Soil water content determination method, prediction model and system based on multispectral data and machine learning inversion

Abstract

The invention discloses a soil water content determination method, prediction model and system based on multispectral data and machine learning inversion. The method comprises the following steps: firstly, constructing a standard sample set containing a plurality of soil types and a plurality of water content gradients; then, collecting a multispectral image of the sample and extracting spectral reflectivity; calculating a plurality of vegetation indexes including a normalized vegetation index NDVI, a soil conditioning vegetation index SAVI, a normalized moisture index NDWI, a ratio vegetation index RVI and a normalized red edge index NDRE based on the reflectivity; and finally, taking the multiple vegetation indexes as input characteristics, taking the actually measured water content as an output target, and adopting a random forest algorithm to train to obtain a soil water content prediction model which can be simultaneously suitable for the multiple soil types. Through multi-feature fusion and ensemble learning, the generalization ability of the model among different soil types is effectively improved, and the problem that a single model is difficult to be universally applied to various soils is solved.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing monitoring and data analysis technology, and in particular to a method, prediction model and system for determining soil moisture content based on multispectral data and machine learning inversion. Background Technology

[0002] Soil moisture content is a key parameter for guiding agricultural irrigation, drought monitoring, and water resource management. Traditional methods for measuring soil moisture content, such as the oven-drying method, while highly accurate, are destructive, time-consuming, labor-intensive, and difficult to implement for large-scale rapid monitoring. Modern probe-based measurement methods, such as the time domain reflectance (TDR), are also costly and unsuitable for regional-scale applications.

[0003] In recent years, remote sensing technology, especially multispectral remote sensing, has made it possible to rapidly and non-contactly estimate soil moisture content. However, existing soil moisture retrieval methods based on remote sensing technology still have significant limitations: First, many models rely on a single vegetation index (such as NDVI), failing to fully utilize the rich information provided by multispectral data, resulting in insufficient prediction accuracy and stability. Second, the spectral characteristics of soils vary significantly depending on their type (such as red soil, black soil, loess, etc.), leading to poor generalization ability and weak universality when models trained on specific soil data are applied to other soil types. Furthermore, traditional statistical methods such as linear regression struggle to capture the nonlinear relationship between soil moisture content and complex spectral characteristics.

[0004] Therefore, there is an urgent need in this field for a soil moisture content measurement scheme that can overcome the influence of soil type differences, has strong generalization ability, and can effectively utilize multispectral information. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a soil moisture content determination scheme based on multispectral data and machine learning inversion. This invention aims to construct a universally applicable soil moisture content prediction model that can be applied to multiple soil types by integrating multiple vegetation index features and utilizing ensemble learning algorithms, thereby solving the problem of weak generalization ability of existing methods when applied across soil types.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for constructing a soil moisture content prediction model, comprising the following steps: 1) Standard sample set preparation step: preparing a standard soil sample set containing at least three different soil types, and each soil type containing multiple different moisture content gradients; 2) Multispectral data acquisition step: using a multispectral camera to photograph the standard soil sample set, acquiring multi-band spectral images of each standard soil sample; 3) Spectral reflectance extraction step: preprocessing the multi-band spectral images, extracting the spectral reflectance of each standard soil sample in the green, red, red-edge, and near-infrared bands; 4) Vegetation index calculation step: based on the spectral reflectance, calculating multiple vegetation indices for each standard soil sample, wherein the multiple vegetation indices include at least the Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Moisture Index (NDWI), Ratio Vegetation Index (RVI), and Normalized Red Edge Index (NDRE). 5) Model training steps: Using the various vegetation indices as input features and the measured water content of the corresponding standard soil samples as the output target, a random forest algorithm is used for training to obtain a soil water content prediction model that can be applied to at least three different soil types.

[0008] Preferably, the at least three different soil types are selected from red soil, black soil, loess, chestnut soil, and purple soil.

[0009] Preferably, the random forest algorithm contains 150 to 200 decision trees.

[0010] Preferably, the multispectral camera includes at least the green light band (495–570 nm), the red light band (600–700 nm), the red edge band (700–730 nm), and the near-infrared band (700–1300 nm).

[0011] Preferably, the calculation formulas for the Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Moisture Index (NDWI), Ratio Vegetation Index (RVI), and Normalized Red Edge Index (NDRE) are as follows:

[0012] The formula for calculating NDVI is:

[0013] The formula for calculating SAVI is: ,

[0014] Where L is the soil adjustment coefficient;

[0015] The formula for calculating NDWI is: ;

[0016] The formula for calculating RVI is: ;

[0017] The formula for calculating NDRE is: ;

[0018] Where R_green, R_red, R_rededge, and R_nir are the spectral reflectances of green light, red light, red edge, and near-infrared bands, respectively.

[0019] Secondly, the present invention provides a soil moisture content prediction model constructed by the method described in the first aspect. This model is configured to receive multiple vegetation indices calculated from multispectral data of the soil to be tested, and output a predicted soil moisture content value applicable to multiple soil types.

[0020] Thirdly, the present invention provides a soil moisture content measuring system, comprising:

[0021] The data acquisition module is used to acquire multi-band spectral images of the soil to be tested using a multispectral camera;

[0022] The data processing module is used to preprocess the multi-band spectral image and extract the spectral reflectance of the green, red, red-edge and near-infrared bands, and then calculate multiple vegetation indices, which include at least NDVI, SAVI, NDWI, RVI and NDRE.

[0023] The moisture content prediction module is embedded with a soil moisture content prediction model according to the second aspect, which is used to receive the multiple vegetation indices and output the prediction results of soil moisture content.

[0024] Preferably, the moisture content prediction module can also output the confidence interval of the prediction result to quantify the uncertainty of the prediction.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1) Outstanding universality and generalization ability: The core advantage of this invention lies in its ability to effectively construct a "general" prediction model by integrating data from multiple soil types into a single model for training and using a combination of multiple vegetation indices that are relatively insensitive to changes in soil background as features. As shown in the examples, this single model demonstrates certain predictive ability (R² ranges from 0.64 to 0.93) on red soil, purple soil, chestnut soil, loess, and black soil, solving the technical problem that a single model is difficult to directly apply to multiple soil types and avoiding the tedious process of modeling each soil type separately. 2) Completeness and Innovation of the Technical Solution: This invention provides a complete technical solution from standardized sample preparation, multispectral data acquisition, multi-feature fusion to model training and application. In particular, the selection of a specific combination of vegetation indices—NDVI, SAVI, NDWI, RVI, and NDRE—can comprehensively reflect soil moisture status from multiple physical dimensions such as vegetation cover, soil background correction, water sensitivity, and vegetation vitality. Combined with random forest, an ensemble learning algorithm that can effectively handle nonlinear relationships, this constitutes a synergistic and effective technical approach, a combination of unconventional methods in this field. 3) Clear practical value: The model and system established by this method can make rapid, non-contact estimation of soil moisture content using widely used multispectral cameras (such as UAV-borne cameras), providing feasible technical tools for applications such as variable irrigation and regional drought monitoring in precision agriculture, and has good prospects for industrial application. Attached Figure Description

[0027] Figure 1 Technical roadmap for constructing a soil moisture content prediction model provided in embodiments of the present invention;

[0028] Figure 2 This is a schematic diagram showing the water content distribution of different soil types in an embodiment of the present invention;

[0029] Figure 3 This is a thermogram showing the correlation between vegetation index and soil moisture content in an embodiment of the present invention.

[0030] Figure 4 This is a comparison chart of the predicted and actual values ​​of the model constructed in this embodiment of the invention on different soil types.

[0031] Figure 5 This is a feature importance analysis diagram of the random forest model in this embodiment of the invention;

[0032] Figure 6 This is a schematic diagram of the human-computer interaction interface of the soil moisture content measurement system provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments, but the scope of protection of the present invention is not limited to the following embodiments.

[0034] The core of this invention in constructing a universal soil moisture content prediction model lies in the construction of the dataset, the selection of features, and the selection of the model. Figure 1 This is a technical roadmap for constructing a soil moisture content prediction model provided in the embodiments of the present invention.

[0035] 1. Preparation of standard sample sets

[0036] To ensure the model's generalization ability, it must include a sufficient diversity of soil types. This example selects five representative soils with significant differences in physicochemical properties: red soil, black soil, loess, chestnut soil, and purple soil.

[0037] (1) Soil sample selection and pretreatment

[0038] Five representative soil types were selected: red soil, black soil, loess, chestnut soil, and purple soil.

[0039] Sample pretreatment process:

[0040] Raw soil → air-dry for 7 days → pass through a 2mm sieve → dry at 105℃ for 6 hours → cool to room temperature → package for later use

[0041] (2) Moisture gradient design and sample preparation

[0042] Three independent replicates were prepared for each moisture content level, and the amount of water added was controlled using a precision electronic balance (accuracy 0.001g).

[0043] Table 1. Moisture gradient design for different soil types

[0044]

[0045] Figure 2 This is a schematic diagram illustrating the moisture content distribution of different soil types in this embodiment of the invention. For each soil type, multiple moisture content gradients were prepared from the air-dried state to near field capacity by precisely controlling the amount of water added (as shown in Table 1, 10 gradients). Each gradient had at least three replicate samples to reduce error. The resulting standard sample set is essentially a standardized database covering the two-dimensional space of "soil type-moisture content," laying the data foundation for training a general model.

[0046] 2. Multispectral data acquisition and feature calculation

[0047] (1) Multispectral data acquisition

[0048] Remote sensing images were acquired using a UAV multispectral camera. The camera included five spectral bands: blue (450 nm ± 16 nm), green (560 nm ± 16 nm), red (650 nm ± 16 nm), red-edge (730 nm ± 16 nm), and near-infrared (840 nm ± 26 nm). This experiment primarily used the green, red, red-edge, and near-infrared bands. Data acquisition was conducted under clear, natural light conditions to ensure sufficient and uniform illumination without cloud cover. The flight altitude was set at 50 meters, achieving a ground resolution of 5.2 cm / pixel. Data acquisition was conducted during midday when the sun's altitude angle was high to minimize the impact of shadows.

[0049] (2) Data processing and feature extraction

[0050] After acquiring multispectral images, data processing was performed using ENVI 5.6 software. First, radiometric correction was performed to convert the numerical values ​​into surface reflectance. Then, the spectral reflectance values ​​of each standard soil sample area were extracted using the software's Region of Interest (ROI) tool, and various vegetation indices were calculated to ensure consistent sampling locations.

[0051] Table 2. Formulas for calculating vegetation index Symbol explanation:

[0052] Rgreen: Reflectance in the green light band (560 nm ± 16 nm)

[0053] Rred: Red light reflectance (650 nm ± 16 nm)

[0054] Rrededge: Reflectivity in the red-edge band (730 nm ± 16 nm)

[0055] Rnir: Near-infrared reflectance (840 nm ± 26 nm)

[0056] L: Soil adjustment coefficient, with a value of 0.5

[0057] These vegetation indices were chosen because they reflect soil moisture status through different physical mechanisms: NDVI assesses vegetation cover, SAVI reduces soil background disturbance, NDRE is sensitive to chlorophyll changes, NDWI directly reflects water content, and RVI provides a simple indicator of vegetation vitality. This combination of indices comprehensively captures the spectral response characteristics of soil moisture content.

[0058] The calculated vegetation index data are shown in Table 3 below:

[0059]

[0060] Note: The table only shows partial data.

[0061] Figure 3 This is a feature correlation heatmap, which visually demonstrates the correlation strength between water content and various vegetation indices (such as NOVI, SAVI, etc.). Darker colors in the graph indicate stronger positive / negative correlations. As shown, some indices exhibit significant correlations with water content (up to 0.593), providing crucial information for feature selection in subsequent models.

[0062] 3. Random Forest Model Optimization

[0063] (1) Hyperparameter optimization: Bayesian optimization method is used to find the optimal parameter combination.

[0064] (2) Training strategy

[0065] Data partitioning: stratified sampling based on soil type, training set:test set = 7:3

[0066] Cross-validation: 5-fold stratified cross-validation

[0067] Feature selection: Based on importance scores, features with importance > 0.05 are retained.

[0068] 4. Model Performance Evaluation

[0069] (1) Overall performance evaluation

[0070] Table 4 Overall Performance Evaluation Table

[0071]

[0072] (2) Performance analysis of each soil type

[0073] Table 5 Performance Analysis of Various Soil Types

[0074]

[0075] Figure 4 This is a comparison chart of predicted and actual soil moisture content values ​​for different soil types provided in embodiments of the present invention. The chart includes sub-charts of the prediction performance for five typical soil types: red soil, chestnut soil, black soil, purple soil, and loess. In each sub-chart, the "actual value" curve and the "predicted value" curve clearly demonstrate the prediction effect of the method or system of the present invention on soil moisture content, and the degree of agreement intuitively reflects the accuracy of the prediction. Each sub-chart is labeled with a coefficient of determination (R²). The predicted values ​​for red soil (R²=0.935) and purple soil (R²=0.922) are highly consistent with the actual values, indicating that the model of the present invention has excellent prediction accuracy; chestnut soil (R²=0.743) and loess (R²=0.743) also show good correlation; even in black soil (R²=0.641), which is relatively difficult to predict, the model can still capture the basic trend of moisture changes. This chart comprehensively demonstrates that the technical solution provided by the present invention has reliable and universal predictive performance across different soil types.

[0076] (3) Feature Importance Analysis

[0077] Figure 5The figure shows the ranking of the importance of various features identified by the random forest model for soil moisture prediction. As shown, the ratio vegetation index (RVI) is the most critical predictor, with its importance far exceeding that of other features (accounting for 33.36%). NDVI and NDWI rank second and third, respectively, while soil type itself has the least influence (3.52%). This analysis demonstrates the dominant role of multispectral vegetation indices in soil moisture retrieval and the effectiveness of the multi-feature fusion strategy adopted in this invention.

[0078] Figure 6 is a schematic diagram of the human-computer interaction interface of the intelligent soil moisture prediction system of the present invention. The human-computer interaction interface of the intelligent soil moisture prediction system of the present invention allows users to select a specific soil type and input a series of vegetation index parameters (such as NDVI, SAVI, NDRE, etc.) to trigger the prediction model built on the random forest algorithm within the system. After the model runs, the interface will intuitively output the quantitative prediction value of soil moisture, confidence interval, qualitative description of water content state and specific agricultural irrigation suggestions, and simultaneously generate a prediction analysis curve, thereby completing the entire process from data input to intelligent prediction and decision support.

[0079] 5. Comparative Experiment Design

[0080] To verify the effectiveness of this invention, a systematic comparative experiment was designed:

[0081] Experimental group: The method of this invention (random forest + multi-feature fusion) Control group:

[0082] Single NDVI linear regression model

[0083] Single RVI linear regression model

[0084] Multiple linear regression model (all vegetation indices)

[0085] Experimental conditions:

[0086] Use the same training and test sets (70% training, 30% testing).

[0087] The same 5-fold stratified cross-validation scheme

[0088] The same performance evaluation metrics (R², RMSE, MAE)

[0089] Same computing environment

[0090] Experimental Results and Analysis

[0091] 1. Overall performance comparison

[0092] Performance evaluation results based on the actual test set (15 samples): See Table 6

[0093] Table 6 Evaluation Results

[0094]

[0095] Accuracy comparison:

[0096] The method of this invention achieved the best performance on the test set (R²=0.2725). Although the overall accuracy is limited, it is still a significant improvement compared to the control group.

[0097] The multiple linear regression model performed the second best (R²=0.2480), validating the effectiveness of multi-feature fusion.

[0098] In the single-index model, the RVI model (R²=0.2050) outperforms the NDVI model (R²=0.1620), consistent with the feature importance analysis.

[0099] Error analysis:

[0100] The method of this invention has the lowest RMSE (0.0596) and MAE (0.0550);

[0101] As model complexity increases, prediction error gradually decreases;

[0102] The error of the single-index model is significantly higher than that of the multi-feature model.

[0103] Efficiency Analysis:

[0104] The single exponential model has the fastest training speed (0.8s), making it suitable for rapid deployment;

[0105] The random forest model has a long training time (15.2s), but it has the best prediction accuracy.

[0106] Multiple linear regression strikes a balance between accuracy and efficiency.

[0107] This invention presents a method for determining soil moisture content based on multispectral data and machine learning inversion. It constructs a random forest model based on multi-feature fusion, effectively improving the prediction accuracy of soil moisture content. The model can automatically identify key predictive factors (RVI importance reaches 33.36%), achieving accurate prediction. A single model is applicable to five different soil types, demonstrating the method's cross-soil universality. The contribution of soil type features is only 3.52%, highlighting the universal value of multispectral indices. The model exhibits good robustness to data noise and good stability.

[0108] This invention presents a method for determining soil moisture content based on multispectral data and machine learning inversion. It establishes a complete standardized experimental procedure and quality control system, achieves effective fusion and optimized configuration of multispectral vegetation indices, and provides interpretable feature importance analysis, pointing the way for subsequent research. It provides reliable technical support for irrigation decision-making in precision agriculture, is applicable to large-area soil moisture monitoring and drought early warning, has a complete method system, and has good prospects for promotion and application and industrialization potential.

[0109] In summary, through the specific embodiments described above, this invention successfully provides a highly universal soil moisture content measurement scheme. Those skilled in the art, inspired by the concept of this invention, can make appropriate adjustments to soil type, moisture content gradient settings, random forest parameters, etc., and all such adjustments should be included within the scope of protection of this invention. Contents not described in detail in this specification are all prior art known to those skilled in the art.

Claims

1. A method for determining soil moisture content based on multispectral data and machine learning inversion, characterized in that, Includes the following steps: S1. Standard sample set preparation steps: Prepare a standard soil sample set containing at least three different soil types, and each soil type contains multiple different moisture content gradients; S2. Multispectral data acquisition steps: Use a multispectral camera to photograph the standard soil sample set and acquire multi-band spectral images of each standard soil sample; S3. Spectral reflectance extraction step: Preprocess the multi-band spectral image to extract the spectral reflectance of each standard soil sample in the green, red, red-edge and near-infrared bands; S4. Vegetation index calculation steps: Based on the spectral reflectance, calculate multiple vegetation indices for each standard soil sample. The multiple vegetation indices include at least the Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Moisture Index (NDWI), Ratio Vegetation Index (RVI), and Normalized Red Edge Index (NDRE). S5. Model training steps: Using the various vegetation indices as input features and the measured water content of the corresponding standard soil samples as the output target, a random forest algorithm is used for training to obtain a soil water content prediction model that can be applied to at least three different soil types.

2. The determination method according to claim 1, characterized in that, The at least three different soil types are selected from the group consisting of red soil, black soil, loess, chestnut soil, and purple soil.

3. The determination method according to claim 1, characterized in that, The random forest algorithm contains 150 to 200 decision trees.

4. The determination method according to claim 1, characterized in that, The multispectral camera includes at least the green band (495–570 nm), the red band (600–700 nm), the red edge band (700–730 nm), and the near-infrared band (700–1300 nm).

5. The determination method according to claim 1, characterized in that, The calculation formulas for the Normalized Difference Vegetation Index (NDVI), Soil-Regulated Vegetation Index (SAVI), Normalized Difference Moisture Index (NDWI), Ratio Vegetation Index (RVI), and Normalized Red Edge Index (NDRE) are as follows: The formula for calculating NDVI is: , The formula for calculating SAVI is: ' Where L is the soil adjustment coefficient; The formula for calculating NDWI is: , The formula for calculating RVI is: ; The formula for calculating NDRE is: ; Where R_green, R_red, R_rededge, and R_nir are the spectral reflectances of green light, red light, red edge, and near-infrared bands, respectively.

6. A soil moisture content prediction model constructed by the method of any one of claims 1 to 5, characterized in that, The model is configured to receive the various vegetation indices calculated from the multispectral data of the soil to be tested, and output a predicted soil moisture content value applicable to various soil types.

7. A soil moisture content determination system based on multispectral data and machine learning inversion, characterized in that, include: The data acquisition module is used to acquire multi-band spectral images of the soil to be tested using a multispectral camera; The data processing module is used to preprocess the multi-band spectral image and extract the spectral reflectance of the green, red, red-edge and near-infrared bands, and then calculate multiple vegetation indices, which include at least NDVI, SAVI, NDWI, RVI and NDRE. The moisture content prediction module is embedded with the soil moisture content prediction model according to claim 6, and is used to receive the multiple vegetation indices and output the prediction results of soil moisture content.

8. The soil moisture content measuring system according to claim 7, characterized in that, The moisture content prediction module is further used to output the confidence interval of the prediction result.

Images