Random forest shallow sea sediment classification method based on multi-temporal remote sensing image fusion

Abstract

The invention provides a random forest shallow sea sediment classification method based on multi-temporal remote sensing image fusion, and relates to the technical field of sediment information extraction. Comprising the following steps: 1, collecting and preprocessing multi-temporal image data to obtain remote sensing reflectivity; 2, the water depth of each single-time-phase image is inverted, and the optimal water depth is obtained; 3, calculating bottom reflectivity characteristics of blue and green wave bands based on the optimal remote sensing image; 4, respectively calculating topographic features and spectral features based on the optimal water depth and the optimal remote sensing image; and 5, in combination with the bottom reflectivity features, the topographic features and the spectral features, carrying out random forest feature optimization and classification model training, and generating a substrate classification result. On the basis, the method solves the problems that an existing remote sensing image substrate classification method is insufficient in feature consideration, noise in a single-time-phase image can cause low classification precision, and therefore negative effects can be generated on accurate acquisition of substrate information.

Description

Technical Field

[0001] This invention relates to the field of seabed information extraction technology, specifically to a random forest shallow seabed sediment classification method based on multi-temporal remote sensing image fusion. Background Technology

[0002] Shallow water sediment detection is a crucial component of marine mapping, and accurate sediment information is a vital foundation for marine engineering, coastal ecological protection, and marine resource development. Traditional sediment acquisition methods, such as gravity sampling and grab sampling, suffer from drawbacks such as high destructiveness, low operational efficiency, high measurement costs, and sparse sampling points, limiting their ability for large-scale synchronous detection and failing to meet the needs of modern marine scientific research and development. Shallow water sediment information inversion based on remote sensing imagery offers advantages such as multi-time-period observation, wide spatial coverage, and low measurement costs. It also easily acquires information on disputed, dangerous, or remote islands and reefs, making it an important means of shallow water sediment acquisition. Currently, domestic and international researchers have conducted extensive studies on seabed sediment classification using remote sensing imagery, and the resulting implementation methods can be broadly categorized into the following two types: Firstly, in-situ seabed data, such as multibeam bathymetry and multibeam backscattering data, is acquired using acoustic techniques. Then, this in-situ acoustic data is combined with remote sensing imagery to extract relevant features and construct a classification model to improve the accuracy of seabed classification. However, in practical applications, the high cost of acquiring in-situ acoustic data, the limited measurement range, and the potential hazards significantly restrict the application of this method for large-scale seabed monitoring in remote sea areas.

[0003] Secondly, by visually interpreting remote sensing images, a limited number of representative areas are selected as substrate samples to construct corresponding classification models, overcoming the limitations of acoustic data acquisition. For example, researchers visually interpret seagrass distribution in shallow waters, generating global-scale seagrass distribution maps through model training; they use object-based image analysis methods to create coral reef distribution maps; they extract derived features from the spectral features of remote sensing images for substrate classification; and they combine water depth inversion results with remote sensing image spectra to conduct substrate classification. While this approach does not rely on in-situ acoustic data, it suffers from an over-reliance on remote sensing image spectral features in its classification framework, insufficient consideration of other features, and inadequate dimensional integration, leading to decreased final classification accuracy. Specifically: In sediment classification, bottom reflectance can eliminate the interference of water column attenuation, directly reflect sediment properties, and accurately distinguish sediment types. Seafloor topography determines the type and spatial distribution characteristics of sediment, providing key spatial constraints for sediment classification. However, current research rarely integrates bottom reflectance and topographic features as core features into classification models, failing to fully explore their potential to improve classification accuracy. Therefore, the sediment classification results obtained by existing remote sensing image-based methods are still not optimal.

[0004] In addition, most current methods for seabed classification using remote sensing images are based on single-temporal images. Image noise may lead to inaccurate capture of spectral and topographic features, and the one-sidedness of single-temporal images cannot provide sufficient classification basis for the model. These factors make single-temporal images less stable in complex sea areas, which will further reduce the overall accuracy of seabed classification.

[0005] In summary, current remote sensing image seabed classification methods suffer from insufficient consideration of features and low classification accuracy due to noise in single-temporal images, which negatively impacts the accurate acquisition of seabed information. Therefore, this invention provides a random forest shallow seabed seabed classification method based on multi-temporal remote sensing image fusion. Summary of the Invention

[0006] The purpose of this invention is to provide a random forest shallow seabed sediment classification method based on multi-temporal remote sensing image fusion, in order to solve the problems mentioned in the background art, that existing remote sensing image seabed sediment classification methods do not fully consider features and that noise in single-temporal images leads to low classification accuracy, thus negatively affecting the accurate acquisition of seabed sediment information.

[0007] This invention is achieved using the following technical solution: A random forest shallow seabed sediment classification method based on multi-temporal remote sensing image fusion includes the following steps: Step 1: Collect and preprocess multi-temporal image data to obtain remote sensing reflectance; Step 2: Invert the water depth of each single-phase image and obtain the optimal water depth; Step 3: Calculate the bottom reflectance characteristics of the blue and green bands based on the best remote sensing image; Step 4: Calculate the topographic features and spectral features based on the optimal water depth and the best remote sensing imagery, respectively; Step 5: Combining bottom reflectance features, terrain features, and spectral features, perform random forest feature optimization and classification model training to generate bottom sediment classification results.

[0008] Furthermore, step 2 includes the following sub-steps: Step 2-1: Use a dual-band log-linear model to invert the water depth of each single-phase image; Step 2-2: Perform adaptive weighted fusion to generate the optimal water depth inversion result.

[0009] Furthermore, step 2-2 includes the following sub-steps: Step 2-2-1: Match the corresponding points at a target pixel location in the multi-temporal images, extract the water depth of the corresponding points in each single temporal phase and calculate the average value; Step 2-2-2: Calculate the difference between the water depth of each corresponding point and the corresponding average value, and then assign fusion weights to the corresponding points in each single time phase based on the proportion of each difference in the sum of all differences. The formula for calculating the fusion weight is as follows: , In the formula, For fusion weights, n is the number of single temporal phases in the multi-temporal image. The water depth for each single time phase; Step 2-2-3: Based on the fusion weights of different time phases, the optimal water depth at the target pixel location is obtained through adaptive weighted fusion; The formula for calculating the optimal water depth is as follows: , In the formula, The optimal water depth; Step 2-2-4: Repeat steps 2-2-1 to 2-2-3 for other target pixel locations to generate the optimal water depth inversion result.

[0010] Furthermore, step 3 includes the following sub-steps: Step 3-1: Construct the bottom reflectivity equation based on the water body radiative transfer model; Step 3-2: Perform median fusion on the multi-temporal images to construct the optimal remote sensing image; Step 3-3: Based on the best remote sensing image, calculate the bottom reflectance characteristics of the blue and green bands using the bottom reflectance equation.

[0011] Furthermore, in step 3-1, the formula for calculating the bottom reflectivity equation is as follows: , , In the formula, , These are the bottom reflectance values ​​for the blue and green bands, respectively. , , , respectively, are the weighted feature vectors for the blue and green bands, and b is the substrate parameter and , The ratio of the attenuation coefficients in the blue and green bands. This represents the sum of the diffuse attenuation coefficients for the green band.

[0012] Furthermore, step 3-1 includes the following sub-steps: Step 3-1-1: Introduce blue and green dual-band models to construct water body radiative transfer models respectively; The calculation formula for the water body radiation transfer model is as follows: , , In the formula, Because of the water depth, This is the logarithm of the difference between the subsurface remote sensing reflectance in the blue band and the deep-water subsurface remote sensing reflectance. This is the logarithm of the difference between the subsurface remote sensing reflectance in the green band and the deep-water subsurface remote sensing reflectance. This represents the sum of diffuse attenuation coefficients in the blue band. This represents the sum of the diffuse attenuation coefficients in the green band. Step 3-1-2: Eliminate water depth in the water radiative transfer model for the blue and green bands to construct a depth-independent bottom reflectivity equation; When eliminating water depth, the calculation formula of the water body radiation transfer model is transformed as follows: , In the formula, ,Depend on Replace each and .

[0013] Furthermore, in step 3-2, the formula for calculating the optimal remote sensing image is as follows: , In the formula, It is the pixel value of the fused image at position (i, j). Let be the pixel value of the original image acquired at time t in the time series at position (i, j), and T be the set of all time points used for fusion.

[0014] Furthermore, in step 4: Topographic features include slope, aspect, topographic relief, roughness, and curvature; The formula for calculating the slope is as follows: , The formula for calculating the slope aspect is as follows: , The formula for calculating the terrain relief is as follows: , The formula for calculating the roughness is as follows: , The formula for calculating the curvature is as follows: , In the formula, ( , () represents spatial coordinates. For the elevation of the target point, For the elevation of a certain adjacent grid point, This represents the horizontal distance between the point where the slope is to be calculated and its adjacent points. , Here, x and y are the components of the normal to the 3D surface plot, and slope is the gradient. The slope between a point and its adjacent points.

[0015] Furthermore, in step 4: Spectral features include water indices and vegetation indices. Water indices include the Normalized Difference Water Index, the Corrected Surface Water Index, and the High Resolution Water Index. Vegetation indices include the Normalized Vegetation Index, the Ratio Vegetation Index, and the Enhanced Vegetation Index. The formula for calculating the Normalized Difference Water Index is as follows: , The formula for calculating the modified surface water index is as follows: , The formula for calculating the high-resolution water index is as follows: , The formula for calculating the normalized vegetation index is as follows: , The formula for calculating the ratio vegetation index is as follows: , The formula for calculating the enhanced vegetation index is as follows: , In the formula, R is the red band remote sensing reflectance in the remote sensing image, G is the green band remote sensing reflectance, B is the blue band remote sensing reflectance, and NIR is the near-infrared band remote sensing reflectance.

[0016] Furthermore, step 5 includes the following sub-steps: Step 5-1: Correspond the multi-dimensional features to the substrate to construct feature vectors; The multi-dimensional features include the three-band features of the remote sensing image, bottom reflectance features, topographic features, and spectral features. Step 5-2: Utilize the feature importance ranking of the random forest model to select the preferred features from the multi-dimensional features; Step 5-3: Input the preferred features into the random forest model for training to generate substrate classification results.

[0017] The beneficial effects achieved by this invention are: This invention provides a method for shallow seabed sediment classification based on multi-temporal remote sensing image fusion. By sequentially collecting and preprocessing multi-temporal image data, retrieving water depth from each single-temporal image, calculating bottom reflectance characteristics, calculating topographic and spectral features, and training a random forest model, accurate shallow seabed sediment classification results can be generated. Compared with existing technologies, this method relies solely on multispectral remote sensing images to obtain topographic and spectral information, avoiding the problems associated with in-situ acoustic data acquisition. Furthermore, this method overcomes the shortcomings of current remote sensing image-based seabed sediment classification methods in insufficient feature consideration and solves the problem of low accuracy in single-temporal image seabed sediment classification, effectively improving the stability and robustness of the seabed sediment classification results. Therefore, this invention provides a new solution for obtaining comprehensive and accurate seabed sediment information using remote sensing images in complex environments. The seabed sediment information obtained based on this method can be directly applied to marine engineering construction, coastal ecological protection, and marine resource development. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of the substrate classification method described in an embodiment of the present invention; Figure 2 This is an example of the feature importance ranking result in the substrate classification method described in this embodiment of the invention. Figure I ; Figure 3 This is an example of the feature importance ranking result in the substrate classification method described in this embodiment of the invention. Figure II ; Figure 4 This is a comparison between the sediment classification results of each single phase and the multi-phase fusion sediment classification results in the sediment classification method described in this embodiment of the invention. Figure I ; Figure 5 This is a comparison between the sediment classification results of each single phase and the multi-phase fusion sediment classification results in the sediment classification method described in this embodiment of the invention. Figure II ; Figure 6 This invention compares the classification performance of the random forest model in the substrate classification method described in this embodiment with that of four other mainstream classification models. Figure I ; Figure 7 This invention compares the classification performance of the random forest model in the substrate classification method described in this embodiment with that of four other mainstream classification models. Figure II . Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] Example 1 This embodiment provides a random forest shallow seabed sediment classification method based on multi-temporal remote sensing image fusion. Please refer to... Figure 1 It includes the following steps: Step 1: Collect and preprocess multi-temporal image data to obtain accurate remote sensing reflectance for water depth inversion. Specifically: Step 1 includes the following sub-steps: Step 1-1: Based on the principles of multi-temporal image selection, collect multi-temporal image data with advantages such as high quality and low random interference, including: Step 1-1-1: In Google Earth Engine (GEE), based on the Scene Classification Map (SCL) bands of the remote sensing image, filter out regions with pixel values ​​of 9 (clouds) and 3 (cloud shadows) to construct a binarization mask (cloud / cloud shadow regions are 1, other regions are 0); then, count the total number of pixels of clouds and cloud shadows in the study area and the total number of pixels in the area, and divide the two to obtain the cloud cover percentage; Step 1-1-2: Sort all remote sensing images of the same season in the study area according to cloud cover, comprehensively evaluate image noise and spectral information richness, and select the remote sensing image with the least cloud cover and the best imaging quality. Step 1-1-3: Based on the cloud cover sorting results, select multi-temporal images with high imaging quality and time close to the best image. The specific interval depends on the availability of image data within a specific time period. It is important to ensure that the acquisition time of the multi-temporal images belongs to the same season to reduce the inversion error caused by seasonal differences in the external environment and the optical properties of water.

[0021] Steps 1-2: Preprocess the multi-temporal image data to obtain accurate remote sensing reflectance, which is the basis for subsequent classification, including: Geometric correction: This process can eliminate geometric distortion of the image, ensure the correct correspondence between pixels and actual geographical locations, and thus guarantee consistency during multi-temporal fusion. Radiometric calibration: This process converts the DN values ​​(pixel brightness values) of the original image into radiance values; Atmospheric correction: This process can eliminate interference from atmospheric scattering, aerosols, and other factors, thereby obtaining a surface reflectance that is close to the true value. Water-Land Separation: Water and land can be identified by calculating the Normalized Difference Water Index (NDWI). The NDWI for seawater is positive, while that for land is negative. Therefore, a threshold of 0 removes land and retains water, while a threshold of 0.05 removes deep water and retains shallow water. The principle is as follows: , In the formula, and These represent the reflectance in the green and near-infrared bands, respectively.

[0022] Step 2: Invert the water depth of each single-phase image and obtain the optimal water depth. Specifically: Step 2 includes the following sub-steps: Step 2-1: Invert the water depth of each single-phase image using the Dual-Band Log-Linear Analysis Model Based on Physics (P-DLA), where: Different seabed materials are identified based on the brightness of multispectral images. Feature pixels with different seabed characteristics are extracted from the images, and water depth is inverted based on the remote sensing reflectance of the feature pixels. The model principle is as follows: , In the formula, It's the water depth. , , , These are the weighted feature vectors for the blue and green bands, respectively. These are substrate parameters. It is the ratio of the attenuation coefficients of the blue and green bands. It is the sum of the diffuse attenuation coefficients of the green band; Then, tidal correction is performed to correct the instantaneous water depth of the multi-time back-evolution results to the steady-state water depth of the same reference surface.

[0023] Step 2-2: Adaptive weighted fusion based on multi-temporal steady-state water depth to generate optimal water depth inversion results, including: Step 2-2-1: Based on the geographic coordinates of the pixels in the multi-temporal remote sensing images, match the corresponding points at the location of a target pixel in the multi-temporal images, extract the water depth of the corresponding points in each single temporal phase, and calculate the average value. Among them, the corresponding points refer to the pixels in different temporal images that correspond to the same actual geographical location. Therefore, in multi-temporal images, the pixels corresponding to a certain target pixel location constitute a group of corresponding points. Step 2-2-2: Calculate the difference between the water depth of each corresponding point and the corresponding average value, and then assign fusion weights to the corresponding points in each single time phase based on the proportion of each difference (reciprocal) in the sum of all differences (reciprocals); The formula for calculating the fusion weight is as follows: , In the formula, For fusion weights, n is the number of single temporal phases in the multi-temporal image. The water depth for each single time phase; Step 2-2-3: Based on the fusion weights of different time phases, the optimal water depth at the target pixel location is obtained through adaptive weighted fusion; The formula for calculating the optimal water depth is as follows: , In the formula, The optimal water depth; Step 2-2-4: Repeat steps 2-2-1 to 2-2-3 for other target pixel locations to generate the optimal water depth inversion result.

[0024] Step 3: Calculate the bottom reflectance characteristics of the blue and green bands based on the best remote sensing image. Specifically: Step 3 includes the following sub-steps: Step 3-1: Construct the bottom reflectivity equation based on the water radiative transfer model, including: Step 3-1-1: Since the blue and green bands have strong penetrating power in water, the blue and green dual bands are introduced to construct water radiative transfer models respectively; The calculation formula for the water body radiative transfer model is as follows: , , In the formula, Because of the water depth, This is the logarithm of the difference between the subsurface remote sensing reflectance in the blue band and the deep-water subsurface remote sensing reflectance. This is the logarithm of the difference between the subsurface remote sensing reflectance in the green band and the deep-water subsurface remote sensing reflectance. This represents the sum of diffuse attenuation coefficients in the blue band. This represents the sum of the diffuse attenuation coefficients for the green band; all parameters can be obtained by extracting feature pixels from remote sensing images using the P-DLA model. Step 3-1-2: In order to eliminate the influence of water depth on bottom reflectivity, water depth is eliminated from the water radiative transfer model of the blue and green bands to construct a bottom reflectivity equation that is independent of depth. When eliminating water depth, the calculation formula for the water body radiation transfer model is transformed as follows: , In the formula, ,Depend on Replace each and The bottom reflectance equations for the blue and green bands can then be obtained, as follows: , , In the formula, , These are the bottom reflectance values ​​for the blue and green bands, respectively. , , , respectively, are the weighted feature vectors for the blue and green bands, and b is the substrate parameter and , The ratio of the attenuation coefficients in the blue and green bands. This represents the sum of the diffuse attenuation coefficients for the green band.

[0025] Step 3-2: Perform median fusion on the multi-temporal images to avoid random errors such as noise in single-temporal images, integrate effective temporal information, construct the best remote sensing image, and obtain accurate remote sensing reflectance. The formula for calculating the best remote sensing image is as follows: , In the formula, It is the pixel value of the fused image at position (i, j). Let be the pixel value of the original image acquired at time t in the time series at position (i, j), and T be the set of all time points used for fusion.

[0026] Step 3-3: Based on the best remote sensing image, calculate the bottom reflectance characteristics of the blue and green bands using the bottom reflectance equation, where: Based on the brightness of shallow water areas in remote sensing images, different bottom materials are identified. Based on the P-DLA model principle, feature pixels with different water depths and bottom materials are extracted from the best remote sensing image after median fusion. Then, the model parameters are solved and substituted into the bottom reflectance calculation formula. This allows for the direct acquisition of accurate blue and green band bottom reflectance without the need for auxiliary data such as water depth and water properties.

[0027] Step 4: Calculate topographic features and spectral features based on the optimal water depth and the best remote sensing imagery. Specifically: Seafloor topography determines the type and spatial distribution of seabed sediment, providing key spatial constraints for sediment classification. Without relying on in-situ water depth data, using the optimal water depth generated by adaptive weighted fusion as the base topographic data, slope, aspect, ruggedness, roughness, and curvature can be calculated, among which: Water depth, as a basic topographic data, is calculated using the following formula: ( , ); Slope is used to represent the degree of inclination of a terrain surface at various points, and the calculation formula is: , Slope aspect is used to indicate the direction of slope at a point on the earth's surface. The calculation formula is: , Topographic relief is used to quantify the undulation and complexity of the seabed surface. The calculation formula is as follows: , Roughness, as an indicator of topographic relief, is calculated using the following formula: , Curvature is used to represent the rate of change of slope and measures the degree of unevenness. The formula for calculation is: , In the formula, ( , () represents spatial coordinates. For the elevation of the target point, For the elevation of a certain adjacent grid point, This represents the horizontal distance between the point where the slope is to be calculated and its adjacent points. , Here, x and y are the components of the normal to the 3D surface plot, and slope is the gradient. The slope between a point and its adjacent points.

[0028] Different substrate types exhibit varying spectral responses, and spectral features can transform these invisible physical differences into quantifiable classification criteria. The spectral signals of remote sensing images are primarily influenced by water bodies and substrates; therefore, the spectral features selected in this method include water indices and vegetation indices. All index features are calculated based on the best remote sensing image data after median fusion processing. Water indices include the Normalized Difference Water Index (NDWI), Modified Surface Water Index (MSWI), and High-Resolution Water Index (HRWI); vegetation indices include the Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), and Enhanced Vegetation Index (EVI), among which: Normalized Difference Water Index (NDDI) is used to extract and monitor water and sediment composition. The calculation formula is as follows: , The corrected surface water index is used to amplify the spectral differences between water and non-water bodies. The calculation formula is as follows: , High-resolution water indexes are used to extract different substrates from high-resolution images. The calculation formula is as follows: , The Normalized Difference Vegetation Index (NDVI) is used to improve the accuracy of seabed habitat mapping. The calculation formula is as follows: , The ratio vegetation index is used to highlight vegetation and quantify vegetation cover. The calculation formula is as follows: , The enhanced vegetation index is used to monitor sediment changes in seagrass and algae. The calculation formula is as follows: .

[0029] In the formula, R is the red band remote sensing reflectance in the remote sensing image, G is the green band remote sensing reflectance, B is the blue band remote sensing reflectance, and NIR is the near-infrared band remote sensing reflectance.

[0030] Step 5: Combining bottom reflectance features, topographic features, and spectral features, perform random forest feature optimization and classification model training to generate bottom sediment classification results. Specifically: Step 5 includes the following sub-steps: Step 5-1: Correspond the multi-dimensional features to the substrate to construct a feature vector, where: Random forest feature optimization calculates the feature importance of individual trees based on the Gini coefficient, normalizes the data, and then takes the average of multiple trees to determine global importance. Key features are then selected by ranking them by importance, reducing random bias in individual trees. This method utilizes multi-dimensional features, including three-band features of remote sensing images (red (R), green (G), and blue (B) bands), bottom reflectance features (blue and green band bottom reflectance features), topographic features (water depth, slope, aspect, topographic relief, roughness, and curvature), and spectral features (normalized difference water index, corrected surface water index, high-resolution water index, normalized vegetation index, ratio vegetation index, and enhanced vegetation index), totaling 17 dimensions. These 17 dimensions are arranged in a layered structure, and feature data at the same coordinate positions in each layer are extracted and correlated with the substrate to construct feature vectors.

[0031] Step 5-2: Utilize the feature importance ranking of the random forest model to select optimal features from multi-dimensional features, and eliminate tail features with low feature contribution; refer to... Figure 2 , Figure 3 Here is an example of the feature importance ranking results.

[0032] Step 5-3: Input the selected features into the random forest model for training, and generate the substrate classification results, where: The Random Forest algorithm uses Bootstrap sampling with replacement to extract different samples for each decision tree, and randomly selects feature subsets when splitting nodes, thus achieving randomness in both samples and features and improving robustness to noise and outliers. Then, each decision tree generates a complete structure and outputs a predicted class based on its own samples and feature set using the CART algorithm. The forest then summarizes the predictions of all trees, and the class with the most votes is the final classification result. This mechanism can effectively compensate for the shortcomings of a single decision tree and reduce overfitting.

[0033] The final prediction result is expressed as follows: , In the formula, For the final prediction result, For characteristic function, For a single decision tree, For output variables, For the category set Choose the category that maximizes the subsequent expression. This represents the total number of trees.

[0034] By performing steps 1 to 5 above, high-precision seabed classification results can be obtained in complex shallow sea environments. Compared with existing technologies, this method uses multi-temporal remote sensing image fusion, solving the problem of low seabed classification accuracy in single-temporal images. Furthermore, the randomness and voting mechanism of the random forest algorithm enable it to exhibit high accuracy and stability in seabed classification. Please refer to... Figure 4 , Figure 5 This section compares the sediment classification results of each single time phase with the fused sediment classification results of multiple time phases; please refer to... Figure 6 , Figure 7 This paper compares the performance of the Random Forest model with four other mainstream classification models (BP classification model, Naive Bayes model, maximum likelihood model, and K-Means model).

[0035] The method provided in this embodiment is applicable to islands and reefs with different offshore distances, different sea areas, and different seabed characteristics, and is also applicable to different satellite remote sensing image data; the seabed information obtained based on this method can be directly used in fields such as marine engineering construction, coastal ecological protection, and marine resource development.

[0036] It should be noted that the parts not described in detail or in detail in the above scheme, such as the specific content of each operation in the preprocessing, the specific method of tidal correction, the specific method of matching corresponding points, and the specific implementation process of uncontrolled water depth inversion of each single-temporal remote sensing image using the P-DLA model, are all existing technologies and do not belong to the improvements made by this invention to the existing technology, nor are they within the protection scope of the technical solution of this invention. Therefore, they will not be elaborated on in this article.

[0037] Of course, the above description is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the embodiments of the present invention. The present invention is also not limited to the above examples, and all equivalent changes and improvements made by those skilled in the art within the scope of the present invention should fall within the patent coverage of the present invention.

Claims

1. A random forest shallow seabed classification method based on multi-temporal remote sensing image fusion, characterized in that, The method comprises the following steps: Step 1: collecting and preprocessing multi-temporal image data to obtain remote sensing reflectivity; Step 2: inverting water depth of each single temporal image and obtaining optimal water depth; Step 3: calculating blue and green band bottom reflectance characteristics based on the best remote sensing image; Step 4: calculating terrain features and spectral features based on the optimal water depth and the best remote sensing image respectively; Step 5: combining bottom reflectance characteristics, terrain features and spectral features to perform random forest feature optimization and classification model training to generate a bottom material classification result.

2. The method according to claim 1, wherein, The step 2 comprises the following sub-steps: Step 2-1: inverting water depth of each single temporal image by using a dual-band logarithmic linear model; Step 2-2: performing adaptive weighted fusion to generate an optimal water depth inversion result.

3. The method according to claim 2, wherein, The step 2-2 comprises the following sub-steps: Step 2-2-1: matching same-named points at a target pixel position in the multi-temporal image, extracting water depths of the same-named points in each single temporal image and calculating an average value; Step 2-2-2: calculating a difference value between each same-named point water depth and the corresponding average value, and then assigning a fusion weight to each same-named point on the single temporal image based on the proportion of each difference value in the sum of all difference values; The calculation formula of the fusion weight is as follows: ， In the formula, is the fusion weight, n is the number of single time images in the multi-time image, is the water depth of each single time image; Step 2-2-3: obtaining the optimal water depth at the target pixel position by adaptive weighted fusion based on the fusion weights of different temporal phases; The calculation formula of the optimal water depth is as follows: ， In the formula, is the optimal water depth; Step 2-2-4: repeating the above steps 2-2-1 to 2-2-3 for other target pixel positions to generate an optimal water depth inversion result.

4. The method according to claim 1, wherein, The step 3 comprises the following sub-steps: Step 3-1: constructing a bottom reflectance equation based on a water body radiation transfer model; Step 3-2: performing median fusion on the multi-temporal image to construct a best remote sensing image; Step 3-3: calculating blue and green band bottom reflectance characteristics by using the bottom reflectance equation based on the best remote sensing image.

5. The method according to claim 4, wherein, In the step 3-1, the calculation formula of the bottom reflectance equation is as follows: ， ， wherein, , Rb and Rg are the bottom reflectance of blue and green band respectively, , Lb is the logarithm of the difference between the subsurface remote sensing reflectance of blue band and the deep water subsurface remote sensing reflectance, Lg is the logarithm of the difference between the subsurface remote sensing reflectance of green band and the deep water subsurface remote sensing reflectance, , Wb and Wg are the weight eigenvector of blue and green band respectively, b is the substrate parameter and , Lb / Lg is the ratio of the attenuation coefficient of blue and green band, Sb is the total diffuse attenuation coefficient of blue band, Sg is the total diffuse attenuation coefficient of green band.

6. The method according to claim 5, wherein, The step 3-1 comprises the following sub-steps: Step 3-1-1: introducing blue and green dual bands to construct a water body radiation transfer model; The calculation formula of the water body radiation transfer model is as follows: ， ， wherein is the water depth, is the logarithm of the difference between the sub-surface remote sensing reflectance in the blue band and the deep water sub-surface remote sensing reflectance, is the logarithm of the difference between the sub-surface remote sensing reflectance in the green band and the deep water sub-surface remote sensing reflectance, is the sum of the diffuse attenuation coefficients in the blue band, is the sum of the diffuse attenuation coefficients in the green band; Step 3-1-2: eliminating water depth from the water body radiation transfer model for blue and green bands to construct a bottom reflectance equation independent of depth; When eliminating water depth, the calculation formula of the water body radiation transfer model is transformed into: ， wherein , by replacing and , respectively.

7. The method according to claim 5, wherein, In the step 3-2, the calculation formula of the best remote sensing image is as follows: ， wherein is the pixel value of the fused image at position (i, j), is the pixel value of the original image at position (i, j) acquired at time t in the time series, T is the set of all time points used for the fusion.

8. The multi-temporal remote sensing image fusion based random forest shallow seabed classification method according to claim 1, characterized in that, In the step 4: The terrain features include slope, aspect, terrain relief, roughness and curvature; The calculation formula of the slope is as follows: ， The calculation formula of the aspect is as follows: ， The calculation formula of the terrain relief is as follows: ， The calculation formula of the roughness is as follows: ， The calculation formula of the curvature is as follows: ， In the formula, , is a spatial coordinate, is a target point elevation, is an elevation of a certain adjacent grid point, is a horizontal distance between a point to be calculated and an adjacent point, , is an x-direction component of a three-dimensional curved surface graph normal, and slope is a slope, is a slope between a certain point and an adjacent point.

9. The method according to claim 1, wherein, In the step 4: The spectral features include water body index and vegetation index, wherein the water body index includes normalized difference water index, modified land water index and high-resolution water index, and the vegetation index includes normalized vegetation index, ratio vegetation index and enhanced vegetation index; The calculation formula of the normalized difference water index is as follows: ， The calculation formula of the modified land water index is as follows: ， The calculation formula of the high-resolution water body index is as follows: ， The calculation formula of the normalized vegetation index is as follows: ， The calculation formula of the ratio vegetation index is as follows: ， The calculation formula of the enhanced vegetation index is as follows: ， In the formula, R is the red band remote sensing reflectivity in the remote sensing image, G is the green band remote sensing reflectivity, B is the blue band remote sensing reflectivity, and NIR is the near-infrared band remote sensing reflectivity.

10. The method according to claim 1, wherein, The step 5 includes the following sub-steps: Step 5-1: Corresponding the multi-dimensional features to the substrate to construct a feature vector; The multi-dimensional features include three-band features, bottom reflectivity features, terrain features and spectral features of the remote sensing image. Step 5-2: Filtering out the preferred features from the multi-dimensional features by using the feature importance ranking of the random forest model; Step 5-3: Inputting the preferred features into the random forest model training to generate the substrate classification result.

Images