Agricultural drought prediction method based on interpretable deep learning

Abstract

The invention relates to an agricultural drought prediction method based on interpretable deep learning. The method comprises the following steps: acquiring agricultural drought related data of a research area; drought event identification is carried out on the volumetric water content of the soil at different depths; independently constructing GRU models for the soil layers with different depths, and training the GRU models based on the agricultural drought related data; based on the trained GRU model, predicting the volumetric water content of the soil; an expected gradient method is adopted to output a feature importance time sequence of meteorological input, soil layer driving differences, seasonal factors and typical dynamic response types are recognized, and an additive decomposition method is combined to analyze an internal state updating mechanism of a gated cycle unit in a GRU model. According to the method, the transparency and credibility of the model can be improved while the prediction precision is kept, and a scientific basis is provided for agricultural drought mechanism research and disaster prevention and reduction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of agricultural hydrology and meteorological disaster monitoring, and particularly relates to an agricultural drought prediction method based on explainable deep learning. BACKGROUND

[0002] As one of the most serious natural disasters, agricultural drought has a significant impact on crop production and ecological system stability for a long time. The existing drought monitoring methods mainly include two categories: one is the precipitation data-based index (such as the standardized precipitation index), and the other is the soil moisture-based index (such as the standardized soil moisture index, SSMI). Among them, the precipitation-based index is difficult to directly reflect the change of soil moisture in the crop root zone, while the soil moisture-based index can more accurately describe the agricultural drought state, but there are still deficiencies in predicting and explaining the formation mechanism of drought. In terms of prediction methods, although the traditional physical process model can simulate the atmospheric and soil moisture process, it is difficult to meet the demand of long-term monitoring in a large range due to its complex calculation and high requirement for data and computing power. In recent years, data-driven models have been gradually applied to agricultural drought prediction, among which deep learning models show good prediction accuracy due to their strong non-linear modeling capability. However, the existing deep learning models generally have the "black box" problem, lacking the explainability of the internal decision mechanism. Although GRU and LSTM are both improved forms of recurrent neural networks, the GRU structure is more concise, relying only on the update gate and reset gate to control the information flow, and it requires fewer parameters and has higher computational efficiency than LSTM. This makes GRU more advantageous in long-term agricultural drought prediction. Therefore, how to reveal the driving factors and formation mechanism of agricultural drought based on the GRU model while ensuring the prediction accuracy has become a problem that needs to be solved in the current technology. SUMMARY

[0003] The purpose of the present application is to provide an agricultural drought prediction method based on explainable deep learning to solve the problems existing in the prior art. The method uses explainable deep learning to not only accurately predict soil volume moisture content but also explain the meteorological driving factors and GRU internal information processing process behind the prediction.

[0004] To achieve the above purpose, the present application provides the following scheme:

[0005] An agricultural drought prediction method based on explainable deep learning, comprising:

[0006] obtaining agricultural drought-related data of a study area; wherein the agricultural drought-related data includes precipitation, air temperature, evapotranspiration, and soil volume moisture content at different depths;

[0007] identifying drought events for soil volume moisture content at different depths;

[0008] The GRU model is constructed for different depths of soil layers respectively, and the GRU model is trained based on the agricultural drought related data; wherein, the input of the model is precipitation, temperature and evapotranspiration, and the output is soil volume water content;

[0009] Based on the trained GRU model, the soil volume water content is predicted.

[0010] The drought event identification result is subjected to explainability analysis: the expected gradient method is used to output the feature importance time series of meteorological input and identify the soil layer driving difference, seasonal factors and typical dynamic response type, and the additive decomposition method is used to analyze the internal state updating mechanism of the gated recurrent unit in the GRU model.

[0011] Optionally, the drought event identification for soil volume water content at different depths comprises:

[0012] Based on the soil volume water content at different depths, the standardized soil moisture index is calculated respectively.

[0013] The calculated standardized soil moisture index is compared with the preset drought threshold value, and if it meets the preset drought threshold value, it is determined as an agricultural drought event; wherein, the preset drought threshold value is that the standardized soil moisture index is less than -1 and the duration is not less than 15 days.

[0014] Optionally, the GRU model is composed of a GRU layer and a fully connected layer.

[0015] The GRU layer is used to encode the long time series of meteorological input.

[0016] The fully connected layer is used to map the time series features extracted by the GRU layer into a scalar output, i.e. to predict the soil volume water content.

[0017] Optionally, during the training of the GRU model, the Adam optimizer is used, and the early stopping strategy is introduced to prevent overfitting.

[0018] Optionally, after training the GRU model, it further comprises:

[0019] The actual soil volume water content value and the predicted value are used to judge the prediction effect of the model, and the prediction performance of the model is evaluated by NSE, MARE and RMSE indexes.

[0020] Optionally, the expected gradient method is used to output the feature importance time series of meteorological input and identify the soil layer driving difference and the typical dynamic response type, which comprises:

[0021] Based on drought event identification results, the marginal contribution of meteorological inputs is calculated and integrated on multiple baselines to obtain the characteristic importance sequence of meteorological factors within a preset time window. The meteorological driving factors of agricultural drought under different soil layers and time periods are identified through the characteristic importance sequence. Furthermore, typical dynamic response types of different meteorological inputs are identified using K-means + DTW. The meteorological inputs are precipitation, temperature, and evapotranspiration within the preset time window. The meteorological driving factors are divided into precipitation factors and temperature-related factors. Temperature-related inputs include temperature and evapotranspiration. K-means + DTW extracts typical dynamic response types by performing pattern clustering on the characteristic importance time series of meteorological inputs.

[0022] Optionally, the expected gradient method can be used to identify seasonal factors, including:

[0023] The thickness-weighted summation method was used to treat all soil layers as a single layer in order to obtain the occurrence of agricultural drought in different seasons. A preset month was selected as the representative month, and the importance scores of meteorological input to the expected gradient features were compared at different time windows.

[0024] Optionally, the additive decomposition method analyzes the internal state update mechanism of the gated recurrent unit in the GRU model, including:

[0025] The predicted soil volumetric water content is decomposed into time steps, and the activation states of the update gate and reset gate of the GRU model are analyzed. The internal state of the gated recurrent unit in the GRU model is visualized, the expected gradient is verified, and the retention and forgetting characteristics of historical meteorological signals within the model are quantified, thereby revealing the model's memory and forgetting mechanisms.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention not only enhances the interpretability and scientific rigor of deep learning models in agricultural drought research, but also provides operable methodological support for the monitoring and early warning of agricultural drought, and has broad prospects for widespread application. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of an interpretable deep learning method for predicting agricultural drought according to an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] like Figure 1 As shown in the figure, this embodiment proposes an interpretable deep learning method for predicting agricultural drought, including:

[0033] Acquire agricultural drought-related data for the study area; wherein, the agricultural drought-related data includes: precipitation, temperature, evapotranspiration, and soil volumetric water content at different depths;

[0034] Drought events were identified by measuring soil volumetric water content at different depths.

[0035] GRU models were constructed separately for soil layers at different depths, and the GRU models were trained based on the agricultural drought-related data. The inputs of the models were precipitation, temperature, and evapotranspiration, and the output was soil volumetric water content.

[0036] Based on the trained GRU model, predict soil volumetric water content;

[0037] Interpretability analysis of drought event identification results: The expected gradient method is used to output the characteristic importance time series of meteorological inputs and identify soil driving differences, seasonal factors and typical dynamic response types. The state update mechanism inside the gated cyclic unit in the GRU model is analyzed by combining the additive decomposition method.

[0038] Specifically, this embodiment not only enhances the interpretability and scientific rigor of deep learning models in agricultural drought research, but also provides operable methodological support for the monitoring and early warning of agricultural drought, demonstrating broad prospects for widespread application. This embodiment discloses an interpretable deep learning-based method for agricultural drought prediction, which specifically includes the following steps:

[0039] Step 1: Determine the study area and collect agricultural drought-related data;

[0040] Step 2: Identify drought events based on soil moisture index;

[0041] Step 3: Build and train the GRU model to predict soil moisture content, and verify the model performance;

[0042] Step 4: Introduce interpretability methods to interpret the model;

[0043] Step 5: Conduct agricultural drought pattern identification and cluster analysis;

[0044] Step 6: Analyze the results and verify the driving mechanism.

[0045] Further, step 1 involves identifying the study area and collecting agricultural drought-related data; specifically including:

[0046] This embodiment selects the Poyang Lake basin as the study area, which has a typical subtropical monsoon climate with significant seasonal variations in precipitation and temperature, and prominent agricultural drought problems. Data is based on ERA5 reanalysis data. A 0.25°×0.25° regular grid with the same resolution as the data was constructed as the spatial analysis unit, dividing the study area into 256 grids. Daily time-series data of total precipitation, 2-meter air temperature, evapotranspiration, and soil volumetric water content in four layers (0-7 cm, 7-28 cm, 28-100 cm, and 100-289 cm) were collected on these grids. The data spans from 2000 to 2024, with a daily time resolution. Precipitation is used as the "precipitation input," and temperature and evapotranspiration are collectively referred to as the "temperature-related input."

[0047] Further, step 2, drought event identification based on soil volumetric water content at different depths, includes:

[0048] Based on the volumetric water content of soil at different depths, the standardized soil moisture index was calculated.

[0049] The calculated standardized soil moisture index is compared with a preset drought threshold. If the standardized soil moisture index meets the preset drought threshold, it is determined to be an agricultural drought event. The preset drought threshold is: the standardized soil moisture index is less than -1 and the duration is not less than 15 days.

[0050] Specifically, in this embodiment, drought event identification includes:

[0051] Based on the soil volumetric water content at different depths (0-7cm, 7-28cm, 28-100cm, 100-289cm), the standardized soil moisture index (SSMI) was used as the drought criterion, and the soil moisture anomaly was calculated through a one-year sliding window.

[0052] ;

[0053] in, This indicates the volumetric water content of the soil at a specific moment. This refers to the average soil volumetric water content during the reference period. This represents the corresponding standard deviation.

[0054] In terms of drought threshold setting, an agricultural drought event is defined as one where SSMI < –1 and the duration is ≥ 15 days.

[0055] Construct a complete SSMI time series to obtain standardized soil moisture values ​​for continuous time periods;

[0056] When a sequence contains consecutive intervals less than -1 that last for 15 days or more, that interval is marked as a drought event.

[0057] The above steps are performed for different soil depths to identify drought events corresponding to each soil layer from shallow to deep.

[0058] The grid cells that showed drought response in all four soil layers and had no other drought events in the preceding 180 days were selected as the final study samples.

[0059] Furthermore, step 3, constructing and training the GRU model to predict soil moisture content, specifically includes:

[0060] After drought event identification is completed, prediction models for different soil layers need to be constructed. In this embodiment, a separate GRU model is built for each soil layer, with the goal of predicting the soil volumetric water content of that layer based on meteorological input data. Specifically, the method is to use the meteorological sequence of the past 90 consecutive days as input features to predict the soil volumetric water content.

[0061] In terms of structural design, the model in this embodiment consists of "a single GRU layer (16-dimensional hidden units) + a fully connected layer output". Specifically, the GRU network encodes long-term meteorological inputs, selectively retaining and forgetting historical information through update and reset gates, thereby capturing short-term and long-term driving factors in the formation of agricultural droughts; the fully connected layer maps the time-series features extracted by the GRU into a scalar output, namely the predicted soil volumetric water content at the corresponding time.

[0062] During training, the Adam optimizer was used with an initial learning rate of 0.05, and an early stopping strategy was introduced to prevent overfitting (training automatically stopped when the validation set loss did not decrease by ≥0.01 within 30 consecutive epochs, with a maximum training epoch of 100). The data was split into training and test sets in a 7:3 ratio, with 10% of the training set randomly selected as the validation set.

[0063] The predictive performance of the final model was evaluated using NSE, MARE, and RMSE metrics.

[0064] In this embodiment, the model training does not directly use "drought events" as input, but rather predicts soil volumetric water content based on full-time-series meteorological driving factors. After training, the model predictions are compared with actual observations, and interpretability methods (expectation gradient and additive decomposition) are used to analyze the driving factors behind the prediction results. Therefore, the identification results of drought events serve more as an important basis for interpreting and validating the model, while the model itself is trained based on full-time-series data.

[0065] Step 2 identifies the time point of the agricultural drought event. Then, the actual soil volumetric water content and the predicted value are used to judge the prediction effect of the model. Based on the agricultural drought time point identified by the actual soil volumetric water content, the contribution of the predictive factors for each day in the 90 days before the occurrence of the agricultural drought is analyzed using interpretability methods.

[0066] Compared to existing technologies, the GRU model has significant advantages: firstly, it requires less data and is more computationally efficient than traditional physical process models; secondly, compared to deep learning models such as LSTM, the GRU structure is simpler, relying on only two gating mechanisms to retain and update time-series information, requiring fewer parameters, training faster, and still maintaining strong long-term memory capabilities. This makes the GRU model both efficient and accurate in long-term series prediction tasks for agricultural drought.

[0067] Further, step 4 involves using the expected gradient method to output the characteristic importance time series of meteorological inputs and identifying soil layer driving differences and typical dynamic response types, including:

[0068] Based on drought event identification results, the marginal contribution of meteorological inputs is calculated and integrated on multiple baselines to obtain the characteristic importance sequence of meteorological factors within a preset time window. The meteorological driving factors of agricultural drought under different soil layers and time periods are identified through the characteristic importance sequence. Furthermore, typical dynamic response types of different meteorological inputs are identified using K-means + DTW. The meteorological inputs are precipitation, temperature, and evapotranspiration within the preset time window. The meteorological driving factors are divided into precipitation factors and temperature-related factors. Temperature-related inputs include temperature and evapotranspiration. K-means + DTW extracts typical dynamic response types by performing pattern clustering on the characteristic importance time series of meteorological inputs.

[0069] Identifying seasonal factors using the expected gradient method includes:

[0070] The thickness-weighted summation method was used to treat all soil layers as a single layer in order to obtain the occurrence of agricultural drought in different seasons. A preset month was selected as the representative month, and the characteristic importance scores of meteorological driving factors were compared in different time windows.

[0071] Specifically, in this embodiment, the thickness of the four soil layers is weighted and summed, and the thickness-weighted summation method is used to treat the four layers as one layer to obtain the occurrence of agricultural drought in different seasons. January (winter) and July (summer) are selected as representative months to compare the importance scores of meteorological driving factors in different time windows.

[0072] The additive decomposition method analyzes the internal state update mechanism of the gated recurrent unit in the GRU model, including:

[0073] The predicted soil volumetric water content is decomposed into time steps, and the activation status of the update gate and reset gate of the GRU model is analyzed. The retention and forgetting characteristics of historical meteorological signals within the model are quantified, thereby revealing the model's memory and forgetting mechanisms.

[0074] Specifically, in this embodiment, the method for introducing interpretability includes:

[0075] After the model is trained and can output soil volumetric water content prediction results, it is necessary to further introduce interpretable methods to analyze the prediction process and its underlying driving mechanisms. The main inputs to this step are: ① the trained GRU model, ② the meteorological driving factor input sequence, and ③ the corresponding prediction output results.

[0076] Expected Gradient (EG): Used to calculate and integrate the marginal contribution of meteorological inputs across multiple baselines to obtain the sequence of characteristic importance of meteorological driving factors within a preset time window. This method can quantify the temporal contribution of different meteorological driving factors in the prediction process and avoid the bias caused by the dependence on a single baseline, thereby improving the robustness of input importance assessment.

[0077] Additive decomposition (AD): This method is used to decompose the prediction results into each time step, analyze the activation state of the update gate and reset gate of the GRU gated recurrent unit, quantify the retention and forgetting characteristics of historical meteorological signals within the model, and thus reveal the model's memory and forgetting mechanisms.

[0078] The two methods are combined to explain the prediction process from the perspectives of "input importance" and "internal memory mechanism".

[0079] The specific calculation process of the expected gradient (hereinafter referred to as EG) can be expressed by the following formula:

[0080]

[0081] in, Indicates to From uniform distribution The expected sampling operation is performed to evaluate the impact of changes in input features on the model output. In the formula... and The first two characters represent the input features and the baseline input, respectively. One element, and It is the model's prediction function, which depends on the input. During the integration process It is in the interval Uniformly distributed random variables are used to sample the path, simulating variations between different inputs. Finally, through weighted integration, the EG method can calculate the contribution of features to the model's prediction results, thus avoiding the dependence on baseline inputs found in traditional methods.

[0082] The goal of the EG method is to quantify the contribution of each feature to a change in a given input (such as precipitation or temperature) and measure the strength of the feature's influence on the model output by calculating scores for these contributions. Larger positive or negative scores indicate that the feature has a significant positive or negative impact on the model output. In this embodiment, the agricultural drought initiation point identified by the SSMI method is used to calculate... These feature importance scores help understand the model's decision-making process and reveal key features influencing prediction results. To further explore the causes of agricultural drought events, this embodiment analyzes the importance of input features in the 90 days prior to drought occurrence, focusing on precipitation and temperature features, which are categorized as precipitation-dominant and temperature-dominant, respectively. Through EG analysis, this embodiment assesses the importance of precipitation and temperature features at different soil depths and explores long-term driving factors of drought events. Furthermore, using a thickness-weighted summation method, treating four soil layers as a single layer, it obtains the occurrence of agricultural drought in different seasons, selecting January (winter) and July (summer) as representative months to compare the feature importance scores of different meteorological driving factors across different time windows. In addition, this embodiment analyzes the temporal changes in feature importance scores during drought occurrence to determine whether short-term driving factors or factors triggering drought exist. Based on these analyses, this embodiment explores the occurrence mechanism of drought events at different soil depths and the dominant factors of drought in different seasons from a deep learning perspective.

[0083] Additive decomposition (AD) provides a more interpretive perspective for understanding the predictions of GRU models. It breaks down the predictions into contributions from individual features, allowing for a clearer understanding of how GRU makes decisions based on time-series data. The advantage of AD lies in its ability to penetrate the "black box" of GRU, revealing the process of hidden state changes over time and analyzing how the model uses the hidden states from previous time steps and the current input to update its internal state and generate the final output. In this way, AD not only provides a more transparent analytical framework for GRU models but also deepens our understanding of their internal mechanisms.

[0084] Output of the GRU model For the final hidden state based on GRU The calculated (i.e.) Its core purpose is to study hidden states. The source of the signal.

[0085] The GRU standard has been updated (element-wise):

[0086]

[0087]

[0088] ,

[0089] ,

[0090] Therefore, the focus of this embodiment is on studying the hidden state. The source of the signal. By analyzing the hidden state update rules of the GRU, we can obtain:

[0091]

[0092] in, Indicates the retention factor. This represents the new information obtained in the current step. Therefore, the hidden state can be represented as:

[0093]

[0094] The proportion of information retained in the hidden state is only affected by the update gate. The impact. Through a step-by-step iterative process, the final time step can be... The process of generating the hidden state can be decomposed into the following formula:

[0095]

[0096] Using this analytical method, this embodiment is able to Decomposed into time steps from the first time step to the second time step. The summation of information provided by each time step across time steps. Specifically, the first... The information contributed by each time step can be regarded as derived from the information at each time step. Time to step The product of the initially acquired information and the partial information retained in subsequent units through the update gate mechanism. Therefore, the formula can be reformulated as:

[0097]

[0098] The above analysis shows that it is not necessary to know the exact time step. Location, hidden state The specific form of the information contained. The key is that we only need to focus on the hidden state. And the update gate Information flow within.

[0099] In the AD method, the final output y of the GRU model can be represented and decomposed into the following form:

[0100]

[0101] The above formula shows that the final output of the model is composed of the contribution to the final output at each time step. These contributions are the new information acquired at that step and are gradually retained in subsequent units through the update gate mechanism. This decomposition algorithm allows for a deeper understanding of the "black box" structure of the GRU model and analysis of how hidden information is processed. This method is highly efficient, requiring only the extraction of the corresponding vectors from the trained GRU network and the performance of a feedforward operation.

[0102] Further, in step 5, cluster analysis is performed on the interpretation results to identify the dynamic response patterns of different soil layers to meteorological driving factors before the drought, including:

[0103] Perform data standardization on the feature importance score sequences;

[0104] A dynamic time warping algorithm is used to construct warped paths to align the standardized sequences, and the similarity between sequences is calculated by minimizing the cumulative distance.

[0105] Based on the calculated sequence similarity, typical drought-driven response patterns are identified through K-means clustering.

[0106] Step 5 specifically includes:

[0107] Based on the cumulative feature importance scores obtained using the EG method, the contributions of precipitation and temperature-related inputs to the prediction results were analyzed at different time windows and in different seasons. By comparing the feature importance scores of shallow and deep soil layers, the differences between different soil layers on different time scales were identified.

[0108] The feature importance score sequences were used as key feature variables, and K-means clustering was employed for classification. To improve the clustering effect, all sequences were standardized before clustering to ensure a mean of 0 and a standard deviation of 1, thereby eliminating scale differences.

[0109] When calculating the similarity between samples, the Dynamic Time Warping (DTW) method is introduced. DTW aligns time series of different lengths by constructing warped paths and aims to minimize the cumulative distance, thereby obtaining the optimal match between time series. Its mathematical expression is:

[0110] ;

[0111] in, Represents a "normalized path", consisting of a set of index pairs. Composition, used to align time series and The goal of finding regular paths is to minimize the cumulative distance, which is the sum of the squared differences of all matching points.

[0112] By using the "K-means + DTW distance metric" approach, typical patterns of response to meteorological inputs in different soil layers can be identified. For example, shallow soils exhibit a rapid response to temperature factors, while deep soils show a cumulative effect of long-term precipitation signals. These patterns provide a reliable basis for a stratified explanation of the formation mechanism of agricultural drought.

[0113] In this embodiment, EG is used to identify driving patterns, and K-means + DTW is used to identify response patterns. A series of time series (curves showing the contribution of meteorological driving factors over time) are obtained based on GRU+EG.

[0114] Specifically, identifying the driving patterns involves: based on the values ​​obtained from the EG method over a specific time period (e.g., within a 90-day time window, an EG value is obtained daily for each predictor), the precipitation-related and temperature-related EG values ​​are accumulated over the 90 days, and their proportions are analyzed. For example, if the precipitation-related EG value is higher than the temperature-related EG value within the 0-90 day period, it can be determined that the driving pattern for that period is mainly dominated by precipitation. In this study, when analyzing the driving patterns at different depths and in different months, the accumulation of EG values ​​is used to determine whether temperature or precipitation is the dominant driving factor.

[0115] K-means + DTW: This method is used to perform pattern clustering on the time series of feature importance scores of meteorological driving factors, thereby extracting typical dynamic response types. Specifically, the K-means algorithm is used to cluster the feature importance scores of meteorological driving factors to identify typical response patterns; for example, it identifies typical patterns such as "rapid increase in temperature signal 10 days before drought" or "gradual accumulation of precipitation signal 30 days before drought." To address the misalignment of time responses between different grids, the DTW method is used to align the time series. The dynamic time warping function of DTW ensures the accuracy of the clustering results.

[0116] Step 6, Result Analysis:

[0117] Seasonal and vertical distribution characteristics: Statistical results of drought events identified by SSMI show that shallow agricultural drought occurs more frequently in summer and autumn, while deep agricultural drought has a significant lag and is mainly concentrated in winter.

[0118] Predictive performance: All soil layer models showed high prediction accuracy, with deep soil showing better predictive performance than shallow soil.

[0119] Typical drought case analysis: The expected gradient evolution curves of different soil layers show the differences in driving factors: shallow soils are more sensitive to rapid changes in temperature factors, while deep soils mainly depend on the long-term accumulation of precipitation signals.

[0120] Comprehensive analysis of driving factors: shallow layer temperature has a significant effect from 0 to 15 days, while deep layer precipitation accumulation from 45 to 90 days has a dominant influence.

[0121] Seasonal differences: Winter is driven by long-term precipitation deficit, while summer is driven by short-term precipitation reduction and high temperature evaporation.

[0122] Internal state evolution (interpretable additive decomposition method): shallow LSTM focuses on the signal of the past 7-10 days, while deep layers can retain precipitation information of 30-45 days or even longer and release it before drought.

[0123] This embodiment achieves a hierarchical analysis of the driving mechanism: shallow soil drought is mainly driven by short-term high temperatures, while deep soil drought is mainly affected by long-term precipitation deficit, revealing the difference in the driving mechanism of agricultural drought with soil depth.

[0124] The seasonal patterns are clear: winter droughts are mainly caused by prolonged insufficient precipitation and have a significant lag effect; summer droughts are triggered by a sudden drop in precipitation in a short period of time and high temperature evaporation, and are characterized by rapid outbreaks.

[0125] The model memory mechanism reveals that the GRU model efficiently captures long-term time-series information by relying on update gates and reset gates. The shallow model mainly responds to meteorological signals 7–10 days before the occurrence of drought, while the deep model can capture and retain precipitation signals 30–45 days in advance, reflecting the layer-by-layer transmission and cumulative response characteristics of drought in the soil profile.

[0126] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An interpretable deep learning method for predicting agricultural drought, characterized in that, include: Acquire agricultural drought-related data for the study area; wherein, the agricultural drought-related data includes: precipitation, temperature, evapotranspiration, and soil volumetric water content at different depths; Drought events were identified by measuring soil volumetric water content at different depths. GRU models were constructed separately for soil layers at different depths, and the GRU models were trained based on the agricultural drought-related data. The inputs of the models were precipitation, temperature, and evapotranspiration, and the output was soil volumetric water content. Based on the trained GRU model, predict soil volumetric water content; Interpretability analysis of drought event identification results: The expected gradient method is used to output the characteristic importance time series of meteorological inputs and identify soil driving differences, seasonal factors and typical dynamic response types. The state update mechanism inside the gated cyclic unit in the GRU model is analyzed by combining the additive decomposition method.

2. The agricultural drought prediction method based on interpretable deep learning according to claim 1, characterized in that, Drought event identification based on soil volumetric water content at different depths includes: Based on the volumetric water content of soil at different depths, the standardized soil moisture index was calculated. The calculated standardized soil moisture index is compared with a preset drought threshold. If the standardized soil moisture index meets the preset drought threshold, it is determined to be an agricultural drought event. The preset drought threshold is: the standardized soil moisture index is less than -1 and the duration is not less than 15 days.

3. The agricultural drought prediction method based on interpretable deep learning according to claim 1, characterized in that, The GRU model consists of a GRU layer and a fully connected layer. The GRU layer is used to encode long-term meteorological inputs; The fully connected layer is used to map the time series features extracted by the GRU layer into a scalar output, i.e., to predict the soil volumetric water content.

4. The agricultural drought prediction method based on interpretable deep learning according to claim 1, characterized in that, During the training of the GRU model, the Adam optimizer is used, and an early stopping strategy is introduced to prevent overfitting.

5. The agricultural drought prediction method based on interpretable deep learning according to claim 1, characterized in that, Training the GRU model further includes: The predictive performance of the model is judged by using actual soil volumetric water content and predicted values. The predictive performance of the model is evaluated by NSE, MARE, and RMSE indices.

6. The agricultural drought prediction method based on interpretable deep learning according to claim 1, characterized in that, The expected gradient method is used to output the characteristic importance time series of meteorological inputs and identify soil driving differences and typical dynamic response types, including: Based on drought event identification results, the marginal contribution of meteorological inputs is calculated and integrated on multiple baselines to obtain the characteristic importance sequence of meteorological factors within a preset time window. The meteorological driving factors of agricultural drought under different soil layers and time periods are identified through the characteristic importance sequence. Furthermore, typical dynamic response types of different meteorological inputs are identified through K-means + DTW. The meteorological inputs are precipitation, temperature, and evapotranspiration within the preset time window. The meteorological driving factors are divided into precipitation factors and temperature-related factors. Temperature-related inputs include temperature and evapotranspiration. K-means + DTW extracts typical dynamic response types by performing pattern clustering on the characteristic importance time series of meteorological inputs.

7. The agricultural drought prediction method based on interpretable deep learning according to claim 6, characterized in that, Identifying seasonal factors using the expected gradient method includes: The thickness-weighted summation method was used to treat all soil layers as a single layer in order to obtain the occurrence of agricultural drought in different seasons. A preset month was selected as the representative month, and the importance scores of meteorological input to the expected gradient features were compared at different time windows.

8. The agricultural drought prediction method based on interpretable deep learning according to claim 1, characterized in that, The additive decomposition method analyzes the internal state update mechanism of the gated recurrent unit in the GRU model, including: The predicted soil volumetric water content is decomposed into time steps, and the activation states of the update gate and reset gate of the GRU model are analyzed. The internal state of the gated recurrent unit in the GRU model is visualized, the expected gradient is verified, and the retention and forgetting characteristics of historical meteorological signals within the model are quantified, thereby revealing the model's memory and forgetting mechanisms.

Images