A method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing

Abstract

The application discloses a kind of based on multi-source remote sensing wild tea tree dynamic monitoring and distribution prediction method, comprising: based on local self-similarity algorithm to optical satellite image and synthetic aperture radar image are registered in space, and the optical feature and synthetic aperture radar feature after registration are spliced into unified remote sensing feature;Based on spatial attention mechanism, spectral attention mechanism and cross-modal attention mechanism, each feature is weighted and fused to generate multi-source fusion feature;Extract initial feature set and carry out correlation and collinearity screening, obtain key feature set and construct time series feature data;Multi-layer Deep-LSTM network of integrated adversarial domain self-adaptive module is constructed, and the optimal prediction model is obtained after training is completed;Based on the optimal prediction model, the distribution density variation trend of wild tea tree in future period is obtained by predicting current time series feature data.The application can improve the accuracy and efficiency of wild tea tree dynamic monitoring and distribution prediction.

Description

Technical Field

[0001] This application relates to the field of wild plant resource monitoring technology, and relates to, but is not limited to, a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing. Background Technology

[0002] Wild tea trees, as the gene pool of the world's origin center of tea, are a core strategic resource for tea variety improvement and sustainable industrial development. Their distribution monitoring and dynamic assessment are of significant strategic value for biodiversity conservation and the sustainable development of the tea industry. Traditional ground-based survey methods are insufficient for large-scale, efficient dynamic monitoring of wild tea tree resources that are scattered and have complex environments. Remote sensing technology, with its advantages of wide coverage and non-contact, periodic observation, has become the mainstream method for vegetation resource surveys.

[0003] In existing technologies, remote sensing monitoring of tea gardens mainly relies on medium-resolution optical satellite imagery such as Landsat 8 and GF-1, combined with models such as random forests and convolutional neural networks for tea garden area extraction and stress assessment. However, this method is only applicable to artificial tea gardens and still has some shortcomings in monitoring wild tea trees. First, existing methods mainly depend on optical remote sensing data. In areas with significant environmental interference, such as cloudy and foggy mountainous areas, the continuity of optical images is poor, resulting in a large amount of missing data and making all-weather monitoring impossible. Second, wild tea trees are often mixed with natural forests, shrubs, and other associated vegetation, and their spectral characteristics are highly similar, making it difficult to accurately distinguish them based on existing spectral and textural features. Furthermore, existing methods mostly focus on static distribution mapping or short-term physiological parameter inversion, lacking the ability to predict the long-term distribution pattern evolution trend of wild tea tree populations. Finally, remote sensing data acquired from different regions and different sensors often exhibit significant domain differences, which can lead to a sharp decline in the predictive performance of models trained on data from specific regions in new regions, i.e., the model's generalization ability is weak.

[0004] Therefore, there is an urgent need for a more comprehensive method for dynamic monitoring and distribution prediction of wild tea trees to solve the problems of incomplete data sources, insufficient feature discrimination, lack of dynamic prediction ability and weak model generalization ability in existing technologies, so as to greatly improve the accuracy and timeliness of wild tea tree resource monitoring and prediction. Summary of the Invention

[0005] This application provides a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing. The method includes: acquiring optical satellite images, synthetic aperture radar (SAR) images, UAV remote sensing images, and environmental data of a target area and extracting features from each to generate optical features, SAR features, biological features, and environmental features; performing sub-pixel-level spatial registration of the optical satellite images and the SAR images based on a local self-similarity algorithm, and stitching the spatially registered optical features and SAR features into a unified remote sensing feature; performing weighted fusion of the optical features, SAR features, and biological features based on spatial attention mechanisms, spectral attention mechanisms, and cross-modal attention mechanisms to generate a multi-source fused feature; and performing weighted fusion of the optical features, SAR features, biological features, and environmental features based on the optical features, SAR features, biological features, and environmental features. An initial feature set is extracted from the features and the multi-source fusion features. The initial feature set is then subjected to correlation and collinearity filtering to obtain a key feature set. Temporal feature data is constructed based on this key feature set. A multi-layer Deep-LSTM network is constructed, and an adversarial domain adaptive module is integrated into the multi-layer Deep-LSTM network. The multi-layer Deep-LSTM network is trained using historical temporal feature data from the temporal feature data to obtain an optimal prediction model. Vegetation areas are determined using the optical features and the synthetic aperture radar features. Tea tree areas are determined using the multi-source fusion features to obtain a spatial distribution map of wild tea trees. Within the spatial range determined by the wild tea tree spatial distribution map, the distribution density change trend of wild tea trees in future periods is predicted based on the optimal prediction model and the current temporal feature data.

[0007] The technical solution provided in this application acquires optical satellite imagery, synthetic aperture radar (SAR) imagery, UAV remote sensing imagery, and environmental data of the target area to overcome the observation blind spots of single optical data in cloudy and foggy mountainous areas, ensuring the continuity and integrity of data acquisition and enabling all-weather, multi-dimensional observation. Feature extraction is performed on each data source to generate optical features, SAR features, biological features, and environmental features. These features characterize tea trees from macroscopic state, microscopic morphology, and growth environment perspectives. Based on this multi-layered, refined feature foundation, a comprehensive data basis is laid for subsequent accurate identification and dynamic analysis. A local self-similarity algorithm is used to perform sub-pixel-level spatial registration of the optical satellite imagery and SAR imagery to ensure… Spatially precise alignment of data from different sources was achieved, and the spatially registered optical features and synthetic aperture radar (SAR) features were stitched together into a unified remote sensing feature. Based on spatial attention, spectral attention, and cross-modal attention mechanisms, optical features, SAR features, and biological features were weighted and fused to generate multi-source fusion features. Fine biological features were dynamically embedded into macroscopic remote sensing features, compensating for the shortcomings of satellite remote sensing in characterizing the internal state of plants. This achieved complementarity and synergy of multimodal information, significantly strengthening the key features that distinguish tea trees from their associated vegetation. An initial feature set was extracted based on optical features, SAR features, biological features, environmental features, and multi-source fusion features. Correlation and collinearity screening were then performed on the initial feature set. The process involves selective filtering to automatically remove redundant and interfering features, resulting in a key feature set. This effectively reduces feature dimensionality and improves the efficiency, stability, and generalization ability of subsequent model training. Based on this key feature set, time-series feature data is constructed to provide standardized input for subsequent prediction models, supporting trend prediction. A multi-layer Deep-LSTM network is built, integrating an adversarial domain adaptation module. This module enables the model to learn domain-invariant features, effectively overcoming parameter differences between different regions and sensor data, thereby improving the model's generalization and robustness in new monitoring areas. Historical time-series feature data from the time-series feature data is then used to train the multi-layer Deep-LSTM network. The optimal prediction model is trained to achieve dynamic and accurate prediction of the long-term evolution trend of tea tree distribution density. Vegetation areas are determined using optical and synthetic aperture radar features; tea tree areas are determined using multi-source fusion features, resulting in a spatial distribution map of wild tea trees; within the spatial range determined by the wild tea tree spatial distribution map, the distribution density change trend of wild tea trees in future periods is predicted based on the optimal prediction model and current time-series feature data. A progressive identification method is used to significantly reduce computational complexity, achieving high-precision and high-efficiency positioning of wild tea trees in complex vegetation backgrounds. This provides reliable forward-looking scientific evidence and decision support for the resource protection, disaster early warning, and sustainable management of wild tea trees.

[0008] Optionally, the step of weightedly fusing the optical features, synthetic aperture radar features, and biological features based on spatial attention mechanisms, spectral attention mechanisms, and cross-modal attention mechanisms to generate multi-source fused features includes: calculating spatial attention weights based on the optical features and performing regional masking on the spatial attention weights based on the vegetation masking matrix of the optical satellite imagery; calculating spectral attention weights based on the optical features; using the unified remote sensing feature as the key and the biological feature as the query, calculating the attention weight of the biological feature relative to the unified remote sensing feature, and embedding the biological feature into the unified remote sensing feature according to the corresponding attention weight to obtain biological information enhancement features; weighting the optical features based on the spatial attention weights, weighting the synthetic aperture radar features based on the spectral attention weights, and weighted fusing them with the biological information enhancement features to obtain multi-source fused features.

[0009] Optionally, the initial feature set includes initial optical features, initial synthetic aperture radar features, initial bio-related features, and initial environmental features. The initial optical features include the normalized vegetation index, enhanced vegetation index, and photochemical reflectance index among optical features, as well as the red-edge band reflectance features derived from the multi-source fusion features. The initial synthetic aperture radar features include the HH / VV polarization ratio, HV / VV polarization ratio, contrast, correlation, energy, and entropy among the synthetic aperture radar features. The initial bio-related features include the bio-information enhancement features. The initial environmental features include the soil moisture index and the comprehensive environmental index among environmental features.

[0010] Optionally, the step of performing correlation screening and collinearity screening on the initial feature set to obtain a key feature set includes: calculating the Pearson correlation coefficient between each initial feature and the target variable of wild tea tree distribution, retaining initial features with an absolute Pearson correlation coefficient greater than 0.7; calculating the variance inflation factor of the initial features with an absolute Pearson correlation coefficient greater than 0.7, retaining initial features with a variance inflation factor less than 3, to obtain a key feature set, wherein the key feature set includes normalized vegetation index, enhanced vegetation index, photochemical reflectance index, HH / VV polarization ratio, HV / VV polarization ratio, contrast, correlation, and soil moisture index.

[0011] Optionally, the multi-layer Deep-LSTM network includes an input layer, at least three LSTM encoder layers, an adaptation layer, an adversarial domain adaptation module, and a prediction layer. After receiving temporal feature data, the multi-layer Deep-LSTM network processes it through the following steps: receiving temporal feature data through the input layer; processing the temporal feature data through the three LSTM encoder layers to generate a comprehensive encoded feature representation; in the adversarial domain adaptation module, processing the comprehensive encoded feature representation through a gradient inversion layer and inputting it into a domain classifier to obtain the discrimination probability of the comprehensive encoded feature representation originating from the source domain or the target domain, and calculating the domain adaptation loss based on the discrimination probability; receiving the comprehensive encoded feature representation and the biological information enhancement features through the adaptation layer, fusing the biological features contained in the biological information enhancement features with the comprehensive encoded feature representation to obtain an enhanced temporal feature representation with fused biological information; receiving the enhanced temporal feature representation through the prediction layer, and calculating the predicted value of the wild tea tree distribution density in the target area through a fully connected network.

[0012] Optionally, the step of determining vegetation regions using the optical features and synthetic aperture radar features; determining tea tree regions using the multi-source fusion features to obtain a spatial distribution map of wild tea trees; and within the spatial range determined by the wild tea tree spatial distribution map, predicting the future distribution density change trend of wild tea trees based on the optimal prediction model and current time-series feature data, includes: inputting the optical features and synthetic aperture radar features into a random forest classifier to determine the predicted probability of each pixel in the optical satellite image and the synthetic aperture radar image belonging to a vegetation category, and identifying regions with a predicted probability greater than 0.2 as vegetation regions, generating a binary classification map of vegetation and non-vegetation; performing morphological post-processing on the binary classification map of vegetation and non-vegetation to remove isolated patches with an area less than 50 pixels in the binary classification map, obtaining an optimized vegetation region mask; within the vegetation region determined by the optimized vegetation region mask, inputting the multi-source fusion features into an improved deep convolutional neural network, the improved deep convolutional neural network including a spatial attention layer, a spectral attention layer, and a distribution layer. The system employs a classification layer; a spatial attention layer calculates the weight of each spatial location in the multi-source fusion feature to generate a spatial attention map; a spectral attention layer calculates the weight of each channel of the multi-source fusion feature to generate a channel attention vector; the spatial attention map and the channel attention vector are used to perform element-wise weighting and channel-wise weighting on the multi-source fusion feature to obtain attention-enhanced features; the attention-enhanced features are input into the classification layer, and the predicted probability of each pixel belonging to the tea tree category is output, with regions having a predicted probability greater than 0.7 designated as tea tree regions, generating a binary classification map of tea trees and associated vegetation; the binary classification map of tea trees and associated vegetation is post-processed to remove isolated patches with an area less than 20 pixels, resulting in a spatial distribution map of wild tea trees; based on the spatial range determined by the spatial distribution map of wild tea trees, the current temporal feature data is input into the optimal prediction model to predict the initial trend of wild tea tree distribution density in future time periods, and the initial trend of distribution density is compared and corrected with real-time environmental data to obtain the distribution density trend.

[0013] Secondly, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the above-mentioned method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing.

[0014] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the above-described method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing.

[0015] The beneficial effects of the technical solutions provided in this application include at least the following: This application provides a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing. It acquires optical satellite imagery, synthetic aperture radar (SAR) imagery, UAV remote sensing imagery, and environmental data of the target area to overcome the observation blind spots of single optical data in cloudy and foggy mountainous areas, ensuring the continuity and integrity of data acquisition and enabling all-weather, multi-dimensional observation. Feature extraction is performed on each data source to generate optical features, SAR features, biological features, and environmental features. These features characterize the tea trees from macroscopic state, microscopic morphology, and growth environment perspectives. Based on this multi-layered, refined feature foundation, a comprehensive data basis is laid for subsequent accurate identification and dynamic analysis. A local self-similarity algorithm is used to analyze the optical satellite imagery and SAR imagery. For example, subpixel-level spatial registration is performed to ensure accurate spatial alignment of data from different sources, and the spatially registered optical features and synthetic aperture radar (SAR) features are stitched together into a unified remote sensing feature. Based on spatial attention, spectral attention, and cross-modal attention mechanisms, optical features, SAR features, and biological features are weighted and fused to generate multi-source fused features. Fine biological features are dynamically embedded into macroscopic remote sensing features, compensating for the shortcomings of satellite remote sensing in characterizing the internal state of plants, thereby achieving complementarity and synergy of multimodal information and significantly strengthening the key features that distinguish tea trees from their companion vegetation. An initial feature set is extracted based on optical features, SAR features, biological features, environmental features, and multi-source fused features. Correlation and collinearity screening are performed to automatically remove redundant and interfering features, resulting in a key feature set. This effectively reduces feature dimensionality and improves the efficiency, stability, and generalization ability of subsequent model training. Based on this key feature set, time-series feature data is constructed to provide standardized input for subsequent prediction models, supporting trend prediction. A multi-layer Deep-LSTM network is built, and an adversarial domain adaptation module is integrated into it. This module enables the model to learn domain-invariant features, effectively overcoming parameter differences between different regions and sensor data, thereby improving the model's generalization and robustness in new monitoring areas. Historical time-series feature data from the time-series feature data is used to refine the multi-layer Deep-LSTM network. The STM network is trained to obtain the optimal prediction model, thereby achieving dynamic and accurate prediction of the long-term evolution trend of tea tree distribution density: vegetation areas are determined by optical features and synthetic aperture radar features; tea tree areas are determined by multi-source fusion features, resulting in a spatial distribution map of wild tea trees; within the spatial range determined by the wild tea tree spatial distribution map, based on the optimal prediction model, the trend of wild tea tree distribution density changes in future periods is predicted using current time-series feature data. By using a progressive identification method, the computational complexity is greatly reduced, achieving high-precision and high-efficiency positioning of wild tea trees in complex vegetation backgrounds. This provides reliable forward-looking scientific basis and decision support for the resource protection, disaster early warning, and sustainable management of wild tea trees. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing, provided for an embodiment of this application; Figure 2 This is a schematic diagram of the hardware entity of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0019] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0020] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0021] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0022] In view of the current problems in the research on dynamic monitoring and distribution prediction of wild tea trees in the field of wild plant resource monitoring technology, this application provides a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing.

[0023] The technical solution of this application is described below, starting with the method embodiments.

[0024] Please refer to Figure 1 It illustrates a flowchart of a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing, as provided in an embodiment of this application. Figure 1 As shown, the method includes at least the following steps S110 to S150.

[0025] Step S110: Acquire optical satellite imagery, synthetic aperture radar imagery, UAV remote sensing imagery, and environmental data of the target area, and extract features from each to generate optical features, synthetic aperture radar features, biological features, and environmental features.

[0026] In this embodiment, multi-platform collaborative observation is used to acquire multi-source heterogeneous data of the target area. Corresponding preprocessing and feature extraction are performed on various types of multi-source heterogeneous data to generate standard features for direct use in subsequent fusion and analysis. Specifically, the multi-source heterogeneous data includes optical satellite imagery, synthetic aperture radar imagery, UAV remote sensing imagery, and environmental data. The optical satellite imagery is acquired through high-resolution optical satellites such as WorldView-3. Multispectral data, including red-edge band (700-740nm) and near-infrared and short-wave infrared bands sensitive to plant chlorophyll, is primarily collected. Atmospheric correction is applied to the optical satellite imagery to eliminate scattering effects. Cloud and cloud shadow pollution are detected and removed using methods based on NDVI thresholding and morphological operations. Key vegetation spectral indices are extracted as optical features from the preprocessed optical satellite imagery, including Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Photochemical Reflectance Index (PRI). Synthetic Aperture Radar (SAR) imagery is acquired from SAR satellites such as ALOS-2 and Cosmo-Skymed, capturing fully polarimetric SAR images of target areas. The polarization modes include HH, VV, and HV. The SAR imagery undergoes speckle noise filtering to improve image quality, polarimetric target decomposition (e.g., Freeman-Durden decomposition) to extract polarimetric scattering features, and texture features are calculated based on the gray-level co-occurrence matrix. SAR features are extracted from the preprocessed SAR imagery, including polarimetric scattering features such as the backscattering coefficient ratio of HH to VV polarization (HH / VV polarization ratio) and the backscattering coefficient ratio of HV to VV polarization (HV / VV polarization ratio), as well as texture features such as contrast, correlation, energy, and entropy. UAV remote sensing images are acquired by UAV platforms equipped with hyperspectral or multispectral cameras during low-altitude flights over key areas. After preprocessing the UAV remote sensing images, such as standard normal variable transformation, biological characteristics that can characterize the state of the tea tree are extracted. These mainly include leaf shape parameters that describe the leaf morphology, such as rectangularity and roundness, as well as biochemical indicators that reflect the internal chemical composition, such as chlorophyll content and tea polyphenol content.Environmental data is collected through a ground sensor network deployed in a regular grid pattern in the target area. This includes soil moisture, temperature, precipitation, and humidity. After imputing missing values ​​and normalizing the raw data collected by the sensors, environmental characteristics are determined based on the standardized environmental data. These characteristics include the soil moisture index and the comprehensive environmental index. The soil moisture index is determined by the ratio between the difference between the real-time measured soil moisture value and the theoretical minimum soil moisture value in the area, and the difference between the theoretical maximum soil moisture value and the theoretical minimum soil moisture value in the area. The comprehensive environmental index is a weighted sum of temperature, precipitation, and relative humidity. The weighting coefficients for temperature, precipitation, and relative humidity need to be set specifically according to the needs of different growth stages of tea trees to comprehensively reflect the synergistic effects of temperature, precipitation, and humidity on tea tree growth. For example, the weighting coefficients for the vegetative growth stage are 0.4, 0.5, and 0.1, respectively.

[0027] Step S120: Based on the local self-similarity algorithm, perform sub-pixel-level spatial registration of the optical satellite image and the synthetic aperture radar image, and stitch the spatially registered optical features and synthetic aperture radar features into a unified remote sensing feature; based on the spatial attention mechanism, spectral attention mechanism and cross-modal attention mechanism, perform weighted fusion of the optical features, the synthetic aperture radar features and the biological features to generate multi-source fusion features.

[0028] In this embodiment, an improved local self-similarity algorithm is used to perform sub-pixel-level spatial registration of optical satellite imagery and synthetic aperture radar (SAR) imagery. The spatially registered optical features and SAR features are then stitched together to form a unified remote sensing feature. Specifically, firstly, multi-scale geometric features that are insensitive to the sensor are extracted to replace traditional grayscale features, overcoming significant texture differences between optical satellite imagery and SAR imagery. Secondly, an illumination normalization network is introduced to process the feature descriptors, eliminating radiation distortion caused by different imaging incident angles and improving feature comparability. Then, a geometric constraint model is used to iteratively filter matching point pairs, eliminating mismatches to obtain a set of accurate matching points. Finally, spatial transformation parameters are calculated based on the accurate matching points, and the SAR imagery is resampled, ultimately achieving sub-pixel-level registration with a registration error of less than 1 meter to ensure spatial consistency. After registration, the spatially registered optical features and SAR features are horizontally stitched together along the feature dimension to construct a unified remote sensing feature.

[0029] Furthermore, based on spatial attention mechanism, spectral attention mechanism and cross-modal attention mechanism, optical features, synthetic aperture radar features and biological features are weighted and fused to generate multi-source fused features. Specifically, spatial attention weights are calculated based on the red-edge band reflectance features of optical characteristics to generate a spatial weight map, which enhances the features of regions with compact structures, such as the tea tree canopy, in the image. A vegetation masking matrix generated after cloud detection from the optical satellite imagery is used to mask the spatial weight map, retaining only the spatial attention weights of vegetation areas and clearing the weights of non-vegetation areas to zero. Spectral attention weights are calculated based on the red-edge band reflectance features of optical characteristics to focus on and amplify the differences in spectral reflectance between tea trees and associated vegetation in sensitive bands. Using unified remote sensing features as the key and biological features as the query, the attention weights of biological features relative to unified remote sensing features are calculated, and biological features are embedded into unified remote sensing features according to the corresponding attention weights to obtain biological information enhancement features. Finally, optical features are weighted based on spatial attention weights, and synthetic aperture radar features are weighted based on spectral attention weights. The weighted optical features and synthetic aperture radar features are then fused with the biological information enhancement features using weighted fusion. The weighting coefficients are learnable parameters or preset parameters. The final result is a multi-source fusion feature that integrates spectral information, spatial structure information, texture information, and biological information, providing a more discriminative feature basis for the subsequent fine classification of tea trees and natural vegetation.

[0030] Step S130: Based on the optical features, synthetic aperture radar features, biological features, environmental features and multi-source fusion features, an initial feature set is extracted. The initial feature set is then subjected to correlation screening and collinearity screening to obtain a key feature set. Time-series feature data is then constructed based on the key feature set.

[0031] In this embodiment, an initial feature set is extracted based on red-edge band reflectance features, synthetic aperture radar features, biological features, environmental features, and multi-source fusion features. This initial feature set includes four major categories of features: initial optical features, initial synthetic aperture radar features, initial biological association features, and initial environmental features, totaling sixteen items. Among them, the initial optical features include the normalized vegetation index, enhanced vegetation index, and photochemical reflectance index, as well as the red-edge band reflectance features derived from the multi-source fusion features (i.e., the mean and standard deviation of the red-edge band reflectance). The initial synthetic aperture radar features include the HH / VV polarization ratio, HV / VV polarization ratio, contrast, correlation, energy, and entropy, as well as the biological association features, including biological information enhancement features. The initial environmental features include the soil moisture index and the comprehensive environmental index, as well as the environmental features.

[0032] Furthermore, the initial feature set was subjected to correlation and collinearity screening to obtain the most concise and effective set of key features. Specifically, firstly, correlation screening was performed by calculating the absolute value of the Pearson correlation coefficient between each initial feature and the target variable of wild tea tree distribution (such as wild tea tree distribution density). Only initial features with an absolute Pearson correlation coefficient greater than 0.7 were retained. The four features with relatively small impact on the target variable—red-edge band reflectance features, tea polyphenols, rectangularity, and roundness—were initially eliminated, leaving twelve features. Then, collinearity screening was performed by calculating the variance inflation factor of the twelve features after correlation screening. Only initial features with a variance inflation factor less than 3 were retained. The four features with high multicollinearity—energy, entropy, environmental comprehensive index, and chlorophyll content—were further eliminated. After two rounds of screening, the set of key features was obtained, including eight key features: normalized vegetation index, enhanced vegetation index, photochemical reflectance index, HH / VV polarization ratio, HV / VV polarization ratio, contrast, correlation, and soil moisture index.

[0033] Furthermore, time-series feature data for model training is constructed based on an eight-dimensional key feature set. Specifically, to address the contradiction between the satellite revisit cycle and the need for continuous daily input for the model, a differentiated data reconstruction strategy is adopted. For the soil moisture index, the daily average value is directly calculated using high-frequency data collected at fixed intervals by the ground sensor network, thus obtaining a complete daily time-series sequence. For the remaining seven features derived from optical satellite imagery and SAR imagery, on days without satellite overpass data, a combination of harmonic analysis and linear interpolation is used to reconstruct the feature values ​​of the day based on the feature values ​​of the preceding and following effective observation days, thereby forming a continuous daily feature sequence. On this basis, a sliding window method is used to sample on this continuous time series, with a window length of 30 days and a fixed step size for sliding truncation, ultimately constructing a time-series feature data set with dimensions of [number of samples, time step 30, number of features 8], providing standardized input for subsequent model training.

[0034] Step S140: Construct a multi-layer Deep-LSTM network and integrate an adversarial domain adaptive module into the multi-layer Deep-LSTM network. Train the multi-layer Deep-LSTM network using historical time-series feature data from the time-series feature data to obtain the optimal prediction model.

[0035] In this embodiment, a multi-layer Deep-LSTM network is constructed. Temporal feature data is received through the input layer, and the network body, including a three-layer LSTM encoder, processes the temporal feature data to generate a comprehensive encoded feature representation. The first layer LSTM unit contains 128 units and is mainly used to capture long-term dependencies from 30 days of temporal feature data, such as the monthly scale of soil moisture and NDVI variation. The second layer LSTM unit contains 64 units and is mainly used to compress and refine high-level features. The last layer LSTM unit contains 1 unit, and its hidden state at the final time step is used as the comprehensive encoded feature representation of the entire network input sequence. The LSTM units inside the network use the tanh activation function and the sigmoid gating mechanism to alleviate the gradient vanishing problem and precisely control the transmission and discarding of information.

[0036] Furthermore, an adaptation layer and an adversarial domain adaptation module are integrated into a multi-layer Deep-LSTM network to construct an improved Deep-LSTM model. The adversarial domain adaptation module, inserted after the LSTM encoder output and before the prediction layer, comprises a gradient inversion layer and a domain classifier. During model training, the gradient inversion layer transparently transmits data during forward propagation and inverts the gradients passed to the domain classifier during backpropagation. This forces the upstream LSTM encoder to learn domain-invariant features that can deceive the domain classifier. The synthesized encoded feature representation, after processing by the gradient inversion layer, is input into the domain classifier. The domain classifier distinguishes whether the synthesized encoded feature representation originates from a known source domain (i.e., the training region) or an unknown target domain (i.e., the new monitoring region), obtaining the discrimination probability of whether the synthesized encoded feature representation originates from the source or target domain. This discrimination probability is used to calculate the domain adaptation loss, which, together with the loss of the main prediction task, optimizes the model, thereby improving the model's generalization ability to data from different regions and with different parameters. The adaptation layer receives the integrated coded feature representation and the biological information enhancement features. It then fuses the biological features contained in the biological information enhancement features with the integrated coded feature representation to strengthen the correlation between remote sensing features and biological states, resulting in an enhanced temporal feature representation that integrates biological information. Finally, this enhanced temporal feature representation is input to the prediction layer, which calculates and outputs the predicted distribution density of wild tea trees in the target area through a fully connected network. The model's overall loss function consists of the mean squared error loss for continuous value density prediction and the cross-entropy loss for potential auxiliary classification tasks. The mean squared error loss effectively reflects the overall level of prediction error, while the cross-entropy loss accurately quantifies the deviation between the classification result and the true category.

[0037] Finally, the improved Deep-LSTM model was trained and optimized. The training data consisted of historical time-series feature data, which was strictly divided into a training set (e.g., data from 2020-2022), a validation set (e.g., data from the first quarter of 2023), and a test set (e.g., data from the second quarter of 2023) according to time-series logic. Training parameters were set to a fixed input sequence length of 30 days, a prediction step size of 7 days, a batch size of 32, and an initial training epoch of 200 epochs. The optimization process used the Adam optimizer with an initial learning rate of 0.001 and a decay mechanism of multiplying the learning rate by 0.9 every 50 training epochs. An early stopping strategy was also introduced, continuously monitoring the loss on the validation set during training. If the loss did not decrease for five consecutive epochs, training was automatically terminated to prevent overfitting. After training, the system saved the model parameters with the best performance (i.e., the smallest loss) on the validation set, thus obtaining the optimal prediction model for subsequent distribution trend prediction.

[0038] Step S150: Determine the vegetation area using the optical features and the synthetic aperture radar features; determine the tea tree area using the multi-source fusion features to obtain a spatial distribution map of wild tea trees; within the spatial range determined by the spatial distribution map of wild tea trees, predict the trend of wild tea tree distribution density changes in the future period based on the optimal prediction model and the current time-series feature data.

[0039] In this embodiment, a three-layer classification strategy is used to perform coarse screening of vegetation areas, fine identification of tea trees, and dynamic prediction of distribution trends. Specifically, the first layer of classification is used to quickly delineate vegetation and non-vegetation areas. Optical features such as NDVI and EVI, and synthetic aperture radar features such as HH / VV polarization ratio are input into an improved random forest classifier to determine the predicted probability of each pixel in the optical satellite image and synthetic aperture radar image belonging to the vegetation category. Based on historical data analysis, a decision threshold is set, and pixel areas with a predicted probability greater than 0.2 are identified as vegetation areas, while those with a probability less than 0.2 are identified as non-vegetation areas, thus generating a preliminary binary classification map of vegetation and non-vegetation. To improve the reliability of the results, morphological post-processing is performed on the binary classification map of vegetation and non-vegetation to remove isolated small patches with an area of ​​less than 50 pixels in the binary classification map. These patches are usually caused by cloud shadows or sensor noise. Finally, an optimized vegetation area mask is obtained, thereby focusing the subsequent analysis on vegetation areas.

[0040] Furthermore, within the vegetation area determined by the optimized vegetation area mask, wild tea trees are further precisely identified. The multi-source fusion feature vector is input into an improved deep convolutional neural network (DCNN), which includes a spatial attention layer, a spectral attention layer, and a classification layer. The spatial attention layer calculates the weight of each spatial location in the multi-source fusion feature to generate a spatial attention map, focusing on the spatial distribution pattern of the tea tree canopy. The spectral attention layer calculates the weight of each channel of the multi-source fusion feature to generate a channel attention vector, enhancing the spectral differences between the tea trees and associated vegetation. The multi-source fusion feature is then weighted element-wise and channel-wise using the spatial attention map and channel attention vector to obtain attention-enhanced features. These attention-enhanced features are input into the classification layer, outputting the predicted probability of each pixel belonging to the tea tree category. A classification threshold is set based on the prior importance of features obtained during model training (e.g., core features such as NDVI and HH / VV polarization ratio contributing over 60%), and regions with a predicted probability greater than 0.7 are classified as tea tree regions, thus generating a binary classification map of tea trees and associated vegetation. Considering that tea trees may be scattered, a more refined post-processing was applied to the binary classification map of the tea trees and associated vegetation. Only extremely small isolated patches with an area of ​​less than 20 pixels in the binary classification map were removed, and a high-precision spatial distribution map of wild tea trees was finally obtained.

[0041] Furthermore, based on the spatial range determined by the wild tea tree spatial distribution map, the future trend of tea tree distribution density is dynamically predicted through third-level classification. The latest 30-day time-series feature data of the target area is input into the optimal prediction model. This optimal prediction model, through its internal adversarial domain adaptive module, ensures its generalization ability to new data. It predicts the future wild tea tree distribution density in 7-day increments, and concatenates these predictions to form an initial change sequence of wild tea tree distribution density for future periods (e.g., 30 days). This change sequence reveals the initial trend of wild tea tree distribution density in the future period, such as growth, stabilization, or degradation. To further improve the reliability of the prediction, the initial trend of distribution density is compared and corrected with real-time environmental data collected by a ground sensor network. For example, if the model predicts a decrease in distribution density in a certain area but the real-time soil moisture is normal, the prediction result is corrected for reasonableness. After correction, a reliable wild tea tree distribution density change trend map and analysis and prediction report are finally obtained.

[0042] In an optional embodiment, the prediction model is incrementally trained and fine-tuned using the latest data at fixed intervals to adapt to environmental changes. Indicators such as the classification Kappa coefficient and trend prediction accuracy are calculated to evaluate prediction performance. Abnormal areas, such as sudden drops in tea tree distribution density, are automatically identified. Targeted on-site verification is conducted on these identified abnormal areas to confirm the causes, such as pests and diseases. The verified abnormal samples and their characteristics are added to the historical training database to expand the dataset. The prediction model is then retrained using the expanded dataset to learn new abnormal patterns. This iterative update mechanism can improve prediction accuracy from the initial 71.2% to over 92.14%. After the update, an updated wild tea tree distribution map and a report on the trend of wild tea tree distribution density changes over the next 7-30 days are automatically output. Abnormal areas are marked in the report, and specific management suggestions, such as water replenishment and pest and disease control, are provided. This completes an automated closed loop from monitoring, evaluation, learning to continuous optimization, ensuring the continuity, accuracy, and self-optimization capabilities of monitoring and prediction.

[0043] In summary, the embodiments of this application provide a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing. This method acquires optical satellite imagery, synthetic aperture radar (SAR) imagery, UAV remote sensing imagery, and environmental data of the target area to overcome the observation blind spots of single optical data in cloudy and foggy mountainous areas, ensuring the continuity and integrity of data acquisition and enabling all-weather, multi-dimensional observation. Feature extraction is performed on each data source to generate optical features, SAR features, biological features, and environmental features. These features characterize the tea trees from macroscopic state, microscopic morphology, and growth environment perspectives. Based on this multi-layered, refined feature foundation, a comprehensive data basis is laid for subsequent accurate identification and dynamic analysis. The method also utilizes a local self-similarity algorithm to analyze the optical satellite imagery and SAR data. Subpixel-level spatial registration was performed on the synthetic aperture radar (SAR) images to ensure accurate spatial alignment of data from different sources. The spatially registered optical features and SAR features were then stitched together to form a unified remote sensing feature. Based on spatial attention, spectral attention, and cross-modal attention mechanisms, optical features, SAR features, and biological features were weighted and fused to generate a multi-source fusion feature. This dynamically embedded fine biological features into the macroscopic remote sensing feature, compensating for the limitations of satellite remote sensing in characterizing the internal state of plants. This achieved complementarity and synergy of multimodal information, significantly enhancing the key features that distinguish tea trees from their associated vegetation. An initial feature set was extracted based on optical features, SAR features, biological features, environmental features, and multi-source fusion features. The feature set undergoes correlation and collinearity filtering to automatically remove redundant and interfering features, resulting in a key feature set. This effectively reduces feature dimensionality and improves the efficiency, stability, and generalization ability of subsequent model training. Based on this key feature set, time-series feature data is constructed to provide standardized input for subsequent prediction models, supporting trend prediction. A multi-layer Deep-LSTM network is built, integrating an adversarial domain adaptation module. This module enables the model to learn domain-invariant features, effectively overcoming parameter differences between different regions and sensor data, thereby improving the model's generalization and robustness in new monitoring areas. Historical time-series feature data from the time-series feature data is used to refine the multi-layer Deep-LSTM network. The LSTM network is trained to obtain the optimal prediction model, thereby achieving dynamic and accurate prediction of the long-term evolution trend of tea tree distribution density: vegetation areas are determined by optical features and synthetic aperture radar features; tea tree areas are determined by multi-source fusion features, resulting in a spatial distribution map of wild tea trees; within the spatial range determined by the wild tea tree spatial distribution map, based on the optimal prediction model, the distribution density change trend of wild tea trees in future periods is predicted by current time-series feature data. By using a progressive identification method, the computational complexity is greatly reduced, achieving high-precision and high-efficiency positioning of wild tea trees in complex vegetation backgrounds, providing reliable forward-looking scientific basis and decision support for the resource protection, disaster early warning, and sustainable management of wild tea trees.

[0044] It should be noted that, in the embodiments of this application, if the above-mentioned method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0045] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements the steps in any of the above embodiments of the method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing. Correspondingly, embodiments of this application also provide a computer program product. When executed by a processor of an electronic device, this computer program product is used to implement the steps in any of the above embodiments of the method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing.

[0046] Based on the same technical concept, this application provides an electronic device for implementing a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing, as described in the above-described method embodiments. Figure 2 This is a hardware entity diagram of an electronic device provided in an embodiment of this application, such as... Figure 2 As shown, the electronic device 200 includes a memory 210 and a processor 220. The memory 210 stores a computer program that can run on the processor 220. When the processor 220 executes the program, it implements the steps in any of the embodiments of this application of a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing.

[0047] The memory 210 is configured to store instructions and applications executable by the processor 220, and can also cache data to be processed or already processed by the processor 220 and various modules in the electronic device (e.g., image data, audio data, voice communication data and video communication data), which can be implemented by flash memory or random access memory (RAM).

[0048] When the processor 220 executes the program, it implements the steps of a method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing, as described above. The processor 220 typically controls the overall operation of the electronic device 200.

[0049] The aforementioned processor can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that other electronic devices can also implement the functions of the aforementioned processor, and this application does not specifically limit the specific implementation.

[0050] The aforementioned computer storage media / memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0051] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0052] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0053] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0054] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0055] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0056] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0057] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0058] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0059] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0060] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic monitoring and distribution prediction of wild tea trees based on multi-source remote sensing, characterized in that, The method includes: Acquire optical satellite imagery, synthetic aperture radar imagery, UAV remote sensing imagery, and environmental data of the target area and extract features from them respectively to generate optical features, synthetic aperture radar features, biological features, and environmental features; The optical satellite imagery and the synthetic aperture radar imagery are spatially registered at the subpixel level using a local self-similarity algorithm, and the spatially registered optical features and synthetic aperture radar features are stitched together into a unified remote sensing feature. Based on spatial attention mechanism, spectral attention mechanism and cross-modal attention mechanism, the optical features, the synthetic aperture radar features and the biological features are weighted and fused to generate multi-source fused features. An initial feature set is extracted based on the optical features, synthetic aperture radar features, biological features, environmental features, and multi-source fusion features. Correlation and collinearity screening are performed on the initial feature set to obtain a key feature set. Time-series feature data is then constructed based on the key feature set. A multi-layer Deep-LSTM network is constructed, and an adversarial domain adaptive module is integrated into the multi-layer Deep-LSTM network. The multi-layer Deep-LSTM network is trained using historical time-series feature data in the time-series feature data to obtain the optimal prediction model. Vegetation areas are determined using the optical features and synthetic aperture radar features; tea tree areas are determined using the multi-source fusion features, resulting in a spatial distribution map of wild tea trees; within the spatial range determined by the spatial distribution map of wild tea trees, the distribution density change trend of wild tea trees in future periods is predicted based on the optimal prediction model and current time-series feature data.

2. The method according to claim 1, characterized in that, The method, based on spatial attention, spectral attention, and cross-modal attention mechanisms, performs weighted fusion of the optical features, synthetic aperture radar features, and biological features to generate multi-source fused features, including: Spatial attention weights are calculated based on the optical features, and the spatial attention weights are regionally masked based on the vegetation masking matrix of the optical satellite imagery; spectral attention weights are calculated based on the optical features. Using the unified remote sensing feature as the key and the biometric feature as the query, the attention weight of the biometric feature relative to the unified remote sensing feature is calculated, and the biometric feature is embedded into the unified remote sensing feature according to the corresponding attention weight to obtain the biometric information enhancement feature. The optical features are weighted based on the spatial attention weight, the synthetic aperture radar features are weighted based on the spectral attention weight, and then weighted and fused with the bio-information enhancement features to obtain multi-source fusion features.

3. The method according to claim 1, characterized in that, The initial feature set includes initial optical features, initial synthetic aperture radar features, initial bio-related features, and initial environmental features. The initial optical features include the normalized vegetation index, enhanced vegetation index, and photochemical reflectance index, as well as the red-edge band reflectance features derived from the multi-source fusion features. The initial synthetic aperture radar features include the HH / VV polarization ratio, HV / VV polarization ratio, contrast, correlation, energy, and entropy, as well as the bio-related features. The initial bio-related features include the bio-information enhancement features. The initial environmental features include the soil moisture index and the comprehensive environmental index, as well as the environmental features.

4. The method according to claim 1, characterized in that, The process of performing correlation screening and collinearity screening on the initial feature set to obtain a key feature set includes: Calculate the Pearson correlation coefficient between each initial feature and the target variable of wild tea tree distribution, and retain the initial features with an absolute Pearson correlation coefficient greater than 0.7; Calculate the variance inflation factor of initial features with an absolute Pearson correlation coefficient greater than 0.7, and retain initial features with a variance inflation factor less than 3 to obtain a key feature set. The key feature set includes normalized vegetation index, enhanced vegetation index, photochemical reflectance index, HH / VV polarization ratio, HV / VV polarization ratio, contrast, correlation, and soil moisture index.

5. The method according to claim 1, characterized in that, The multi-layer Deep-LSTM network includes an input layer, at least three LSTM encoder layers, an adaptation layer, an adversarial domain adaptive module, and a prediction layer. After receiving temporal feature data, the multi-layer Deep-LSTM network processes it through the following process: Temporal feature data is received through the input layer; the temporal feature data is processed by the three-layer LSTM encoder to generate a comprehensive encoded feature representation. In the adversarial domain adaptation module, the comprehensive encoded feature representation is processed by the gradient inversion layer and then input into the domain classifier to obtain the discrimination probability that the comprehensive encoded feature representation originates from the source domain or the target domain, and the domain adaptation loss is calculated based on the discrimination probability. The adaptation layer receives the integrated coding feature representation and the biological information enhancement feature, and fuses the biological features contained in the biological information enhancement feature with the integrated coding feature representation to obtain an enhanced temporal feature representation with fused biological information. The enhanced temporal feature representation is received by the prediction layer, and the predicted distribution density of wild tea trees in the target area is calculated by a fully connected network.

6. The method according to claim 1, characterized in that, The process involves determining vegetation areas using the optical features and synthetic aperture radar features; determining tea tree areas using the multi-source fusion features to obtain a spatial distribution map of wild tea trees; and within the spatial range determined by the wild tea tree spatial distribution map, predicting the future distribution density change trend of wild tea trees based on the optimal prediction model and current time-series feature data, including: The optical features and synthetic aperture radar features are input into a random forest classifier to determine the predicted probability of each pixel in the optical satellite image and the synthetic aperture radar image belonging to the vegetation category. Regions with a predicted probability greater than 0.2 are identified as vegetation regions, generating a binary classification map of vegetation and non-vegetation. Morphological post-processing is performed on the binary classification map of vegetation and non-vegetation to remove isolated patches with an area of ​​less than 50 pixels in the binary classification map, resulting in an optimized vegetation region mask. Within the vegetation area defined by the optimized vegetation area mask, the multi-source fusion features are input into an improved deep convolutional neural network, which includes a spatial attention layer, a spectral attention layer, and a classification layer. The spatial attention layer calculates the weight of each spatial location in the multi-source fusion features to generate a spatial attention map. The spectral attention layer calculates the weight of each channel of the multi-source fusion features to generate a channel attention vector. The multi-source fusion features are then weighted element-wise and channel-wise using the spatial attention map and the channel attention vector to obtain attention-enhanced features. These attention-enhanced features are input into the classification layer, which outputs the predicted probability of each pixel belonging to the tea tree category. Regions with a predicted probability greater than 0.7 are designated as tea tree regions, generating a binary classification map of tea trees and associated vegetation. The binary classification map of tea trees and associated vegetation is then post-processed to remove isolated patches with an area less than 20 pixels, resulting in a spatial distribution map of wild tea trees. Based on the spatial range determined by the spatial distribution map of wild tea trees, the current time-series feature data is input into the optimal prediction model to predict the initial trend of wild tea tree distribution density in the future period. The initial trend of distribution density is then compared and corrected with real-time environmental data to obtain the distribution density change trend.

7. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 6.

Images