Inland lake water depth inversion method based on lake bottom classification

Abstract

The application discloses an inland lake water depth inversion method based on lake bottom classification, comprising the following steps: S1, acquiring data required for lake water depth inversion and processing; S2, classifying lake bottom topography based on a support vector machine classification algorithm, and dividing shallow water areas and deep water areas; S3, constructing a lake shallow water area water depth inversion model by using a deep neural network algorithm; S4, inversing lake deep water area water depth based on the support vector machine algorithm; and S5, combining the shallow water area and deep water area water depths obtained in S3 and S4 to obtain inland lake water depth information. The method constructs different models for different lake bottom topographies, and inverses water depths in different regions, so that the lake water depth inversion precision is greatly improved.

Description

Technical Field

[0001] This invention belongs to the technical field of lake depth inversion methods, specifically relating to a method for inverting the depth of inland lakes based on lake bottom classification. Background Technology

[0002] Accurate mapping of lake depth is fundamental to understanding lake level changes and water storage capacity, and is of great significance for assessing lake ecosystems and hydrological cycles.

[0003] Currently, traditional methods for measuring water depth include shipborne sonar sounding systems, airborne laser sounding systems, and manual underwater measurement. These methods are time-consuming, labor-intensive, costly, and have limited coverage. Compared to traditional methods, remote sensing monitoring of water depth is widely used due to its advantages of real-time performance, wide coverage, and low cost. Therefore, optical satellites have become one of the important data sources for water depth inversion. Optical satellite water depth inversion models mainly include physical models, semi-theoretical / semi-empirical models, and statistical models. Among them, physical models are based on the principle of water attenuation of light and the spectral characteristics of water bodies, but their parameters are complex and difficult to obtain. Semi-theoretical / semi-empirical models have the advantages of theoretical simplification and ease of operation. These models invert water depth by constructing a linear regression relationship between wavebands and measured data. However, due to the complexity and heterogeneity of water optics, simple linear models cannot accurately reflect water depth conditions. Therefore, to solve the complex nonlinear regression problem, machine learning algorithms are introduced into water depth inversion, improving the accuracy of water depth monitoring. Support Vector Machine (SVM) models have good generalization ability and strong robustness, making them suitable for inversion of small sample datasets; while Deep Neural Network (DNN) models automatically learn the features in the dataset through multi-layer hidden neural networks, giving them a significant advantage in handling large samples and non-linear datasets.

[0004] However, existing water depth inversion methods mainly focus on large, homogeneous islands, ports, estuaries, and coastal areas, while paying less attention to complex lakebeds and inland lakes heavily affected by human activity. The lakebed topography differs significantly between shallow and deep water areas of inland lakes, and different types of lakebeds have varying impacts on the accuracy of water depth inversion. Using limited measured water depth data makes it difficult to monitor the complex changes in lakebed topography, while increasing sampling points requires substantial human and financial resources. In particular, due to the topographical limitations of shallow lake areas or the inaccessibility of lake regions, obtaining more measured data is even more challenging. Therefore, efficiently acquiring lakebed topographic data in shallow water areas has become a key focus.

[0005] In recent years, unmanned aerial vehicles (UAVs) have been widely used in water depth monitoring due to their advantages such as high spatial resolution, low cost, high timeliness, good flexibility, and immunity to atmospheric interference. Compared with satellite imagery, UAVs can not only provide high-precision spatial resolution but also acquire inversion results that closely match measured data, and can also provide ground-based data for satellite-scale inversion. Therefore, UAV technology has become a new method for rapidly collecting ground-based measured data. Currently, existing water depth inversion studies assume that the lake bottom is uniformly distributed, neglecting the influence of different lake topography on the accuracy of water depth inversion, and comparing the inversion capabilities of various empirical models or single machine learning models, but lacking methods for water depth inversion based on lake bottom classification and utilizing different machine learning algorithms. Summary of the Invention

[0006] The purpose of this invention is to provide a method for inland lake water depth inversion based on lake bottom classification. Different models are constructed for different lake bottom topography, and then the water depth of different regions is inverted, which greatly improves the accuracy of lake water depth inversion.

[0007] The technical solution adopted in this invention is an inland lake water depth inversion method based on lake bottom classification, which is implemented according to the following steps:

[0008] S1, Obtain and process the data required for lake depth inversion;

[0009] S2, based on the support vector machine classification algorithm, classifies the lake bottom terrain and divides it into shallow water area and deep water area;

[0010] S3, using deep neural network algorithms, constructs a water depth inversion model for shallow lake areas;

[0011] S4, Inverting the water depth of deep water areas in lakes based on the support vector machine algorithm;

[0012] S5, combined with the shallow and deep water depths obtained from S3 and S4, yields the water depth information of inland lakes.

[0013] The invention is further characterized by:

[0014] The data to be acquired in step S1 includes: measured water depth data, UAV multispectral data, Sentinel-2 data, and lake boundary data; among which, the UAV multispectral data collection area is evenly distributed in the shallow water area of ​​the lake.

[0015] Step S2 is as follows:

[0016] S21, Determine the optimal threshold for dividing the shallow-deep regions of the lake:

[0017] Select lake water depths of 1-3m, with 0.1m intervals. Based on lake boundary data, Sentinel-2 data and measured water depth data, the support vector machine algorithm is used to divide the lake into shallow and deep water areas. The area of ​​the deep and shallow water areas of the lake is counted each time, and the threshold with the largest change is used as the optimal threshold for dividing the shallow-deep areas of the lake.

[0018] S22, Under the optimal threshold, the support vector machine algorithm is used to divide the shallow water area and deep water area of ​​the lake.

[0019] Step S3 is as follows:

[0020] S31, using formula (7) single-band model and formula (8) band ratio model to transform the UAV data bands, increasing the sensitive bands:

[0021] (7)

[0022] (8)

[0023] In equations (7) and (8), H To obtain the water depth through inversion, X n For band reflectivity, A 0 and A 1 represents an undetermined coefficient;

[0024] S32. The SHAP framework was used to select the top nine variables with the highest correlation to the measured water depth data of the UAV area. Based on the obtained nine variables, the water depth of the UAV area was inverted using SVM. The inversion result was then resampled to 10m and used as the measured water depth data of shallow water area in the field.

[0025] S33, using single-band model and band ratio model to transform the Sentinel-2 data bands;

[0026] S34. Using the SHAP framework, the top nine variables with the highest correlation to the measured shallow water depth data are selected. Based on the top nine optimal band combinations, a shallow water depth inversion model is constructed using DNN.

[0027] Step S4 is as follows:

[0028] S41, using single-band model and band ratio model to transform the Sentinel-2 data bands;

[0029] S42 uses the SHAP framework to select the top nine variables that are most correlated with the measured deep water depth data. Based on the top nine most important bands, SVM is used to estimate the deep water depth.

[0030] In the SHAP framework used in steps S32, S34, and S42, the features jThe SHAP value is defined as follows:

[0031] (9)

[0032] In equation (9), For the original feature set, express Any feature subset in, Features j A subset of all elements in the previous sequence. The output of the machine learning model for the feature subset. Features j The cumulative contribution value.

[0033] Step S5 is as follows:

[0034] S51, using the optimal threshold obtained in step S1 as the standard, assign the threshold number to the pixels in the shallow water area that are greater than the threshold.

[0035] S52, using the optimal threshold obtained in step S1 as the standard, assign the threshold number to the pixels in the deep water area that are smaller than the threshold.

[0036] S53, by stitching together the water depth inversion results from the shallow and deep water areas, lake bottom elevation information with a spatial resolution of 10 meters is obtained.

[0037] The beneficial effects of this invention are:

[0038] (1) The method of the present invention fully considers the influence of different lake topography on the accuracy of water depth inversion, and avoids the problem of reduced water depth inversion accuracy caused by inverting water depth based on a single lake topography;

[0039] (2) The method of the present invention divides the lake bottom into shallow water area and deep water area for different types of lake bottom, and uses DNN and SVM to construct the optimal model to invert the water depth of different areas, which improves the accuracy of lake water depth inversion and has robustness;

[0040] (3) In view of the problems of time-consuming, labor-intensive and costly traditional methods of acquiring measured water depth data, the method of this invention rematches the UAV inversion results to the satellite scale and replaces the measured water depth data in the field, which greatly increases the accuracy of the measured water depth data and improves the accuracy of lake water depth inversion. It has important research significance for the improvement of water depth inversion methods. Attached Figure Description

[0041] Figure 1 This is a schematic flowchart of the method of the present invention;

[0042] Figure 2 This is a spatial distribution map of the lake bottom topography obtained in Embodiment 1 of the present invention;

[0043] Figure 3 This is a spatial distribution map of the water depth inversion in the UAV region obtained in Embodiment 1 of the present invention;

[0044] Figure 4 This is a schematic diagram of the water depth inversion results obtained in Embodiment 1 of the present invention;

[0045] Figure 5 This is a schematic diagram of the water depth inversion result using only SVM in Embodiment 2 of the present invention;

[0046] Figure 6 This is a schematic diagram of the water depth inversion result using only DNN in Embodiment 2 of the present invention;

[0047] Figure 7 This is a schematic diagram showing the result of water depth inversion using the method of the present invention in Embodiment 2 of the present invention;

[0048] Figure 8 This is a comparison chart of the accuracy of the inversion results of the three methods in Embodiment 2 of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0050] This invention relates to an inland lake depth retrieval method based on lake bottom classification. It utilizes UAV multispectral and Sentinel-2 data to jointly retrieve inland lake depths. The process is as follows: Figure 1 As shown, please follow these steps:

[0051] S1, Data Preparation and Processing:

[0052] The prepared data included: measured water depth, UAV multispectral data, Sentinel-2 data, and lake boundary data. Measured water depth data was obtained using the HOBO water level and temperature automatic recorder; UAV multispectral data was preprocessed using Pix4D software; Sentinel-2 data was obtained through radiometric calibration and atmospheric correction of multispectral remote sensing images; and lake boundary data was obtained through Normalized Difference Water Index (NDWI) and visual interpretation. The specific methods for obtaining each data point were as follows:

[0053] (1) Measured water depth data:

[0054] Field-measured water depth data was collected using a HOBO water level and temperature automatic recorder. First, the HOBO instrument was set to a sampling time of 30 seconds, connected to a steel cable, and atmospheric pressure was acquired. The instrument was then slowly submerged to the lake bottom and allowed to remain stationary for 5 minutes. Subsequently, the instrument was retrieved, and the lake bottom pressure was converted into water depth using HOBOware software. Simultaneously, GPS-RTK was used to record the sampling point information.

[0055] (2) UAV multispectral data:

[0056] This invention distributes the drone data collection area evenly in the shallow waters of the lake, making the collected data more representative.

[0057] Field multispectral data acquisition: Data acquisition will be conducted under clear, cloudless or partly cloudy, and windless conditions, with acquisition times selected from 8:00-11:00 and 15:00-18:00. Before flight, the white reference plate will be measured using an ASD-2000 spectrometer, and the multispectral sensor will be radiometrically calibrated using known reflectance values. Flight altitude, heading, and azimuth overlap will be set.

[0058] The main steps for indoor drone data processing are: 1. Add all drone images; 2. Import POS coordinate files; 3. Select the output coordinate system, WGS84, click Edit, and select Advanced Coordinate System; 4. Output the results and view the quality report.

[0059] (3) Sentinel 2 data:

[0060] Select Sentinel 2 Level 2 product data (good image quality, no cloud cover) that overlaps with or is a few days apart from the field sampling time, and use SNAP software to resample the images to 10m, and convert the coordinate system to WGS-84 geographic coordinate system.

[0061] (4) Lake boundary data:

[0062] Using the normalized difference water index ( NDWI The lake is separated into land and water zones. A threshold is set to 0, and visual interpretation is used to correct the lake boundary. The calculation formula is as follows:

[0063] (1)

[0064] In the formula: Green , NIR These are the reflectances in the green light and near-infrared bands, respectively.

[0065] S2, Classification of Lake Bottom Topography

[0066] S21, Determine the optimal threshold for dividing the shallow-deep regions of the lake:

[0067] Select lake water depths of 1-3m, with 0.1m intervals. Based on lake boundary data, Sentinel-2 data, and measured water depth data, the support vector machine algorithm is used to divide the lake into shallow and deep water areas. The areas of the deep and shallow water areas of the lake are counted each time, and the threshold with the largest change is taken as the optimal threshold for dividing the shallow-deep areas of the lake.

[0068] S22, Under the optimal threshold, the support vector machine algorithm is used to divide the shallow water area and deep water area of ​​the lake.

[0069] The formula for classifying the bottom of lakes based on SVM is:

[0070] (2)

[0071] In equation (2), x For sample input, f ( x ) is a nonlinear mapping, w For weight vectors, b The threshold value is used.

[0072] For a given training dataset D={( x 1, y 1), ( x 2, y 2), …( x n , y n Using an insensitive loss function Then the optimization objective of SVM can be expressed as:

[0073] (3)

[0074] (4)

[0075] In equation (3), i =1, 2… n , C This is called the penalty parameter. C The larger the value, the greater the penalty for classification. By introducing the Lagrange function, the optimization problem in equation (3) can be transformed into a dual problem for solution.

[0076] (5)

[0077] In equation (5), , For Lagrange multipliers; nsv The number of support vector machines; This is the kernel function.

[0078] (6)

[0079] In equation (6), For kernel parameters.

[0080] S3, based on a deep neural network algorithm, retrieves the water depth in the shallow water area of ​​a lake. Specifically:

[0081] S31. Use formula (7) single-band model and formula (8) band ratio model to transform the UAV data bands, increase the sensitive bands, and improve the inversion accuracy.

[0082] (7)

[0083] (8)

[0084] In equations (7) and (8), H To obtain the water depth through inversion, X n For band reflectivity, A 0 and A 1 is an undetermined coefficient.

[0085] S32, using the SHAP framework, selected the top nine variables with the strongest correlation to the UAV-measured water depth data in the region, and identified the following features. j The SHAP value is defined as follows:

[0086] (9)

[0087] In equation (9), For the original feature set, express Any feature subset in, Features j A subset of all elements in the previous sequence. The output of the machine learning model for the feature subset. Features j The cumulative contribution value.

[0088] Based on the nine variables obtained, the water depth in the UAV area was retrieved using SVM, and the retrieved result was resampled to 10m to serve as the field measured water depth data for shallow water areas.

[0089] S33, use formula (7) single-band model and formula (8) band ratio model to transform the Sentinel 2 data band;

[0090] S34. Using the SHAP formula (9), the top nine variables with the greatest correlation to the measured shallow water depth data are selected. Based on the top nine optimal band combinations, a shallow water depth inversion model is constructed using DNN.

[0091] The basic principle of SHAP:

[0092] SHAP is a method for interpreting predictions from machine learning models. It's based on the Shapley value concept from game theory and applies it to the calculation of feature importance. The SHAP value calculation process includes the following steps: 1. Construct a set of participants containing all feature subsets; 2. For each feature subset, calculate the marginal contribution of each participant. The marginal contribution represents the impact of each feature within the feature subset; 3. Calculate the SHAP value of each feature based on its marginal contribution. The SHAP value represents the degree to which each feature contributes to the prediction result; 4. Sum the SHAP values ​​of all features to obtain the final prediction interpretation.

[0093] The SHAP value of feature j is defined by the above formula (9). The following can be obtained by calculating the SHAP value: 1. Feature importance ranking: By calculating the SHAP value of all samples and taking the average, the global importance ranking of each feature can be obtained; 2. Feature importance of a single sample: For a single sample, the local importance of each feature can be calculated, so as to understand the specific influence of the feature on the prediction result in the sample; 3. Interaction between features: The calculation of the SHAP value takes into account the interaction between features, so it can reveal the results of different feature combinations on the prediction result.

[0094] The basic principles of deep neural networks (DNNs):

[0095] A DNN can be understood as a neural network with many hidden layers; it can also be called a multilayer perceptron model. Multilayer networks learn the relationship between input features and output values ​​to obtain a linear function, expressed as:

[0096] (10)

[0097] The function model is followed by an activation function that corresponds to the output:

[0098] (11)

[0099] A deep neural network (DNN) consists of three layers: the first layer is the input layer; the middle layers are hidden layers, which can have one or more; and the last layer is the output layer, which can have one or more output results. Each node in each layer is fully connected to nodes in adjacent layers. l Any node in the layer must be related to the first node. l Connect any node in the ±1 layer.

[0100] Assume the activation function is The output values ​​of the hidden layer and the output layer are a For a three-layer DNN model, the output of the second layer... a 1 2 , a 22 , a 3 2 have:

[0101] (12)

[0102] The output of the last third layer a 1 3 have:

[0103] (13)

[0104] Therefore, it can be assumed that if the first l -1 floor has m For the nth neuron, then for the nth neuron... l The first layer j The output of each neuron a j l have:

[0105] (14)

[0106] Represented by a matrix, the first l The layer output is:

[0107] (15)

[0108] In equation (15), L This represents the total number of layers in a deep neural network. W This is a matrix consisting of the weight parameters between the hidden layer nodes and the output layer nodes. b The matrix represents the node biases, and the error exists between the initial predictions and the true values, i.e., the loss function:

[0109] (16)

[0110] In equation (16), a L For predicted values, y Substituting the true value into formula (15), the loss function becomes:

[0111] (17)

[0112] Therefore, we can obtain W , b gradient:

[0113] (18)

[0114] (19)

[0115] The common parts in formulas (18) and (19) are denoted as :

[0116] (20)

[0117] Then for the first l Inactive data of the layer z l The gradient can be expressed as:

[0118] (twenty one)

[0119] Calculated l Layer W l and b l The gradient is as follows:

[0120] (twenty two)

[0121] (twenty three)

[0122] Using mathematical induction to obtain :

[0123] (twenty four)

[0124] (25)

[0125] Then any layer can be obtained. W and b :

[0126] (26)

[0127] (27)

[0128] S4, based on the support vector machine algorithm, retrieves the water depth in deep water areas. Specifically:

[0129] S41, use formula (7) single-band model and formula (8) band ratio model to transform the Sentinel 2 data bands;

[0130] S42, using the SHAP formula (9), the top nine variables with the greatest correlation to the measured deep water depth data are selected. Based on the top nine most important bands, the deep water depth is estimated using SVM.

[0131] S5 combines the water depths of shallow and deep water areas to obtain inland lake water depth information. Specifically:

[0132] S51, using the optimal threshold obtained in step 1 as the standard, assign the threshold number to the pixels in the shallow water area that are greater than the threshold.

[0133] S52, using the optimal threshold obtained in step 1 as the standard, assign the threshold number to the pixels in the deep water area that are smaller than the threshold.

[0134] S53, by stitching together the water depth inversion results from the shallow and deep water areas, lake bottom elevation information with a spatial resolution of 10 meters is obtained.

[0135] Example 1

[0136] In this embodiment, Hongjiannao Lake was selected as the research object, and the results obtained according to the above method are as follows:

[0137] (1) Classification of lake bottom

[0138] Multiple experiments showed that the lake area decreased most significantly in deep water when the threshold was greater than 2.2m. Therefore, 2.2m was chosen as the optimal threshold for dividing the shallow and deep regions of the lake. Based on the SVM algorithm, the deep and shallow water areas of the lake were divided. The overall classification accuracy of both the training and test datasets was above 94%, the Kappa coefficient was above 0.88, and the user accuracy and mapping accuracy were above 90%. The classification accuracy of the lake bottom was high and can be used for subsequent research. The spatial distribution of the lake bottom topography obtained is shown in the figure. Figure 2 As shown.

[0139] (2) Inversion of water depth in UAV area based on SVM algorithm

[0140] Using SVM to perform water depth inversion in the UAV region, the inversion accuracy was calculated to be: R 2 =0.95, RMSE=0.27, MAE=0.20, RPD=4.86, the spatial distribution obtained by inversion is as follows: Figure 3 As shown.

[0141] (3) Retrieve lake depth in shallow water areas based on DNN algorithm

[0142] Using DNN to perform water depth inversion in shallow water areas, the inversion accuracy was calculated to be: R 2 =0.87, RMSE=0.19, MAE=0.11, RPD=2.78.

[0143] (4) Inversion of lake depth in deep water areas based on SVM algorithm

[0144] Using SVM to perform deep-water depth inversion, the inversion accuracy was calculated to be: R 2 =0.89, RMSE=0.39, MAE=0.29, RPD=3.07.

[0145] (5) Combining the water depth inversion results of shallow and deep water areas, the final lakebed elevation information with a spatial resolution of 10 meters is obtained, such as... Figure 4As shown, the maximum water depth of Hongjiannao is 5.2m, and the average depth is 3.87m.

[0146] Example 2

[0147] This embodiment compares the method of the present invention with the support vector machine algorithm and the deep neural network algorithm. Hongjiannao Lake was selected as the research object, and a total of 194 measured water depth data points were collected, including 100 in shallow water and 94 in deep water. To ensure the reliability of the experiment, 10 water depth data points were selected to verify the inversion results, and the remaining water depth data were used for the experiment. The lake water depth inversion was performed using the method of the present invention, the support vector machine algorithm, and the deep neural network algorithm, respectively. The results are as follows:

[0148] (1) Based on 174 measured water depth data points, the lake water depth was inverted using the support vector machine algorithm, yielding a maximum water depth of 5.2 m and an average depth of 3.91 m. The water depth inversion results are as follows: Figure 5 As shown.

[0149] (2) Based on 174 measured water depth data points, the lake's water depth was retrieved using a deep neural network algorithm. The results showed a maximum water depth of 5.6 m and an average depth of 3.81 m. The water depth retrieval results are as follows: Figure 6 As shown.

[0150] (3) Using the method of this invention (based on lake bottom classification, utilizing support vector machine and deep neural network algorithms) to invert the lake depth, the maximum lake depth was found to be 5.2m, and the average depth was 3.88m. The depth inversion results are as follows: Figure 7 As shown.

[0151] (4) Accuracy comparison

[0152] Using 20 measured water depth data points, the water depth inversion results of the three methods were verified, and the accuracy of the three methods was compared as follows: Figure 8 As shown in the figure, the method of the present invention and the verification data R 2 =0.99, the fitting method Y=1.01X+0.09 has better accuracy than the other two methods.

[0153] Example 3

[0154] The method for inverting inland lake depth based on lakebed classification is implemented according to the following steps:

[0155] S1, Obtain and process the data required for lake depth inversion;

[0156] S2, based on the support vector machine classification algorithm, classifies the lake bottom terrain and divides it into shallow water area and deep water area;

[0157] S3, using deep neural network algorithms, constructs a water depth inversion model for shallow lake areas;

[0158] S4, Inverting the water depth of deep water areas in lakes based on the support vector machine algorithm;

[0159] S5, combined with the shallow and deep water depths obtained from S3 and S4, yields the water depth information of inland lakes.

Claims

1. A method for inverting inland lake water depth based on lakebed classification, characterized in that, The specific steps are as follows: S1, Obtain and process the data required for lake depth inversion; S2, based on the support vector machine classification algorithm, classifies the lake bottom terrain and divides it into shallow water area and deep water area; S3, using deep neural network algorithms, constructs a water depth inversion model for shallow lake areas; S4, Inverting the water depth of deep water areas in lakes based on the support vector machine algorithm; S5, combined with the shallow and deep water depths obtained from S3 and S4, yields the water depth information of inland lakes; Step S2 is as follows: S21, Determine the optimal threshold for dividing the shallow-deep regions of the lake: Select lake water depths of 1-3m, with 0.1m intervals. Based on lake boundary data, Sentinel-2 data and measured water depth data, the support vector machine algorithm is used to divide the lake into shallow and deep water areas. The area of ​​the deep and shallow water areas of the lake is counted each time, and the threshold with the largest change is used as the optimal threshold for dividing the shallow-deep areas of the lake. S22, Under the optimal threshold, the support vector machine algorithm is used to divide the shallow water area and deep water area of ​​the lake; Step S3 is as follows: S31, using formula (7) single-band model and formula (8) band ratio model to transform the UAV data bands, increasing the sensitive bands: （7） （8） In equations (7) and (8), H To obtain the water depth through inversion, X n For band reflectivity, A 0 and A 1 represents an undetermined coefficient; S32. The SHAP framework was used to select the top nine variables with the highest correlation to the measured water depth data of the UAV area. Based on the obtained nine variables, the water depth of the UAV area was inverted using SVM. The inversion result was then resampled to 10m and used as the measured water depth data of shallow water area in the field. S33, using single-band model and band ratio model to transform the Sentinel-2 data bands; S34. The SHAP framework was used to select the top nine variables that were most correlated with the measured shallow water depth data. Based on the top nine optimal band combinations, a shallow water depth inversion model was constructed using DNN. Step S4 is as follows: S41, using single-band model and band ratio model to transform the Sentinel-2 data bands; S42, using the SHAP framework to select the top nine variables with the highest correlation to the measured deep water depth data, and using SVM to estimate the deep water depth based on the top nine most important bands; In the SHAP framework used in steps S32, S34, and S42, the features j The SHAP value is defined as follows: （9） In equation (9), For the original feature set, express Any feature subset in, Features j A subset of all elements in the previous sequence. The output of the machine learning model for the feature subset. Features j The cumulative contribution value.

2. The inland lake water depth inversion method based on lake bottom classification according to claim 1, characterized in that, The data to be acquired in step S1 includes: measured water depth data, UAV multispectral data, Sentinel-2 data, and lake boundary data; among which, the UAV multispectral data collection area is evenly distributed in the shallow water area of ​​the lake.

3. The inland lake water depth inversion method based on lake bottom classification according to claim 1, characterized in that, Step S5 is as follows: S51, using the optimal threshold obtained in step S1 as the standard, assign the threshold number to the pixels in the shallow water area that are greater than the threshold. S52, using the optimal threshold obtained in step S1 as the standard, assign the threshold number to the pixels in the deep water area that are smaller than the threshold. S53, by stitching together the water depth inversion results from the shallow and deep water areas, lake bottom elevation information with a spatial resolution of 10 meters is obtained.

Images