Ecological hydrological simulation method based on machine learning model

Abstract

The invention provides an ecological hydrological simulation method based on a machine learning model, and the method comprises the steps: obtaining ecological hydrological data, carrying out the preprocessing of the obtained ecological hydrological data, obtaining a data set, building an ecological hydrological model based on an improved ConvLSTM network structure, and training the ecological hydrological model based on the data set. And obtaining a trained ecological hydrological model, inputting to-be-measured data into the trained ecological hydrological model, and performing evapotranspiration prediction. According to the method, evapotranspiration prediction is realized based on various data, advantages of remote sensing data and ground meteorological data are fully utilized, evapotranspiration prediction precision is remarkably improved, real-time data can be processed, real-time performance is high, different applications can be performed according to different areas, flexible adjustment of model input can be realized, expandability is high, and the method is suitable for popularization and application. The method can be applied to multiple fields, and provides real-time support for agricultural and ecological monitoring.

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring and meteorological technology, and in particular to an ecohydrological simulation method based on a machine learning model. Background Art

[0002] Evapotranspiration (ET) refers to the process by which plants release water vapor into the atmosphere through transpiration and evaporation from the soil surface. It is an important component of the hydrological cycle and a key factor in water resources management and agricultural production. Accurately estimating ET is helpful for the scientific management of water resources and the promotion of water-saving agriculture.

[0003] Existing evapotranspiration estimation methods are mainly divided into direct measurement methods and indirect estimation methods. The direct measurement method uses evapotranspiration instruments (such as sensors, meteorological towers, etc.) for measurement, but its practical application is limited due to its expensive equipment, complex operation, and significant impact from environmental factors. The indirect estimation method mainly relies on indirect information such as meteorological data, remote sensing data, and soil data. Commonly used models include radiation energy balance method, meteorological method, remote sensing method, etc. However, existing estimation methods usually have the following shortcomings: 1. Data uniformity: Many methods rely on only one type of data, such as using only meteorological data or remote sensing data, which results in the estimation accuracy being limited by a single data source.

[0004] 2. Poor real-time performance: The calculation process of the existing methods is relatively complicated and cannot achieve real-time estimation.

[0005] 3. Low accuracy: Under complex terrain, different vegetation coverage or changing climate conditions, the estimation accuracy of existing methods cannot meet high-precision requirements.

[0006] Therefore, it is necessary to design an ecohydrological simulation method based on machine learning models. Summary of the invention

[0007] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide an eco-hydrological simulation method based on a machine learning model.

[0008] To achieve the above object, the present invention provides the following solutions: The present invention provides an ecohydrological simulation method based on a machine learning model, comprising: Obtain ecohydrological data; Preprocess the acquired ecohydrological data to obtain a data set; Constructing eco-hydrological model based on improved ConvLSTM network structure; The eco-hydrological model is trained based on the data set to obtain a trained eco-hydrological model; The data to be tested are input into the trained eco-hydrological model to predict evapotranspiration.

[0009] Preferably, the eco-hydrological data include remote sensing data, surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data.

[0010] Preferably, the acquired eco-hydrological data is preprocessed, specifically including: Preprocess remote sensing data; Data cleaning is performed on surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data.

[0011] Preferably, the remote sensing data is preprocessed, specifically: Acquire remote sensing images and perform noise reduction processing to obtain images to be corrected; Perform geometric correction on the corrected image, resample the geometrically corrected image, and complete image registration; Perform radiometric calibration and atmospheric correction on the registered images; Obtain a full-color image, and fuse the image after radiometric calibration and atmospheric correction with the full-color image to obtain a fused image; The fused image is cropped using the vector boundary to obtain the preprocessed remote sensing image.

[0012] Preferably, data cleaning is performed on the surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data, specifically: Data anomaly detection is performed on surface temperature data, surface humidity data, wind speed data, air pressure data, and measured evapotranspiration data based on correlation; Perform anomaly repair on the anomaly points detected by the anomaly.

[0013] Preferably, the improved ConvLSTM network structure is specifically as follows: The traditional ConvLSTM network structure is used as the backbone network, the CBAM attention module is integrated, and the SWATS optimization algorithm is used in the back propagation process of network training to obtain the improved ConvLSTM network structure.

[0014] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention provides an ecohydrological simulation method based on a machine learning model, the method comprising: obtaining ecohydrological data, preprocessing the obtained ecohydrological data to obtain a data set, constructing an ecohydrological model based on an improved ConvLSTM network structure, training the ecohydrological model based on the data set to obtain a trained ecohydrological model, inputting the data to be tested into the trained ecohydrological model, and predicting evapotranspiration. The present invention realizes the prediction of evapotranspiration based on multiple data, makes full use of the advantages of remote sensing data and ground meteorological data, significantly improves the accuracy of evapotranspiration prediction, can process real-time data, has strong real-time performance, can be used for different applications according to different regions, can realize flexible adjustment of model input, has strong scalability, can be applied to multiple fields, and provides real-time support for agriculture and ecological monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0016] Figure 1 A flow chart of a method provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the anomaly detection process; Figure 3 This is a schematic diagram of the abnormality repair process; Figure 4 This is a schematic diagram of the internal structure of the traditional ConvLSTM network; Figure 5 This is a schematic diagram of the improved ConvLSTM network structure; Figure 6 It is a schematic diagram of the overall structure of the CBAM module; Figure 7 Schematic diagram of the internal structure of the channel attention module; Figure 8 The internal structure of the spatial attention module. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The purpose of the present invention is to provide an eco-hydrological simulation method based on a machine learning model, which realizes the prediction of evapotranspiration based on multiple data, fully utilizes the advantages of remote sensing data and ground meteorological data, significantly improves the accuracy of evapotranspiration prediction, can process real-time data, has strong real-time performance, can be used in different applications according to different regions, can realize flexible adjustment of the model input, has strong scalability, can be applied to multiple fields, and provides real-time support for agricultural and ecological monitoring.

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides an ecohydrological simulation method based on a machine learning model, comprising: Step 100: Obtain ecohydrological data; Step 200: preprocessing the acquired eco-hydrological data to obtain a data set; Step 300: constructing an eco-hydrological model based on the improved ConvLSTM network structure; Step 400: training the eco-hydrological model based on the data set to obtain a trained eco-hydrological model; Step 500: Input the data to be tested into the trained eco-hydrological model to predict evapotranspiration.

[0021] In step 100, the eco-hydrological data includes remote sensing data, surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data.

[0022] In step 200, the acquired eco-hydrological data is preprocessed, specifically including: Step 201: preprocessing remote sensing data; Step 202: Clean the surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data.

[0023] In step 201, the remote sensing data is preprocessed, specifically: Acquire the remote sensing image and perform noise reduction to obtain the image to be corrected, specifically: Choose the appropriate noise reduction method based on the noise type and intensity in the remote sensing image: For the peaks or bright spots formed by periodic noise, use bandpass or slot filtering to eliminate them; For the spike noise that is not parallel to the scanning direction, Fourier transform is used for filtering; For the stripes parallel to the scanning direction and the bad lines irrelevant to the radiation signal, Fourier transform and low-pass filtering are used to eliminate or weaken them; After noise reduction, the area, characteristics and influence of thin clouds and shadows in the remote sensing image are determined, filtering technology is used to reduce the noise and blur caused by thin clouds, and methods including ratio method, interpolation method and model method are used to remove the shadow area to obtain the image to be corrected; Perform geometric correction on the corrected image, resample the geometrically corrected image, and complete image registration, specifically: Determine the reference image, select corresponding ground control points from the image to be corrected and the reference image, where the ground control points have the same position and obvious features on the image to be corrected and the reference image, and have obvious and clear point marks, including road intersections and river intersections. The objects on the ground control points do not change over time. The ground control points are evenly distributed in the entire image and a certain number is guaranteed. The ground control points can be selected from the topographic map as a reference, or obtained through field GPS measurement, or obtained from the corrected image. Determine the coordinate transformation relationship between the corresponding ground control points in the reference image and the image to be corrected, and obtain the geometric correction model. Based on the geometric correction Model, transform the pixel coordinates in the image to be corrected into the corrected coordinate system, obtain the new coordinates of the pixels in the image to be corrected, and judge whether the new coordinates of the ground control points in the reference image are within the set error range based on the coordinate positions of the ground control points in the reference image. If so, complete the geometric correction. If not, readjust the ground control points, iterate to obtain an accurate geometric correction model, and apply it to the entire image to be corrected. Through interpolation calculation, estimate the pixel values ​​at non-integer coordinates in the image after geometric correction according to the values ​​of the original adjacent pixels in the image to be corrected, and complete image resampling. The calculation methods of interpolation calculation include nearest neighbor method, bilinear interpolation method and cubic convolution method, etc.; The registered images are subjected to radiometric calibration and atmospheric correction, specifically: Convert the brightness grayscale value of the registered image into an absolute radiation brightness value, eliminate the radiation value difference of the unchanged ground objects, complete the radiation calibration, and invert the real reflectivity of the surface objects according to the preset radiation brightness value and atmospheric parameters to complete the atmospheric correction; Obtain a full-color image, fuse the image after radiometric calibration and atmospheric correction with the full-color image to obtain a fused image, specifically: A panchromatic image is a black-and-white image of the entire visible light wavelength range (generally defined as 0.4μm-0.7μm) acquired by a remote sensor. It is a mixed image of this wavelength range, generally a grayscale image. Panchromatic images and multispectral images are parallel and both belong to remote sensing images. The images that have been calibrated by radiation and corrected by the atmosphere are fused into a high-resolution multispectral image to obtain a fused image. The mosaicking and cropping of the fused image is completed, including the following steps: Image fusion based on Fourier transform or decomposition method to generate a high-resolution multispectral image to provide more comprehensive information; The mosaic image and the fused image are geometrically registered so that they are in the same spatial coordinate system. Then, a certain method is used to color match the adjacent images so that the images at different phases are coordinated and consistent in color. Determine the overlapping area, select a stitching line connecting the two images in the overlapping area, complete the image mosaic, and then perform color uniformity processing to ensure the color balance of the mosaicked image; The image is cropped according to the division boundary, and the area outside the study area is removed to obtain the preprocessed remote sensing image.

[0024] In step 202, data cleaning is performed on the surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data, specifically: Data anomaly detection is performed on surface temperature data, surface humidity data, wind speed data, air pressure data, and measured evapotranspiration data based on correlation; Perform anomaly repair on the anomaly points detected by the anomaly.

[0025] The above two steps are introduced, in which data anomaly detection is performed on the surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data based on correlation, specifically: The surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data are all regarded as multidimensional time series, and data anomaly detection is performed on them. The detailed introduction is as follows: With regard to the existing problems in the anomaly detection of multidimensional time series data, considering that the time series data itself is computable, a solution is proposed from the perspective of calculating the correlation of the original multidimensional time series data itself. Since the anomaly situations in multidimensional time series data are often complex and diverse, for example: only a single dimension in the multidimensional data always produces anomalies at the same time, multiple dimensions all have errors at the same time, only one dimension of the multidimensional data has anomalies at different times, and multiple dimensions all have errors at different times, etc., the present invention only targets the scenario where only a single dimension of the multidimensional data has anomalies at different times, and provides a basic framework for detecting anomalies, such as Figure 2 As shown, the framework can roughly divide the anomaly detection method into three parts: correlation calculation, determination of anomaly dimension, and location of anomaly points; Correlation calculation is the first step to express the multidimensional time series data to be cleaned in a form that can be recognized and processed by the computer. Since the meaning of each dimension in the multidimensional time series data varies greatly and its numerical units are completely different, it is not feasible to calculate some causal relationships between them. Therefore, it is considered that a certain statistic can be used to measure the relationship between values ​​with different meanings. Through investigation, the similarity measurement method is used to describe the correlation between different dimensions. Finally, it is believed that the correlation coefficient is more appropriate for the calculation of the correlation of multidimensional time series data. The correlation coefficient formula is used to calculate the correlation between each two dimensions in the sliding window. The size of the correlation coefficient indicates the strength of the association relationship. The abnormal dimension is determined mainly by calculating the correlation between the multidimensional time series data of the historical sliding window size and the multidimensional time series data in the first sliding window and comparing them one by one according to their locations. In this operation, whether the correlation of each dimension is stable is evaluated. Once the correlation coefficient measuring the correlation of a certain dimension is significantly reduced, there is reason to believe that there is an abnormal phenomenon in the dimension and the dimension is regarded as an abnormal dimension. The location of outliers is based on the operation of determining the abnormal dimension. After clarifying which dimension has an abnormality, we should pay attention to the time point at which the outlier is located in the abnormal dimension. In order to solve this problem, it is considered and believed that this is no longer handled by relying solely on the relationship between multi-dimensional time series data, but needs to be combined with the speed change law on the single-dimensional time series data to assist in judging the abnormal point. Therefore, the present invention proposes a strategy for locating abnormal points based on the speed change rate of single-dimensional time series data and the stability of the correlation relationship between dimensions; Different from the single-dimensional time series data anomaly detection method, the anomaly detection framework proposed in the present invention is based on the correlation coefficient for anomaly judgment. It is highly sensitive to data changes. Even if the deviation between the abnormal data and its real data is small, the corresponding correlation coefficient will have a significant difference, thereby avoiding the inability to identify those abnormal values ​​that meet the detection rules. At the same time, the detection framework does not use machine learning technology, but only uses mathematical calculations to complete the entire process, which saves a lot of time and helps to clean data online in real time. In addition, the window size is specified according to the data characteristics, and the detection of abnormal points in the current window only depends on other dimensions and the previous historical window, which basically eliminates the influence of time on anomaly judgment. Therefore, the entire detection framework relies on multiple dimensions in the same environment to jointly assist in the judgment of anomalies. In theory, the detection standards obtained based on these dimensions are more reasonable than the anomaly constraints predicted by outdated and invalid historical data in the single-dimensional time series data detection method. In summary, the multi-dimensional time series data anomaly detection framework based on correlation solves the problems existing in the constraint-based single-dimensional time series data anomaly detection method to a certain extent, and obtains stronger detection capabilities; Next, we will introduce them in detail: 1. Correlation calculation First, the multidimensional time series data needs to be represented and stored in a logical form through a data structure to facilitate the subsequent calculation of the correlation coefficient; The relationship between multidimensional time series data describing the same scene or phenomenon is complex and unknown. Without prior knowledge of which dimensions each dimensional data is correlated with, the correlation coefficient is calculated for each dimensional data and other dimensions except itself, so as to rely on the most relevant dimensional data to detect anomalies. i As a vertex V i , any two vertices (dimensions) V i and V j The correlation between them can be represented by an edge E ij It is particularly important to note that V i V j With V j V i The correlation coefficient r ij and r ii are equal, so E ij should be an undirected edge, and r ijis the weight on the edge. After the above analysis, an undirected graph is used to characterize the logical structure of the collected multidimensional time series data. At the same time, this undirected graph can be stored in the form of an adjacency matrix Cor Matrix. In order to realize real-time anomaly detection, a sliding window is used to segment the stored complete multidimensional time series data, trying to transform the global anomaly detection problem into a local anomaly detection problem. Each segmented local multidimensional time series data is called a double sliding window, and the correlation calculation is performed on the data represented as an array in each double sliding window in chronological order. In the correlation calculation process, since the correlation coefficient ranges from [-1, 1], a negative value represents that the two-dimensional data is negatively correlated, and a positive number represents that the two-dimensional data is positively correlated. However, whether the obtained correlation coefficient is positively correlated or negatively correlated, it will not affect the effect of anomaly detection. 2. Determination of abnormal dimensions For anomaly detection of multi-dimensional time series data, the idea is to first determine the anomaly dimension and then locate the anomaly point. The calculation of correlation is the basis for anomaly discovery. Considering that the correlation between multi-dimensional data will change over time, but this correlation will be stable to a certain extent over a period of time, that is, the correlation will produce small fluctuations within a certain range. The correlation coefficient of the data in the current window is obtained through the correlation calculation stage, and the correlation coefficient adjacency matrix of the current window is compared with the correlation coefficient adjacency matrix of the historical window to determine the anomaly sequence; 3. Locate abnormal points This consistency can be described by the similarity of the speed change rate ratio between the data on the abnormal dimension and the dimension most related to it. When the speed change rate ratio between the data in the abnormal dimension is significantly lower or higher than the speed change rate ratio of the data in another dimension (exceeding or lower than the threshold), the abnormal point can be located. Among them, the threshold for judging whether the speed change rate fluctuates significantly needs to be set by yourself, and its value is set to 0.5 according to experience.

[0026] Perform abnormal repair on the abnormal points detected by the anomaly, specifically: The abnormal repair behavior is divided into two steps, such as Figure 3 As shown, it includes generating a candidate set of repair values ​​and determining a final repair value; Generating a candidate set of repair values ​​is the first step in anomaly repair. It first receives the anomalies found in the anomaly detection stage in the order of the double sliding windows, and stores and represents the anomalies and their related dimensions in each double sliding window. Secondly, it defines the correlation between the multi-dimensional time series data in the current double sliding window, and combines the total impact of other dimensions on the dimensions to be repaired and calculates the impact value of each dimension. Finally, it randomly extracts several dimensions to assist the dimensions to be repaired in generating a candidate set of repair values. Determining the final repair value is the second and last step in anomaly repair. It first uses a clustering-based method to optimally classify the candidate set of repair values ​​generated in anomaly repair to obtain several clusters. Secondly, according to the "minority obeys majority" rule (that is, points with low data density are outliers or anomalies, and points with high density form a class or cluster), the cluster with the highest data density is selected. For such a cluster, we believe that its cluster center can best represent all data, so the cluster center is often used as the final repair value. However, it may happen that the cluster center does not belong to the candidate set of repair values. Therefore, the data that exists in the candidate set of repair values ​​and is closest to the cluster center is generally selected as the final repair value.

[0027] In step 300, the improved ConvLSTM network structure is specifically: First, the traditional ConvLSTM network structure is introduced: ConvLSTM is an excellent algorithm proposed by Dr. Shi Xingjian of the Hong Kong University of Science and Technology in 2017 for spatiotemporal sequence prediction. It is also an improved algorithm of LSTM. The LSTM algorithm has a very strong ability to process spatiotemporal sequence data, and CNN is also very capable of extracting image features. ConvLSTM combines CNN and LSTM modules at the bottom layer, and adds convolution operations on the basis of LSTM. The internal structure of ConvLSTM is as follows Figure 4 As shown; The calculation method of ConvLSTM state at each moment is shown in the following formula:

[0028] In the formula, * represents the convolution operation, ∘ represents the Hadamard operation (matrix corresponding element multiplication). Unlike LSTM, the connection between each gate and the input in ConvLSTM becomes a convolution operation, and the connection between states also becomes a convolution operation; The traditional ConvLSTM network structure is used as the backbone network, the CBAM attention module is integrated, and the SWATS optimization algorithm is used in the back propagation process of network training to obtain the improved ConvLSTM network structure. The model structure is as follows: Figure 5 As shown; The CBAM convolutional attention module includes a spatial attention module and a channel attention module. It is a dual-channel attention mechanism. The CBAM attention module infers the attention weights in two independent dimensions, space and channel, and integrates them with the input feature map to perform adaptive feature optimization. Then, different weights are assigned according to the importance of the features. The Adam algorithm and the SGD algorithm are often used in the back propagation stage of network training. The Adam algorithm converges quickly and does not require the specification of hyperparameters, but its generalization performance is not as good as the SGD algorithm. The SGD algorithm converges slowly and requires the specification of hyperparameters, but its generalization ability is worse than the Adam algorithm. The SWATS optimization algorithm is used. The SWATS optimization algorithm integrates the Adam and SGD algorithms, and can use different optimization algorithms at different stages of model training, integrating the advantages of the two algorithms and making up for their disadvantages. It can speed up network convergence and achieve lower losses, reduce the gap between network prediction values ​​and real data, and thus obtain better generalization effects. The CBAM convolutional attention module is introduced. The convolutional attention module (Convolutional Block Attention Module, CBAM) is an attention module for feedforward convolutional neural networks. Compared with the SEN et attention module that only focuses on channel information, CBAM combines spatial and channel dimensions and achieves better results in many fields. The overall structure of the CBAM module is shown in the figure. Figure 6 As shown; from Figure 6 It can be seen that the input features obtained by the convolutional layer will first pass through a channel attention module to generate channel attention weights. The channel attention weights perform weighted operations on the input features to obtain features suitable for processing by the spatial attention module. The features are then passed through a spatial attention module to generate spatial attention weights. The spatial attention weights perform weighted operations on the features to obtain the final calculation results. The internal structure of the channel attention module is as follows Figure 7 As shown; The feature map input to the channel attention module will pass through the maximum pooling layer and average pooling layer based on Width and Height respectively, and then enter the shared fully connected layer. The features output by the fully connected layer are then accumulated by multiplying the corresponding elements one by one, and then activated by the Sigmoid function to obtain the channel attention weight, which is then element-wise multiplied with the input features to generate the input feature map required by the spatial attention module; The role of the channel attention mechanism is to compress the input feature map in the spatial dimension, obtain a one-dimensional vector and then perform subsequent operations. Average pooling and maximum pooling can aggregate the spatial information of the feature map and pass it to the shared fully connected layer. The spatial dimension of the input feature map is then compressed, element by element, and then the sum is calculated and merged to finally generate a channel attention feature map. The calculation method of the channel attention mechanism is shown in the following formula:

[0029] Figure 8 The internal structure diagram of the spatial attention module. The feature map output by the channel attention module is the input of the spatial attention module. First, the input feature map is subjected to maximum pooling and average pooling operations based on the channel. Then, the calculation results of the two are concatenated based on the channel. Then, the calculation result is reduced to a channel through a 7×7 convolution operation. The spatial attention weight is generated through the activation operation of the Sigmoid function. Finally, the output spatial attention weight is multiplied by the input feature map of the spatial attention module to obtain the output feature of the CBAM attention module. Similar to the channel attention mechanism, the spatial attention mechanism also compresses the channel of the feature, performs average pooling and maximum pooling operations on the channel dimension, and then merges it with the previously extracted feature map with channel 1 to obtain a feature map with channel number 2. The calculation formula of the spatial attention mechanism is shown in the following formula:

[0030] In the formula, is the Sigmoid function, 7×7 represents the size of the convolution kernel, and the convolution kernel of this size has a better effect; Introduction to the SWATS optimization algorithm: Under the same circumstances, the Adam algorithm is better than the SGD algorithm in terms of network convergence speed and generalization ability in the initial stage of training, but these two indicators of the Adam algorithm stagnate in the subsequent stage. If generalization ability is used as a measurement indicator, the SGD algorithm is better than the Adam algorithm. SWATS, like the Adam algorithm, does not require the specification of initial hyperparameters, and the default hyperparameters have good performance. The network converges faster in the early stage of training, but the Adam algorithm has the limitation of poor generalization performance due to uneven gradient scaling. The SWATS algorithm combines the characteristics of the Adam and SGD algorithms. The Adam algorithm is used in the early stage of network training to ensure a faster network convergence speed, and switches to the SGD algorithm at the appropriate time to obtain better generalization performance. The conversion from Adam to SGD algorithm is automatic and does not introduce any hyperparameters. The selection of switching points and the learning rate of the SGD algorithm are obtained through learning as part of the training process. The SWATS algorithm combines the characteristics of the Adam and SGD algorithms. The advantages of the Adam algorithm and the SGD algorithm overcome the shortcomings of the weak generalization ability of the Adam algorithm and the slow convergence speed of the SGD algorithm, and achieve better results than the two in terms of neural network convergence speed, generalization ability and robustness. Therefore, the present invention uses the SWATS algorithm in the back propagation stage of network training; During specific operation, the data set is input into the neural network, and after extraction and processing by the convolution layer and the maximum pooling layer, feature data suitable for the input form of the ConvLSTM module is obtained. The ConvLSTM module extracts and calculates the input features, and the output features are processed by the CBAM convolutional attention module. The channel attention module in CBAM first assigns weights to the ConvLSTM input data to obtain feature data suitable for processing by the spatial attention module. The spatial attention module calculates and outputs the input feature data.

[0031] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0032] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An ecohydrological simulation method based on a machine learning model, characterized in that: include: Obtain ecohydrological data; Preprocess the acquired ecohydrological data to obtain a data set; Constructing eco-hydrological model based on improved ConvLSTM network structure; The eco-hydrological model is trained based on the data set to obtain a trained eco-hydrological model; The data to be tested are input into the trained eco-hydrological model to predict evapotranspiration.

2. The ecohydrological simulation method based on a machine learning model according to claim 1, characterized in that: The eco-hydrological data include remote sensing data, surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data.

3. The ecohydrological simulation method based on machine learning model according to claim 2 is characterized in that: Preprocess the acquired ecohydrological data, including: Preprocess remote sensing data; Data cleaning is performed on surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data.

4. The ecohydrological simulation method based on a machine learning model according to claim 3 is characterized in that: Preprocess the remote sensing data, specifically: Acquire remote sensing images and perform noise reduction processing to obtain images to be corrected; Perform geometric correction on the corrected image, resample the geometrically corrected image, and complete image registration; Perform radiometric calibration and atmospheric correction on the registered images; Obtain a full-color image, and fuse the image after radiometric calibration and atmospheric correction with the full-color image to obtain a fused image; The fused image is cropped using the vector boundary to obtain the preprocessed remote sensing image.

5. The ecohydrological simulation method based on a machine learning model according to claim 4, characterized in that: Data cleaning is performed on the surface temperature data, surface humidity data, wind speed data, air pressure data and measured evapotranspiration data, specifically: Data anomaly detection is performed on surface temperature data, surface humidity data, wind speed data, air pressure data, and measured evapotranspiration data based on correlation; Perform anomaly repair on the anomaly points detected by the anomaly.

6. The ecohydrological simulation method based on a machine learning model according to claim 5, characterized in that: The improved ConvLSTM network structure is as follows: The traditional ConvLSTM network structure is used as the backbone network, the CBAM attention module is integrated, and the SWATS optimization algorithm is used in the back propagation process of network training to obtain the improved ConvLSTM network structure.

Images