A method for downscaling microwave remote sensing soil moisture data images

Abstract

The present invention relates to the field of remote sensing technology, and more specifically to a method for downscaling microwave remote sensing soil moisture data images. The method utilizes a machine learning algorithm to perform nonlinear fitting on the relationship between soil moisture data images and downscaling factors, thereby establishing a nonlinear relationship model. The model has fast training speed, strong generalization capabilities, is less susceptible to uncertainty factors such as model parameters, considers a large number of independent variables, and is more adaptable to complex regions. Thus, downscaling methods based on complex physical mechanisms are avoided while leveraging the advantages of microwave remote sensing data to improve the spatial resolution of microwave remote sensing soil moisture data images.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a microwave remote sensing soil moisture data image downscaling method. Background Art

[0002] Soil moisture is a critical physical quantity in the Earth system, controlling environmental and climate change, connecting the surface water cycle and energy cycle, and consequently influencing ecosystem balance. Furthermore, soil moisture determines soil thermal properties and evapotranspiration, influencing hydrological and agricultural processes such as drought and crop yield. Accurately determining the spatiotemporal distribution of soil moisture is crucial for research in areas such as drought detection and climate change prediction.

[0003] Currently, soil moisture monitoring primarily relies on site-based observations and satellite remote sensing inversion. Traditional site-based observation methods are susceptible to the impact of site distribution, resulting in poor temporal and spatial contrast in their results, making them unsuitable for large-scale monitoring. Compared to site-based observations, satellite remote sensing inversion offers a wider range and greater timeliness, enabling rapid acquisition of spatiotemporal variations in soil moisture. Remote sensing technologies for macro-scale soil moisture monitoring at the regional scale include optical and microwave remote sensing. Optical remote sensing is significantly affected by weather and is only suitable for detecting visible surface materials. Microwave remote sensing, on the other hand, is unaffected by weather and can acquire information around the clock, possessing a high degree of penetration into ground objects, making it widely used in soil moisture monitoring. However, existing microwave remote sensing products, such as SMAP (Soil Moisture Active Passive), SMOS (Soil Moisture and Ocean Salinity), and the Fengyun-3 meteorological satellite, have short temporal ranges and low spatial resolution, making them difficult to monitor the spatiotemporal dynamics of soil moisture data and images at local or regional scales.

[0004] To address the low spatial resolution of existing microwave remote sensing products, which are insufficient for local or regional scale studies, downscaling methods can be used to improve the spatial resolution of soil moisture data images and obtain clearer ground details. Currently, the main downscaling methods for microwave remote sensing soil moisture products include empirical downscaling, semi-empirical downscaling, and physical-based downscaling. Empirical downscaling methods, including those based on empirical statistics and machine learning, rely on constructing linear or nonlinear relationship models between microwave remote sensing soil moisture data images and downscaling factors at a coarse spatial resolution, ultimately calculating soil moisture data images at a high spatial resolution. For example, machine learning algorithms explore the nonlinear relationship between soil moisture and downscaling factors to construct downscaling relationship models. However, the accuracy of downscaled soil moisture is significantly affected by varying the downscaling factors. Semi-empirical and physical-based downscaling methods are relatively complex and susceptible to uncertainty in model parameters, making them difficult to generalize to large-scale regions.

[0005] It can be seen that there is an urgent need to explore a spatial downscaling method, establish a downscaling model, and improve the spatial resolution of soil moisture data images. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention aims to provide a method for downscaling microwave remote sensing soil moisture data images.

[0007] In order to achieve the above objectives, the present invention adopts the following technical solutions to achieve them.

[0008] A microwave remote sensing soil moisture data image downscaling method includes the following steps:

[0009] Step 1: Obtain large-scale soil moisture data images from microwave remote sensing inversion products and calculate the grid features within the source large-scale grid in the form of average values; select auxiliary data from the microwave remote sensing inversion product data and resample all types of auxiliary data to the source large-scale;

[0010] Step 2: Establish a prediction model based on the nonlinear relationship between large-scale soil moisture data images and a variety of large-scale auxiliary data, and use the XGBoost algorithm to train the model to obtain the optimal large-scale prediction model;

[0011] Step 3: The large-scale auxiliary data is used as the input of the large-scale optimal prediction model, and the large-scale optimal prediction model outputs the pixel value of the predicted large-scale soil moisture data image; then, based on the predicted soil moisture data image and the source large-scale soil moisture data image, the residual of the soil moisture data image at the source large-scale is calculated;

[0012] Step 4: Perform Kriging interpolation and resampling on the residual of the soil moisture data image at the source large scale to a small scale to obtain the residual of the soil moisture data image at the small scale;

[0013] Step 5: resample the large-scale auxiliary data to a small scale to obtain small-scale auxiliary data; input the small-scale auxiliary data into the large-scale optimal prediction model to obtain the predicted value of the soil moisture data image at the small scale without residual correction; then add the predicted value of the small-scale soil moisture data image to the residual of the soil moisture data image at the small scale to obtain the small-scale soil moisture data image after residual correction.

[0014] Compared with the existing technology, the beneficial effects of the present invention are: using a machine learning algorithm to perform nonlinear fitting on the relationship between the soil moisture data image and the downscaling factor to establish a nonlinear relationship model; the model training speed is fast, the generalization ability is strong, it is not easily affected by uncertainty factors such as model parameters, more independent variables are considered, and it is more adaptable to complex areas; thereby avoiding the downscaling method of complex physical mechanisms, and at the same time utilizing the advantages of microwave remote sensing data to improve the spatial resolution of microwave remote sensing soil moisture data images. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] Figure 1 Schematic diagram of the process of the present invention;

[0017] Figure 2 This is a schematic diagram of cutting out the SMAP within the Chongqing area in an embodiment of the present invention;

[0018] Figure 3 A schematic diagram of the process of generating a tree for the XGBoost algorithm;

[0019] Figure 4 is a large-scale soil moisture data image with a resolution of 9 km;

[0020] FIG5 is a small-scale soil moisture data image with a resolution of 1 km after being downscaled by the method of the present invention. DETAILED DESCRIPTION

[0021] The embodiments of the present invention will be described in detail below with reference to examples. However, those skilled in the art will understand that the following examples are only used to illustrate the present invention and should not be construed as limiting the scope of the present invention.

[0022] refer to Figure 1 , a microwave remote sensing soil moisture data image downscaling method, comprising the following steps:

[0023] Step 1: Obtain the 9KM large-scale soil moisture data image in the microwave remote sensing inversion product, and calculate the grid features within the source 9KM large-scale grid in the form of average value; select auxiliary data from the microwave remote sensing inversion product data, and resample all types of auxiliary data to the source 9KM large scale;

[0024] Obtain soil moisture data covering the Chongqing area in 2021, while expanding some boundary areas to ensure complete coverage of the Chongqing area. In this embodiment, the daily average soil moisture data is obtained from the microwave remote sensing soil moisture product (L3_SMAP_E) with a spatial resolution of 9 kilometers provided by the SMAP enhanced L3 radiometer, generating 9-kilometer grid data. The characteristics are single-band, active and passive. The time resolution of L3_SMAP_E is one day. The morning data is used and processed into tif format with spatial reference and coordinates. The soil moisture data within the Chongqing area is clipped using ArcGIS. The clipping idea is as follows: Figure 2 As shown in the figure, the Chongqing area can be fully covered, and finally an effective soil moisture data image at a scale of 9 kilometers with a period of 8-10 days can be obtained.

[0025] Dynamic data information is aggregated at the 9KM scale, and the grid features within the 9KM scale grid are calculated as average values. When selecting auxiliary data, various auxiliary data can be selected based on the topographic and geomorphological characteristics of the area to be downscaled. Auxiliary data suitable for complex mountainous areas are selected from the product data. Chongqing is mountainous, so LU / LC types such as slope and altitude are selected. Since soil texture affects the water holding capacity of the soil and has a significant impact on soil moisture distribution, soil with a lower sand content is more likely to remain moist. Auxiliary data of soil types such as soil clay content, soil sand content, and soil organic carbon content are selected. A total of 58 auxiliary data are selected in this embodiment.

[0026] Step 2: Establish a prediction model based on the nonlinear relationship between the 9km large-scale soil moisture data image and 58 types of auxiliary data at the 9km large scale, and use the XGBoost algorithm to train the model to obtain the optimal prediction model at the 9km large scale.

[0027] The nonlinear relationship model between soil moisture data and 58 features is established as follows:

[0028] f(SMAP)=f(Mean,GLCM,Elev,Slope,Soil_CEC,Soil_N,…,SAW)

[0029] At the 9km scale, the XGBoost algorithm is used for training and learning. The independent variables are the data of each feature, with a total of 58 features and 3,000 training samples. The dependent variable is the pixel value converted from the soil moisture data.

[0030] The XGBoost algorithm trains the model, and the flowchart of the generated tree is as follows Figure 3 As shown, it contains the following sub-steps:

[0031] Sub-step 2.1, initialize the predicted value of each sample;

[0032] Specifically, XGBoost is an integration of multiple weak classifiers, consisting of t base models, and the additive operation formula is:

[0033]

[0034] Where, f j (x i ) is the value of sample i in the jth tree, is the predicted value of the i-th sample. The model has t trees in total, and the predicted value of sample i is the sum of the values ​​on all trees;

[0035] The loss function is composed of the predicted value and the true value y i Defined as:

[0036]

[0037] Sub-step 2.2, define the objective function;

[0038] Specifically, the objective function is defined as:

[0039]

[0040] Where y i and are the true value and predicted value at the tth iteration; f j is the base model generated in round j; is the loss function; Ω(f j ) is a regularization term; the regularization term is used to measure the complexity of the model to avoid overfitting caused by excessive model complexity;

[0041] The objective function is further expressed as:

[0042]

[0043] Where,

[0044] Sub-step 2.3, simplify the second-order Taylor expansion of the objective function;

[0045] Specifically, considering the second-order Taylor expansion, we can further simplify the objective function to obtain:

[0046]

[0047] Where,

[0048] In the above formula It is irrelevant to this round of iteration and can be regarded as a constant term together with C. After omitting the constant term, it can be simplified to:

[0049]

[0050] Sub-step 2.4, simplification of the objective function based on the decision tree;

[0051] Specifically, the decision tree is selected as the base model, which is determined by its basic structure q and leaf weight ω. For the decision tree, there are: t (x) = ω q(x) ;

[0052] q(x) represents the function that maps samples to leaf nodes, T is the number of leaf nodes, and the complexity of the decision tree is expressed as:

[0053]

[0054] The original objective function can be expressed as:

[0055]

[0056] After simplification, we get:

[0057]

[0058] Where, I j ={i|q(x i )=j} is the sample set corresponding to the jth leaf node, and is defined as Substituting and simplifying, we get:

[0059]

[0060] Sub-step 2.5, building a decision tree based on the optimal split point partitioning algorithm;

[0061] Specifically, take the derivative of the above formula with respect to ω, set the derivative to 0, and find the corresponding value when the objective function is minimized. for:

[0062]

[0063] The minimum value of the objective function is:

[0064]

[0065] Normally, γ is 0. When the information gain is greater than 0, the leaf node can be split. The left node I is obtained after the leaf node is split. Land right node I R , the change value of the objective function after splitting is:

[0066]

[0067] In the formula, the definition

[0068] Sub-step 2.6, use the new decision tree to predict the sample and add it to the original value, and create the decision tree in a loop until the stopping condition is met.

[0069] The XGBoost algorithm constructs a new decision tree based on the residual between the predicted value and the true value at each iteration. Only one tree is trained each time, and the final prediction result is the sum of all trees. When the number of iterations reaches the upper limit or the residual no longer decreases, it stops and obtains a strong classifier with multiple decision trees, thus obtaining an optimal prediction model.

[0070] Step 3: The 9KM large-scale auxiliary data is used as the input of the 9KM large-scale optimal prediction model, and the 9KM large-scale optimal prediction model outputs the pixel value of the predicted 9KM large-scale soil moisture data image; then, based on the predicted soil moisture data image and the source 9KM large-scale soil moisture data image, the residual of the soil moisture data image on the source 9KM large-scale is calculated;

[0071] Step 4: Perform Kriging interpolation and resampling on the residual of the soil moisture data image at the source 9 km large scale to the 1 km small scale to obtain the residual of the soil moisture data image at the 1 km small scale;

[0072] Step 5: According to the principle of downscaling, the functional relationship between soil moisture data and its auxiliary data does not change with the change of spatial scale. The auxiliary data of the 9KM large scale is resampled to the 1KM small scale to obtain the auxiliary data of the 1KM small scale; the auxiliary data of the 1KM small scale is input into the optimal prediction model of the 9KM large scale to obtain the predicted value of the soil moisture data image at the 1KM small scale without residual correction; the predicted value of the 1KM small scale soil moisture data image is then added to the residual of the soil moisture data image at the 1KM small scale to obtain the soil moisture data image at the 1KM small scale after residual correction, thereby realizing the downscaling processing of the soil moisture data image.

[0073] Simulation test results

[0074] By comparing Figure 4(a) with Figure 5(a), Figure 4(b) with Figure 5(b), Figure 4(c) with Figure 5(c), and Figure 4(d) with Figure 5(d), it can be seen that the downscaling method of the present invention reduces the resolution of the soil moisture data image from a large scale of 9 km to a small scale of 1 km, achieving downscaling of the spatial resolution of the soil moisture data image in complex mountainous areas and obtaining clearer ground details.

[0075] Although this specification has provided a detailed description of the present invention using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made based on the present invention. Therefore, such modifications and improvements, which do not depart from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A microwave remote sensing soil moisture data image downscaling method, characterized in that: The following steps are involved: Step 1: Obtain large-scale soil moisture data images from microwave remote sensing inversion products and calculate the grid features within the source large-scale grid in the form of average values; Select auxiliary data from microwave remote sensing inversion product data and resample all types of auxiliary data to the source large scale; Step 2: Establish a prediction model based on the nonlinear relationship between large-scale soil moisture data images and a variety of large-scale auxiliary data, and use the XGBoos t algorithm to train the model to obtain the optimal large-scale prediction model; Step 3: The large-scale auxiliary data is used as the input of the large-scale optimal prediction model, and the large-scale optimal prediction model outputs the pixel value of the predicted large-scale soil moisture data image; then, based on the predicted soil moisture data image and the source large-scale soil moisture data image, the residual of the soil moisture data image at the source large-scale is calculated; Step 4: Perform Kriging interpolation and resampling on the residual of the soil moisture data image at the source large scale to a small scale to obtain the residual of the soil moisture data image at the small scale; Step 5: resample the large-scale auxiliary data to a small scale to obtain small-scale auxiliary data; input the small-scale auxiliary data into the large-scale optimal prediction model to obtain the predicted value of the soil moisture data image at the small scale without residual correction; then add the predicted value of the small-scale soil moisture data image to the residual of the soil moisture data image at the small scale to obtain the small-scale soil moisture data image after residual correction.

2. The microwave remote sensing soil moisture data image downscaling method according to claim 1, characterized in that: The XGBoost algorithm trains the model, which includes the following sub-steps: Sub-step 2.1, initialize the predicted value of each sample; Sub-step 2.5, define the objective function; Sub-step 2.3, simplify the second-order Taylor expansion of the objective function; Sub-step 2.4, simplification of the objective function based on the decision tree; Sub-step 2.5, building a decision tree based on the optimal split point partitioning algorithm; Sub-step 2.6, use the new decision tree to predict the sample and add it to the original value, and create the decision tree in a loop until the stopping condition is met.

3. The microwave remote sensing soil moisture data image downscaling method according to claim 2, characterized in that: Specifically, in sub-step 2.1, XGBoost is an additive formula consisting of t base models: Where, f j (x i ) is the value of sample i in the jth tree, is the predicted value of the i-th sample. The model has t trees in total, and the predicted value of sample i is the sum of the values ​​on all trees; The loss function is composed of the predicted value and the true value y i Defined as:

4. The microwave remote sensing soil moisture data image downscaling method according to claim 2, characterized in that: Specifically, in sub-step 2.2, the objective function is defined as: Where y i and are the true value and predicted value at the tth iteration; f j is the base model generated in round j; is the loss function; Ω(f j ) is the regularization term; The objective function is further expressed as: Where, 5. The microwave remote sensing soil moisture data image downscaling method according to claim 2, characterized in that: Specifically, in sub-step 2.3, consider the second-order Taylor expansion and further simplify the objective function to obtain: Where, In the above formula It is irrelevant to this round of iteration and can be regarded as a constant term together with C. After omitting the constant term, it can be simplified to:

6. The microwave remote sensing soil moisture data image downscaling method according to claim 2, characterized in that: Sub-step 2.4 Specifically, the decision tree is selected as the base model, which is determined by its basic structure q and leaf weight ω. For the decision tree, f t (x) = ω q(x) ; q(x) represents the function that maps samples to leaf nodes, T is the number of leaf nodes, and the complexity of the decision tree is expressed as: The original objective function can be expressed as: After simplification, we get: Where, I j ={i|q(x i )=j} is the sample set corresponding to the jth leaf node, and is defined as Substituting and simplifying, we get:

7. The microwave remote sensing soil moisture data image downscaling method according to claim 2, characterized in that: Sub-step 2.5 Specifically, take the derivative of the above formula with respect to ω, set the derivative to 0, and find the value corresponding to the minimum of the objective function. for: The minimum value of the objective function is: Normally, γ is 0. When the information gain is greater than 0, the leaf node can be split. The left node I is obtained after the leaf node is split. L and right node I R , the change value of the objective function after splitting is: In the formula, the definition

Images