Soil humidity high-precision extraction method supporting cross-regional migration in mountain land surface environment

Abstract

The invention relates to the technical field of soil humidity prediction, in particular to a high-precision soil humidity extraction method capable of supporting cross-regional migration in a mountain land surface environment, which comprises the following steps: training a 3D-CNN-LSTM model in a source domain by using multi-source environmental factor data, constructing a basic prediction framework, predicting frame by frame through a target domain sliding window mechanism, and obtaining a prediction result. Acquiring a soil humidity prediction map with the resolution ratio of 30 meters, and evaluating the performance of the model by adopting MSE, MAE and R2 indexes; through monthly time sequence images and spatial statistical analysis, annual scale and monthly scale space-time distribution laws are extracted; performing population initialization, crossover variation and selection on the candidate feature set through a genetic algorithm; and constructing a 3D-CNN-LSTM fusion framework, extracting local spatial-temporal characteristics through a three-dimensional convolution kernel, capturing time dependence in combination with LSTM, and outputting a predicted value to obtain a cross-regional soil humidity prediction model. Existing research mostly focuses on regional scale or short-term dynamic, and time-space change characteristics are difficult to consider under a complex terrain background.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of soil moisture prediction, and particularly relates to a high-precision soil moisture extraction method supporting cross-regional migration under mountain surface environment. BACKGROUND

[0002] Soil moisture, as a key medium of terrestrial hydrological cycle and ecosystem function, not only has a significant impact on vegetation growth, biogeochemical processes and agricultural production, but also plays an important role in drought monitoring, water resources management and geological disaster warning. Soil moisture is defined as the ratio of free water volume to total volume in soil, i.e. volumetric water content, and its accurate acquisition is crucial for regional environmental research.

[0003] With the rapid development of remote sensing technology, geographic information system (GIS) and machine learning (ML), multi-source spatio-temporal data has been widely used in soil moisture monitoring and prediction. Although traditional empirical models and statistical regression methods have achieved certain success in small-scale research, it is often difficult to capture the complex nonlinear relationship in massive multi-dimensional data, and the core challenge of high-precision extraction of soil moisture supporting cross-regional migration in mountain environment lies in the multi-dimensional heterogeneity of mountain environment. The mountain terrain is rugged, with large differences in slope and orientation, and high spatial heterogeneity of surface roughness, which makes it difficult for traditional microwave radiation transfer models relying on smooth surface assumption to accurately describe the scattering process. At the same time, terrain also interferes with water vapor transport and precipitation distribution, exacerbating the spatial difference of soil moisture, and the extraction method cannot fully capture the spatio-temporal heterogeneity of soil moisture and quickly acquire large-scale data. These all constitute a natural barrier for model cross-regional migration, and the traditional extraction method has obvious migration bottleneck, while the cross-regional migration method has the key value of breaking the "no station dilemma" in mountainous areas and clear technical breakthrough direction. From the perspective of mountain environment, the micro-topography in karst mountainous areas in southwest China can make the soil moisture fluctuate by 30%-50%, and also interfere with the remote sensing signal. The difference in terrain humidity index in different regions also changes the soil water transport mechanism; the vertical differentiation of climate makes the driving factors such as precipitation and evapotranspiration change with altitude gradient. From arid mountainous areas to humid mountainous areas, the dominant mechanism of soil moisture changes, and the weight of precipitation characteristics decreases by more than 30%, and the influence of surface humidity increases by nearly 2 times. In terms of traditional methods, physical models rely on the assumption of smooth surface, and the roughness of mountain surface makes the inversion error exceed 0.1cm 3 cm -3 , and the regional dependence of core parameters is strong, with the vegetation extinction coefficient in humid mountainous areas being 1.5-2 times that in arid areas; in single remote sensing inversion, optical remote sensing is disturbed by clouds and fog, and the RMSE of cross-regional inversion increases by 0.03m 3 / m 3Passive microwave resolution is low, multi-source data fusion lacks terrain normalization processing, and scale extrapolation deviation will also occur, while cross-region migration method is of great significance, which can provide high-precision data for monitoring sparse mountainous areas through data-rich area model pre-training, and plays a key role in the fields of hydrology, agriculture, disaster warning, etc. Its technical breakthrough direction is clear, including eliminating remote sensing signal interference by implementing SCS+C terrain correction method through DEM data, and improving the inversion accuracy by 20%-30%; fusion of hyperspectral specific waveband, microwave remote sensing and terrain factor to construct cross-region general feature set, such as unmanned aerial vehicle scale precision verification in karst area in southwest China.

[0004] In recent years, deep learning technology has also been widely used in meteorology, hydrology and ecology due to its automatic feature extraction and nonlinear fitting capability; however, common two-dimensional convolutional neural network (2D-CNN) and long short-term memory network (LSTM) still face challenges in processing spatio-temporal coupling information, such as dimension redundancy, information interference and insufficient model interpretability, and due to the scarcity and distribution difference of target domain samples, cross-region application deviation is easy to occur, resulting in overfitting and degradation of generalization ability.

[0005] Existing researches focus on the construction of single-scale or single-data-type prediction models. Researches based on support vector machine (SVM) and random forest (RF) have advantages in feature selection and model simplification, but it is difficult to obtain spatial consistency and temporal evolution characteristics at the same time; methods based on convolutional neural network (CNN) to extract spatial features or using long short-term memory network (LSTM) to model time series have their own advantages, but they cannot consider the spatio-temporal characteristics; a few models combining CNN and LSTM usually train directly in the target domain, ignoring the differences between source and target domains, and it is difficult to achieve efficient cross-domain generalization. Therefore, it is urgent to construct a soil moisture prediction framework that can efficiently process high-dimensional multi-source data, consider spatio-temporal dependence and overcome cross-region differences. SUMMARY

[0006] In view of the above status, the present application provides a soil moisture prediction method of spatio-temporal deep learning model fused with transfer learning, which can solve the problem that existing researches mostly focus on regional scale or short-term dynamics and are difficult to consider spatio-temporal variation characteristics in complex terrain background. In order to achieve the above purpose, the present application adopts the following technical solutions:

[0007] The soil moisture prediction method of the fusion migration learning spatio-temporal deep learning model comprises the following steps: training a 3D-CNN-LSTM model in a source domain by using multi-source environmental factor data, constructing a basic prediction framework through spatio-temporal feature extraction and parameter optimization, and obtaining the 3D-CNN-LSTM model; freezing the 3D-CNN feature extraction layer of the pre-trained model, fine-tuning the LSTM time sequence layer and the output layer by using a small amount of samples in a target domain, realizing adaptive updating of parameters, and obtaining the 3D-CNN-LSTM model fine-tuned in the target domain; obtaining a 30-meter resolution soil moisture prediction map through frame-by-frame prediction by using a target domain sliding window mechanism, and using MSE, MAE and R 2 index to evaluate the performance of the model, to obtain a high-resolution prediction map and a performance evaluation report; taking soil moisture grid data as a response variable, extracting annual and monthly spatio-temporal distribution rules through monthly time series images and spatial statistical analysis, obtaining spatio-temporal characteristics of soil moisture; initializing, crossing, mutating and selecting a population of candidate feature sets through a genetic algorithm, extracting factors that significantly contribute to soil moisture prediction, and obtaining a key feature subset; constructing a 3D-CNN-LSTM fusion framework, extracting local spatio-temporal features through a three-dimensional convolution kernel, combining LSTM to capture time dependence, outputting prediction values, and obtaining a cross-regional soil moisture prediction model.

[0008] Further, the 3D-CNN-LSTM model is pre-trained in the source domain by using multi-source environmental factor data; spatial features are extracted through a 3D convolutional neural network, and time series dynamic characteristics are captured through a long short-term memory network to construct a spatio-temporal fusion deep learning framework; model parameters are trained using large-scale source domain data, so that the model learns the general rules of soil moisture changes, and the 3D-CNN-LSTM model is obtained.

[0009] Further, the 3D-CNN feature extraction layer of the pre-trained model is frozen, the LSTM time sequence layer and the output layer are fine-tuned by using a small amount of samples in the target domain, adaptive updating of parameters is realized, and the 3D-CNN-LSTM model fine-tuned in the target domain is obtained, comprising the following steps: loading the 3D-CNN-LSTM model pre-trained in the source domain, freezing the 3D-CNN feature extraction layer weight; a small amount of samples are randomly extracted from the target domain, and only the LSTM time sequence modeling layer and the fully connected output layer are unfrozen; adaptive updating of parameters is realized by minimizing the mean square error loss function, and the loss function formula is: wherein y i and Respectively represent the observation value and model prediction value of the i-th sample, and N is the total number of samples; the adaptive update of the model parameters is realized, and a 3D-CNN-LSTM model after fine-tuning in the target domain is obtained.

[0010] Further, the soil moisture prediction map of 30m resolution is obtained by frame-by-frame prediction through the target domain sliding window mechanism, and the MSE, MAE and R 2 The performance of the model is evaluated by the indicators, and the high-resolution prediction map and performance evaluation report are obtained, including the following steps: using a sliding window mechanism with a window size of 5*5 and a step size of 1 pixel, the soil moisture is predicted frame by frame on the target domain data, and the soil moisture prediction map of 30m resolution is output; the mean square error MSE, the mean absolute error MAE and the determination coefficient R 2 The performance of the model is evaluated by the indicators; the MSE is 3.989, the MAE is 1.455, and the R 2 =0.903; compared with the LST, CNN, CNN-LSTM benchmark models, the advantages of the 3D-CNN-LSTM model are verified, and the high-resolution prediction map and performance evaluation report are obtained.

[0011] Further, the soil moisture spatio-temporal characteristics are obtained by taking the soil moisture grid data as the response variable, and through monthly time series image and spatial statistical analysis, the annual and monthly spatio-temporal distribution rules are extracted, including the following steps: collecting monthly time series remote sensing images of the target domain, and generating soil moisture grid data; analyzing the annual scale characteristics, and identifying the spatial distribution rules of the high-humidity area and the low-humidity area; analyzing the monthly scale characteristics, and exploring the dynamic characteristics of the soil moisture change with the seasons, to obtain the soil moisture spatio-temporal characteristics.

[0012] Further, the key feature subset is obtained by using genetic algorithm to initialize, cross, mutate and select the candidate feature set, and extracting the factors that significantly contribute to the soil moisture prediction, including the following steps: initializing the population of the candidate feature set to generate a feature subset; globally optimizing the feature subset through cross, mutation and selection operations; using the determination coefficient R 2 as the fitness function to evaluate the prediction performance of each feature subset, and the formula is: Wherein, is the mean value of the observation value; iteratively optimizing the subset combination, removing redundant or weakly correlated features, and retaining the factors that significantly contribute to the soil moisture prediction, to obtain the key feature subset.

[0013] Further, the 3D-CNN-LSTM fusion framework is constructed, the local spatio-temporal features are extracted through the three-dimensional convolution kernel, the time dependence is captured by combining LSTM, and the prediction value is output, to obtain the cross-regional soil moisture prediction model, including the following steps: a 3D convolutional neural network 3D-CNN is constructed, which is used to extract spatial features from spatio-temporal data, and the convolution formula is: wherein x represents input data, w(i,j,k) represents a convolution kernel, and y(x,y,z) is an output feature map; a long short-term memory network (LSTM) is constructed to capture dynamic characteristics of a time sequence, and core formulas include a forgetting gate, an input gate, an output gate, and cell state updating; a fusion framework is constructed, and a three-dimensional convolution kernel is used to extract local space-time features in a time-row-column three-dimensional space; after pooling, an output sequence enters an LSTM layer to capture time dependence, and a soil moisture prediction value is output through a fully connected layer, a mean square error is used as a loss function of the model, and a cross-regional soil moisture prediction model is obtained.

[0014] Further, the long short-term memory network (LSTM) is constructed to capture dynamic characteristics of a time sequence, and core formulas include a forgetting gate, an input gate, an output gate, and cell state updating, and include the following steps: the forgetting gate formula is f t =σ(w f ·[h t-1 ,x t ]+b f ); the input gate formula is i t =σ(w i ·[h t-1 ,x t ]+b i ); the output gate formula is o t =σ(w o ·[h t-1 ,x t ]+b o ); the cell state updating formula is C t =f t *C t-1 +i t *tanh(w c ·[h t-1 ,x t ]+b c ; the hidden state output formula is h t =o t *tanh(c t ); wherein f t is the forgetting gate, i t is the input gate, o t is the output gate, c t is the cell state, h t is the hidden state, x t is the current input data, w f , w i , w o , w c are weight matrices, b f , b i , b o , bc σ is the bias term, and σ is the activation function.

[0015] Furthermore, the construction of the fusion framework, which uses a three-dimensional convolutional kernel to extract local spatiotemporal features in the time-row-column three-dimensional space, includes the following steps: The construction of the fusion framework, which uses a three-dimensional convolutional kernel to extract local spatiotemporal features in the time-row-column three-dimensional space, is calculated using the following formula: Where, x (l-1) w represents the input feature map of layer l-1. (l) Let l be the kernel weights of convolution, and f() be the activation function.

[0016] In the technical solution provided by this invention, a 3D-CNN-LSTM model is trained in the source domain using multi-source environmental factor data. A basic prediction framework is constructed through spatiotemporal feature extraction and parameter optimization to obtain the 3D-CNN-LSTM model. The 3D-CNN feature extraction layer of the pre-trained model is frozen, and the temporal and output layers of the LSTM are fine-tuned using a small number of samples from the target domain to achieve adaptive parameter updates, resulting in a fine-tuned 3D-CNN-LSTM model in the target domain. A 30-meter resolution soil moisture prediction map is obtained through frame-by-frame prediction using a target domain sliding window mechanism, and MSE, MAE, and R... 2 The model performance is evaluated using indicators to obtain high-resolution prediction maps and performance evaluation reports. Using soil moisture raster data as the response variable, monthly time-series images and spatial statistical analysis are used to extract annual and monthly spatiotemporal distribution patterns, yielding the spatiotemporal characteristics of soil moisture. A genetic algorithm is used to initialize, cross-mutate, and select candidate feature sets, extracting factors that significantly contribute to soil moisture prediction, resulting in a key feature subset. A 3D-CNN-LSTM fusion framework is constructed, using three-dimensional convolutional kernels to extract local spatiotemporal features and combining LSTM to capture time dependencies, outputting predicted values ​​to obtain a cross-regional soil moisture prediction model. Existing research in this invention mostly focuses on regional scales or short-term dynamics, making it difficult to consider spatiotemporal variation characteristics in complex terrain contexts. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0018] Figure 1 This is a schematic diagram of the first embodiment of a high-precision soil moisture extraction method supporting cross-regional migration in a mountainous surface environment according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of a second embodiment of a high-precision soil moisture extraction method supporting cross-regional migration in a mountainous surface environment according to an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the third embodiment of a method for high-precision extraction of soil moisture in a mountainous surface environment that supports cross-regional migration, as described in this invention.

[0021] Figure 4 This is a schematic diagram of the fourth embodiment of a high-precision soil moisture extraction method supporting cross-regional migration in a mountainous surface environment according to an embodiment of the present invention. Detailed Implementation

[0022] 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. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0024] A high-precision method for extracting soil moisture to support cross-regional migration in mountainous surface environments, such as... Figure 1 As shown, the process includes the following steps: A 3D-CNN-LSTM model is trained in the source domain using multi-source environmental factor data. A basic prediction framework is constructed through spatiotemporal feature extraction and parameter optimization to obtain the 3D-CNN-LSTM model. The 3D-CNN feature extraction layer of the pre-trained model is frozen, and the temporal and output layers of the LSTM are fine-tuned using a small number of samples from the target domain to achieve adaptive parameter updates, resulting in a fine-tuned 3D-CNN-LSTM model for the target domain. A 30-meter resolution soil moisture prediction map is obtained through frame-by-frame prediction using a sliding window mechanism in the target domain. MSE, MAE, and R... 2 The performance of the model is evaluated using indicators to obtain high-resolution prediction maps and performance evaluation reports. Using soil moisture raster data as the response variable, the spatiotemporal distribution patterns at the annual and monthly scales are extracted through monthly time-series images and spatial statistical analysis to obtain the spatiotemporal characteristics of soil moisture. The candidate feature set is initialized, crossover, mutated, and selected using a genetic algorithm to extract factors that significantly contribute to soil moisture prediction, resulting in a key feature subset. A 3D-CNN-LSTM fusion framework is constructed, which extracts local spatiotemporal features through three-dimensional convolutional kernels and combines them with LSTM to capture time dependencies, outputting predicted values ​​to obtain a cross-regional soil moisture prediction model.

[0025] likeFigure 2 As shown, in this embodiment, multi-source environmental factor data is used to pre-train the 3D-CNN-LSTM model in the source domain (e.g., southern China, approximately 18.°N-35.°N, 100.°E-120.°E); spatial features are extracted through a 3D convolutional neural network, and the dynamic characteristics of time series are captured by a long short-term memory network to construct a spatiotemporal fusion deep learning framework; the model parameters are trained using large-scale source domain data, enabling the model to learn the general laws of soil moisture change, thus obtaining the 3D-CNN-LSTM model.

[0026] By employing multi-source environmental factor data and pre-training a 3D-CNN-LSTM model in the source domain, a spatiotemporal fusion deep learning framework is constructed. This fully leverages the advantages of 3D convolutional neural networks in extracting spatial features and long short-term memory networks in capturing the dynamic characteristics of time series data. Training with large-scale source domain data allows the model to learn the general patterns of soil moisture changes, improving not only its ability to capture soil moisture variation characteristics but also its generalization ability, providing a solid and reliable foundation for subsequent accurate soil moisture prediction.

[0027] like Figure 3 As shown, in this embodiment, a pre-trained 3D-CNN-LSTM model based on the source domain (e.g., the nine main urban districts of Chongqing) is loaded, and the weights of the 3D-CNN feature extraction layer are frozen. A small number of samples are randomly selected from the target domain (the nine main urban districts of Chongqing, approximately 28.5°N-30.5°N, 105.5°E-107.5°E), and only the LSTM temporal modeling layer and the fully connected output layer are unfrozen. The parameters are adaptively updated by minimizing the mean squared error loss function, the formula of which is: Among them, y i and Let N represent the observed value and the model prediction value of the i-th sample, respectively, and N be the total number of samples. This enables adaptive updating of the model parameters, resulting in a 3D-CNN-LSTM model fine-tuned for the target domain.

[0028] Loading a pre-trained 3D-CNN-LSTM model from the source domain and freezing some weights can accelerate the convergence of the target domain model training by leveraging knowledge from the source domain. Unfreezing some layers from a small number of randomly selected samples from the target domain for fine-tuning not only utilizes the characteristics of the target domain data to optimize the model but also reduces computational costs. By minimizing the mean squared error loss function to achieve adaptive parameter updates, the model's predictions become closer to the observed values, effectively improving the model's accuracy in predicting relevant data in the target domain (the nine main urban districts of Chongqing) and enhancing the model's applicability.

[0029] In this embodiment, a sliding window mechanism with a window size of 5×5 and a step size of 1 pixel is used to predict soil moisture frame by frame on the target domain data, outputting a soil moisture prediction map with a resolution of 30 meters; the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) are used. 2 The performance of the evaluation model is assessed by the following indicators: MSE = 3.989, MAE = 1.455, R... 2 =0.903; Compared with the benchmark models LST, CNN, and CNN-LSTM, the advantages of the 3D-CNN-LSTM model are verified, and a high-resolution prediction map and performance evaluation report are obtained.

[0030] Employing a specific sliding window mechanism for frame-by-frame prediction, this method can meticulously process target domain data and output high-precision 30-meter resolution soil moisture prediction maps, providing detailed data support for related research. (The last sentence appears to be incomplete and possibly refers to a different method or approach.) 2 The comprehensive evaluation of model performance provides a clear picture of the model's predictive accuracy. Comparison with various benchmark models highlights the advantages of the 3D-CNN-LSTM model, validating its effectiveness and advancement in soil moisture prediction. This provides reliable high-resolution prediction results and a scientific basis for performance evaluation in soil moisture monitoring and related fields.

[0031] like Figure 4 As shown, in this embodiment, monthly time-series remote sensing images of the target domain are collected to generate soil moisture raster data; annual scale features are analyzed to identify the spatial distribution patterns of high and low humidity areas; monthly scale features are analyzed to explore the dynamic characteristics of soil moisture changes with the seasons, thus obtaining the spatiotemporal characteristics of soil moisture.

[0032] Collecting monthly time-series remote sensing images and generating soil moisture raster data provides rich and continuous basic data for research. Annual-scale feature analysis can accurately identify the spatial distribution of high and low humidity zones, helping to grasp the macroscopic pattern of soil moisture. Monthly-scale feature analysis can explore its dynamic changes with the seasons, revealing its inherent patterns of change. Through these operations, the spatiotemporal characteristics of soil moisture are comprehensively acquired, providing a detailed and reliable data foundation and scientific basis for scientific research on soil moisture, agricultural planning, and water resource management.

[0033] Annual-scale characteristic analysis: High humidity areas are mainly distributed along the banks of the Jialing River, the Yangtze River and their tributaries, as well as in river valley depressions (controlled by topographic-hydrological coupling) and on the shady slopes of mountains (insufficient sunshine and weak evaporation); Low humidity areas are concentrated on mountain ridges (steep slopes and rapid runoff) and urban core areas (hardened surface and low infiltration), with soil moisture values ​​ranging from 0.5989 to 24.3305.

[0034] Lunar-scale feature analysis:

[0035] January to March (winter to spring): Soil moisture generally remains at a high level, with high-value areas remaining stable, while low-value areas in urban areas expand slightly with the arrival of spring;

[0036] April to June (late spring and early summer): The highest humidity value rises from 25.39 in January to 29.57 in June. In April, the high value area shrinks due to rising temperature. In May, the high value area extends along the river due to increased precipitation. In June, the high value area is scattered in the tributary basin.

[0037] July-September (late summer and early autumn): The lowest humidity value in August drops to 0.22. Strong convective precipitation triggers localized high humidity patches, and subsequent high temperatures lead to a decrease in humidity.

[0038] October to December (late autumn and early winter): Rainfall decreases and evaporation weakens, humidity distribution tends to be stable, high values ​​are concentrated near the main water system, and low values ​​are distributed in highlands and hardened ground.

[0039] In this embodiment, the candidate feature set is initialized to generate a feature subset; the feature subset is globally optimized through crossover, mutation, and selection operations; and the coefficient of determination R is used. 2 As a fitness function, the predictive performance of each feature subset is evaluated using the following formula: in, The mean of the observed values ​​is used; the subset combination is iteratively optimized, redundant or weakly correlated features are removed, and factors that contribute significantly to soil moisture prediction are retained to obtain the key feature subset.

[0040] Performing operations such as population initialization on the candidate feature set can comprehensively and systematically uncover the relationships between features, generating a more representative feature subset. The coefficient of determination R0 is used as the basis for this analysis. 2 As a fitness function to evaluate prediction performance, it scientifically and reasonably measures the merits of each feature subset. By iteratively optimizing and eliminating redundant or weakly correlated features, factors that significantly contribute to soil moisture prediction can be accurately retained, resulting in a key feature subset, which effectively improves the quality of features and the relevance of predictions.

[0041] In this embodiment, a 3D convolutional neural network (3D-CNN) is constructed to extract spatial features from spatiotemporal data. The convolution formula is: Where x represents the input data, w(i,j,k) represents the convolution kernel, and y(x,y,z) is the output feature map; a Long Short-Term Memory (LSTM) network is constructed to capture the dynamic characteristics of time series, and the core formulas include the forget gate, input gate, output gate, and unit state update; a fusion framework is constructed, and a three-dimensional convolution kernel is used to extract local spatiotemporal features in the time-row-column three-dimensional space; after pooling, the output sequence enters the LSTM layer to capture time dependence, and the soil moisture prediction value is output through a fully connected layer. The model uses mean squared error as the loss function to obtain a cross-regional soil moisture prediction model.

[0042] 3D convolutional neural networks can fully extract spatial features from spatiotemporal data, effectively capturing the spatial distribution patterns of soil moisture. Long short-term memory networks can accurately capture the dynamic characteristics of time series data, grasping the changing trends of soil moisture over time. The fusion framework combines the two to comprehensively extract local spatiotemporal features. Mean squared error is used as the loss function to optimize the model, making the model's predictions more accurate and providing an efficient and accurate tool for cross-regional soil moisture prediction.

[0043] In this embodiment, the forgetting gate formula is: f t =σ(w f ·[h t-1 ,x t ]+b f The input gate formula is: i t =σ(w i ·[h t-1 ,x t ]+b i The output gate formula is: o t =σ(w o ·[h t-1 ,x t ]+b o The cell state update formula is: C t =f t *C t-1 +i t *tanh(w c ·[h t-1 ,x t ]+b c The hidden state output formula is: h t =o t *tanh(c t ); where f t For the Gate of Oblivion, i t For the input gate, o t For output gate, c t For the unit state, h t In the hidden state, x t For the current input data, w f w i w o w c Let b be the weight matrix. f b i b o b c σ is the bias term, and σ is the activation function.

[0044] The forgetting gate formula determines the retention and forgetting of information from the previous hidden state and the current input data, effectively filtering key historical information. The input gate formula controls the amount of new information added to the hidden state, helping the model learn new features. The output gate and unit state update formulas work together to adjust the hidden state, making it better reflect the current situation. These formulas work synergistically, enabling the Long Short-Term Memory network to accurately capture the dynamic characteristics of time series data, providing a solid foundation for cross-regional soil moisture prediction models.

[0045] In this embodiment, a fusion framework is constructed, and a three-dimensional convolutional kernel is used to extract local spatiotemporal features in the time-row-column three-dimensional space. The calculation formula is as follows: Where, x (l-1) w represents the input feature map of layer l-1. (l) Let l be the kernel weights of convolution, and f() be the activation function.

[0046] Three-dimensional convolutional kernels can extract local spatiotemporal features from the time-row-column three-dimensional space, breaking through the limitations of traditional two-dimensional convolution in spatiotemporal data processing and more comprehensively capturing the complex correlations of soil moisture data in the spatiotemporal dimension. By processing the input feature map through specific calculation formulas, the model can deeply mine the inherent information of the data. This provides a more accurate feature extraction method for cross-regional soil moisture prediction models, helping to improve the model's perception and prediction capabilities of soil moisture changes and enhance the reliability of prediction results.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A high-precision method for extracting soil moisture to support cross-regional migration in mountainous surface environments, characterized in that, The method for high-precision extraction of soil moisture supporting cross-regional migration in mountainous surface environments includes the following steps: The 3D-CNN-LSTM model was trained in the source domain using multi-source environmental factor data. A basic prediction framework was constructed through spatiotemporal feature extraction and parameter optimization to obtain the 3D-CNN-LSTM model. The 3D-CNN feature extraction layer of the pre-trained model is frozen, and the LSTM temporal layer and output layer are fine-tuned using a small number of samples in the target domain to achieve adaptive parameter updates, resulting in a 3D-CNN-LSTM model fine-tuned in the target domain. Soil moisture prediction maps with a 30-meter resolution were obtained by frame-by-frame prediction using a target domain sliding window mechanism. MSE, MAE, and R... 2 The performance of the model is evaluated using indicators, resulting in high-resolution prediction graphs and performance evaluation reports. Using soil moisture raster data as the response variable, the spatiotemporal distribution patterns at the annual and monthly scales were extracted through monthly time-series images and spatial statistical analysis to obtain the spatiotemporal characteristics of soil moisture. The candidate feature set is initialized, crossover, and selected using a genetic algorithm to extract factors that significantly contribute to soil moisture prediction, thus obtaining a subset of key features. A 3D-CNN-LSTM fusion framework is constructed. Local spatiotemporal features are extracted by three-dimensional convolutional kernels, and LSTM is combined to capture time dependence and output predicted values ​​to obtain a cross-regional soil moisture prediction model.

2. The method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 1, characterized in that, The method of training a 3D-CNN-LSTM model using multi-source environmental factor data in the source domain, constructing a basic prediction framework through spatiotemporal feature extraction and parameter optimization, and obtaining the 3D-CNN-LSTM model includes the following steps: The 3D-CNN-LSTM model is pre-trained in the source domain using multi-source environmental factor data; Spatial features are extracted by 3D convolutional neural networks and time series dynamic characteristics are captured by long short-term memory networks, thus constructing a spatiotemporal fusion deep learning framework. By training the model parameters using large-scale source domain data, the model learns the general laws of soil moisture change, resulting in a 3D-CNN-LSTM model.

3. The method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 1, characterized in that, The process of freezing the 3D-CNN feature extraction layer of the pre-trained model, fine-tuning the LSTM temporal layer and output layer using a small number of samples from the target domain to achieve adaptive parameter updates, and obtaining the target domain-fine-tuned 3D-CNN-LSTM model includes the following steps: Load the source domain pre-trained 3D-CNN-LSTM model and freeze the weights of the 3D-CNN feature extraction layer; A small number of samples are randomly drawn from the target domain, and only the LSTM temporal modeling layer and the fully connected output layer are unfrozen; Adaptive parameter updating is achieved by minimizing the mean squared error loss function, which is formulated as follows: Among them, y i and Let represent the observed value and the model predicted value of the i-th sample, respectively, and N be the total number of samples; Adaptive updating of model parameters is achieved to obtain a 3D-CNN-LSTM model with fine-tuned target domain.

4. The method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 1, characterized in that, The method involves frame-by-frame prediction using a target domain sliding window mechanism to obtain a 30-meter resolution soil moisture prediction map, employing MSE, MAE, and R... 2 The performance of the model is evaluated using metrics to obtain a high-resolution prediction graph and a performance evaluation report, including the following steps: A sliding window mechanism with a window size of 5×5 and a step size of 1 pixel is used to predict soil moisture on the target domain data frame by frame, and output a soil moisture prediction map with a resolution of 30 meters. The mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) were used. 2 Indicators are used to evaluate model performance; The MSE = 3.989, MAE = 1.455, R 2 =0.903; The advantages of the 3D-CNN-LSTM model are verified by comparing it with LST, CNN, and CNN-LSTM benchmark models, and high-resolution prediction graphs and performance evaluation reports are obtained.

5. The method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 1, characterized in that, The method of using soil moisture raster data as the response variable, and extracting the spatiotemporal distribution patterns at the annual and monthly scales through monthly time-series imagery and spatial statistical analysis to obtain the spatiotemporal characteristics of soil moisture includes the following steps: Collect monthly time-series remote sensing images of the target area to generate soil moisture raster data; Analyze the annual-scale characteristics to identify the spatial distribution patterns of high-humidity and low-humidity areas; We analyzed the monthly scale characteristics to explore the dynamic characteristics of soil moisture variation with the seasons and obtained the spatiotemporal characteristics of soil moisture.

6. The method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 1, characterized in that, The process of initializing, crossovering, and selecting the candidate feature set using a genetic algorithm to extract factors that significantly contribute to soil moisture prediction and obtain a key feature subset includes the following steps: Population initialization is performed on the candidate feature set to generate a feature subset; The feature subset is globally optimized through crossover, mutation, and selection operations; Using the coefficient of determination R 2 As a fitness function, the predictive performance of each feature subset is evaluated using the following formula: in, The mean of the observed values; The subset combination is iteratively optimized, redundant or weakly correlated features are removed, and factors that contribute significantly to soil moisture prediction are retained to obtain the key feature subset.

7. The method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 1, characterized in that, The construction of the 3D-CNN-LSTM fusion framework, which extracts local spatiotemporal features through three-dimensional convolutional kernels and combines LSTM to capture temporal dependencies, outputs predicted values ​​to obtain a cross-regional soil moisture prediction model, includes the following steps: A 3D convolutional neural network (3D-CNN) is constructed to extract spatial features from spatiotemporal data. The convolution formula is as follows: Where x represents the input data, w(i,j,k) represents the convolution kernel, and y(x,y,z) is the output feature map; A Long Short-Term Memory (LSTM) network is constructed to capture the dynamic characteristics of time series. The core formulas include the forget gate, input gate, output gate, and cell state update. A fusion framework is constructed, and three-dimensional convolutional kernels are used to extract local spatiotemporal features in the time-row-column three-dimensional space; After pooling, the output sequence enters the LSTM layer to capture time dependence, and the soil moisture prediction value is output through the fully connected layer. The model uses the mean squared error as the loss function to obtain a cross-regional soil moisture prediction model.

8. A method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 7, characterized in that, The construction of a Long Short-Term Memory (LSTM) network to capture the dynamic characteristics of time series data includes the following core formulas: forget gate, input gate, output gate, and cell state update: The forget gate formula is: f t =σ(w f ·[h t-1 ,x t ]+b f ); The input gate formula is: i t =σ(w i ·[h t-1 ,x t ]+b i ); The output gate formula is: o t =σ(w o ·[h t-1 ,x t ]+b o ); The unit state update formula is: C t =f t *C t-1 +i t *tanh(w c ·[h t-1 ,x t ]+b c ; The hidden state output formula is: h t =o t *tanh(c t ); Where f t For the Gate of Oblivion, i t For the input gate, o t For output gate, c t For the unit state, h t In the hidden state, x t For the current input data, w f w i w o w c Let b be the weight matrix. f b i b o b c σ is the bias term, and σ is the activation function.

9. A method for high-precision extraction of soil moisture supporting cross-regional migration in a mountainous surface environment according to claim 1, characterized in that, The construction of the fusion framework, which uses three-dimensional convolutional kernels to extract local spatiotemporal features in the time-row-column three-dimensional space, includes the following steps: The constructed fusion framework uses a three-dimensional convolutional kernel to extract local spatiotemporal features in the time-row-column three-dimensional space. The calculation formula is as follows: Where, x (l-1) w represents the input feature map of layer l-1. (l) Let l be the kernel weights of the convolution, and f() be the activation function.

Images