CN114119618A - Remote sensing extraction method for inland salt lake artemia strips based on deep learning - Google Patents
Remote sensing extraction method for inland salt lake artemia strips based on deep learning Download PDFInfo
- Publication number
- CN114119618A CN114119618A CN202111336862.8A CN202111336862A CN114119618A CN 114119618 A CN114119618 A CN 114119618A CN 202111336862 A CN202111336862 A CN 202111336862A CN 114119618 A CN114119618 A CN 114119618A
- Authority
- CN
- China
- Prior art keywords
- artemia
- data
- remote sensing
- water body
- deep learning
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Geometry (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a remote sensing extraction method of inland salt lake artemia strips based on deep learning. Firstly, preprocessing acquired remote sensing data to obtain surface reflectivity data, determining a water body range, preliminarily obtaining artemia-water body data, then selecting typical data, cutting and amplifying to generate a sample, establishing an artemia-water body data set, then constructing and training an artemia extraction deep learning model, evaluating the precision and robustness of the trained model, simulating abundant samples through data, further generalizing the application range of the model, and finally extracting artemia strips by using the generalized model. The method identifies and extracts the artemia strips in the salt lake by using a deep learning method, does not need to consider the problem of threshold, can basically realize automatic processing, can more accurately obtain the artemia strips, and has a business popularization prospect.
Description
Technical Field
The invention belongs to the technical field of deep learning semantic segmentation and remote sensing image processing, and particularly relates to a remote sensing extraction method for artemia salina strips in inland salt lake based on deep learning.
Background
Artemia (Artemia) is a small crustacean living in high salinity water and has very important economic and ecological values. The artemia contains abundant protein and fat and is easy to store and circulate, so that the artemia not only becomes a live bait feed widely used by the aquaculture industry at home and abroad, but also is an important component of carbon flux and biological chain in salt lake. According to statistics, the global annual output of the artemia is about 3000-4000 tons (dry quality of finished products), which relates to 25 hundred million yuan RMB, so the artemia is also called 'soft gold'. Artemia resources are widely distributed throughout the world, with a total of about 350 artemia producing areas. Wherein the artemia resource amount of the Eibei lake occupies the leaderboard of the inland salt lake in China, the annual output of the artemia cysts is about 200-400 tons, and the international market price is about 60 ten thousand yuan/ton. However, due to deterioration of the ecological environment of salt lakes and increasing demand for artemia in the global aquaculture industry, the quantity of artemia biomass is gradually shrinking.
At present, the statistics of artemia resources and yield are mostly based on-site sampling, and the scale is small; the wide distribution range and high salinity living environment of the artemia make the traditional on-site sampling method difficult to realize. In addition, the artemia mature period is short, the strips are easily affected by wind and lake salinity to generate seasonal variation, and the uncertainty of field investigation is large. In contrast, satellite remote sensing has the technical advantages of wide observation range, fast information acquisition, short period, low cost and the like, and becomes an important means for detecting floating objects on water, particularly algae, plankton, oil spill, floating garbage and the like. The artemia are poor in autonomous movement capability, strips with the length of thousands of meters and the width of dozens of meters are formed under the drive of wind in the growth and reproduction period of the artemia, pigments such as carotene and heme are contained in the artemia, and spectral characteristics which are obviously different from other water color parameters and floating objects in water particularly in short-wave infrared and near-infrared bands are achieved, so that the artemia can be detected through a remote sensing means. Meanwhile, deep learning is used as a novel algorithm which can interpret a large amount of data and has strong adaptability, extracted features can be made to be more robust, and the method is successfully applied to various remote sensing fields such as image classification, change detection, target identification and cloud shadow removal.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a remote sensing extraction method of artemia strips in inland salt lake based on deep learning. The method identifies and extracts the artemia strips in the salt lake by using a deep learning method, does not need to consider the problem of threshold, can basically realize automatic treatment, and can more accurately obtain the artemia strips.
In order to achieve the aim, the technical scheme provided by the invention is a remote sensing extraction method of artemia salina strips in inland salt lake based on deep learning, which comprises the following steps:
step 1, acquiring a multispectral remote sensing image of a typical salt lake, preprocessing the multispectral remote sensing image to obtain surface reflectivity data, determining a water body range, and preliminarily obtaining artemia-water body data;
step 1.1, acquiring a multispectral data source;
step 1.2, carrying out radiometric calibration, geometric correction and atmospheric correction on the remote sensing image data;
step 1.3, preprocessing a cloud and cloud shadow mask;
step 1.4, extracting in a water body range to obtain artemia-water body data preliminarily;
step 2, further selecting artemia-water body data from the preprocessed surface reflection data in the step 1, cutting and amplifying the data to generate a sample, and establishing a typical and comprehensive artemia-water body data set;
step 2.1, selecting typical and comprehensive artemia-water body data;
step 2.2, data cutting and sample augmentation;
step 3, constructing and training a artemia extraction deep learning model;
step 3.1, normalizing the artemia-water body sample data obtained in the step 2;
step 3.2, selecting a proper deep learning semantic segmentation model as an artemia extracting deep learning model;
step 3.3, configuring parameters and carrying out model training;
step 4, evaluating the precision and robustness of the artemia extraction deep learning model trained in the step 3;
step 4.1, performing precision evaluation on the trained model by using the precision rate, the recall rate and the F1 score;
4.2, carrying out robustness evaluation on the trained model;
step 5, enriching samples through data simulation, and further generalizing the artemia to extract a deep learning model;
and 6, extracting artemia strips from the multispectral remote sensing image preprocessed in the step 1 by using the artemia extraction deep learning model generalized in the step 5.
And moreover, selecting an image of the Landsat-8 OLI and HY-1C CZI multispectral lake with the artemia salina stripe from the multispectral sensor in the step 1.1, further selecting a remote sensing image with the 2019 adulthood 2021 year image time range of 4-11 months per year in consideration of the growth period and the icing condition of the artemia in the Eiby lake, and then screening the multispectral remote sensing image according to the area of the water area of the salina lake, the density, the length and the width of the artemia stripe and other water color parameter conditions of the salina.
Moreover, the Landsat-8 OLI Level-2 product data selected in the step 1.2 is earth surface reflectivity data which is subjected to radiation calibration, precise geometric correction and atmospheric correction, HY-1C CZI Level-1C is an atmospheric apparent reflectivity product which is subjected to radiation correction and geometric correction, and HY-1C CZI data atmospheric correction is carried out through a 6S atmospheric correction model or FLAASH atmospheric correction in ENVI on the basis of acquiring a CZI spectral response function and satellite parameters.
In step 1.3, the 3 rd position cloud, the 4 th position cloud shadow and the 6 th position rolling cloud of the Landsat-8 QA wave band are used for realizing the cloud mask coding of Landsat-8 OLI data, and the HY-1C CZI data is subjected to the cloud mask through the haze _ tool of ENVI, so that invalid data such as cloud, cloud shadow and the like are removed.
And in the step 1.4, firstly, a global surface water data set GSW is applied to carry out land mask, then a buffer area is arranged to be corroded inwards, the influence of water body-land mixed pixels is eliminated, the proximity effect of a lake bank is reduced, a water body range of a salt lake is obtained, then the mask is manufactured by utilizing the water body range of the salt lake, and non-water body pixels are removed, so that artemia-water body data are obtained.
In the step 2.1, artemia data are selected from the preprocessed ground surface reflection data in the step 1, the artemia gathering form of the selected data comprises a slender strip shape and a cluster shape, and the artemia density is displayed as a dark reddish brown to light red characteristic in a true color image from high to low; selecting a water body background of data considering different water colors due to different contents of chlorophyll, suspended matters and CDOM and influence of lake sediment; the selected data also comprises artemia-water body data under the influence of adverse conditions such as thin clouds, flares and the like.
And 2.2, cutting the primarily selected artemia-water body data into samples with the size of NxN by a sliding window method, and performing sample augmentation by using rotation and mirror image to expand the sample amount and meet the subsequent model training requirements.
In step 3.1, the data type is firstly converted from UInt16 to UInt8, the numerical range of each wave band is between 0 and 255, and then linear normalization is performed to map the data to a [0,1] interval, so that the order difference is eliminated, the neglect of small data is avoided, and the effects of unifying the sample evaluation indexes and accelerating the training convergence speed are achieved.
In the step 3.2, a full convolution U-shaped neural network is selected, the network takes Tensorflow as a frame, the depth of the network is selected to be F according to the complexity of an extraction target, normalized NxN artemia-water HY-1C CZI sample spectrum information is input, and the two segmentation results are output; wherein, the input HY-1C CZI sample comprises spectral information of blue band 0.42-0.50 μm, green band 0.52-0.60 μm, red band 0.61-0.69 μm, near infrared 0.76-0.89 μm, and four bands with central wavelengths of 0.460 μm, 0.560 μm, 0.650 μm and 0.825 μm in sequence.
Moreover, the basic structural unit of the full convolution U-shaped neural network in the step 3.3 is a convolution-batch normalization layer-ReLU activation function, the batch size is set according to the GPU condition, and the initial learning rate is set as d; in the iterative training process of the model, the maximum iterative round number is set as M, and an early-stopping mechanism is configured, namely, if the training precision does not rise continuously for 10 rounds or the error does not fall continuously for 10 rounds, the iteration is terminated early.
In step 4.1, precision evaluation is performed according to three precision evaluation indexes, namely, precision rate, recall rate and F1 score, that is:
in the formula, Precision is Precision; recall is Recall; F1-Score is F1 Score; TP is the number of positive samples of artemia in the two categories; FP is the number of false detection samples; FN is the number of missed samples, and is obtained by comparing the extraction result of the U-Net model with the reference result of manual interpretation.
And in the step 4.2, robustness evaluation is to select n samples with complicated background signal interference conditions of thin clouds, vortexes, shallow water, near-shore substrate, turbid water, glaring and dark water pixels in the artemia-water body data set, compare the extraction result of the U-Net model with the reference result of visual interpretation, and analyze the applicability and stability of the model in the application process.
In step 5, Landsat-8 OLI data is simulated as HY-1CCZI data by the spectrum matching factor, and the formula for performing spectrum adjustment on the OLI data by the spectrum matching factor is as follows:
in the formula, ρSR(lambda) corresponds to artemia or waterThe specific spectrum of the body, the spectrum of the artemia is acquired by an SVC HR-1024 spectrometer through a water tank control experiment; lambda [ alpha ]0、SRFλRespectively, the center wavelength and the spectral response function.
And resampling the OLI sample after spectrum adjustment to be consistent with the CZI spatial resolution, realizing data simulation from OLI to CZI, enriching the time sequence range of the sample, and finally inputting the simulated data serving as a fresh sample into a trained U-Net model, so that the adaptability of the algorithm to multi-platform multi-source samples is enhanced, and the application range of the model is further expanded and generalized.
Compared with the prior art, the invention has the following advantages:
1) based on artemia visible light, near infrared spectrum information and a multispectral remote sensing data source, remote sensing extraction of the artemia strips in the inland salt lake is achieved by applying a full-convolution U-shaped neural network (U-Net) deep learning model, and the method is high in precision and strong in reliability.
2) Under the conditions of complex salt lake water body background signal interference and different observation conditions, the method provided by the invention has certain robustness and generalization capability, can basically realize automation, and is suitable for wide inland salt lake artemia strip extraction.
3) The method has the potential of space-time distribution analysis of artemia resources and multi-source multi-platform remote sensing large data processing, and can provide practical basis for reasonable management and ecological protection of the artemia resources.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention.
Detailed Description
The invention provides a remote sensing extraction method of artemia salina strips in inland salt lake based on deep learning, which comprises the steps of preprocessing obtained remote sensing data to obtain surface reflectivity data, determining a water body range, preliminarily obtaining artemia-water body data, selecting typical data, generating a sample through cutting and expanding, establishing an artemia-water body data set, constructing and training an artemia extraction deep learning model, carrying out precision and robustness evaluation on the trained model, simulating abundant samples through data, further generalizing the application range of the model, and finally extracting artemia strips by using the generalized model.
The technical scheme of the invention is further explained by selecting a typical inland salt lake Xinjiang Aibi lake with artemia as a research area. The area of the Aibi lake is about 650km2Average water depth 1.4m, lake elevation 189m, average annual wind speed 165 days>8m/s, the artemia resource amount occupies the leaderboard of inland salt lake in China. As shown in fig. 1, the process of the embodiment of the present invention includes the following steps:
step 1, acquiring a multispectral remote sensing image of a typical salt lake, preprocessing the multispectral remote sensing image to obtain surface reflectivity data, determining a water body range, and primarily obtaining artemia-water body data.
Step 1.1, acquiring a multispectral data source.
Firstly, selecting an image of a salt lake with artemia strips on Landsat-8 OLI and HY-1C CZI multispectral sensors, further selecting a remote sensing image with the image time range of 2019 plus 2021 year in 4-11 months per year in consideration of the growth period and the icing condition of artemia in the Eiby lake, and then screening the remote sensing multispectral image according to the area of a water area of the salt lake, the density, the length and the width of the artemia strips and other water color parameter conditions of the salt lake.
Step 1.2, carrying out radiometric calibration, geometric correction and atmospheric correction on the remote sensing image data.
The Landsat-8 OLI Level-2 product data used in this example is the surface reflectance data that has been processed by radiometric calibration, precision geometry correction, and atmospheric correction, and the 21 selected scene image data originates from the usgs official website (https:// earth scanner. usgs. gov /). The selected HY-1C CZI Level-1C data meeting the requirements of 10 scenes is from a Chinese marine satellite data service system (https:// osdds. nsoas. org. cn /), is an atmospheric apparent reflectivity product which is subjected to radiation correction and geometric correction, and HY-1C CZI data atmospheric correction can be carried out through a 6S atmospheric correction model or FLAASH atmospheric correction in ENVI on the basis of obtaining a CZI spectral response function and satellite parameters.
And 1.3, preprocessing the cloud and cloud shadow masks.
In the embodiment, the 3 rd bit (cloud), the 4 th bit (cloud shadow) and the 6 th bit (rolling cloud) of the Landsat-8 QA waveband are used for realizing the cloud mask coding of Landsat-8 OLI data, and the HY-1C CZI data is subjected to the cloud mask through the haze _ tool of ENVI, so that invalid data such as cloud and cloud shadow are removed. The quality image naming rules for OLI are shown in table 1:
TABLE 1 Landsat-8 OLI quality image naming
And step 1.4, extracting in a water body range, and primarily obtaining artemia-water body data.
The salt lake artemia stripe near-infrared band spectral reflectivity is high, the characteristics are similar to those of land, water body extraction by using a conventional water body index method can misjudge artemia pixels as land, so that part of the artemia pixels are removed from the water body, and the establishment of a subsequent artemia-water body data set and the training of a deep learning model are influenced. This example uses the global surface water (GSW, https:// www.nature.com/articules/nature 20584) dataset to accomplish salt lake water body range extraction. Supposing that the pixels covered by the water body all the year round are not water puddles formed by seasonal shallow water or precipitation, extracting the pixels with the incidence rate of 87% -93% according to the probability of the pixels of the water body provided by the GSW to obtain a primary water body boundary, then setting an inward corrosion buffer zone of 300m, eliminating the influence of mixed pixels of the water body and the land, reducing the proximity effect of the lake bank and obtaining the water body range of the salt lake. And (4) making a mask by utilizing the water body range of the salt lake, and removing the non-water body pixels to obtain artemia-water body data.
And 2, further selecting artemia-water body data from the preprocessed surface reflection data in the step 1, cutting and amplifying to generate a sample, and establishing a typical and comprehensive artemia-water body data set.
And 2.1, selecting typical and comprehensive artemia data.
Selecting artemia data from the preprocessed ground surface reflection data in the step 1, wherein the artemia aggregation form of the selected data comprises a slender strip shape and a cluster shape, and the artemia density is from high to low, namely the artemia is displayed as a dark reddish brown to light red characteristic in a true color image. The water background of the selected data needs to consider different water colors due to different contents of chlorophyll, suspended matters, CDOM and the like and the influence of lake sediment. The selected data also comprises artemia-water body data under the influence of adverse conditions such as thin clouds, flares and the like.
And 2.2, data cutting and sample augmentation.
The artemia-water body data selected in the step 2.1 are cut into samples with the size of 64 multiplied by 64 by a sliding window method with 40% of coincidence rate, the samples are expanded by means of rotation, mirror image and the like, the sample size is expanded, and the requirement of subsequent model training is met.
And 3, constructing and training a artemia extraction deep learning model.
Selecting a proper deep learning semantic segmentation model, taking spectral information of the artemia-water body sample as the input of a network, performing iterative training, and automatically extracting artemia strips.
And 3.1, normalizing the artemia-water body sample data obtained in the step 2.
In order to ensure that the spectral information contained in the sample can be correctly received by the deep learning network, firstly, the data type of the sample is converted into UInt8 from UInt16, namely, the numerical range of each wave band is between 0 and 255, and then, the data is mapped to a [0,1] interval through linear normalization so as to eliminate the order difference and avoid overlooking the undersized data; the method can also play a role in unifying the sample evaluation indexes and accelerating the training convergence speed.
And 3.2, selecting a proper deep learning semantic segmentation model as the artemia extracting deep learning model.
In this embodiment, a full convolution U-network (U-Net) is selected, the network uses the tensrflow as a frame, the depth of the network is selected to be 5 according to the complexity of an extraction target, normalized spectrum information of a 64 × 64 artemia-water body HY-1C CZI sample is input, and the result of the two segmentation is output. Wherein, the input HY-1C CZI sample comprises spectral information of blue band 0.42-0.50 μm, green band 0.52-0.60 μm, red band 0.61-0.69 μm, near infrared 0.76-0.89 μm, and four bands with central wavelengths of 0.460 μm, 0.560 μm, 0.650 μm and 0.825 μm in sequence.
And 3.3, configuring parameters and carrying out model training.
The basic structural unit of the U-Net network is '3 multiplied by 3 convolution-Batch Normalization (BN) layer-ReLU activation function', the batch size is set according to the GPU condition, the initial learning rate is set to be 1 multiplied by e-4. In the iterative training process of the model, the maximum iterative round number is set to 1000, and an early-stopping mechanism is configured, namely, if the training precision does not rise continuously for 10 rounds or the error does not fall continuously for 10 rounds, the iteration is terminated early.
And 4, evaluating the precision and robustness of the artemia extraction deep learning model trained in the step 3.
And 4.1, evaluating the precision of the trained model by using the precision rate, the recall rate and the F1 score.
The normalized 64 x 64 artemia-water body sample data is used for model training, and in order to verify the applicability of the model trained in the step 3 to the remote sensing images with the same sensor type and larger size, the HY-1C CZI image preprocessed in the step 1 and with complex water body conditions in the same area of the Eibei lake in the two scenes of 2019, 4 and 28 days and 2019, 5 and 1 day is used for evaluating the precision of the trained model. Precision evaluation is carried out by using three Precision evaluation indexes of Precision (Precision), Recall (Recall) and F1 Score (F1-Score), namely:
wherein TP is the number of positive samples of artemia in the two categories; FP is the number of false detection samples; FN is the number of missed samples, and is obtained by comparing the extraction result of the U-Net model with the reference result of manual interpretation.
The positive detection number of the U-Net algorithm to the two-scene HY-1C CZI image reaches 1337, 163 missing detection pixels and 116 false detection pixels; the accuracy rate, the recall rate and the F1 score of the algorithm obtained by calculation are 92.02%, 89.13% and 90.55%, the overall accuracy index is close to 90%, and the U-Net algorithm can accurately extract the artemia strips.
And 4.2, carrying out robustness evaluation on the trained model.
Selecting 50 samples with complex background signal interference conditions such as thin clouds, vortexes, shallow water, near-shore bottom materials, turbid water, glaring and dark water pixels in the artemia-water body data set, comparing the extraction result of the U-Net model with the reference result of visual interpretation, and analyzing the applicability and stability of the model in the application process. In general, the U-Net algorithm has certain impedance to cloud, flare and turbid water body reflection, is sensitive to artemia pixel reaction, is widely suitable for artemia extraction under the interference of complex water body background signals, and has high robustness.
And 5, enriching samples through data simulation, and further generalizing the application range of the artemia extraction deep learning model.
When the model is trained, only HY-1C CZI data is used, and different sensor data can be mixed after spectral response conversion. And simulating Landsat-8 OLI data into HY-1C CZI data by using a spectrum matching factor (SABF), wherein the HY-1C CZI data is used as a fresh sample input into a trained U-Net model, the adaptability of the algorithm to multi-platform multi-source samples is enhanced, and the application range of the generalized model is expanded.
For a U-Net deep learning model obtained by training HY-1C CZI data, visible light and near infrared bands shared by preprocessed Landsat-8 OLI sample data and the CZI sample data can be selected, wherein the visible light and near infrared bands comprise a blue band of 0.45-0.51 μm, a green band of 0.53-0.59 μm, a red band of 0.64-0.67 μm and a near infrared band of 0.85-0.88 μm, the central wavelengths are 0.483 μm, 0.563 μm, 0.655 μm and 0.865 μm in sequence, and OLI data are subjected to spectrum adjustment through a spectrum matching factor (SABF), namely:
in the formula, ρSR(lambda) corresponds to the specific spectrum of the artemia or the water body, and the spectrum of the artemia is acquired by an SVC HR-1024 spectrometer through a water tank control experiment; lambda [ alpha ]0、SRFλRespectively, the center wavelength and the spectral response function.
And resampling the OLI sample after spectrum adjustment to 50m resolution, realizing data simulation from OLI to CZI, enriching the time sequence range of the sample, inputting the simulated data serving as a fresh sample into a trained U-Net model, enhancing the adaptability of the algorithm to multi-platform multi-source samples, and further expanding and generalizing the application range of the model.
And 6, extracting artemia strips from the multispectral remote sensing image preprocessed in the step 1 by using the artemia extraction deep learning model generalized in the step 5.
In specific implementation, the above process can adopt computer software technology to realize automatic operation process.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.
Claims (10)
1. A remote sensing extraction method for artemia salina strips in inland salt lake based on deep learning is characterized by comprising the following steps:
step 1, acquiring a multispectral remote sensing image of a typical salt lake, preprocessing the multispectral remote sensing image to obtain surface reflectivity data, determining a water body range, and preliminarily obtaining artemia-water body data;
step 2, further selecting artemia-water body data from the preprocessed surface reflection data in the step 1, cutting and amplifying the data to generate a sample, and establishing a typical and comprehensive artemia-water body data set;
step 3, constructing and training a artemia extraction deep learning model;
step 4, evaluating the precision and robustness of the artemia extraction deep learning model trained in the step 3;
step 4.1, performing precision evaluation on the trained model by using the precision rate, the recall rate and the F1 score;
4.2, carrying out robustness evaluation on the trained model;
step 5, enriching samples through data simulation, and further generalizing the application range of the artemia extraction deep learning model;
and 6, extracting artemia strips from the multispectral remote sensing image preprocessed in the step 1 by using the artemia extraction deep learning model generalized in the step 5.
2. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 1, wherein the remote sensing extraction method comprises the following steps: the step 1 comprises the following steps:
step 1.1, acquiring a multispectral data source;
step 1.2, carrying out radiometric calibration, geometric correction and atmospheric correction on the remote sensing image data;
step 1.3, preprocessing a cloud and cloud shadow mask;
step 1.4, extracting in a water body range to obtain artemia-water body data preliminarily;
firstly, a global surface water data set GSW is used for land masking, then a buffer area is set to be corroded inwards, the influence of water-land mixed pixels is eliminated, the proximity effect of a lake bank is reduced, the water body range of a salt lake is obtained, then a mask is manufactured by utilizing the water body range of the salt lake, and non-water body pixels are removed, so that artemia-water body data are obtained.
3. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 2, wherein the remote sensing extraction method comprises the following steps: in the step 1.1, selecting an image of a salt lake with artemia strips on Landsat-8 OLI and HY-1C CZI multispectral sensors, further selecting a remote sensing image with the image time range of 2019-plus 2021 year in 4-11 months per year according to the growth period and the icing condition of the artemia in the Eiby lake, and then screening the multispectral remote sensing image according to the water area of the salt lake, the density, the length and the width of the artemia strips and other water color parameter conditions of the salt lake.
4. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 2, wherein the remote sensing extraction method comprises the following steps: the Landsat-8 OLI Level-2 product data selected in the step 1.2 is earth surface reflectivity data which is subjected to radiation calibration, accurate geometric correction and atmospheric correction, HY-1C CZI Level-1C is an atmospheric apparent reflectivity product which is subjected to radiation correction and geometric correction, and HY-1C CZI data atmospheric correction is carried out through a 6S atmospheric correction model or FLAASH atmospheric correction in ENVI on the basis of obtaining a CZI spectral response function and satellite parameters.
5. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 2, wherein the remote sensing extraction method comprises the following steps: step 1.3, the 3 rd position cloud, the 4 th position cloud shadow and the 6 th position rolling cloud of the Landsat-8 QA wave band are used for realizing the cloud mask coding of Landsat-8 OLI data, and the cloud mask of HY-1C CZI data is realized through a haze _ tool of ENVI, so that the invalid data of cloud and cloud shadow are removed.
6. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 1, wherein the remote sensing extraction method comprises the following steps: the establishing of the artemia-water body data set in the step 2 comprises the following steps:
step 2.1, selecting typical and comprehensive artemia-water body data;
selecting artemia data from the preprocessed ground surface reflection data in the step 1, wherein the artemia gathering form of the selected data comprises a slender strip shape and a cluster shape, and the artemia density is displayed as a dark reddish brown to light red characteristic in a true color image from high to low; selecting a water body background of data considering different water colors due to different contents of chlorophyll, suspended matters and CDOM and influence of lake sediment; the selected data also comprises artemia-water body data under the adverse conditions of thin clouds and dazzling;
step 2.2, data cutting and sample augmentation;
the primarily selected artemia-water body data are cut into samples with the size of NxN by a sliding window method, rotation and mirror image are used for sample augmentation, the sample size is expanded, and the requirements of subsequent model training are met.
7. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 1, wherein the remote sensing extraction method comprises the following steps: the construction and training of the artemia extraction deep learning model in the step 3 comprise the following steps:
step 3.1, normalizing the artemia-water body sample data obtained in the step 2;
firstly, the data type is converted into UInt8 from UInt16, the numerical range of each wave band is between 0 and 255, and then linear normalization is carried out, so that the data are mapped to a [0,1] interval, the order difference is eliminated, the phenomenon that undersized data are ignored is avoided, and the effects of unifying sample evaluation indexes and accelerating the training convergence speed are achieved;
step 3.2, selecting a proper deep learning semantic segmentation model as an artemia extracting deep learning model;
selecting a full convolution U-shaped neural network, wherein the network takes Tensorflow as a frame, the depth of the network is selected to be F according to the complexity of an extraction target, normalized NxN artemia-water body HY-1C CZI sample spectrum information is input, and the result is output as a result of the division of the NxN artemia-water body HY-1C CZI sample spectrum information; wherein, the input HY-1C CZI sample comprises spectral information of a blue band of 0.42-0.50 μm, a green band of 0.52-0.60 μm, a red band of 0.61-0.69 μm, a near infrared band of 0.76-0.89 μm, and four bands with central wavelengths of 0.460 μm, 0.560 μm, 0.650 μm and 0.825 μm in sequence;
step 3.3, configuring parameters and carrying out model training;
the basic structural unit of the full convolution U-shaped neural network is a convolution-batch normalization layer-ReLU activation function, the batch size is set according to the GPU condition, and the initial learning rate is set as d; in the iterative training process of the model, the maximum iterative round number is set as M, and an early-stopping mechanism is configured, namely, if the training precision does not rise continuously for 10 rounds or the error does not fall continuously for 10 rounds, the iteration is terminated early.
8. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 1, wherein the remote sensing extraction method comprises the following steps: in step 4.1, precision evaluation is performed according to three precision evaluation indexes of precision rate, recall rate and F1 score, namely:
in the formula, Precision is Precision; recall is Recall; F1-Score is F1 Score; TP is the number of positive samples of artemia in the two categories; FP is the number of false detection samples; FN is the number of missed samples, and is obtained by comparing the extraction result of the full convolution U-shaped neural network model with the reference result of manual interpretation.
9. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 7, wherein the remote sensing extraction method comprises the following steps: and 4.2, robustness evaluation is to select n samples with thin clouds, vortexes, shallow water, near-shore substrate, turbid water, glaring and dark water pixel complex background signal interference conditions in the artemia-water body data set, compare the extraction result of the full convolution U-shaped neural network model with the reference result of visual interpretation, and analyze the applicability and stability of the model in the application process.
10. The remote sensing extraction method of the artemia salina stripe based on deep learning of claim 7, wherein the remote sensing extraction method comprises the following steps: in step 5, Landsat-8 OLI data is simulated into HY-1CCZI data through the spectrum matching factor, and the formula for performing spectrum adjustment on the OLI data through the spectrum matching factor is as follows:
ρλ0,CZI=β(λ)×ρλ0,OLI (5)
in the formula, ρSR(lambda) corresponds to the specific spectrum of the artemia or the water body, and the spectrum of the artemia is acquired by an SVC HR-1024 spectrometer through a water tank control experiment; lambda [ alpha ]0、SRFλCentral wavelength and spectral response function, respectively;
and resampling the OLI sample after spectrum adjustment to be consistent with the CZI spatial resolution, realizing data simulation from OLI to CZI, enriching the time sequence range of the sample, and finally inputting the simulation data serving as a fresh sample into a trained full-convolution U-shaped neural network model, so that the adaptability of the algorithm to multi-platform multi-source samples is enhanced, and the application range of the generalized model is further expanded.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111336862.8A CN114119618A (en) | 2021-11-12 | 2021-11-12 | Remote sensing extraction method for inland salt lake artemia strips based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111336862.8A CN114119618A (en) | 2021-11-12 | 2021-11-12 | Remote sensing extraction method for inland salt lake artemia strips based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114119618A true CN114119618A (en) | 2022-03-01 |
Family
ID=80378788
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111336862.8A Pending CN114119618A (en) | 2021-11-12 | 2021-11-12 | Remote sensing extraction method for inland salt lake artemia strips based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114119618A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116699096A (en) * | 2023-08-08 | 2023-09-05 | 凯德技术长沙股份有限公司 | Water quality detection method and system based on deep learning |
-
2021
- 2021-11-12 CN CN202111336862.8A patent/CN114119618A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116699096A (en) * | 2023-08-08 | 2023-09-05 | 凯德技术长沙股份有限公司 | Water quality detection method and system based on deep learning |
CN116699096B (en) * | 2023-08-08 | 2023-11-03 | 凯德技术长沙股份有限公司 | Water quality detection method and system based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109781626B (en) | Near-shore high-suspended sand water body green tide remote sensing identification method based on spectral analysis | |
CN113592882B (en) | Crown extraction method based on multi-source remote sensing of unmanned aerial vehicle | |
CN112183209A (en) | Regional crop classification method and system based on multi-dimensional feature fusion | |
CN110082298B (en) | Hyperspectral image-based wheat variety gibberellic disease comprehensive resistance identification method | |
CN113591766B (en) | Multi-source remote sensing tree species identification method for unmanned aerial vehicle | |
CN109635765B (en) | Automatic extraction method for remote sensing information of shallow sea coral reef | |
CN107504923B (en) | Kelp culture area monitoring method integrating remote sensing image and extension rope information | |
CN111340826A (en) | Single tree crown segmentation algorithm for aerial image based on superpixels and topological features | |
CN112699756B (en) | Hyperspectral image-based tea origin identification method and system | |
CN105181912A (en) | Method for detection of freshness during rice storage | |
Lang et al. | Detection of chlorophyll content in maize canopy from UAV imagery | |
CN112669363B (en) | Method for measuring three-dimensional green space of urban green space | |
CN116883853B (en) | Crop space-time information remote sensing classification method based on transfer learning | |
Xiao et al. | A random forest-based algorithm to distinguish Ulva prolifera and Sargassum from multispectral satellite images | |
CN114778483A (en) | Method for correcting terrain shadow of remote sensing image near-infrared wave band for monitoring mountainous region | |
CN116012726A (en) | Deep learning-based black and odorous water body remote sensing monitoring and grading method | |
CN114119618A (en) | Remote sensing extraction method for inland salt lake artemia strips based on deep learning | |
CN109740645A (en) | A kind of CART Decision-Tree Method suitable for high score No.1 image | |
CN106568730B (en) | A kind of rice yin-yang leaf fringe recognition methods based on Hyperspectral imaging near the ground | |
CN116559111A (en) | Sorghum variety identification method based on hyperspectral imaging technology | |
CN111339959A (en) | Method for extracting offshore buoyant raft culture area based on SAR and optical image fusion | |
CN114724035A (en) | Early water bloom detection method based on remote sensing technology | |
CN113538559B (en) | Extraction method of offshore aquaculture raft extraction index based on hyperspectral remote sensing image | |
CN114119617A (en) | Method for extracting inland salt lake artemia zone of multispectral satellite remote sensing image | |
Suarez et al. | Learning image vegetation index through a conditional generative adversarial network |
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 |