Artificial forest land deep soil moisture utilization vertical distribution estimation method

Abstract

The invention discloses a vertical distribution estimation method for water utilization of deep soil of a man-made forest land, and belongs to the technical field of ecological hydrological modeling. The invention aims to solve the problems of high field mass sampling cost, large physical model simulation error or incapability of obtaining a soil moisture utilization vertical distribution rule in the prior art. The core innovation of the invention lies in that depth information is used as a key numerical input feature, and the depth information and input factors of soil, weather, vegetation and the like jointly construct an input quantity of a machine learning model; and taking the soil water deficit amount representing water utilization as an output target of the corresponding depth. According to the method, the accurate mapping relation between the input factors and the specific depth deficit amount is established, so that direct prediction of the soil water deficit amount under different conditions and at different depths is achieved, fine vertical distribution of deep soil water utilization is obtained, and technical support is provided for sustainable evaluation of regional ecological restoration.

Description

Technical Field

[0001] This invention relates to the field of eco-hydrological modeling and machine learning application technology, and in particular to a data-driven method for estimating the vertical distribution of deep soil moisture use in artificial forests. Background Technology

[0002] Plantations, as an important means of ecological restoration, have played a significant role in mitigating soil erosion in ecologically fragile areas such as the Loess Plateau. However, the rapid expansion of plantations has also led to the continuous depletion of deep soil moisture, resulting in soil drying and posing a potential threat to the long-term stability of regional ecosystems. Deep soil water is a critical water source for deep-rooted woody plants in arid and semi-arid regions; therefore, accurately quantifying the utilization of deep soil moisture by plantations is of great significance for assessing the sustainability of ecological restoration and optimizing forestry management.

[0003] When quantifying the use of deep soil moisture by vegetation, due to the difficulty of directly measuring root water absorption, the academic community usually adopts an indirect but effective method: that is, by comparing the soil water storage of artificial forest land with that of control land (such as wasteland) that can reflect the natural background state, the difference is defined as soil water deficit, which is used as the core quantitative indicator for measuring soil water use. Existing studies are mostly based on this indicator.

[0004] Currently, there are significant limitations in the technical means of obtaining information on deep soil moisture use. Traditional field drilling and sampling is the main method for obtaining deep soil moisture data, but this method is costly, time-consuming, and difficult to deploy over a large area and at multiple time points. This results in extremely scarce deep soil moisture data (e.g., at depths of 20 meters or more), making it very difficult to obtain relevant information. Soil hydrological models based on physical processes are another commonly used tool, but their application is often limited by difficulties in obtaining parameters, complex model structures, and insufficient ability to handle spatial heterogeneity, which restricts the accuracy and regional applicability of simulation results.

[0005] To overcome the shortcomings of traditional methods, machine learning methods, represented by random forests (RF), have begun to be applied to soil moisture estimation. Existing research shows that machine learning methods have advantages in handling nonlinear relationships and have been successfully used to estimate surface or rhizosphere soil moisture. However, directly applying existing machine learning methods to the refined estimation of deep soil moisture use still faces technical challenges.

[0006] First, most existing machine learning applications focus on surface or shallow root zone soils, and the reliability of inferring deep soil water storage from shallow measurement data remains unclear; simple extrapolation may introduce significant errors. Second, even when studies extend machine learning methods to estimate deep soil water deficit, the model construction typically aims to predict the total deficit across the entire profile, outputting a single macroscopic indicator. This method fails to provide information on the vertical distribution of deficit within the soil profile, i.e., it cannot answer how much water was consumed at which depth. This lack of information on the vertical distribution of deficit limits a deeper understanding of specific hydrological mechanisms such as vegetation water use strategies, key water-absorbing layers, and the development process of soil desiccation, and also makes data-driven decisions on refined water resource management for different soil layers lacking.

[0007] Therefore, there is an urgent need in this field for a technical solution that can accurately and efficiently estimate the vertical distribution of deep soil moisture utilization in planted forests, in order to address the shortcomings of existing technologies in providing information on deep and stratified soil moisture deficit. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies in estimating the vertical distribution of deep soil moisture use in planted forests. Specifically, traditional methods (such as drilling) are costly, and the physical model parameters are complex. Although existing studies have used machine learning methods to estimate soil moisture, there are two main limitations: first, the indirect calculation method of modeling forest land and control land separately and then calculating the difference accumulates errors; second, even if there are methods to directly model soil water deficit, they are often estimated as a whole, a single value, which only yields an overall water use figure and cannot reveal the vertical distribution pattern of water use with soil depth.

[0009] Therefore, the core technical problem to be solved by this invention is: how to establish a model that can directly, accurately and efficiently estimate the vertical distribution of water deficit in deep soil of artificial forests, and thereby refine the vertical pattern of soil water use.

[0010] To address the aforementioned technical problems, this invention proposes a method for estimating the vertical distribution of deep soil moisture utilization in artificial forests.

[0011] The core innovation of this invention lies in the fact that instead of modeling soil water deficit as a single, overall value, it directly trains a machine learning model using sequential data containing vertical dimension information—the soil water deficit from the initial layer to different depth layers—as the dependent variable. This allows the model to directly learn the mapping relationship between environmental factors and the vertical distribution of water use.

[0012] Specifically, the present invention is achieved through the following technical solution: A method for estimating the vertical distribution of deep soil moisture use in planted forests includes the following steps: S1) Obtain soil moisture content data of deep profiles of artificial forest land and control land as reference in the study area, and calculate a set of soil water deficits from the initial layer to different depth layers of artificial forest land representing the vertical profile. S2) Construct a dataset where each data point in the dataset includes: a) Using each depth and at least one environmental factor associated with that depth as a set of input features, wherein both the depth and the at least one environmental factor are input factors; b) The soil water deficit corresponding to the depth is used as the dependent variable; S3) A machine learning model is used to train the dataset to establish a mapping relationship between the input factors and the dependent variable, thereby obtaining a prediction model that can directly predict the soil water deficit at that depth based on environmental factors and depth.

[0013] Furthermore, the target depth in S2) is input into the machine learning model as a numerical feature.

[0014] Furthermore, the machine learning model is an ensemble learning model.

[0015] Furthermore, the ensemble learning model is a random forest model.

[0016] Furthermore, the control area in S1) is a wasteland or farmland that can reflect the natural background state of local soil moisture.

[0017] Furthermore, the soil water deficit in S1) is calculated using the following formula:

[0018] in the formula ΔSWS d It is the soil water deficit from the initial layer to different depth layers of the artificial forest; θ f , θ c These are the soil moisture contents of the artificial forest and the control area, respectively. z It is the sampling interval; i It is the sampling sequence number.

[0019] Furthermore, the environmental factors in S2) include at least one of soil data, meteorological data, and vegetation data.

[0020] Furthermore, the soil data includes sand, silt, and clay particles at different depths and their corresponding geographical location information; the meteorological data includes precipitation, potential evapotranspiration and their corresponding geographical location information; and the vegetation data includes forest age, tree species and their corresponding geographical location information.

[0021] Compared with the prior art, the present invention has the following significant advantages: (1) Achieves direct and accurate estimation of the "vertical distribution" of water use: This invention uses the deficit sequence containing depth information as the modeling target, enabling the model to directly capture and predict the pattern of water use variation with depth. This is fundamentally different in technical concept from methods that can only estimate the "total deficit," providing richer and more refined decision-making information. It solves the problem that existing technologies cannot effectively obtain the vertical profile of soil water deficit, providing key data support for the refined assessment of vegetation water use.

[0022] (2) The technical solution is novel, and the model has high efficiency and reliability: This invention avoids the cumbersome steps and error accumulation problems of the traditional "separate modeling and then subtraction". By directly modeling the vertically distributed data, the model can learn the inherent laws more efficiently. The results of the embodiments show that the vertically distributed data predicted by the model constructed by this method is in high agreement with the actual observation values, which reflects the feasibility and effectiveness of this technical solution. Attached Figure Description

[0023] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described 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.

[0024] Figure 1 This is a schematic diagram of the research area and sample point distribution used in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram showing the sensitivity analysis results of key hyperparameters of the random forest model in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram illustrating the changes in model performance with the ratio of training set to test set in an embodiment of the present invention.

[0027] Figure 4 This is a comparison and evaluation chart of the soil water deficit predicted by the model and the measured value in an embodiment of the present invention.

[0028] Figure 5 This is a SHAP analysis diagram showing the impact of each input factor on the model output in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram illustrating the variation of the importance of each input factor with depth in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments described are merely illustrative of the invention and are not intended to limit its scope. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this invention should be included within the protection scope of this invention.

[0031] This embodiment provides a specific way to implement the method described in this invention. This embodiment is carried out against the background of the Loess Plateau in China (100°54′-114°33′ E, 33°43′-41°16′ N). Figure 1 This diagram illustrates the study area and the spatial distribution of the study sampling points in this embodiment. This embodiment involves 72 pairs of soil profiles from forest and control areas, which are derived from… Figure 1 The geographical locations shown include some locations that contain multiple sets of paired profile data. In this embodiment, the difference in deep soil water storage between the forest land and the control land is calculated, i.e., "soil water deficit". ΔSWS d This indicator is used as a core indicator for quantifying and estimating "deep soil moisture use in vegetation". The "soil water deficit" mentioned below refers to this calculated indicator.

[0032] Step S1: First, basic data for calculating soil water deficit is obtained. This process employs a comparative analysis method, quantifying the impact of plantations by selecting representative plantations and paired control sites (e.g., farmland or grassland) that reflect the original soil moisture conditions. The paired plantations and control sites have similar topographical conditions. Data sources may include actual field sampling and / or publicly available literature. In this embodiment, 33 pairs of plantations and control sites were selected for actual field sampling. Soil profiles ranging from 3 to 25 meters in depth were collected using a soil drill, with sampling intervals of 20 cm, layer by layer. The mass moisture content of each soil layer was determined by drying in a 105°C oven. In addition, 16 published data that met the requirements for paired experiments were screened from academic databases, totaling 57 pairs of profile data, to expand the dataset.

[0033] The formula for calculating the forest profile deficit below the infiltration depth is as follows:

[0034] in the formula ΔSWSd It is the soil water deficit from the initial layer to different depth layers of the artificial forest; θ f , θ c These are the soil moisture contents of the forest land and the control land, respectively. z The sampling interval is 20cm in this embodiment. i This is the sampling sequence number. Therefore, each sampling point yields a series of soil water deficit values ​​that correspond one-to-one with the depth information.

[0035] Step S2: Construct the dataset for model training. First, obtain the input factor data. Soil data can be obtained from the World Soil Coordination Database (HWSD2.0), extracting the average content of sand, silt, and clay particles in the 0-200cm soil layer at each sampling point. Meteorological data can be obtained from the data sharing service system of the Institute of Atmospheric Physics, Chinese Academy of Sciences, extracting the multi-year average precipitation and multi-year average potential evapotranspiration data corresponding to each sampling point. Vegetation data can be obtained from field surveys and literature records, including the age and tree species information of plantations at each sampling point.

[0036] Subsequently, the aforementioned data were integrated to construct the final dataset. Each data point in this dataset represents the state of a specific sample point at a specific depth and includes the following fields: depth information as input factors, geographical location information (latitude and longitude), sand content, silt content, clay content, average annual precipitation, average annual potential evapotranspiration, forest age, and tree species; and soil water deficit as the dependent variable, which strictly corresponds to the depth information. ΔSWS d In this embodiment, depth information is used as a numerical feature input to the model. Sand content, silt content, clay content, average annual precipitation, average annual potential evapotranspiration, forest age, and tree species are used as independent variables. Geographical location information (latitude and longitude) is mainly used in this embodiment to accurately match and correlate data from different sources spatially, ensuring that the data for each sample point corresponds accurately. It is not used as an independent variable to be directly input into the random forest model for training. The structure of this dataset allows the machine learning model to learn how depth, as a continuous variable, works in conjunction with other independent variables (sand content, silt content, clay content, average annual precipitation, average annual potential evapotranspiration, forest age, and tree species) to affect soil water deficit during training.

[0037] Step S3 involves constructing, training, and evaluating a machine learning model using the constructed dataset. The technical solution of this invention can employ various machine learning models capable of handling structured regression problems. This embodiment uses a Random Forest model for illustration. To optimize model performance, a grid search method combined with a 10-fold cross-validation strategy can be used, with the root mean square error (RMSE) as the evaluation criterion, to optimize the model's key hyperparameters. Figure 2 The results of sensitivity analysis are shown for maximum depth (max_depth), maximum number of features (max_features), minimum number of samples per leaf node (min_samples_leaf), minimum number of samples per node split (min_samples_split), and number of base estimators (n_estimators). Figure 2 (a) shows the impact of maximum depth (max_depth) on model performance; Figure 2 (b) The impact of the number of base estimators (n_estimators) on model performance; Figure 2 (c) shows the impact of different parameters (min_samples_leaf, min_samples_split, max_features) on model performance; Note: The vertical axis represents the grid search score, which indicates the root mean square error (RMSE). The lower the value, the better the model performance.

[0038] Based on the analysis results, this embodiment selects a set of parameter combinations as follows: max_depth=6, max_features=4, min_samples_leaf=2, min_samples_split=3, n_estimators=150.

[0039] Before model training, the total dataset is randomly divided into training and test sets. Figure 3 The effect of different training set ratios on model performance is shown. Figure 3 (a) and Figure 3 (b) The proportion of training set data is used as the horizontal axis to show its impact on model performance. Figure 3 (a) focuses on performance changes during the training process, while Figure 3(b) focuses on the model's generalization performance on the corresponding test set. As the proportion of the training set increases (i.e., the proportion of the test set decreases), the model performance shows a gradual improvement trend. When the proportion of the training set is below 20% (i.e., the proportion of the test set is above 80%), the performance metrics exhibit unstable fluctuations. When the proportion of the training set reaches approximately 80% (i.e., using 20% ​​or approximately 20% of the data as the test set), the performance evaluation results are relatively stable. Therefore, in this embodiment, 70% of the data is selected as the training set, and 30% as the test set. The training set is used to train the configured random forest model with hyperparameters.

[0040] The trained model is used to make predictions on 30% of the test set data, and the predictions are compared with the actual observations to evaluate the model performance. Evaluation metrics may include the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Figure 4 This is a graph showing the performance evaluation results of the random forest model in this embodiment for predicting soil water deficit from the initial layer to different depth layers in the plantation. Figure 4 (a), 4(d), and 4(g) demonstrate the performance in the depth range of 3.4–10 meters, respectively. Figure 4 (b), 4(e), and 4(h) demonstrate the performance in the 10–16 meter depth range; Figure 4 (c), 4(f), and 4(i) demonstrate performance in the depth range of 16–25.2 meters; specifically, Figure 4 (a)-(c) are violin illustrations. Figure 4 (d)-(f) are scatter plots. Figure 4 (g)-(i) are density distribution maps; Figure 4 The results shown indicate that the model is effective in predicting soil water deficit at different depths. ΔSWS d It exhibits good performance at depths of 3.4-10 meters, with an R² of 0.890, MAE of 56.822 mm, and RMSE of 79.917 mm; at 10-16 meters, R² is 0.937, MAE is 60.599 mm, and RMSE is 78.154 mm; and at 16-25.2 meters, R² is 0.982, MAE is 46.289 mm, and RMSE is 65.925 mm. Figure 4 The violin plot, scatter plot, and density distribution plot visually demonstrate the consistency between the model predictions and the measured values.

[0041] Furthermore, after the model is built, the results can be analyzed using model interpretation tools. This step is an interpretation or application of the results obtained by the method of this invention, and is not a necessary component of the estimation method. In this embodiment, the SHAP (SHapley Additive exPlanations) method can be used to analyze the trained random forest model. Specifically, in order to quantitatively evaluate the impact of different input factors on deep soil moisture of vegetation utilization, this embodiment uses a random forest model and combines it with SHAP (SHapley Additive exPlanations) plots for analysis. This model uses eight input factors as independent variables for training, namely: depth, forest age, tree species, potential evapotranspiration, annual precipitation, clay content, sand content, and silt content. Figure 5 and Figure 6 The impact of these input factors is shown from different dimensions. The different numbers of input factors shown in the figures are due to the different analytical purposes, as explained below: Figure 5 (a) shows a SHAP summary plot of all 8 input factors, which sorts the overall importance of each factor by the average absolute value of SHAP. Figure 5 (b) to Figure 5 (f) is the SHAP dependency graph, showing only a portion of it. Figure 5 (a) The top five most important input factors (depth, forest age, tree species, potential evapotranspiration and annual precipitation) are included to analyze the marginal effects of the most important factors in detail, while the three soil texture factors with relatively lower importance are omitted to keep the chart concise. Figure 6 This illustrates the trends in the importance of seven key input factors (forest age, tree species, etc.) as a function of depth. In this analysis, "depth" itself is used as the dimension of analysis (horizontal axis), so the figure shows the changes in the importance of the remaining seven factors across different depth ranges. The results indicate that the importance of different environmental factors varies with soil depth, suggesting that the model captures complex relationships along the vertical dimension.

[0042] The embodiments of the present invention have been described in detail above. The description of the embodiments above is only for the purpose of helping to understand the method and core idea of ​​the present invention, and is not intended to limit the scope of implementation of the present invention. Therefore, all equivalent changes or modifications made to the structure, features and principles described in the patent claims of the present invention should be included in the patent application scope of the present invention. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for estimating the vertical distribution of deep soil moisture use in planted forests, characterized in that, Includes the following steps: S1) Obtain soil moisture content data of deep profiles of artificial forest land and control land as reference in the study area, and calculate a set of soil water deficits from the initial layer to different depth layers of artificial forest land representing the vertical profile. S2) Construct a dataset where each data point in the dataset includes: a) Using each depth and at least one environmental factor associated with that depth as a set of input features, wherein both the depth and the at least one environmental factor are input factors; b) The soil water deficit corresponding to the depth is used as the dependent variable; S3) A machine learning model is used to train the dataset to establish a mapping relationship between the input factors and the dependent variable, thereby obtaining a prediction model that can directly predict the soil water deficit at that depth based on environmental factors and depth.

2. The method according to claim 1, characterized in that, The depth mentioned in S2) is input into the machine learning model as a numerical feature.

3. The method according to claim 1, characterized in that, The machine learning model is an ensemble learning model.

4. The method according to claim 3, characterized in that, The ensemble learning model is a random forest model.

5. The method according to claim 1, characterized in that, The control area in S1) is a wasteland or farmland that can reflect the natural background state of local soil moisture.

6. The method according to claim 1, characterized in that, The soil water deficit in S1) is calculated using the following formula: , in the formula ΔSWS d It is the soil water deficit from the initial layer to different depth layers of the artificial forest; θ f , θ c These are the soil moisture contents of the artificial forest and the control area, respectively. z It is the sampling interval; i It is the sampling sequence number.

7. The method according to claim 1, characterized in that, The environmental factors in S2) include at least one of soil data, meteorological data, and vegetation data.

8. The method according to claim 7, characterized in that, The soil data includes sand, silt, and clay particles at different depths and their corresponding geographical location information; the meteorological data includes precipitation, potential evapotranspiration and their corresponding geographical location information; the vegetation data includes forest age, tree species and their corresponding geographical location information.

Images