Vegetation growth prediction method and system based on climate data and geographic information

Abstract

The invention provides a vegetation growth prediction method and system based on climate data and geographic information, and is applied to the technical field of vegetation growth prediction.The method comprises the steps that climate data and auxiliary feature data of a target area are acquired; the climate data and the auxiliary feature data are input into a vegetation growth prediction model, a vegetation growth prediction value, output by the vegetation growth prediction model, of the target region is obtained, and the vegetation growth prediction model comprises a dynamic feature processing module used for performing dynamic feature processing based on the climate data to obtain dynamic features; the static feature fusion module is used for performing static feature processing based on the auxiliary feature data to obtain static features; fusing the static features and the dynamic features through a gating mechanism to obtain fused features; and the prediction module is used for determining a vegetation growth prediction value of the target region based on a fusion result of the dynamic features and the static features. According to the method, the vegetation growth prediction precision can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of vegetation growth prediction, and in particular to a vegetation growth prediction method based on climate data and geographic information. Background Art

[0002] Vegetation is a core component of terrestrial ecosystems, playing an irreplaceable role in regulating the water cycle, driving carbon fluxes, and supporting biodiversity. Accurately predicting vegetation growth dynamics is crucial for early warning of ecosystem anomalies, optimizing crop management decisions, and mitigating the risks of vegetation degradation and tree mortality caused by extreme climate events.

[0003] Current research has employed statistical models and neural network methods for vegetation growth prediction. Deep learning methods, with their excellent nonlinear modeling capabilities and robustness to interference, have demonstrated advantages in short- and medium-term forecasting tasks. However, most models still rely primarily on historical vegetation growth indices (such as NDVI or LAI) as input, lacking consideration for environmental factors. These models often overlook the time-lagged effects of climate on vegetation growth and the influence of surrounding landforms, and fail to fully account for interregional ecological coupling, limiting their generalization and adaptability.

[0004] It can be seen from this that the vegetation growth prediction method in the relevant technology has technical problems with great limitations. Summary of the Invention

[0005] The present invention provides a vegetation growth prediction method and system based on climate data and geographic information, which is used to solve the defects of the vegetation growth prediction methods in the existing technology, and achieve a significant improvement in the accuracy of vegetation growth prediction.

[0006] The present invention provides a vegetation growth prediction method based on climate data and geographic information, comprising the following steps.

[0007] Acquire climate data and auxiliary feature data of a target area, wherein the auxiliary feature data is used to represent static geographic information; input the climate data and the auxiliary feature data into a vegetation growth prediction model to obtain a vegetation growth prediction value of the target area output by the vegetation growth prediction model, wherein the vegetation growth prediction model includes: a dynamic feature processing module, used to perform dynamic feature processing based on the climate data to obtain dynamic features; a static feature fusion module, used to perform static feature processing based on the auxiliary feature data to obtain static features; and fuse the static features with the dynamic features through a gating mechanism to obtain fused features; a prediction module, used to determine the vegetation growth prediction value of the target area based on the fusion result of the dynamic features and the static features.

[0008] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, the dynamic feature processing module includes: a multi-scale time-lag attention module, a residual convolutional network, a Transformer encoder and a time attention module; the dynamic feature processing based on the climate data to obtain dynamic features includes: dividing the climate data into lag periods and weighting the attention through the multi-scale time-lag attention module to obtain fused time-lag features of different lag periods; extracting local features of the fused time-lag features through the residual convolutional network to obtain local time series features; extracting global dependencies of the local time series features through the Transformer encoder to obtain global long time series features; and adjusting the time step weight distribution of the global long time series features through the time attention module to obtain dynamic features.

[0009] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, the climate data is divided into lag periods and attention weighted to obtain fused time-lag features of different lag periods, including: splicing the climate data of the current time step with the climate data of multiple consecutive historical time steps to obtain a time-lag information input sequence; using a target number of convolution branches, local feature extraction is performed on the time-lag information input sequence according to the lag period to obtain multiple local time series features; and using a multi-layer perceptron, attention weighting is performed on the multiple spliced ​​local time series features to obtain fused time-lag features of different lag periods.

[0010] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, the time step weight distribution of the global long time series features is adjusted to obtain dynamic features, including: inputting the global long time series features into a time step sliding window, and determining the mean and standard deviation of the predicted values ​​within the time step sliding window; determining the standard residual feedback item of the time step sliding window based on the mean and the standard deviation; generating a modified attention weight of the time step sliding window based on the basic attention score of the time step sliding window and the standard residual feedback item of the time step sliding window; and adjusting the time step weight distribution of the global long time series features based on the modified attention weight of each time step sliding window to obtain dynamic features.

[0011] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, static feature processing is performed based on the auxiliary feature data to obtain static features, including: nonlinear mapping of the auxiliary feature data through a gated recurrent unit and a multilayer perceptron to obtain static features.

[0012] According to the present invention, a vegetation growth prediction method based on climate data and geographic information is provided, wherein the climate data includes at least: average temperature, maximum temperature, minimum temperature, precipitation, maximum precipitation, solar radiation and wind speed; the auxiliary feature data includes at least: digital elevation model, slope, forest canopy height, land cover type and the Euclidean distance between the target area and a specific region.

[0013] The present invention also provides a vegetation growth prediction system based on climate time lag effects and the influence of neighboring land features, comprising the following modules: an acquisition module, for acquiring climate data and auxiliary feature data of a target area, wherein the auxiliary feature data is used to represent static geographic information; a vegetation growth prediction module, for inputting the climate data and the auxiliary feature data into a vegetation growth prediction model, and obtaining a vegetation growth prediction value of the target area output by the vegetation growth prediction model, wherein the vegetation growth prediction model comprises: a dynamic feature processing module, for performing dynamic feature processing based on the climate data to obtain dynamic features; a static feature fusion module, for performing static feature processing based on the auxiliary feature data to obtain static features; and fusing the static features with the dynamic features through a gating mechanism to obtain fused features; a prediction module, for determining the vegetation growth prediction value of the target area based on the fusion result of the dynamic features and the static features.

[0014] The present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for predicting vegetation growth based on climate data and geographic information as described above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting vegetation growth based on climate data and geographic information as described above is implemented.

[0016] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned vegetation growth prediction methods based on climate data and geographic information.

[0017] The vegetation growth prediction method and system based on climate data and geographic information provided by the present invention ensure data comprehensiveness by obtaining climate data and static geographic information as input; the dynamic feature processing module extracts dynamic features from climate data to capture the time-lagged impact of climate on vegetation growth; the static feature fusion module processes static geographic information and fuses it with dynamic features, using a gating mechanism to balance the impact of the two and taking into account geographic neighborhood dependencies; the prediction module outputs vegetation growth prediction values ​​based on the fused features, integrating multi-scale time-lagged features and geographic information to improve prediction accuracy and generalization capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 It is a flow chart of the vegetation growth prediction method based on climate data and geographic information provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of the data-driven vegetation growth prediction model provided by the present invention, which takes into account the climate time lag effect and the influence of neighboring land objects.

[0021] Figure 3 It is a box plot comparing the prediction performance of the model provided by the present invention and five baseline models at 23 consecutive prediction time steps.

[0022] Figure 4 It is a structural diagram of the vegetation growth prediction system based on climate time lag effect and the influence of neighboring land objects provided by the present invention.

[0023] Figure 5 It is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0025] Vegetation is a core component of terrestrial ecosystems and plays an irreplaceable role in maintaining energy balance, regulating water cycle, driving carbon flux changes, and supporting biodiversity. Its spatial distribution and temporal dynamic changes can reflect the health of the ecosystem as well as the regional climate, geological and topographic characteristics.

[0026] Vegetation growth is driven by multiple factors, including climate, topography, and human activities, and exhibits significant climate-induced time-lag responses and spatial dependence. Accurately modeling these complex mechanisms is key to achieving dynamic predictions of vegetation growth.

[0027] Existing studies have used ecological process models, statistical modeling, and neural network methods to predict vegetation growth. Deep learning methods, owing to their excellent nonlinear modeling capabilities and robustness to interference, have demonstrated advantages in short- and medium-term forecasting tasks. However, most models still rely primarily on historical vegetation growth indicator series (such as NDVI or LAI) as input, lacking consideration of environmental factors. These models often overlook the time-lagged effects of climate on vegetation growth and the influence of surrounding landforms, and fail to fully account for interregional ecological coupling, limiting the models' generalization and adaptability. Therefore, it is necessary to develop a vegetation growth prediction method that comprehensively considers multi-scale time-lag characteristics and geographic neighborhood dependencies, in order to develop a widely applicable, high-precision, large-scale vegetation growth prediction capability.

[0028] Optionally, the vegetation growth prediction method based on climate data and geographic information of an embodiment of the present invention can be executed by a server, or by a terminal device, or jointly by a server and a terminal device, taking the example of the vegetation growth prediction method based on climate data and geographic information in this embodiment being executed by a server.

[0029] Figure 1 This is a flow chart of the vegetation growth prediction method based on climate data and geographic information provided by the present invention. Figure 1 As shown, the method includes the following: Step 101: Acquire climate data and auxiliary feature data of a target area, wherein the auxiliary feature data is used to represent static geographic information.

[0030] In an embodiment of the present invention, three main types of data are used: vegetation physiological and ecological parameters or index data, climate data and auxiliary feature data, among which physiological and ecological parameters are used as target values ​​to quantify vegetation growth conditions, climate data are used as dynamic input variables of the model, and auxiliary feature data are used as static input variables of the model.

[0031] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, the climate data includes at least: average temperature, maximum temperature, minimum temperature, precipitation, maximum precipitation, solar radiation and wind speed; The auxiliary feature data include at least: digital elevation model, slope, forest canopy height, land cover type and Euclidean distance between the target area and the specific area.

[0032] Vegetation index data are derived from the Leaf Area Index (LAI) data in the MODIS MOD15A2H product, covering the period from 2002 to 2020, with a temporal resolution of 16 days and a spatial resolution of 500 m. The raw data were declouded and retrieved from the Google Earth Engine platform. They were resampled to a spatial resolution of 0.1° (approximately 10 km), consistent with mainstream climate data such as CMIP6 and ERA5, facilitating spatial scale uniformity in model training and analysis. Furthermore, given that tropical forest vegetation dynamics often exhibit significant spatial correlations at scales of several kilometers, ecological processes such as canopy growth, evapotranspiration, and interactions between vegetation and the atmosphere primarily occur at this scale, the choice of 0.1° strikes a balance between preserving ecological change information and reducing spatial noise, making it suitable for robust modeling in complex tropical environments.

[0033] Climate data comes from the ERA5-Land land reanalysis product, covering seven variables: mean temperature, maximum temperature, minimum temperature, precipitation, maximum precipitation, solar radiation, and wind speed. Spatial resolution is 0.1°, and temporal resolution is daily. To align with model input, all daily variables are aggregated into 16 days of data. Average variables are averaged, while extreme variables are taken as the maximum / minimum values. Units are unified and outliers are removed.

[0034] Ancillary feature data included distance to water bodies, cities, and protected areas, forest canopy height, digital elevation model (DEM), slope, and land cover type. Distance to water bodies and cities was calculated using the Euclidean distance algorithm based on the MCD12Q1 land cover product for the mid-year of the study period (2010). Distance to protected areas was calculated using the World Database on Protected Areas (WDPA) using the same method. Forest canopy height data was derived from the global forest canopy height dataset at 30-meter spatial resolution. DEM data was derived from the Global Terrain Elevation Model, and slope data was extracted from the DEM using the Slope tool in ArcGIS. Land cover type data was derived from the MCD12Q1 product. Because all ancillary feature data vary slowly over time, they were considered static and resampled to 0.1° resolution.

[0035] Through the embodiments of the present invention, by integrating multidimensional climate data (covering core vegetation driving factors such as temperature, precipitation, wind speed, and radiation) with static geographic features (such as topography, land cover type, and neighborhood spatial relationships), it is convenient to subsequently fine-tune modeling of multi-scale time lag effects (such as soil temperature hysteresis) and spatial heterogeneity (such as the impact of slope / distance from water bodies on vegetation distribution) of vegetation growth.

[0036] Step 102: Input the climate data and the auxiliary feature data into a vegetation growth prediction model to obtain a vegetation growth prediction value of the target area output by the vegetation growth prediction model. The vegetation growth prediction model includes: A dynamic feature processing module is used to process dynamic features based on climate data to obtain dynamic features; The static feature fusion module is used to process static features based on auxiliary feature data to obtain static features; and to fuse static features with dynamic features through a gating mechanism to obtain fused features; The prediction module is used to determine the vegetation growth prediction value of the target area based on the fusion results of dynamic features and static features.

[0037] In this embodiment of the present invention, a multi-scale time-lag attention module is used to extract information from different lag periods for dynamic features. The attention mechanism then performs a weighted fusion of these time-lag features. A residual convolutional network is used to further extract local temporal features. A Transformer encoder is then used to capture global long-term temporal dependencies, enhancing the ability to learn nonlinear temporal patterns. To further improve the model's adaptive adjustment capabilities, a temporal attention module with error feedback is designed to dynamically adjust the weight distribution of time steps.

[0038] Furthermore, to effectively utilize static geographic information, the model employs gated recurrent units (GRUs) and multi-layer perceptrons (MLPs) to perform nonlinear mapping of static features. This gating mechanism fuses static and dynamic features, thereby enhancing the model's ability to express spatial information. The following details the design principles and implementation of the main modules.

[0039] In some embodiments, climate data is input into a dynamic feature processing module, and the climate data is divided into lag periods and attention weighted to obtain fused time-lag features of different lag periods; local feature extraction is performed on the fused time-lag features to obtain local time series features; global dependency extraction is performed on the local time series features to obtain global long time series features; and time step weight distribution adjustment is performed on the global long time series features to obtain dynamic features.

[0040] In some embodiments, auxiliary feature data (static geographic information) is input into a static feature fusion module, and the auxiliary feature data is nonlinearly mapped through a gated recurrent unit and a multi-layer perceptron to obtain static features; dynamic features are concatenated with static features; static features are fused with dynamic features through a gating mechanism to generate dynamic-static fusion weights; dynamic features and static features are weighted and summed according to the dynamic-static fusion weights to generate joint spatiotemporal features (i.e., fusion features).

[0041] In some embodiments, the fused features are input into a fully connected layer, and the high-dimensional fused features are compressed by the fully connected layer to match the dimension of the vegetation growth quantitative indicator (such as a single-dimensional LAI).

[0042] Through the above steps of the embodiment of the present invention, by obtaining climate data and static geographic information as input, the comprehensiveness of the data is ensured; the dynamic feature processing module extracts dynamic features from the climate data to capture the time-lagged impact of climate on vegetation growth; the static feature fusion module processes the static geographic information and fuses it with the dynamic features, using a gating mechanism to balance the impact of the two, and taking into account the geographic neighborhood dependency; the prediction module outputs a vegetation growth prediction value based on the fused features, integrating multi-scale time-lagged features and geographic information, and improving the prediction accuracy and generalization ability.

[0043] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, the dynamic feature processing module includes: a multi-scale time-delay attention module, a residual convolutional network, a Transformer encoder and a temporal attention module; Dynamic feature processing is performed based on climate data to obtain dynamic features, including: Through the multi-scale time-lag attention module, the climate data is divided into lag periods and attention weighted to obtain the fused time-lag features of different lag periods; Through the residual convolutional network, local features are extracted from the fused time-delay features to obtain local time series features; Through the Transformer encoder, global dependencies are extracted from local time series features to obtain global long time series features; Through the temporal attention module, the time step weight distribution of the global long time series features is adjusted to obtain dynamic features.

[0044] In this embodiment of the present invention, the multi-scale time-lag attention module concatenates the current year's data with the last 12 time steps of the previous year for each climate variable (such as temperature and precipitation) in the time series data to form an input sequence containing historical lagged information. Each variable independently generates L+1 lagged branches (L corresponds to the number of historical time steps), which are divided into different lagged periods through a right shift operation. For example, a branch with a lag order of 1 shifts the original sequence right by 1 time step. A one-dimensional convolution operation is applied to each lagged branch with a kernel size of 3, a stride of 1, and padding to maintain the sequence length. Local features of each lagged period are extracted. The output features of the L+1 branches are concatenated and input into an MLP (containing two fully connected layers with a Reluctant Unit (ReLU) activation function) to generate an attention weight matrix. The branch features are weighted and summed after softmax normalization to obtain a fused time-lag feature.

[0045] The residual convolutional network consists of four residual blocks connected in series. Each block contains a one-dimensional convolutional layer (with kernel sizes of 3, 7, 5, and 3, respectively, and channel numbers of 64, 64, 256, and hidden_dim, respectively), batch normalization, leaky ReLU activation (negative slope 0.06), dropout, and a skip connection mechanism. When the number of input and output channels is inconsistent, a 1×1 convolution is used for mapping to implement the skip connection; otherwise, the identity mapping is used directly. The convolution output in each block is normalized and added to the skip input. The output is then processed through an activation function and dropout to output local temporal features. This structure is used to extract the fused multivariate time-lag features through layer-by-layer convolution, effectively capturing pattern information within the local time window and outputting a high-dimensional temporal representation suitable for subsequent Transformer encoding.

[0046] The Transformer encoder linearly maps local temporal features to a multidimensional embedding space, superimposes sinusoidal positional encoding to preserve temporal information, and ultimately outputs global long-term temporal features.

[0047] The temporal attention module uses a sliding window (length 7) to calculate the mean and standard deviation of historical prediction errors and normalize the absolute error at the current time step. Global long-term temporal features are fed into the basic attention layer to calculate an initial attention score. The normalized error is mapped into a feedback term using a learnable parameter matrix. This is then added to the basic score and processed through a softmax function to generate the final weights, adjusting the importance of the time step features. The global long-term temporal features are weighted and summed according to the adjusted weights, and the output is a dynamic feature for use in subsequent fusion modules.

[0048] Through the embodiments of the present invention, the convolution operation of the time-delay branch independently processes each climate variable, avoiding feature confusion. The residual convolution alleviates the degradation problem of deep networks through skip connections. The position encoding and multi-head mechanism of the Transformer encoder ensure global dependency modeling, and the error feedback of the temporal attention module enables dynamic weight correction for adaptive prediction error.

[0049] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, climate data is divided into lag periods and attention weighted to obtain fused time-lag features of different lag periods, including: The climate data of the current time step is concatenated with the climate data of multiple consecutive historical time steps to obtain a time-lag information input sequence; Through the target number of convolution branches, local features of the time-delay information input sequence are extracted according to the lag period to obtain multiple local time series features; Through the multi-layer perceptron, attention weighting is performed on the spliced ​​multiple local time series features to obtain the fused time lag features of different lag periods.

[0050] In this embodiment of the present invention, time lags of up to six months (12 time steps) are comprehensively considered. During the data preprocessing stage, the current year's data is concatenated with the last 12 time steps of the previous year to form a time-lagged input sequence. To extract local features at different lag periods, L+1 convolution branches (here L=12) are constructed for each single climate variable, each corresponding to a specific lag order l = 0, 1, …, L. For the lth branch, its input sequence is right-shifted by l time steps. A one-dimensional convolution operation is used for each branch to extract local temporal features. Because different lags have varying importance for prediction, the concatenated features are weighted by attention using a multi-layer perceptron (MLP), and a weighted sum is performed across all lag branches.

[0051] Through the embodiments of the present invention, by splicing current and historical climate data to construct a multi-time-lag input sequence, using multiple convolution branches to independently extract the local time series features of each lag period, and combining with the attention mechanism for dynamic weighted fusion, the multi-scale lagged response of vegetation to different climate factors (such as a 2-month lag in precipitation and the immediate impact of temperature) is effectively captured, and the complex time-lag relationship that traditional linear methods cannot handle is solved through nonlinear modeling.

[0052] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, the time step weight distribution of global long time series features is adjusted to obtain dynamic features, including: Input the global long time series features into the time step sliding window and determine the mean and standard deviation of the predicted values ​​within the time step sliding window; Based on the mean and standard deviation, determine the standard residual feedback term of the time step sliding window; Generate the modified attention weight of the time-step sliding window based on the basic attention score of the time-step sliding window and the standard residual feedback term of the time-step sliding window; Based on the modified attention weight of the sliding window at each time step, the time step weight distribution of the global long time series features is adjusted to obtain dynamic features.

[0053] In an embodiment of the present invention, in order to improve the prediction performance of the model in the time dimension, the present invention proposes a temporal attention module with an error feedback mechanism.

[0054] Traditional attention mechanisms typically rely solely on input features to calculate attention weights, ignoring important information about historical prediction errors. To better calibrate the model's focus on time series information and improve its adaptive adjustment capabilities in time series prediction, this paper designs a dynamic error feedback attention mechanism.

[0055] During the forward propagation process, a sliding window is used to calculate the mean and standard deviation of the predicted values ​​within the local time window, normalizing the absolute prediction error. Softmax is then used to combine the basic attention score and the normalized residual feedback term to generate the final attention weight. This attention weight is dynamically modified to drive dynamic feature calibration. This approach is real-time, with independent calculations at each time step, eliminating the need to wait for the entire sequence to be calculated.

[0056] Through the embodiments of the present invention, the sliding window counts the mean and standard deviation of the prediction error in real time, generates a standardized residual feedback term, and dynamically superimposes it on the basic attention score, so that the model can adaptively correct the time step weight distribution based on historical prediction errors (such as reducing the weight proportion of time periods with larger errors), thereby calibrating feature attention in real time and suppressing error propagation.

[0057] According to a vegetation growth prediction method based on climate data and geographic information provided by the present invention, static feature processing is performed based on auxiliary feature data to obtain static features, including: The auxiliary feature data are nonlinearly mapped through the gated recurrent unit and the multi-layer perceptron to obtain static features.

[0058] In this embodiment of the present invention, in addition to dynamic climate variables, auxiliary feature data (static geographic information) also plays an important role in vegetation growth. Because these static features vary little in the short term, they can provide spatial constraints and context for dynamic features. This invention uses MLP and GRU networks to perform nonlinear mapping on static features, respectively, and then fuses them with dynamic time series features through a gating mechanism. This achieves adaptive feature selection, enabling flexible and effective fusion of static and dynamic information, further improving the model's efficiency in utilizing spatially heterogeneous information.

[0059] Through the embodiments of the present invention, the potential spatial sequence relationship of static features (such as terrain and land cover type) is captured through a gated recurrent unit (such as the synergistic effect of slope and water body distance), and its gating mechanism is used to screen key geographic information (such as the impact of forest canopy height on evapotranspiration). At the same time, a multi-layer perceptron (MLP) is used to perform high-order nonlinear combinations (such as the product effect of DEM and slope) on discrete features (such as land type coding and protected area distance), and the two types of mapping results are fused to form static features.

[0060] The vegetation growth prediction model provided by this invention was trained using data from 2002–2015 as the training set, data from 2016–2019 as the validation set, and data from 2020 as the test set. The Adam optimizer was used, with an initial learning rate of 0.0021, a batch size of 1024, and 100 training epochs. The model employed a two-layer Transformer encoder, each with four attention heads, an embedding dimension of 256, and a dropout probability of 0.2.

[0061] To comprehensively evaluate model performance, we use MAE, MSE, RMSE, and R² to measure mean error, variance sensitivity, error magnitude, and goodness of fit, respectively. All metrics are evaluated spatially, based on pixel dimensions and time step averages. The specific formulas are shown below.

[0062] in, represents the mean absolute error, represents the mean square error, represents the root mean square error, represents the coefficient of determination, represents the number of samples, represents the sample index, represents the predicted value, Indicates the actual value, Indicates the average of the actual values.

[0063] To validate the effectiveness of our model, we compared four classic deep learning models (RNN, LSTM, BiLSTM, and CNN-LSTM) with a statistical prediction model (ARIMAX). While the traditional models used only seven dynamic climate variables, the ARIMAX model used vegetation indices from 2002 to 2019 and fitted these seven dynamic variables.

[0064] In addition, to verify the independent contribution of each module, an incremental ablation experiment based on CNN-Transformer (CT) was designed, introducing the time-delay attention module, the temporal attention module and the geographic information fusion module in sequence, keeping the remaining hyperparameters consistent with the training process.

[0065] refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of the data-driven vegetation growth prediction model provided by the present invention, which takes into account the climate time lag effect and the influence of neighboring land objects.

[0066] like Figure 2 As shown in the figure, the dynamic variable input contains 7 climate time series variables (i.e., climate data, such as temperature and precipitation). The current time step is spliced ​​with the previous 12 historical time steps through the lag period division (Lag of 12 time steps) to form a multi-time lag input sequence (e.g., input dimension: 13×7).

[0067] The static variable input contains seven geographic static variables (such as elevation and land type), which are temporarily expanded to align with the dynamic variable dimensions (such as copying the time dimension) to facilitate subsequent fusion.

[0068] Multi-scale Lag Attention: Use convolution branches (Corn 1d) to independently extract local features for each lag branch, generate attention weights (Softmax normalization) through MLP (W1), and output fused lag features after weighted fusion.

[0069] Residual Conv Block: Contains four layers of one-dimensional convolution, which retains the original information through skip connections and enhances the ability to extract local temporal patterns.

[0070] Transformer encoder: After positional encoding of input features, global dependencies are modeled through multi-head self-attention and feedforward networks (LeakyReLU activation), and global long-term temporal features are output.

[0071] Time Attention Calibration: Introduces an error feedback mechanism to dynamically modify attention weights (matrix addition) based on the prediction residual of a sliding window (e.g., E = Y - Ŷ), suppressing the propagation of historical errors.

[0072] GRU and MLP mapping: Static variables are processed by a gated recurrent unit (GRU) to capture spatial sequence relationships (such as terrain continuity), and then processed by a multi-layer perceptron (MLP) for nonlinear transformation (Tanh / Sigmoid function) to output static features.

[0073] Gated Fusion: Dynamic and static features are adaptively fused through Hadamard product (⊙) and gating weight (G), where the gating value G is generated by the dynamic feature through the Sigmoid function.

[0074] The final output is a vegetation index (Output Layer) used to quantify vegetation growth. It is normalized to a reasonable range using the Sigmoid function. The baseline models (RNN, LSTM, BiLSTM, CNN-LSTM, and ARIMAX) are listed for performance comparison.

[0075] MAE, MSE, RMSE, and R² are used to quantify the prediction accuracy (marked in the light blue area) for horizontal comparison.

[0076] refer to Figure 3 , Figure 3 It is a box plot comparing the prediction performance of the model provided by the present invention and five baseline models at 23 consecutive prediction time steps.

[0077] like Figure 3 Figures (a)–(d) show the distributions of MAE, MSE, RMSE, and R², respectively. The upper and lower bounds of each box represent the 75th and 25th percentiles, respectively, and the horizontal line in the middle of the box represents the median. The whiskers indicate the observed maximum and minimum values, and the scattered points indicate the specific values ​​corresponding to each prediction time step. A tighter distribution of boxes and scattered points indicates greater model stability, smaller prediction errors, and better fitting accuracy. This shows that the proposed vegetation growth prediction model has superior predictive performance and is much more stable than other baseline models.

[0078] The vegetation growth prediction system based on climate time lag effect and influence of neighboring land objects provided by the present invention is described below. The vegetation growth prediction system based on climate time lag effect and influence of neighboring land objects described below and the vegetation growth prediction method based on climate time lag effect and influence of neighboring land objects described above can be referenced to each other.

[0079] refer to Figure 4 , Figure 4 It is a structural diagram of the vegetation growth prediction system based on climate time lag effect and the influence of neighboring land objects provided by the present invention.

[0080] An acquisition module 401 acquires climate data and auxiliary feature data of a target area, wherein the auxiliary feature data is used to represent static geographic information; The vegetation growth prediction module 402 inputs the climate data and the auxiliary feature data into the vegetation growth prediction model to obtain the vegetation growth prediction value of the target area output by the vegetation growth prediction model, wherein the vegetation growth prediction model includes: A dynamic feature processing module is used to process dynamic features based on climate data to obtain dynamic features; The static feature fusion module is used to process static features based on auxiliary feature data to obtain static features; and to fuse static features with dynamic features through a gating mechanism to obtain fused features; The prediction module is used to determine the vegetation growth prediction value of the target area based on the fusion results of dynamic features and static features.

[0081] Specifically, the vegetation growth prediction system based on climate time lag effects and the influence of neighboring land objects provided by the present invention can implement all the method steps implemented in the above-mentioned vegetation growth prediction method embodiment based on climate time lag effects and the influence of neighboring land objects, and can achieve the same technical effect. The parts and beneficial effects of this embodiment that are the same as those of the method embodiment will not be described in detail here.

[0082] Figure 5 This is a schematic diagram of the physical structure of the electronic device provided by the present invention, such as Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communications bus 540. The processor 510 may call logic instructions in the memory 530 to execute a vegetation growth prediction method based on climate data and geographic information. The method includes: obtaining climate data and auxiliary feature data of a target area, wherein the auxiliary feature data is used to represent static geographic information; inputting the climate data and the auxiliary feature data into a vegetation growth prediction model to obtain a vegetation growth prediction value for the target area output by the vegetation growth prediction model; wherein the vegetation growth prediction model includes: a dynamic feature processing module for performing dynamic feature processing based on the climate data to obtain dynamic features; a static feature fusion module for performing static feature processing based on the auxiliary feature data to obtain static features; and fusing the static features with the dynamic features through a gating mechanism to obtain fused features; and a prediction module for determining the vegetation growth prediction value for the target area based on the fusion result of the dynamic and static features.

[0083] Furthermore, the logic instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0084] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the vegetation growth prediction method based on climate data and geographic information provided by the above methods, the method including: obtaining climate data and auxiliary feature data of the target area, wherein the auxiliary feature data is used to represent static geographic information; inputting the climate data and auxiliary feature data into a vegetation growth prediction model to obtain a vegetation growth prediction value of the target area output by the vegetation growth prediction model, wherein the vegetation growth prediction model includes: a dynamic feature processing module, which is used to perform dynamic feature processing based on climate data to obtain dynamic features; a static feature fusion module, which is used to perform static feature processing based on auxiliary feature data to obtain static features; and fusing static features and dynamic features through a gating mechanism to obtain fused features; a prediction module, which is used to determine the vegetation growth prediction value of the target area based on the fusion result of dynamic features and static features.

[0085] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the vegetation growth prediction method based on climate data and geographic information provided by the above-mentioned methods, the method comprising: obtaining climate data and auxiliary feature data of the target area, wherein the auxiliary feature data is used to represent static geographic information; inputting the climate data and auxiliary feature data into a vegetation growth prediction model to obtain a vegetation growth prediction value of the target area output by the vegetation growth prediction model, wherein the vegetation growth prediction model comprises: a dynamic feature processing module for performing dynamic feature processing based on climate data to obtain dynamic features; a static feature fusion module for performing static feature processing based on auxiliary feature data to obtain static features; and fusing static features with dynamic features through a gating mechanism to obtain fused features; a prediction module for determining the vegetation growth prediction value of the target area based on the fusion result of dynamic features and static features.

[0086] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0087] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A vegetation growth prediction method based on climate data and geographic information, characterized in that: include: Acquiring climate data and auxiliary feature data of a target area, wherein the auxiliary feature data is used to represent static geographic information; Inputting the climate data and the auxiliary feature data into a vegetation growth prediction model to obtain a vegetation growth prediction value of the target area output by the vegetation growth prediction model, wherein the vegetation growth prediction model includes: A dynamic feature processing module, configured to perform dynamic feature processing based on the climate data to obtain dynamic features; A static feature fusion module is used to perform static feature processing based on the auxiliary feature data to obtain static features; and to fuse the static features with the dynamic features through a gating mechanism to obtain fused features; The prediction module is used to determine the vegetation growth prediction value of the target area based on the fusion result of the dynamic features and the static features.

2. The vegetation growth prediction method based on climate data and geographic information according to claim 1, characterized in that: The dynamic feature processing module includes: a multi-scale time-delay attention module, a residual convolutional network, a Transformer encoder and a temporal attention module; The performing dynamic feature processing based on the climate data to obtain dynamic features includes: The climate data is divided into lag periods and attention weighted by the multi-scale time-lag attention module to obtain fused time-lag features of different lag periods; Performing local feature extraction on the fused time-lag features through a residual convolutional network to obtain local time series features; Performing global dependency extraction on the local temporal features through the Transformer encoder to obtain global long-term temporal features; Through the time attention module, the time step weight distribution of the global long time series features is adjusted to obtain dynamic features.

3. The vegetation growth prediction method based on climate data and geographic information according to claim 2, characterized in that: The lag period division and attention weighting of the climate data are performed to obtain fusion time lag features of different lag periods, including: The climate data of the current time step is concatenated with the climate data of multiple consecutive historical time steps to obtain a time-lag information input sequence; By using a target number of convolution branches, local features are extracted from the time-delay information input sequence according to the lag period to obtain multiple local time series features; Through a multi-layer perceptron, attention weighting is performed on the multiple local time series features after splicing to obtain fused time lag features of different lag periods.

4. The vegetation growth prediction method based on climate data and geographic information according to claim 2, characterized in that: The step of adjusting the time step weight distribution of the global long time series features to obtain dynamic features includes: Inputting the global long time series features into a time step sliding window, and determining the mean and standard deviation of the predicted values ​​within the time step sliding window; Determining a standard residual feedback item of the time step sliding window based on the mean and the standard deviation; Generate a modified attention weight for the time-step sliding window based on the base attention score of the time-step sliding window and the standard residual feedback term of the time-step sliding window; Based on the modified attention weight of each time-step sliding window, the time-step weight distribution of the global long-term time series feature is adjusted to obtain a dynamic feature.

5. The vegetation growth prediction method based on climate data and geographic information according to claim 1, characterized in that: The performing static feature processing based on the auxiliary feature data to obtain static features includes: The auxiliary feature data are nonlinearly mapped by a gated recurrent unit and a multilayer perceptron to obtain static features.

6. The vegetation growth prediction method based on climate data and geographic information according to claim 1, characterized in that: The climate data includes at least: average temperature, maximum temperature, minimum temperature, precipitation, maximum precipitation, solar radiation and wind speed; The auxiliary feature data at least include: digital elevation model, slope, forest canopy height, land cover type and the Euclidean distance between the target area and a specific region.

7. A vegetation growth prediction system based on climate time lag effect and the influence of neighboring land features, characterized by: include: An acquisition module, which acquires climate data and auxiliary feature data of a target area, wherein the auxiliary feature data is used to represent static geographic information; The vegetation growth prediction module inputs the climate data and the auxiliary feature data into a vegetation growth prediction model to obtain a vegetation growth prediction value of the target area output by the vegetation growth prediction model, wherein the vegetation growth prediction model includes: A dynamic feature processing module, configured to perform dynamic feature processing based on the climate data to obtain dynamic features; A static feature fusion module is used to perform static feature processing based on the auxiliary feature data to obtain static features; and to fuse the static features with the dynamic features through a gating mechanism to obtain fused features; The prediction module is used to determine the vegetation growth prediction value of the target area based on the fusion result of the dynamic features and the static features.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the vegetation growth prediction method based on climate data and geographic information as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting vegetation growth based on climate data and geographic information as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for predicting vegetation growth based on climate data and geographic information as claimed in any one of claims 1 to 6 is implemented.

Images