Tobacco quality time sequence prediction method and system fusing geographic information and LSTN

Abstract

The present application relates to the field of computer technology, more particularly, the present application relates to a tobacco quality time sequence prediction method and system fusing geographic information and LSTN, the method comprises the following steps: collecting the slope, slope direction and relative elevation index of the monitoring plot, inputting the embedding network to generate a spatial constraint vector; projecting the meteorological data to the hidden space step by step, and calculating the response deviation combined with the last hidden state; inputting the response deviation into two correction networks to map into the correction coefficient vector of the forget gate and the input gate; using the correction coefficient vector to modulate the forget gate and the input gate of the LSTM, updating the cell state and the hidden state; repeating the processing until the end of the growing season, and predicting the tobacco quality with the last hidden state vector. The present application drives the gating correction through the spatial constraint vector and the response deviation, so that the LSTM can dynamically adjust the memory update strategy according to the microclimate difference of the terrain, and improves the accuracy and robustness of the tobacco quality prediction under complex terrain.

Description

Technical Field

[0001] This invention relates to the field of computer technology. More specifically, this invention relates to a method and system for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN. Background Technology

[0002] In tobacco-producing areas with complex terrain, differences in slope, aspect, and relative elevation among different plots result in significant differentiation in solar radiation, temperature decrease, and water redistribution, forming highly heterogeneous microclimate environments. The temporal response patterns of tobacco quality to daily meteorological factors during the growing season exhibit fundamental differences between plots.

[0003] Existing technologies often concatenate static geographic attributes of land parcels with daily meteorological features at each time step and then input the result into a Long Short-Term Memory (LSTM) network to capture temporal dependencies using its gating mechanism. However, static attributes only participate in linear combination as input features, and the gating weights are shared globally, making it impossible to form differentiated memory update strategies for each land parcel. This results in insufficient ability to express spatial heterogeneity and poor prediction robustness under anomalous climatic conditions. Summary of the Invention

[0004] To address the technical problems of existing technologies that fail to develop differentiated memory update strategies for different land parcels, resulting in insufficient spatial heterogeneity representation and poor prediction robustness under abnormal climatic conditions, this invention provides solutions in the following aspects.

[0005] In the first aspect, the method for time-series prediction of tobacco leaf quality that integrates geographic information and LSTN includes: The slope, aspect, and relative elevation index of the monitored plots are collected and spliced ​​into a static terrain attribute vector. The static terrain attribute vector is then input into the embedded network, and a spatial constraint vector is output to characterize the ideal microclimate response mode of each monitored plot. Meteorological observation data is processed into meteorological feature vectors step by step, and then projected onto the same latent space as the spatial constraint vector to obtain meteorological projection features. The hidden state vector of the previous time step is obtained, and the response deviation is calculated based on the meteorological projection features, the spatial constraint vector, and the hidden state vector of the previous time step. The response deviation is input into two correction networks respectively, and mapped to the correction coefficient vector of the forget gate and the correction coefficient vector of the input gate; The original forget gate activation value, the original input gate activation value, and the candidate cell state are calculated based on the meteorological feature vector of the current time step and the hidden state vector of the previous time step. The correction coefficient vector of the forget gate is multiplied element-wise with the original forget gate activation value to obtain the corrected forget gate activation value. The correction coefficient vector of the input gate is multiplied element-wise with the original input gate activation value to obtain the corrected input gate activation value. The cell state vector is updated with the corrected forget gate activation value and the corrected input gate activation value. The hidden state vector of the current time step is calculated based on the updated cell state vector. Repeat the above time-step processing until the end of the growing season, and output the tobacco quality prediction using the hidden state vector of the last time step.

[0006] Optionally, obtaining the relative elevation index includes: Calculate the first difference between the elevation value of the center point of the monitored plot and the lowest elevation value in the study area, calculate the second difference between the highest elevation value and the lowest elevation value in the study area, and use the ratio of the first difference to the second difference as the relative elevation index.

[0007] Optionally, the calculation of the response deviation includes: Calculate the squared Euclidean distance between the meteorological projection feature and the spatial constraint vector, calculate the dot product between the hidden state vector of the previous time step and the spatial constraint vector, map the dot product using the sigmoid function to obtain the modulation factor, and use the product of the squared Euclidean distance and the modulation factor as the response bias.

[0008] Optionally, the embedded network adopts a fully connected layer structure, maps the static terrain attribute vector to a feature space of the same dimension as the LSTM hidden state, and outputs the spatial constraint vector.

[0009] Optionally, the meteorological observation data is processed into a meteorological feature vector, including: standardizing the meteorological observation data and concatenating the standardized meteorological elements in a fixed order to form the meteorological feature vector.

[0010] Optionally, the two correction networks are a forget gate correction network and an input gate correction network, respectively. The parameters of the two networks are independent of each other. Both networks use a fully connected layer structure to map the input response deviation to a correction coefficient vector with the same dimension as the LSTM hidden state vector. The last layer is activated by the sigmoid function so that each element of the correction coefficient vector is in the range of 0 to 1.

[0011] Optionally, updating the cell state vector with the modified gate activation value includes: multiplying the modified forget gate activation value element-wise with the cell state vector of the previous time step, multiplying the modified input gate activation value element-wise with the candidate cell state vector, and then adding the two together to obtain the cell state vector updated at the current time step.

[0012] Optionally, the tobacco quality prediction is output as the hidden state vector of the last time step, including: inputting the hidden state vector of the last time step into the fully connected output layer and obtaining the prediction vector of the tobacco quality index through linear transformation.

[0013] Secondly, a tobacco leaf quality time-series prediction system integrating geographic information and LSTN includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the tobacco leaf quality time-series prediction method integrating geographic information and LSTN described in any one of the above claims is implemented.

[0014] The present invention has the following beneficial effects: 1. This invention introduces a plot-specific spatial constraint vector, encoding topographic attributes such as slope, aspect, and relative elevation as latent spatial anchors that reflect their ideal microclimate response patterns. This enables the model to effectively distinguish the long-term climate response characteristics of different plots from a spatial perspective, avoiding the limitation of traditional methods where static attributes only participate in linear combinations as input features and cannot affect the gating mechanism. This significantly improves the spatial discriminability and plot adaptability of quality prediction for different plots in tobacco planting areas with complex terrain.

[0015] 2. This invention projects real-time meteorological features into a latent space identical to the spatial constraint vector and calculates the response deviation by combining the hidden state of the previous time step, thus quantitatively characterizing the degree of deviation of the current meteorological conditions from the ideal model of each plot. At the same time, it uses the dot product of the hidden state and the spatial constraint vector to obtain the modulation factor through sigmoid mapping, and weights the square of the Euclidean distance, effectively suppressing the interference introduced by small fluctuations in the hidden state, so that the deviation signal remains stable and reliable during the time-series transmission process, and enhances the model's anti-interference ability against daily random meteorological fluctuations.

[0016] 3. This invention utilizes response bias to drive two independent lightweight correction networks, which dynamically generate correction coefficient vectors for the forget gate and the input gate, respectively. The value range of each element of the correction coefficient is between 0 and 1, which realizes element-wise and plot-specific adaptive adjustment of the standard LSTM gate activation value. This enables the model to differentiate the forgetting intensity of historical information and the writing speed of new candidate information under different terrain conditions and meteorological deviations, thus making up for the defect that the uniform gate weight cannot adapt to the microclimate differences between plots.

[0017] 4. This invention applies the modified forget gate and input gate to the cell state update process, enabling the accumulation of long-term memory to be dynamically adjusted in real time according to the inherent topographic conditions of the plot and daily meteorological deviations. After the time-series processing of the entire growing season, the quality prediction is output in the hidden state of the last time step, ensuring the effective aggregation of climate signals throughout the season and the full modeling of the plot-specific response process. Thus, under a unified model framework, the prediction accuracy of normal years and the prediction robustness of abnormal climate years are improved simultaneously. Attached Figure Description

[0018] Figure 1 This is a flowchart of steps S1-S4 in the method for predicting tobacco quality time series by integrating geographic information and LSTN according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] Reference Figure 1 The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN includes steps S1-S4, as follows: S1: Collect static topographic attribute data of the monitored plots and generate spatial constraint vectors to characterize their ideal microclimate response patterns.

[0021] This step focuses on the entire tobacco planting study area, which includes several monitoring plots, each representing an independent plot unit. Subsequent operations use individual plots as the basic processing unit, ultimately generating corresponding spatial constraint vectors for each plot. These vectors characterize the anchor point in the latent state space of the long-term ideal microclimate response pattern determined by the plot's topographic conditions.

[0022] In tobacco-producing areas with complex topography, differences in slope, aspect, and elevation among different plots lead to significant variations in solar radiation, temperature decrease, and moisture redistribution, creating highly heterogeneous microclimate environments. Consequently, the temporal response patterns of tobacco quality to daily meteorological factors such as temperature and precipitation during the growing season exhibit fundamental differences between plots. These differences are deeply correlated with the unchanging topographic attributes of the plots over time; therefore, it is necessary to extract features from static topographic attributes that can reflect the long-term ideal microclimate response patterns of the plots.

[0023] First, high-precision digital elevation model (DEM) data covering the tobacco planting area is used, with a grid resolution of 5 meters or 10 meters. Based on this DEM data, the slope and aspect of each grid are calculated one by one using the 3×3 moving window fitting surface method. For a single monitoring plot, the slope and aspect values ​​of the grid containing the center point of the monitoring plot are extracted.

[0024] Simultaneously, the lowest and highest elevation values ​​of all DEM grids within the entire study area were statistically analyzed, and the elevation value of the center point of each individual monitoring plot was obtained. The difference between the elevation value of the center point of the individual monitoring plot and the lowest elevation value in the entire study area was calculated as the first difference, and the difference between the highest and lowest elevation values ​​in the entire study area was calculated as the second difference. The ratio of the first difference to the second difference was used as the relative elevation index of the individual monitoring plot. The closer the relative elevation index is to 1, the higher the position of the monitoring plot on the regional vertical gradient.

[0025] The slope, aspect, and relative elevation index of each individual monitoring plot are then sequentially concatenated to form a three-dimensional static terrain attribute vector for that plot. This vector is only related to the spatial location of the plot and does not change with time. The above extraction and calculation operations are performed on each monitoring plot within the entire study area to obtain the static terrain attribute vector corresponding to each monitoring plot.

[0026] Next, the static terrain attribute vectors of each monitored plot are input into a learnable embedding network to generate their respective spatial constraint vectors. The embedding network adopts a conventional two-layer fully connected layer structure, the specific implementation of which is a well-known technique in the field and will not be described in detail here: the first layer maps the input from 3D to a preset intermediate dimension and activates it with ReLU; the last layer further maps the intermediate dimension features to the same dimension as the LSTM hidden state and activates the output with tanh.

[0027] In the above-described embedded network, the intermediate dimension is a preset hyperparameter, which is taken as an empirical value of 64 in this embodiment of the invention; the hidden state dimension is taken as an empirical value of 256. The weight matrix and bias vector of the embedded network are both trainable parameters, which are jointly optimized through backpropagation during the overall model training phase.

[0028] The resulting spatial constraint vector is used to characterize the anchor point of the long-term ideal microclimate response model of each monitored plot under its specific topographic conditions in the LSTM hidden state space.

[0029] S2: Extract real-time meteorological features step by step and project them into the latent space, then calculate the response deviation from the spatial constraint vector.

[0030] This step uses each day of the growing season as a time step. For each time step, the current meteorological observation data is extracted and processed into a standardized feature vector. This vector is then projected onto a latent space identical to the spatial constraint vector to calculate the degree of deviation of the real-time meteorological state from the ideal microclimate response model of the plot. This degree of deviation is used to quantitatively characterize the deviation of the current meteorological conditions from the ideal response model of the plot.

[0031] First, at the current time step, read the values ​​of four raw meteorological elements for the corresponding date from the meteorological data table: daily average temperature, daily cumulative rainfall, daily average relative humidity, and daily saturated water vapor pressure difference.

[0032] Furthermore, the Z-score standardization method is used to standardize the above four original meteorological element values, resulting in four dimensionless standardized values. These four standardized values ​​are then concatenated in a fixed order to form a real-time multidimensional meteorological feature vector. This standardization method is a well-known technique in this field and will not be described in detail here.

[0033] Since the aforementioned multidimensional meteorological feature vector is 4-dimensional, while the spatial constraint vector obtained in S1 has a dimension of 256, the two have different dimensions and reside in different feature spaces, making direct comparison impossible. Therefore, a projection matrix is ​​obtained by cropping the weight matrix of the last fully connected layer of the embedded network in S1. This weight matrix has a size of 256×64, and its first four columns are used to form a projection matrix with a dimension of 256×4. The parameters of this projection matrix are optimized along with the embedded network during the overall model training phase and do not need to be set separately.

[0034] Furthermore, a linear transformation is performed on the aforementioned multidimensional meteorological matrix using the projection matrix to obtain the meteorological projection matrix, which is obtained by multiplying the projection matrix with the multidimensional meteorological matrix. Since the projection matrix originates from the weight matrix of the last layer of the embedded network, this linear transformation maps the multidimensional meteorological matrix to a latent space of the same dimension as the spatial constraint vector, while maintaining the same linear basis. This ensures that the meteorological projection matrix and the spatial constraint vector reside in the same feature space, thus achieving physical comparability.

[0035] Next, at the current time step, the meteorological projection features and the spatial constraint vector of the monitored plot have been obtained. At the same time, according to the standard time-series update process of the LSTM network, the hidden state vector of the previous time step has been calculated and saved in the forward propagation. Therefore, the hidden state of the previous time step is used as the current representation of historical memory to participate in the bias calculation.

[0036] Both meteorological projection features and spatial constraint vectors belong to the same dimensional latent space. The squared Euclidean distance between them in this space directly measures the static deviation between the current combination of meteorological elements and the long-term ideal microclimate response model of the plot. However, since the hidden state vector of the previous time step may fluctuate slightly due to gating updates in consecutive time steps, if only the squared Euclidean distance is used as the deviation quantification value, this fluctuation will be directly transmitted to the deviation result, causing unnecessary jitter in the deviation signal between time steps. To suppress this interference introduced by state fluctuations, this step calculates the dot product of the hidden state vector and the spatial constraint vector of the previous time step, and uses the modulation factor obtained by nonlinear mapping of this dot product to weight the squared Euclidean distance.

[0037] Specifically, the dot product of the hidden state vector and the spatial constraint vector of the previous time step is first calculated. This dot product reflects the degree of fit between the model's historical memory and the spatial constraints of the monitored plot in the previous time step: when the dot product is positive and the absolute value is large, it indicates that the direction of the hidden state vector in the previous time step is relatively consistent with the spatial constraint vector, that is, the historical memory fits the ideal pattern of the plot well; when the dot product approaches zero or is negative, the degree of fit is low.

[0038] Subsequently, the dot product is mapped to the (0,1) interval to obtain the modulation factor. Based on this, the response deviation at the current time step is calculated, which satisfies the following relationship: In the formula, For the current time step The response deviation between meteorological projection characteristics and spatial constraint vectors. For the current time step Meteorological projection characteristics, For spatial constraint vectors, This represents the square of the Euclidean distance between the two, used to quantify the absolute deviation of the current meteorological characteristics from the ideal model of the plot in the latent space. For the previous time step The hidden state vector, This represents the vector dot product operation, used to measure the consistency of historical memory states with spatial constraint vectors in direction. For the sigmoid function, is the modulation factor.

[0039] The above Provides a basic metric for static deviation. The weight of this metric is dynamically adjusted based on the degree of fit between historical memory and the ideal model of the plot: when the degree of fit is high, the current deviation metric is retained; when the degree of fit is low, the deviation metric is suppressed, thereby effectively suppressing the interference introduced by the small fluctuations of the hidden state in the previous time step. The larger the value, the more the current meteorological conditions deviate from the normal ideal pattern of the plot, and the model should incorporate more new meteorological information. The smaller the value, the closer the current weather is to the long-term ideal microclimate of the site, and the model should maintain the continuity of historical memory.

[0040] S3: Construct two correction networks to map the response bias scalar to the correction coefficient vectors of the forget gate and the input gate, respectively.

[0041] In existing LSTM architectures, static terrain attributes only participate in linear combination as input features and do not directly affect the forget gate and input gate. The gating weights tend to be globally compromised during training, making it impossible to form differentiated memory update strategies for each plot. To address this, this step constructs two lightweight correction networks that dynamically generate correction coefficients for the forget gate and input gate based on the response deviation of the current weather conditions relative to the ideal pattern for each plot, thus enabling plot-specific gating adjustments.

[0042] After obtaining the corresponding response bias at each time step, this bias is fed into two lightweight correction networks with identical structures but independent parameters: one for generating the forget gate correction coefficient vector, and the other for generating the input gate correction coefficient vector. Since the response bias is a single real-valued scalar, directly adjusting all gating dimensions with this scalar would not allow for differentiated processing of each dimension of the gating vector. Therefore, the response bias needs to be input into the correction network and mapped to a correction coefficient vector of the same dimension as the gating vector, enabling each dimension to obtain independent adjustment coefficients.

[0043] Both the forget gate correction network and the input gate correction network employ a conventional two-layer fully connected structure, the specific implementation of which is well-known in the field and will not be elaborated here: the first layer maps the input response bias to a preset intermediate dimension and activates it via ReLU; the last layer further maps the intermediate dimension to a dimension consistent with the LSTM hidden state vector dimension and activates it via the sigmoid function, outputting a correction coefficient vector where each element's value is within the (0,1) interval. The intermediate dimension of both networks is taken as an empirical value of 16, and the output dimension is taken as 256. The weight matrices and bias vectors of each correction network are trainable parameters, jointly optimized through backpropagation during the end-to-end training of the overall model, without the need for manual preset of specific values.

[0044] Finally, at each time step, the correction coefficient vectors for the forget gate and the input gate are obtained. Both have a dimension of 256, and each element is in the range of (0,1) and is dimensionless. They are used to perform plot-specific adaptive adjustment of the subsequent LSTM standard gating activation values.

[0045] S4: Modulate the LSTM forget gate and input gate using the correction coefficient vectors of the forget gate and input gate and update the cell state to output the tobacco quality prediction with the hidden state at the last time step of the growing season.

[0046] In existing LSTM architectures, the forget gate weight matrix and input gate weight matrix are shared by all monitored plots. After training, these weight parameters can only express the average memory strategy after compromise among plots, and cannot differentiate the information erasure and writing intensity according to the microclimate characteristics of a single plot. Therefore, after obtaining the original gate activation values ​​through standard computation, this step uses the correction coefficient vectors of the forget gate and input gate generated in the previous steps to perform element-wise secondary modulation on the forget gate and input gate, enabling the model to obtain adaptive memory update behavior under the current meteorological deviation state of the current plot. Meanwhile, the standard forward computation of LSTM is a well-known technique in this field, and its operation is only briefly described here.

[0047] First, at each time step, the correction coefficient vectors for the forget gate and the input gate are obtained. The meteorological feature vector of the current time step is concatenated with the hidden state vector of the previous time step along the feature dimension. This concatenation is then passed through the weight matrices and bias vectors of the forget gate, input gate, and output gate, respectively, and activated by the sigmoid function to obtain the original forget gate activation value, the original input gate activation value, and the output gate activation value. Simultaneously, the concatenated vector is activated by the hyperbolic tangent function tanh(⋅) to obtain the candidate cell state. All the aforementioned weight matrices and bias vectors are trainable parameters, jointly optimized through backpropagation during the overall model training phase.

[0048] Next, the correction coefficient vector of the forget gate is multiplied element-wise with the original forget gate activation value to obtain the corrected forget gate activation value. Since both the original forget gate activation value and the correction coefficient vector of the forget gate are non-negative, the values ​​in each dimension after multiplication are less than or equal to the original forget gate activation value, meaning that the erasure intensity of historical information in that dimension is weakened, thereby relatively enhancing the retention of past information. Similarly, the correction coefficient vector of the input gate is multiplied element-wise with the original input gate activation value to obtain the corrected input gate activation value, and the intensity of new information writing is adjusted accordingly. In this step, the output gate activation value is not corrected and the original output gate value calculated by the standard is used directly.

[0049] Then, based on the modified forget gate and input gate, the cell state is updated: In the formula, For the current time step The updated cell state vector, with a dimension of 256, carries the long-term memory information accumulated by the model up to the current time step. For the previous time step The cell state vector stores historical information accumulated from all previous time steps. For the current time step The modified forget gate activation vector is used to control the cell state vector of the previous time step. The proportion of information from each dimension is retained; the closer the value is to 1, the more information in that dimension is retained; the closer the value is to 0, the more information is forgotten. For the current time step The candidate cell state vector is obtained by mapping the current input using the hyperbolic tangent function. Each element takes a value in the interval (-1, 1), is dimensionless, and represents the number of candidate updates to the cell state based on the input information at the current time step. For the current time step The modified input gate activation vector is used to control the candidate cell state. The proportion of information from each dimension being written to the current cell state; a value closer to 1 indicates more new information being written, while a value closer to 0 indicates less being written. This represents the element-wise multiplication operation of vectors.

[0050] This update process determines the long-term memory of the current time step by combining the retained portion of historical information from the previous moment after selective forgetting with the newly added portion of the current input after selective writing. The correction coefficient affects... and It indirectly controls the degree of forgetting historical information and the speed of accumulation of new information.

[0051] Then, based on the updated cell state vector and the original output gate activation values, the hidden state at the current time step is calculated: In the formula, For the current time step The hidden state vector has a dimension of 256 and is dimensionless. It carries the short-term memory information of the current time step and also serves as the input for the next time step in subsequent calculations. This is the original output gate activation vector, with a dimension of 256. Each element takes a value in the interval (0,1), is dimensionless, and is obtained by mapping the current input to the previous hidden state using the sigmoid function. It is used to control... The amount of information from each dimension is output to the current hidden state. The closer the value is to 1, the more information from that dimension is output, and the closer the value is to 0, the less information is output. To convert the cell state vector The result after each element is compressed to the interval (-1,1) by the hyperbolic tangent function is dimensionless; ⊙ represents the element-wise multiplication of vectors. It will be used as the hidden state input for the next time step and will be passed and updated in the time-series loop until it covers all time steps of the entire growing season.

[0052] Since tobacco quality is the result of the combined effects of climate conditions throughout the season, the hidden state vector at the final time step is used as an aggregated representation of the entire process for final prediction. After processing the meteorological feature sequences of all time steps, the hidden state vector output by the LSTM at the final time step has been iteratively updated through the aforementioned steps, accumulating meteorological time-series information related to tobacco quality formation throughout the entire growing season. Because the cell state update and hidden state calculation at each time step have incorporated gating correction based on the plot spatial constraint vector, the features contained in the hidden state vector carry the differentiated response process of the monitored plot under different micrometeorological conditions.

[0053] The hidden state vector at the last time step is input into the fully connected output layer, and a linear transformation is used to obtain the prediction vector of tobacco quality indicators: In the formula, For the calculated prediction vector, each component provides a predicted value for the corresponding quality index. This is the weight matrix of the fully connected output layer, with a size of 256 × the number of tobacco quality indicators to be predicted. In this embodiment, an empirical value of 2 is used, corresponding to nicotine content and total sugar content. For the final time step The hidden state vector, The output bias vector has a length of 2, and is related to the weight matrix. The trainable parameters that together constitute the fully connected layer are jointly optimized through backpropagation during training.

[0054] In the forward prediction phase, the complete daily meteorological sequence of the target plot during the growing season is input into the trained model. After one forward propagation, the predicted values ​​of tobacco quality indicators for the corresponding plot are output by the fully connected output layer.

[0055] For example, suppose there is a monitoring plot A in a tobacco planting research area, with a growing season of 153 days. The daily meteorological data of this plot (daily average temperature, daily cumulative rainfall, daily average relative humidity, and daily saturated vapor pressure difference) are input into a pre-trained tobacco quality prediction model in chronological order.

[0056] The model processes data sequentially along time steps. At each time step, the forget gate and input gate of the LSTM are dynamically adjusted based on the terrain attributes of the monitored plot A (the spatial constraint vector generated by S1) and the real-time response bias (calculated by S2). After 153 iterative updates, the hidden state vector of the last time step has aggregated the meteorological information and its differentiated response process throughout the growing season.

[0057] The hidden state vector of the last time step is input into the fully connected output layer to obtain the prediction vector. Each component in the prediction vector represents the predicted value of the corresponding tobacco quality index that needs to be predicted.

[0058] During the training phase, the model achieves joint optimization of all trainable parameters of the embedding network, projection matrix, lightweight correction network, and LSTM backbone by minimizing the loss function between the predicted and true values.

[0059] The system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the method for time-series prediction of tobacco quality by fusing geographic information and LSTN according to the first aspect of the present invention.

[0060] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0061] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for time-series prediction of tobacco leaf quality integrating geographic information and LSTN, characterized in that, include: The slope, aspect, and relative elevation index of the monitored plots are collected and spliced ​​into a static terrain attribute vector. The static terrain attribute vector is then input into the embedded network, and a spatial constraint vector is output to characterize the ideal microclimate response mode of each monitored plot. Meteorological observation data is processed into meteorological feature vectors step by step, and then projected onto the same latent space as the spatial constraint vector to obtain meteorological projection features. The hidden state vector of the previous time step is obtained, and the response deviation is calculated based on the meteorological projection features, the spatial constraint vector, and the hidden state vector of the previous time step. The response deviation is input into two correction networks respectively, and mapped to the correction coefficient vector of the forget gate and the correction coefficient vector of the input gate; The original forget gate activation value, the original input gate activation value, and the candidate cell state are calculated based on the meteorological feature vector of the current time step and the hidden state vector of the previous time step. The correction coefficient vector of the forget gate is multiplied element-wise with the original forget gate activation value to obtain the corrected forget gate activation value. The correction coefficient vector of the input gate is multiplied element-wise with the original input gate activation value to obtain the corrected input gate activation value. The cell state vector is updated with the corrected forget gate activation value and the corrected input gate activation value. The hidden state vector of the current time step is calculated based on the updated cell state vector. Repeat the above time-step processing until the end of the growing season, and output the tobacco quality prediction using the hidden state vector of the last time step.

2. The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN according to claim 1, characterized in that, The acquisition of the relative elevation index includes: Calculate the first difference between the elevation value of the center point of the monitored plot and the lowest elevation value in the study area, calculate the second difference between the highest elevation value and the lowest elevation value in the study area, and use the ratio of the first difference to the second difference as the relative elevation index.

3. The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN according to claim 1, characterized in that, The calculation of the response deviation includes: Calculate the squared Euclidean distance between the meteorological projection feature and the spatial constraint vector, calculate the dot product between the hidden state vector of the previous time step and the spatial constraint vector, map the dot product using the sigmoid function to obtain the modulation factor, and use the product of the squared Euclidean distance and the modulation factor as the response bias.

4. The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN according to claim 1, characterized in that, The embedded network adopts a fully connected layer structure, which maps the static terrain attribute vector to a feature space of the same dimension as the LSTM hidden state, and outputs the spatial constraint vector.

5. The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN according to claim 1, characterized in that, Processing meteorological observation data into meteorological feature vectors includes: standardizing the meteorological observation data and concatenating the standardized meteorological elements in a fixed order to form the meteorological feature vector.

6. The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN according to claim 1, characterized in that, The two correction networks are the forget gate correction network and the input gate correction network, respectively. Their parameters are independent of each other. Both adopt a fully connected layer structure to map the input response bias to a correction coefficient vector with the same dimension as the LSTM hidden state vector. The last layer is activated by the sigmoid function so that each element of the correction coefficient vector is in the range of 0 to 1.

7. The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN according to claim 1, characterized in that, Updating the cell state vector with the corrected gate activation value includes: multiplying the corrected forget gate activation value element-wise with the cell state vector of the previous time step, multiplying the corrected input gate activation value element-wise with the candidate cell state vector, and then adding the two together to obtain the cell state vector updated at the current time step.

8. The method for time-series prediction of tobacco leaf quality by integrating geographic information and LSTN according to claim 1, characterized in that, The method for outputting tobacco quality predictions using the hidden state vector of the last time step includes: inputting the hidden state vector of the last time step into a fully connected output layer and obtaining the prediction vector of tobacco quality indicators through linear transformation.

9. A time-series prediction system for tobacco leaf quality integrating geographic information and LSTN, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the time-series prediction method for tobacco quality based on the fusion of geographic information and LSTN as described in any one of claims 1-8.

Images