CN115711838A - Method for inverting suspended sediment in water body based on artificial neural network and high-resolution No. 1 satellite and application of method - Google Patents
Method for inverting suspended sediment in water body based on artificial neural network and high-resolution No. 1 satellite and application of method Download PDFInfo
- Publication number
- CN115711838A CN115711838A CN202210392343.1A CN202210392343A CN115711838A CN 115711838 A CN115711838 A CN 115711838A CN 202210392343 A CN202210392343 A CN 202210392343A CN 115711838 A CN115711838 A CN 115711838A
- Authority
- CN
- China
- Prior art keywords
- model
- neural network
- inversion
- water body
- water
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 title claims abstract description 96
- 239000013049 sediment Substances 0.000 title claims abstract description 65
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 24
- 238000003062 neural network model Methods 0.000 claims abstract description 23
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 21
- 238000007637 random forest analysis Methods 0.000 claims abstract description 8
- 238000010801 machine learning Methods 0.000 claims abstract description 7
- 238000012706 support-vector machine Methods 0.000 claims abstract description 7
- 230000000694 effects Effects 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 19
- 238000005070 sampling Methods 0.000 claims description 12
- 238000012360 testing method Methods 0.000 claims description 12
- 238000004458 analytical method Methods 0.000 claims description 10
- 238000002310 reflectometry Methods 0.000 claims description 10
- 238000012546 transfer Methods 0.000 claims description 10
- 238000001228 spectrum Methods 0.000 claims description 9
- 238000012937 correction Methods 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 7
- 238000010276 construction Methods 0.000 claims description 6
- 238000011160 research Methods 0.000 claims description 6
- 238000011156 evaluation Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 230000003595 spectral effect Effects 0.000 claims description 5
- 230000005540 biological transmission Effects 0.000 claims description 4
- 230000005855 radiation Effects 0.000 claims description 4
- 238000010521 absorption reaction Methods 0.000 claims description 3
- 238000012625 in-situ measurement Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 239000002352 surface water Substances 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 230000002411 adverse Effects 0.000 claims 1
- 238000010835 comparative analysis Methods 0.000 claims 1
- 239000000725 suspension Substances 0.000 claims 1
- 239000010410 layer Substances 0.000 description 55
- 238000005259 measurement Methods 0.000 description 7
- 238000012544 monitoring process Methods 0.000 description 6
- 210000002569 neuron Anatomy 0.000 description 5
- 238000005303 weighing Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 238000000643 oven drying Methods 0.000 description 4
- 239000000126 substance Substances 0.000 description 4
- 238000010200 validation analysis Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 230000007423 decrease Effects 0.000 description 3
- 235000015097 nutrients Nutrition 0.000 description 3
- 239000004576 sand Substances 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000001035 drying Methods 0.000 description 2
- 239000003344 environmental pollutant Substances 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 239000012528 membrane Substances 0.000 description 2
- 231100000719 pollutant Toxicity 0.000 description 2
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 229930002875 chlorophyll Natural products 0.000 description 1
- 235000019804 chlorophyll Nutrition 0.000 description 1
- ATNHDLDRLWWWCB-AENOIHSZSA-M chlorophyll a Chemical compound C1([C@@H](C(=O)OC)C(=O)C2=C3C)=C2N2C3=CC(C(CC)=C3C)=[N+]4C3=CC3=C(C=C)C(C)=C5N3[Mg-2]42[N+]2=C1[C@@H](CCC(=O)OC\C=C(/C)CCC[C@H](C)CCC[C@H](C)CCCC(C)C)[C@H](C)C2=C5 ATNHDLDRLWWWCB-AENOIHSZSA-M 0.000 description 1
- 239000003653 coastal water Substances 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 229910052802 copper Inorganic materials 0.000 description 1
- 239000010949 copper Substances 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000005553 drilling Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000012851 eutrophication Methods 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 239000013535 sea water Substances 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 239000002344 surface layer Substances 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Images
Landscapes
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention relates to a method for inverting suspended sediment in a water body based on an artificial neural network and a high-resolution No. 1 satellite and application thereof. The invention selects four layers (including a double-layer hidden layer) of artificial neural network models to construct a BP neural network inversion model. The establishment of the training sample set is a key link for inversion by utilizing a neural network. The invention improves the inversion accuracy and the fitting degree of the suspended sediment in the turbid water body. Compared with other machine learning algorithms (such as a random forest model, a gradient lifting regression tree model and a support vector machine model), the BPNN neural network model constructed by the method has a better inversion effect on the suspended sediment concentration of the turbid water body.
Description
Technical Field
The invention relates to a method for inverting suspended sediment in a water body based on an artificial neural network and a high-resolution No. 1 satellite and application thereof.
Background
The suspended silt is one of important water quality parameters in inland and near-shore class II water bodies, carries various nutrient substances and pollutants in the water bodies, the transparency of the water bodies is influenced by the content of the nutrient substances and the pollutants, and further the radiation quantity of incident light of water bodies at different depths is also influenced by the nutrient substances, so that the suspended silt influences the living environment of aquatic organisms [1] . Meanwhile, for estuary region, silt expansion of the peripheral mudflat is directly influenced by suspended silt transmission and migration, and engineering such as mudflat reclamation, wharf, bridge construction and the like is also influenced by the silt expansion [2] . Therefore, the concentration of the suspended sediment in the water body and the space-time distribution pattern thereof have important significance for researching the influence of the water ecological environment, wading engineering and landform. The chemical analysis of indoor water quality by the water sample collected by the actual measurement site is a conventional water environment monitoring method, and the method is time-consuming and labor-consuming and has limited number of monitoring sites, so that the water area condition cannot be completely reflected, and the continuity in time and space is insufficient.
The remote sensing technology has the advantage of obtaining information, can monitor the water environment condition of a wide area dynamically in a macroscopic, real-time and real-time manner, and is an important means for monitoring the water environment. The high-resolution one-number satellite (GF-1) is the first satellite of a Chinese high-resolution earth observation system, and has the following perfect characteristics: (1) high spatial and temporal resolution; (2) the repetition period was only 4 days. Therefore, the method has extremely high practical value in the aspect of monitoring water quality change, the spatial resolution of the WFV sensor is 16 meters, and the high precision can meet the precision requirement of inversion of suspended sediment [3] 。
In recent years, research work is mainly based on measured data and remote sensing data, and modeling is carried out after data analysis. The accuracy of AVHRR measurement of the concentration of suspended sediment in Hangzhou bay is high by the aid of the Lijing based on quasi-synchronous actual measurement data and the AVHRR data [4] . AVHRR data are utilized at the same depth of the Lisihai, and a quantitative mode of suspended sediment at the Yangtze river mouth is established by a gray scale method, a slope method and a sediment index method, so that the result shows that the sediment index method is more suitable for a high-concentration sediment area [5] . The remote sensing estimation model is established by combining the Chengxiang and the like with FY-1D remote sensing data and the actual measurement of suspended sediment value at the quasi-synchronous Zhujiang mouth, and the applicability of FY-1D to offshore sediment monitoring is verified [6] . Liuwang soldiers and the like establish an index model based on the ratio of HJ-CCDB4 to B3 by analyzing the correlation between actually measured spectral data of Hangzhou bay and suspended sediment concentration [7] . The drystroke and the like simulate the data of the satellite I with the highest score according to the measured data, and apply the data to inversion of the silt in the Bay of Hangzhou province to obtain a model with the correlation as high as 0.8 and good inversion effect [8] . Although the related inversion technology is mature at present, the near-shore water body spectrum has complexity, and the established model has regional limitation and lacks universality.
Disclosure of Invention
The invention aims to provide a method for inverting the concentration of suspended sediment in a water body based on an artificial neural network and a high-resolution No. 1 satellite, which has high accuracy and strong practicability and is more suitable for inverting the concentration of suspended sediment in turbid water bodies in Hangzhou gulf.
In order to achieve the purpose, the invention adopts the following technical scheme:
a method for inverting the concentration of suspended sediment in a water body based on an artificial neural network and a high-resolution No. 1 satellite comprises the following steps:
the method comprises the following steps: in-situ measurement
And carrying out field sampling on the water body to obtain the water surface remote sensing reflectivity. The water body is inland and near shore class II water body.
Step two: data pre-processing
Preprocessing the acquired original high-resolution No. 1 satellite remote sensing image, and adopting a FLAASH image atmospheric correction method based on a radiation transmission model.
Step three: algorithm architecture and model construction
In recent years, the application of the artificial neural network algorithm is a big hotspot, has the advantages of strong fault tolerance, optimal solution calculation instead of accurate solution, self-organization and the like, and is a new star in the field of geophysical inversion [9] . The input layer, the hidden layer and the output layer are the basic components of BP neural network, and are all connected with each other, but the neurons of each layer are not connected with each other. The characteristics of the hidden layer determine the network performance, and the increase of the number of the hidden layers can improve the network performance, however, excessive number of the hidden layers can cause problems of overfitting, low network convergence rate and the like [10] . In view of the characteristics of the artificial neural network, the invention selects the artificial neural network model with four layers (including the two hidden layers) to construct the BP neural network inversion model. The establishment of the training sample set is a key link for inversion by using a neural network, and the inversion precision and the inversion efficiency are directly influenced. The method selects more sensitive independent variables, brings the wave band combination value and the actually measured suspended sediment concentration into the selected neural network model (set as 2 hidden layers and 1 output layer), and finally determines the specific parameter setting of the BP neural network after training. Setting the number of neurons of the hidden layer is very important, and too few neurons can cause insufficient network freedom and lack of learning space; too many numbers can result in over-training. The final specific parameters of the BP neural network can be obtained through continuous training and trying.
Step four: accuracy test and evaluation
The present invention finally determines the coefficient (R) 2 ) The three indices Root Mean Square Error (RMSE) and Mean Absolute Percent Error (MAPE) determine the best combination. R of optimal neural network model 2 0.9398 and 0.9198 for MAPE, 0.34 and 15.53 for RMSE, 0.35 and 0.16M for RMSE -1 The result shows that the performance of the algorithm can be used for inverting the suspended sediment concentration of the Hangzhou bay.
Step five: compared with other common model analysis
And evaluating the superiority of the BP neural network model in inversion of suspended sediment concentration of the turbid water body. And respectively selecting other machine learning algorithms commonly applied to inversion of the suspended sediment in the water body, such as a random forest model, a gradient lifting regression tree model and a support vector machine model, and comparing and evaluating inversion accuracy of the random forest model, the gradient lifting regression tree model and the support vector machine model.
Step six: inversion results and analysis
And applying the trained BP neural network model to a high-resolution No. 1 remote sensing image (8, 21 days in 2019) to estimate the suspended sediment concentration of the Bay water body in Hangzhou.
Wherein, the step of measuring in the field of step one is:
and recording information such as wind speed, wind direction, longitude and latitude and the like of each sampling point. And (5) refrigerating the water sample at low temperature, and conveying the water sample to a laboratory on the same day.
Determining total suspended matter concentration by drying, baking and weighing method (GB 11901-89 standard), oven drying Whatman GFF filter membrane at 105 deg.C, weighing, filtering water sample, oven drying at 105 deg.C, weighing again, and calculating to obtain total suspended matter concentration [11] 。
Surface water spectra were measured at each sampling location and the instrument used for spectral collection was an ASD fieldspec 4 field portable spectrometer. The water body spectrum measurement adopts a measurement method above the water surface, an instrument observation plane is arranged to form an included angle of 135 degrees with a solar incidence plane, and an instrument is arranged to form an included angle of 40 degrees with the normal direction of the water surface. And (4) replacing the water surface irradiation brightness formula and the water surface incident irradiance formula to obtain the water surface remote sensing reflectivity.
The data preprocessing step in the second step is as follows: and (4) adopting a FLAASH image atmosphere correction method based on a radiation transmission model. In ENVI software, a FLAASH Atmospheric Correction tool is used, image head file information is combined, relevant parameters are set according to the reflection and absorption characteristics of a wave spectrum curve of a water body, and Atmospheric calibration processing is carried out on an image.
The algorithm architecture and model construction in the third step means that: algorithm background based on neural network models. The method comprises the following steps of selecting a BP neural network model containing a double-layer hidden layer to realize inversion of suspended sediment concentration of the water body, selecting a most sensitive inversion factor and a sensitive independent variable in order to establish a specific model and determine specific parameter setting of the model, and performing the following steps:
1) Calculating values of two single wave bands of wave band 2 and wave band 3 of the remote sensing data corresponding to the measured points, and the addition, subtraction, multiplication and division combinations of the two single wave bands are b2, b3, b2/b3, b3/b2, b2 x b3, b2+ b3, b2-b3 and b3-b2 respectively, wherein the total two single wave bands and the 6 wave band combinations are combined.
2) And combining the 6 wave band combinations in pairs, and then respectively carrying out addition, subtraction, multiplication and division operations to obtain 60 wave band combinations in total, wherein the former 2 single wave band combinations and 6 wave band combinations are calculated to be 68 wave band and wave band combinations in total.
3) And (3) solving correlation coefficients of the values of the 68 wave bands and the wave band combinations and the actually measured suspended sediment concentration, finding out the wave band combinations with strong correlation, and taking the wave band combinations as independent variables of the prediction model.
4) And (3) bringing independent variables and actually measured suspended sediment concentration of the prediction model into a neural network model based on the following steps: randomly drawn from 40 sets of experimental data, 30 sets of data were used as training samples, and the remaining 10 sets of samples were model test samples. Setting the number of hidden layers in the BP neural network as two layers. The transfer function of the first hidden layer adopts tansig; the transfer function of the hidden layer of the second layer adopts logsig, the transfer function of the output layer is purelin, the training function adopts trailing, the learning function of the weight and the threshold adopts learngdm, and the performance function adopts mse.
5) Training an artificial neural network model, determining the optimal parameters of the model by continuously trying, and continuously adjusting the number of layer nodes in the training process by referring to the following formula:
l is the number of hidden layer nodes; m and n are the number of nodes of the input layer and the output layer respectively; α is a constant of 0 to 10. And finally determining the specific parameter setting of the BP neural network model.
The accuracy test and evaluation in the fourth step are as follows: by determining the coefficient (R) 2 ) The three indices Root Mean Square Error (RMSE) and Mean Absolute Percent Error (MAPE) determine the best combination. In order to ensure that the model is not influenced by factors such as data spatiality, a training sample set and a testing sample set are randomly selected. R of optimal neural network model 2 0.9398 and 0.9198 for MAPE, 0.34% and 15.53% for RMSE, 0.35 and 0.16M for RMSE -1 . The result shows that the performance of the algorithm can be used for inverting the suspended sediment concentration of the Hangzhou gulf.
The analysis and comparison with other common models in the step five means that: and aiming at the 10 verification samples, respectively selecting other machine learning algorithms commonly applied to inversion of the suspended sediment concentration of the water body, such as a random forest model, a gradient lifting regression tree model and a support vector machine model, and comparing and evaluating inversion accuracy of different inversion models.
And sixthly, applying the trained BP neural network model to the GF-1 image (8, 21/8 in 2019) to estimate the suspended sediment concentration of the Hangzhou bay water body.
The invention further provides an application for inverting the suspended sediment in the water body.
Compared with the prior art, the invention has the beneficial effects that: by adopting the technical proposal, the utility model has the advantages that,
1. the BPNN model can be verified using GF-1 images synchronized with field measurements using domestic GF-1 multi-spectral camera image data of GF-1.
2. The hidden layer of the BPNN model is set into two layers, so that the inversion accuracy and the fitting degree of the suspended sediment of the turbid water body are improved. Compared with other machine learning algorithms (such as a random forest model, a gradient lifting regression tree model and a support vector machine model), the BPNN neural network model constructed by the method has a better inversion effect on the suspended sediment concentration of the turbid water body.
One of the important indexes for evaluating the water quality of the water body is a suspended sediment concentration value, a neural network model is used based on GF-1 satellite remote sensing images, a double-layer hidden layer arrangement is selected, a turbid water body suspended sediment concentration inversion model is constructed, the time-space characteristics of the turbid water body suspended sediment concentration inversion model are analyzed, and the turbid water body suspended sediment concentration inversion model is compared with the accuracy of other conventional inversion methods. Research results show that the BP neural network model can be used for inverting the suspended sediment concentration value of the turbid water body, and compared with other conventional methods, the model has higher correlation (R) 2 ＝0.9398)。
Drawings
FIG. 1 is a schematic diagram of the distribution of research zones and sampling points;
FIG. 2 is a graph of the remote reflectance of water at each sample point;
FIG. 3 is a schematic diagram of the BPNN model construction;
FIG. 4 is a graph of the remote reflectance for each band;
FIG. 5 is a model validation result;
fig. 6 is the inversion result of the suspended sediment concentration of the water body in the gulf of hangzhou.
Detailed Description
Now, the example of inverting the suspended sediment in the water body of the Bay in Hangzhou province based on the artificial neural network and the high-resolution No. 1 satellite is taken as a specific explanation. The artificial neural network is an operational model and is formed by connecting a large number of nodes. Each node represents a particular output function, called the excitation function. Every connection between two nodes represents a weighted value, called weight, for the signal passing through the connection, which is equivalent to the memory of the artificial neural network. The high-resolution No. 1 satellite is a high-resolution No. 1 satellite in China, and the latest data of the high-resolution No. 1 satellite is 8 meters of multispectral and 2 meters of panchromatic.
As shown in fig. 1 to 5, a method for inverting suspended sediment concentration in a gulf of hangzhou based on an artificial neural network and a high-resolution No. 1 satellite includes the following steps:
the method comprises the following steps: in situ measurement
Field sampling was performed in the gulf of hangzhou on days 8, 5, 2013 and 21, 2013, for a total of 40 samples collected, as shown in fig. 1. And recording the information (wind speed, wind direction, longitude and latitude and the like) of each sampling point. And (5) refrigerating the water sample at a low temperature, and conveying the water sample to a laboratory on the same day. Determining total suspended matter concentration by conventional drying, baking and weighing method (GB 11901-89 standard), oven drying Whatman GFF filter membrane at 105 deg.C, weighing at 105 deg.C, oven drying again, filtering to obtain water sample with certain volume, and calculating to obtain total suspended matter concentration [10] 。
Surface water spectra were measured at each sampling location. Spectral acquisition an ASD fieldspec 4 field portable spectrometer was used; the water body spectrum is measured by a measuring method above the water surface, the included angle between the observation plane of the instrument and the solar incident plane is set to be 135 degrees, and the included angle between the instrument and the normal direction of the water surface is 40 degrees.
And (3) substituting the water leaving irradiation brightness formula and the water surface incident irradiance formula to obtain a water surface remote sensing reflectivity formula:
Step two: data pre-processing
The main data of the invention is domestic WFV multispectral camera image data of GF-1, and a FLAASH image atmospheric correction method is adopted. In ENVI software, a FLAASH Atmospheric Correction tool is used, image head file information is combined, relevant parameters are set according to the reflection and absorption characteristics of a wave spectrum curve of a water body, and the images are subjected to Atmospheric calibration processing.
Step three: algorithm architecture and model construction
The invention selects a BP neural network model containing a double-layer hidden layer to realize inversion of suspended sediment concentration. The BP neural network model enables a model output value and an expected output value to have the minimum mean square error through a gradient descent method and a gradient search technology. The input layer, the hidden layer and the output layer are basic components of a neural network, each layer is composed of a certain number of neurons, the method is similar to a human nerve cell correlation mode, the input layer transmits external input information to the hidden layer, the hidden layer processes the information, meanwhile, a learning training result is transmitted to the output layer, and therefore the output result is processed [12] . The output of each node is the result of a weighted summation of the inputs delivered by a transfer function, usually an S-type hyperbolic tangent function (hidden layer nodes) or a linear function (output layer nodes), the connections between the nodes being weights, the input layer distributing the input signal only into the network without processing it, the association between the input and output ultimately depending on the weight values associated with each connection [13] . The basic structure of the artificial neural network model used in the present invention is shown in fig. 3. In order to establish a specific model and determine specific parameter setting of the BPNN model, the invention selects the most sensitive inversion factor, selects more sensitive independent variables: (1) calculating values of two single wave bands (wave band 2 and wave band 3) of the remote sensing data corresponding to the measured points, and addition, subtraction, multiplication and division combinations thereof, wherein the values are respectively b2, b3, b2/b3, b3/b2, b2 x b3, b2+ b3, b2-b3 and b3-b2, and the total number of the two single wave bands and the 6 wave band combinations are respectively calculated; (2) to pairEvery two phase groups of 6 wave band combinations are subjected to addition, subtraction, multiplication and division operation respectively to obtain 60 wave band combinations in total, the previous wave band combinations are added to obtain 68 wave band and wave band combinations in total, and the correlation coefficients of the values of the wave bands and the wave band combinations and the actually measured suspended sediment concentration are obtained as shown in table 1:
TABLE 1 correlation analysis of 68 band combinations
(3) The following band combinations are removed: selecting a band combination with strong correlation from invalid band combinations, repeating band combinations and low correlation coefficient values, and finally selecting a band combination with 35 reflectivity as the inversion factor shown in table 2:
table 2 selected 35 band combinations and their correlation coefficients
(4) Randomly drawn from 40 experimental data sets, wherein 30 data sets are used as training samples, and the remaining 10 sample points are model test samples. Setting the number of hidden layers of the BP neural network into two layers, wherein the transfer function of the hidden layer of the first layer adopts tansig; the transfer function of the hidden layer of the second layer is logsig, and the transfer function of the output layer is purelin; the training function adopts trailing, the learning function of the weight and the threshold adopts leanngdm, and the performance function adopts mse. The number of layer nodes is continuously adjusted during network training with reference to the formula L = √ (m + n + α), (L is the number of hidden layer nodes; m and n are the number of input and output layer nodes, respectively; α is a constant from 0 to 10). Through repeated parameter adjustment tests, when the number of the nodes of the first layer hidden layer is 35, the number of the nodes of the second layer hidden layer is 18, the allowed maximum training times is 1000, the learning rate is 0.05, and the minimum mean square error is le -6 The network converges, as shown in table 3:
TABLE 3 BP neural network parameter settings
Step four: accuracy testing and evaluation
The present invention is based on determining a coefficient (R) 2 ) The average absolute percentage error (MAPE) and the Root Mean Square Error (RMSE) determine the optimal combination, as shown in formulas (7) and (8):
before training the neural network model, normalization processing of input data is carried out, i.e. all data are converted into [ -1,1]The value of the 35 wave band combinations is input data, and the suspended silt degree is output data. And ensuring that the model is not influenced by factors such as data spatiality and the like, and randomly extracting a training sample set and a testing sample set. Fig. 5 is a result of Matlab modeling output of the BP neural network, which is a comparison between a model predicted value and an actual measurement value of a training set and a test set, respectively. R of optimal neural network model 2 0.9398 and 0.9198 for MAPE, 0.34% and 15.53% for RMSE, 0.35 and 0.16M for RMSE -1 . The result shows that the performance of the algorithm can be used for inverting the suspended sediment concentration of the Hangzhou gulf.
Step five: compared with other common model analysis
Based on GF-1 satellite remote sensing images, aiming at 10 verification samples, other machine learning algorithms commonly applied to inversion of the concentration of suspended sediment in the water body are respectively selected, such as a random forest model, a gradient lifting regression tree model and a support vector machine model, and the inversion accuracy of the concentration of suspended sediment in the turbid water body in the Hangzhou gulf is compared and evaluated by different inversion models. Table 4 shows the results of the statistical analysis. The results show that support vector regression is applied to the training data set (R) 2 =0.44, mape = 64.5%) and a validation data set (R) 2 =0.27,mape =68.3%) were underperforming, while gradient lifting regression yielded overfitting (R) to the validation dataset 2 =0.02, mape = 86.3%), R of artificial neural network for validation dataset compared to random forest regression 2 The increase was 19% and the decrease in MAPE was 26.17%, as shown in table 4:
TABLE 4 comparison of four exemplary machine learning algorithms
Therefore, compared with other common inversion models, the BPNN neural network model constructed by the invention has better effect on inverting turbid water bodies with higher suspended sediment concentration.
Step six: inversion results and analysis
The trained BP neural network model was applied to the GF-1 image of 21/8/2019 to estimate the suspended sediment concentration in the gulf of hangzhou, as shown in fig. 5.
According to effective pixel estimation, the average suspended sediment concentration of the Hangzhou bay water body is 679mg/L, and inversion results show that the suspended sediment concentration of the Hangzhou bay water body is generally higher, the internal difference is obvious, and the middle area is highest, and gradually decreases towards the gulf mouth and the open sea respectively. A high sand-containing area appears near the mouth of the south-south China, and gradually decreases towards the southwest, and the high sand-containing area reflects the influence of sediment in the sea of the Yangtze river on the Bay of Hangzhou. The cross section from Jinshan mountain in Hangzhou Bay to Fenggu has high suspended sediment concentration in south and low sediment concentration in north, and the great area of the east shoal is higher, and the deep groove in north is a low concentration area. Because the tidal range is great, the trend effect gathers to the gulf top for this regional suspended sediment concentration is higher. The deep groove near the gold mountain north bank has less influence of resuspension due to the deeper water body, so that the sediment content is lower, and a low-concentration sand-containing area is formed.
References to the invention:
[1] liu obviously, plum copper base, chen Qing Lian, influence of main components of water body on apparent optical quantity [ J ] ocean technique, 2004,23 (1): 85-89.
[2] Plum, huangjia group, weyuchun, etc. the remote sensing evaluation of the ground hyperspectral spectrum in the eutrophication state of lakes [ J ] environmental science 2006,27 (9): 1170-1175.
[3] The international research progress of the remote sensing water inversion algorithm of the ocean water satellite with Hierochlan, yangwei and ocean water satellite is J. Ocean notification, 2002, 21 (2): 77-83.
[4] The remote sensing mode research on the suspended sediment concentration of the surface layers of second class water bodies near shore [ J ] is developed in Liu Shi, zhou Yun Xuan, jiang Xue and so on [ 2006, 21 (1): 321-326 ].
[5] Satellite remote sensing monitoring of coastal water color environment of China [ J ]. Forth study, 2000, 20 (3): 240-246.
[6]Li Wenkai,Zeng Qun,Tian Liqiao,et al.Eistimation of total suspended matter concentration from HJ-1B CCD1imagery with the assistance of MODIS imagery[J].Journal of Central China Norrmal University(Natural Sciences),2016,50(04):619-623.
[7]Yin Yuwei,Tang Danling,Liu Yupeng.Remote Sensing of spatial-temporal distribution of suspended sedimentsurrounding islands and reefs in the South China Sea[J.Remote Sensing Technology and Application,2019,34(02):435-444.
[8] Chengqian, liubo, litting and Zhuli are constructed and applied based on a high-grade No. 1 Hangzhou Bay estuary suspended sediment concentration remote sensing inversion model [ J ]. Ocean environmental science 2015,34 (04): 558-563+577.
[9] Queen Yiming, song Xian Hai, zhang Qiang, seismic surface wave nonlinear inversion [ J ] using artificial neural network algorithm for geophysical prospecting of petroleum 2021,56 (05): 979-991+924.
[10] LIAOJU, zhangjian, penqingyuan drilling internal thermal resistance calculation model based on artificial neural network [ J ]. Chongqing university of technology (Nature science), 2021,35 (09): 256-261.
[11] The neural network method is used for inverting the analysis of the chlorophyll concentration of seawater [ J ] in the field of terrestrial information science 2004,6 (2): 31-36.
[12] Liu Da Zhen, fu Dong Yang, shenchun Yan, etc. remote sensing research on suspended silt in estuary and near bank water is developed [ J ] ocean environmental science 2010,29 (4): 611-616.
[13] The research on the spatial-temporal dynamic law of suspended silt based on remote sensing inversion results, for example, zhujiang estuary and adjacent sea area [ J ] information science edition, 2005,30 (8): 677-681.
Claims (10)
1. A method for inverting suspended sediment in a water body based on an artificial neural network and a high-resolution No. 1 satellite is characterized by comprising the following steps: the method comprises the following steps:
the method comprises the following steps: in-situ measurement
Sampling the water body in the field to obtain the water surface remote sensing reflectivity;
step two: data pre-processing
Preprocessing the collected original high-resolution No. 1 satellite remote sensing image;
step three: algorithm architecture and model construction
Constructing a BPNN model based on an artificial neural network algorithm background; the algorithm framework and the model are constructed, and a BP neural network model containing a double-layer hidden layer is selected to realize inversion of suspended sediment concentration in a research area; constructing a specific model and determining specific parameter setting of a BPNN model by selecting the most sensitive inversion factor;
step four: accuracy testing and evaluation
The optimal combination is determined by three indexes of the decision coefficient, the root mean square error and the average absolute percentage error.
2. The method of claim 1, wherein the body of water is inland and near shore class ii bodies of water.
3. The method of claim 1, wherein the step of measuring in the field of step one comprises: recording longitude and latitude, wind speed and wind direction information of each sampling point during sampling; refrigerating the water sample at low temperature, and sending the water sample to a laboratory on the same day; measuring the concentration of total suspended matters; measuring a surface water spectrum at each sampling location; and (4) replacing the water irradiation brightness formula and the remote sensing reflectivity formula to obtain the water surface remote sensing reflectivity, and calculating the reflectivity of each measuring point to obtain a reflectivity curve graph.
4. The method as claimed in claim 1, wherein the data preprocessing of step two is performed by FLAASH image atmosphere correction method based on radiation transmission model; in ENVI software, a FLAASH Atmospheric Correction tool is used for setting relevant parameters according to reflection and absorption characteristics of spectral curves of vegetation and water bodies and by combining image header file information, so that Atmospheric calibration processing of images is realized.
5. The method according to claim 1, wherein the step of constructing the algorithm architecture and model in the step three is:
1) Spectral reflectivity data of a waveband 2 and a waveband 3 are mainly acquired to eliminate adverse effects of an insensitive waveband on modeling;
2) Calculating two single-waveband values of a waveband 2 and a waveband 3 of the remote sensing data corresponding to the measured point, and addition, subtraction, multiplication and division combinations of the two single-waveband values, wherein the two single-waveband values are b2, b3, b2/b3, b3/b2, b2 x b3, b2+ b3, b2-b3 and b3-b2, and the two single-waveband values and the 6 waveband combinations are combined;
3) Carrying out pairwise combination addition, subtraction, multiplication and division operation on the 6 wave band combinations to obtain 60 wave band combinations in total, and adding the previous 2 wave bands and 6 wave band combinations to obtain 68 wave band and wave band combinations in total;
4) Calculating correlation coefficients of the values of the 68 wave bands and the wave band combinations and actually measured suspended sediment concentration, finding out wave band combinations with strong correlation, taking the selected wave band combinations with strong correlation as independent variables of a prediction model, and eliminating invalid wave band combinations, repeated wave band combinations and wave band combinations with low correlation coefficient values;
5) Randomly taking 30 groups of data from 40 groups of experimental data as training samples, and taking the remaining 10 groups of sampling points as model test samples; setting the BP neural network into 2 hidden layers and 1 output layer; the transfer function of the first hidden layer adopts tansig; the transfer function of the second layer hidden layer is logsig, the transfer function of the output layer is purelin, the training function adopts trailing, the learning function of the weight and the threshold adopts leargdm, and the performance function adopts mse;
6) Before training the model by the neural network, normalizing the input data, and converting all data into the range of [ -1,1 ]; the input data is a value of 35 wave band combinations, and the output data is suspended sediment density;
7) The number of layer nodes is continuously adjusted in the network training process by referring to the following formula:
l is the number of hidden layer nodes; m and n are the number of nodes of the input layer and the output layer respectively; α is a constant from 0 to 10; through continuous parameter adjustment tests, when the number of the nodes of the first hidden layer is set to be 35, the number of the nodes of the second hidden layer is set to be 18, the allowed maximum training times is 1000, the learning rate is 0.05, and the minimum mean square error is le -6 When, the network converges; and finally determining specific parameters of the neural network.
6. The method of claim 1, further comprising:
step five: compared with other common model analysis
Carrying out comparative evaluation on inversion accuracy of suspended sediment concentration of the turbid water body of the Hangzhou bay on different inversion models;
step six: inversion results and analysis
And applying the trained BP neural network model to the high-resolution No. 1 remote sensing image to obtain an inversion result.
7. The method as claimed in claim 6, wherein the analysis and comparison with other common models in the fifth step are based on top of the top-grade first satellite remote sensing image, and other machine learning algorithms commonly applied to inversion of the suspended sediment concentration of the water body, such as a random forest model, a gradient lifting regression tree model and a support vector machine model, are respectively selected for 10 verification samples, so as to compare and evaluate the inversion accuracy of the suspended sediment concentration of the turbid water body of the Bay of Hangzhou province for different inversion models; and sixthly, applying the trained BP neural network model to the GF-1 image to estimate the suspended sediment concentration of the Bay water body in Hangzhou.
8. Use of the method of claim 1 for inverting sediment in suspension in a body of water.
9. Use according to claim 8, characterized in that: the water body is inland and near shore class II water body.
10. Use according to claim 8, characterized in that: the water body is Hangzhou bay water body.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210392343.1A CN115711838A (en) | 2022-04-13 | 2022-04-13 | Method for inverting suspended sediment in water body based on artificial neural network and high-resolution No. 1 satellite and application of method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210392343.1A CN115711838A (en) | 2022-04-13 | 2022-04-13 | Method for inverting suspended sediment in water body based on artificial neural network and high-resolution No. 1 satellite and application of method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115711838A true CN115711838A (en) | 2023-02-24 |
Family
ID=85230491
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210392343.1A Withdrawn CN115711838A (en) | 2022-04-13 | 2022-04-13 | Method for inverting suspended sediment in water body based on artificial neural network and high-resolution No. 1 satellite and application of method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115711838A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116415508A (en) * | 2023-06-12 | 2023-07-11 | 珠江水利委员会珠江水利科学研究院 | Method and system for generating two-dimensional sediment model of estuary |
-
2022
- 2022-04-13 CN CN202210392343.1A patent/CN115711838A/en not_active Withdrawn
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116415508A (en) * | 2023-06-12 | 2023-07-11 | 珠江水利委员会珠江水利科学研究院 | Method and system for generating two-dimensional sediment model of estuary |
| CN116415508B (en) * | 2023-06-12 | 2023-10-13 | 珠江水利委员会珠江水利科学研究院 | Method and system for generating two-dimensional sediment model of estuary |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112070234B (en) | Water chlorophyll and phycocyanin land-based remote sensing machine learning algorithm under complex scene | |
| CN112213287B (en) | Coastal beach salinity inversion method based on remote sensing satellite image | |
| Pozdnyakov et al. | Operational algorithm for the retrieval of water quality in the Great Lakes | |
| CN106767687A (en) | A kind of method of utilization remote sensing moisture measurement beach elevation | |
| Cao et al. | A deep learning method for cyanobacterial harmful algae blooms prediction in Taihu Lake, China | |
| CN111597756A (en) | Water quality parameter inversion method based on multispectral data of unmanned aerial vehicle | |
| CN111460681B (en) | Satellite remote sensing method for monitoring depth of real light layer of offshore water body | |
| CN114239422A (en) | Method for improving marine chlorophyll a concentration prediction accuracy based on machine learning | |
| CN114460013B (en) | Coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method | |
| CN113408644B (en) | Satellite data reconstruction method and method for detecting response of upper ocean to typhoon | |
| CN114297938B (en) | Inversion method of optical shallow water depth based on neural network | |
| CN114201732A (en) | Sentinel-2A image-based shallow sea water depth inversion method | |
| Hang et al. | Estimation of chlorophyll-a concentration in Lake Taihu from Gaofen-1 wide-field-of-view data through a machine learning trained algorithm | |
| Cui et al. | Deep learning–based remote sensing estimation of water transparency in shallow lakes by combining Landsat 8 and Sentinel 2 images | |
| CN112816421A (en) | Land-based remote sensing monitoring method for nutritive salt and chemical oxygen demand of water body | |
| Matsui et al. | Improving the resolution of UAV-based remote sensing data of water quality of Lake Hachiroko, Japan by neural networks | |
| CN115711838A (en) | Method for inverting suspended sediment in water body based on artificial neural network and high-resolution No. 1 satellite and application of method | |
| CN115561176A (en) | Water quality inversion method based on feature adaptive operation and machine learning fusion | |
| CN113763272B (en) | Remote sensing inversion method for photosynthetic effective radiation attenuation coefficient of eutrophic lake | |
| Li et al. | High-accuracy remote sensing water depth retrieval for coral islands and reefs based on LSTM neural network | |
| CN111723522A (en) | Method for calculating exchange flux of organic carbon dissolved in lake and river | |
| CN114965300B (en) | Lake turbidity drawing method for constructing BP-TURB based on optical water body type and BP neural network algorithm | |
| Wang et al. | Study on Remote Sensing of Water Depths Based on BP Artificial Neural Network. | |
| CN115359365B (en) | Suaeda salsa identification and coverage estimation method for coastal intertidal zone | |
| Liu et al. | Application of GIS and Remote Sensing in Spatial Distribution of Nitrogen and Phosphorus Pollutant in Urban Rivers: A Case Study of Linyi Economic Development Zone, China |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20230224 |