garden area estimation, the specific calculation formula is as follows:

[0103] In the formula: This represents the revised total area of ​​the tea plantation. The horizontal projected area of ​​a single pixel; The pixel locations of the tea garden area in the segmentation results The corresponding radian slope value is extracted from the preprocessed slope raster data. This represents the set of pixels identified as tea garden areas in the segmentation results. Terrain correction is introduced to eliminate the impact of projection distortion caused by terrain undulations on area estimation.

[0104] This embodiment addresses the issues of spectral distortion, projection distortion, and insufficient feature representation in remote sensing images of tea gardens with complex terrain. By introducing a learnable C-correction coefficient to adaptively adapt to the tea garden scene, it completes high-resolution image terrain correction. Simultaneously, the terrain data is encoded, and an attention mechanism is used to fuse terrain features with high-resolution spatial features, thereby enhancing the features of complex terrain areas and improving the segmentation accuracy of tea gardens.

[0105] This embodiment overcomes the shortcomings of traditional dataset construction that does not conform to the phenological patterns of tea gardens. By using time-series window division, normalized difference vegetation index maximum value screening, and scale alignment technology, it integrates high-resolution and medium-resolution time-series images to fully capture information on key phenological periods of tea gardens and adapt to the special needs of tea garden segmentation.

[0106] This embodiment overcomes the shortcomings of traditional temporal modeling, such as homogenized weights and loss of key phenological information. It utilizes a lightweight temporal attention encoder combined with a multi-head attention mechanism to quantify the weights of different temporal windows, and combines sine and cosine function position encoding to avoid temporal confusion. While ensuring a small number of parameters, it accurately captures the key phenological characteristics of tea gardens and improves robustness in complex scenarios.

[0107] This embodiment addresses the problem of poor fusion results caused by fixed weights in existing multimodal fusion methods by designing an adaptive attention fusion mechanism. This mechanism integrates topographic fusion spatial features, temporal phenological features, and spectral enhancement features, dynamically adjusts the weights of each modal feature channel, and achieves effective feature enhancement and ineffective feature suppression, thereby improving the ability to distinguish tea gardens from similar vegetation such as woodlands and orchards.

[0108] This embodiment addresses the problem that existing area estimation methods do not consider topographic projection distortion and have large errors. By combining the topographically corrected image scale with multi-scale accurate segmentation results, a slope correction formula is introduced to eliminate the influence of projection distortion, thereby achieving accurate statistics on tea garden area.

[0109] This invention provides an integrated solution to the core problems of tea garden segmentation in complex terrain, such as insufficient suppression of terrain interference, defects in multimodal feature fusion, inaccurate expression of key phenological features, and large area estimation errors. Compared with existing tea garden segmentation technologies based on remote sensing images, the technical advantages of this invention are as follows: This invention introduces terrain correction technology to adapt to complex tea garden segmentation scenarios. Existing tea garden segmentation techniques generally lack consideration for terrain projection distortion: Publication No. CN 117011702 B only introduces digital elevation models as terrain features, failing to effectively integrate terrain features with spatial features, and also neglecting the impact of terrain projection distortion on tea garden area estimation. This invention utilizes an optimized C-correction algorithm with learnable coefficients to adaptively adapt to tea garden scenarios and complete terrain correction. Simultaneously, it fuses the encoded terrain features with high-resolution spatial features through an attention mechanism, strengthening the representation of complex terrain features, effectively suppressing interference from terrain undulations, and significantly improving the segmentation accuracy of tea gardens in complex terrain.

[0110] Introducing multimodal feature fusion enhances vegetation feature representation. Existing multimodal fusion techniques suffer from problems such as "insufficient feature representation and brute-force feature fusion." The schemes proposed in publications CN 117011702 B and CN 113221806 B simply splice spatial and spectral features without fully integrating temporal phenological and topographic features. Furthermore, they employ fixed weight fusion, lacking consideration for the varying importance of features across different regions, making it difficult to distinguish tea gardens from similar vegetation such as woodlands and orchards. This invention constructs a multimodal temporal dataset encompassing spatial, temporal, spectral, and topographic features. It employs an adaptive attention fusion mechanism, dynamically adjusting channel weight allocation based on the feature importance of different topographic regions and phenological periods, achieving efficient complementarity of multimodal features and significantly improving the ability to distinguish tea gardens from similar vegetation.

[0111] This invention leverages attention-weighted temporal features to enhance the model's ability to capture key phenological periods. Existing temporal feature extraction techniques suffer from problems such as "loss of key phenological information and high computational complexity." For example, the scheme proposed in patent application CN 115512218A uses simple averaging and stacking methods to extract temporal features, failing to quantify the importance of different temporal windows, leading to the averaging of key phenological features in tea gardens. The paper "Research on Accurate Identification of Tea Gardens Based on Satellite Remote Sensing Imagery" employs multiple complex temporal models, resulting in a large number of parameters and low computational efficiency. This invention uses a lightweight temporal attention encoder, quantifying the weights of different temporal windows through a multi-head attention mechanism to accurately capture key phenological features. It also incorporates sine and cosine function positional encoding to avoid temporal confusion, improving the effectiveness and specificity of temporal feature extraction while maintaining a lightweight model and low computational complexity.

[0112] To improve the accuracy of tea garden area estimation, existing tea garden segmentation techniques often focus only on segmentation precision, resulting in large errors due to the lack of consideration for terrain influence in area estimation. Publication number CN 113221806 B proposes directly estimating area based on classification results, but this fails to eliminate projection distortion caused by terrain undulations, leading to significant estimation errors under complex terrain conditions and failing to meet the needs of accurate area statistics in industry management. This invention combines the terrain-corrected image scale with multi-scale accurate segmentation results, introducing a slope correction formula to eliminate the influence of projection distortion, achieving accurate tea garden area statistics and effectively mitigating the estimation errors caused by terrain influence.

[0113] Example 2: This embodiment is a further optimization based on Embodiment 1. In this embodiment, the step of correcting the topographic projection distortion of the tea garden area according to the following formula further includes: The brightness spatial gradient is extracted using the grayscale distribution of the high-resolution remote sensing image data, and a local morphology restoration function is constructed. Using the brightness spatial gradient as an adjustment variable for micro-topographic undulation, subpixel-level spatial downscaling correction is performed on the initial slope field provided by the digital elevation model data to generate a subpixel-level slope value matrix that corresponds one-to-one with the pixel coordinates of the high-resolution remote sensing image data. Based on the subpixel-level slope value matrix, slope normal projection compensation is performed on the horizontal projection area of ​​each pixel in the tea garden area, and the corrected total area of ​​the tea garden is obtained by differential accumulation.

[0114] The steps for generating the sub-pixel level slope value matrix include: performing grayscale processing on the high-resolution remote sensing image data and calculating the pixel-level brightness spatial gradient; standardizing the brightness spatial gradient to eliminate differences in illumination intensity; introducing a sensitivity adjustment coefficient to linearly map the standardized brightness spatial gradient to a correction amount for the initial slope value, and achieving local topological smoothing or refinement adjustment of the initial slope value by superimposing the correction amount.

[0115] The experimental area was a terraced tea garden in a mountainous region. The terrain of this area is characterized by an extremely fragmented, stepped distribution. The width of the terraces is usually between 2 and 5 meters. However, the spatial resolution of the ALOS digital elevation model data is 12.5 meters. This results in a single digital elevation model cell covering multiple terrace levels, causing the initial slope calculation to appear as a sloping surface, which cannot reflect the true physical surface area of ​​the "stepped" terrain.

[0116] (1) Subpixel-level morphology feature extraction First, acquire high-resolution remote sensing image data with a spatial resolution of 0.5 meters. The brightness spatial gradient is extracted using the Sobel operator. The mathematical description is as follows:

[0117] In the formula: pixel position The brightness spatial gradient of high-resolution remote sensing image data at that location; These are the grayscale pixel values ​​converted from high-resolution remote sensing image data. This is the partial derivative of the grayscale pixel value in the horizontal direction; This is the partial derivative of the grayscale pixel value in the vertical direction.

[0118] (2) Downscaling spatial correction of the slope field Using the luminance spatial gradient to determine the initial slope value at a resolution of 12.5 meters. Micro-level corrections are performed. Terrace edges are identified by sharp changes in the brightness spatial gradient, while the terrace plane is identified by gradual changes. Sub-pixel level slope values ​​are used. The calculation formula is as follows:

[0119] In the formula: The subpixel-level radian slope value after spatial downscaling adjustment; To map digital elevation model data to pixel locations using bilinear interpolation The initial radian slope value; The sensitivity adjustment coefficient of the brightness spatial gradient to terrain roughness is set to 0.15 in this embodiment based on the lighting conditions. This represents the mean of the spatial gradient of brightness within the high-resolution remote sensing image data. The standard deviation of the spatial gradient of brightness within high-resolution remote sensing image data; To prevent extremely small constants with a denominator of zero, the value is taken as... .

[0120] In the above formula, the correction term The dimensionless normalized gradient value is obtained by comparing it with the sensitivity coefficient. After multiplication, the result is added to a constant 1 to form a relative adjustment factor for the initial slope, ensuring that the corrected slope value maintains the radian dimension and has a clear physical meaning. At locations where the brightness gradient changes drastically, such as the edge of the terraces, this factor amplifies the initial slope to simulate the steepening effect of the facade; at the level of the terraces, it essentially maintains the original slope.

[0121] (3) Area differential projection integral A set of pixels representing the tea garden area generated based on the segmentation results. The area is then calculated by combining the corrected sub-pixel level slope value. The formula is as follows:

[0122] In the formula: This represents the revised total area of ​​the tea plantation. The horizontal projection area of ​​a single pixel in high-resolution remote sensing image data is 0.25 square meters. This is the set of pixels in the segmentation result that are all identified as tea garden areas.

[0123] In Example 1, due to the limitation of the spatial resolution of the digital elevation model data, it was impossible to capture the additional area of ​​the terraced fields. Through the sub-pixel-level correction in this example, the area estimation accuracy in steep terraced fields with a slope greater than 25 degrees is improved by about 12% compared to traditional planar statistical methods, and by 4.5% compared to methods that rely solely on coarse-resolution digital elevation models for correction, thus alleviating the problem of "understated statistical values" for the area of ​​mountain tea gardens.

[0124] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal remote sensing method for tea garden segmentation and area estimation based on terrain correction, characterized in that, Includes the following steps: A multimodal time-series dataset is constructed, which includes high-resolution remote sensing image data, medium-resolution time-series remote sensing image data, and digital elevation model data for the same target area. The multimodal time-series dataset is input into a pre-built multimodal tea garden segmentation model based on terrain correction to obtain tea garden area segmentation results; Based on the tea garden area segmentation results and combined with the slope information in the digital elevation model data, the tea garden area is corrected for topographic projection distortion according to the following formula to obtain the corrected total tea garden area: In the formula: This represents the revised total area of ​​the tea plantation. The horizontal projected area of ​​a single pixel in high-resolution remote sensing image data; pixel location of the tea garden area The corresponding radian slope value; This is the set of pixels in the segmentation result that are all identified as tea garden areas.

2. The multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction according to claim 1, characterized in that, The terrain-corrected multimodal tea garden segmentation model processes the input data in the following manner: The core spatial features of high-resolution remote sensing image data are extracted using a high-resolution image spatial feature extraction module. The medium-resolution image multi-temporal phenological feature extraction module uses a lightweight temporal attention encoder to perform temporal modeling on the medium-resolution temporal remote sensing image data. The lightweight temporal attention encoder uses a multi-head attention mechanism to weight and aggregate features from different temporal windows, and outputs temporal phenological aggregated features. The terrain attention correction module calculates the slope and aspect based on the digital elevation model data, and calculates the cosine of the solar incidence angle by combining the solar zenith angle and solar azimuth angle; a learnable C correction coefficient is introduced, and an optimized C correction algorithm is used to perform band-by-band terrain correction on the high-resolution remote sensing image data; the extracted terrain features after correction are fused with the core spatial features by attention to obtain high-resolution core features. The high-resolution core features, the temporal phenological aggregate features, and the spectral information enhancement features are spliced ​​together through the multimodal feature fusion module, and the weights of each channel are dynamically adjusted through the channel attention gating mechanism to obtain the final fused features. The multi-scale segmentation module decodes the final fused features and the multi-scale features output by the high-resolution image spatial feature extraction module to output the segmentation result of the tea garden area.

3. The multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction according to claim 1, characterized in that, The lightweight temporal attention encoder is processed according to the following steps: For each frame of the medium-resolution time-series remote sensing image data, spectral features are initially extracted and then attention weighted by the lightweight three-dimensional attention module SimAM to obtain a time-series spectral feature sequence. The temporal spectral feature sequence is position-encoded using sine and cosine functions to obtain temporal features after injecting position information; The multi-head attention module maps the time-series features after the injection of location information into query vectors, key vectors, and value vectors, and calculates the time-series features after multi-head attention weighting. The temporal features weighted by multi-head attention are aggregated into temporal aggregated features using a multilayer perceptron, and then adjusted to the same spatial scale as the core spatial features through bilinear interpolation to obtain the temporal phenological aggregated features.

4. The multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction according to claim 2, characterized in that, The terrain attention correction module processes the data according to the following steps: Obtain the solar zenith angle and solar azimuth angle from high-resolution remote sensing image data, and calculate the cosine of the solar incidence angle using the following formula, combined with the slope and aspect. : In the formula: The solar zenith angle; This is the solar azimuth angle; Slope in radians; The slope is in radians; band-by-band adaptive C-correction is performed according to the following formula: In the formula: For the first Image data after band correction; For the first Raw image data for each band; The learnable C-correction coefficient is initialized to 0.

1. For band identification.

5. The multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction according to claim 2, characterized in that, The step of performing attention fusion between the corrected extracted terrain features and the core spatial features in the terrain attention correction module includes: The core spatial features and the terrain features are concatenated in the channel dimension, and the attention weights of the terrain features are calculated by adaptive average pooling, 1×1 convolution and S-shaped growth curve activation function. The terrain features are multiplied element-wise with their attention weights to obtain weighted terrain features; After concatenating the weighted terrain features with the core spatial features, the high-resolution core features are output through 3×3 convolution, batch normalization, and linear rectified activation function.

6. The multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction according to claim 2, characterized in that, In the multimodal feature fusion module, the channel attention gating mechanism specifically includes: taking the intermediate fused features obtained after initial convolutional fusion and then sequentially applying adaptive average pooling, 1×1 convolutional dimensionality reduction, linear rectified activation function, 1×1 convolutional dimensionality increase, and S-curve activation function to calculate the channel attention weight vector; and multiplying the channel attention weight vector element-wise with the intermediate fused features to obtain the attention-weighted fused features.

7. The multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction according to claim 2, characterized in that, The multi-scale segmentation module specifically includes: constructing a feature pyramid based on the spatial features at multiple scales output by the high-resolution image spatial feature extraction module and the final fused features; processing the feature levels in the feature pyramid through spatial decoding branch and edge decoding branch respectively, and splicing and fusing the outputs of the two branches in the channel dimension to obtain decoded features; upsampling the decoded features to the scale of the original high-resolution remote sensing image, and processing them through a segmentation head containing 3×3 dilated convolution to output binary segmentation results for the tea garden area and non-tea garden area.

8. The multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction according to claim 2, characterized in that, The step of constructing a multimodal time-series dataset further includes: acquiring high-resolution remote sensing image data of the target area during key phenological periods of tea, wherein the high-resolution remote sensing image data is a 4-channel image containing red, green, blue and near-infrared bands; acquiring medium-resolution remote sensing images of the target area at multiple time points throughout the entire growth cycle of tea, and constructing medium-resolution time-series remote sensing image data by dividing the time series window and filtering by the maximum value of the normalized difference vegetation index; acquiring digital elevation model data of the target area, and calculating slope and aspect information based on the digital elevation model data.

9. A multimodal remote sensing tea garden segmentation and area estimation system based on terrain correction, used to execute the multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction as described in any one of claims 1-8, characterized in that, include: The dataset construction unit is used to construct a multimodal time-series dataset, which includes high-resolution remote sensing image data, medium-resolution time-series remote sensing image data, and digital elevation model data for the same target area. The model segmentation unit is used to input the multimodal time-series dataset into a pre-constructed terrain-corrected multimodal tea garden segmentation model to obtain tea garden area segmentation results; wherein, the terrain-corrected multimodal tea garden segmentation model processes the input data in the following manner: The core spatial features of high-resolution remote sensing image data are extracted using a high-resolution image spatial feature extraction module. The medium-resolution image multi-temporal phenological feature extraction module uses a lightweight temporal attention encoder to perform temporal modeling on the medium-resolution temporal remote sensing image data. The lightweight temporal attention encoder uses a multi-head attention mechanism to weight and aggregate features from different temporal windows, and outputs temporal phenological aggregated features. The terrain attention correction module calculates the slope and aspect based on the digital elevation model data, and calculates the cosine of the solar incidence angle by combining the solar zenith angle and solar azimuth angle; a learnable C correction coefficient is introduced, and an optimized C correction algorithm is used to perform band-by-band terrain correction on the high-resolution remote sensing image data; the extracted terrain features after correction are fused with the core spatial features by attention to obtain high-resolution core features. The high-resolution core features, the temporal phenological aggregate features, and the spectral information enhancement features are spliced ​​together through the multimodal feature fusion module, and the weights of each channel are dynamically adjusted through the channel attention gating mechanism to obtain the final fused features. The multi-scale segmentation module decodes the final fused features and the multi-scale features output by the high-resolution image spatial feature extraction module to output the segmentation result of the tea garden area. The area correction unit is used to correct the terrain projection distortion of the tea garden area based on the tea garden area segmentation results and the slope information in the digital elevation model data, according to the following formula, to obtain the corrected total area of ​​the tea garden: In the formula: This represents the revised total area of ​​the tea plantation. The horizontal projected area of ​​a single pixel in high-resolution remote sensing image data; pixel location of the tea garden area The corresponding radian slope value; This is the set of pixels in the segmentation result that are all identified as tea garden areas.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multimodal remote sensing tea garden segmentation and area estimation method based on terrain correction as described in any one of claims 1 to 8.

Images