CN110059758B - Remote sensing image culture pond detection method based on semantic segmentation - Google Patents
Remote sensing image culture pond detection method based on semantic segmentation Download PDFInfo
- Publication number
- CN110059758B CN110059758B CN201910333358.9A CN201910333358A CN110059758B CN 110059758 B CN110059758 B CN 110059758B CN 201910333358 A CN201910333358 A CN 201910333358A CN 110059758 B CN110059758 B CN 110059758B
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- culture pond
- sensing image
- image
- model
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a remote sensing image culture pond detection method based on semantic segmentation. The method is completed in four steps, firstly, the sub-meter-level remote sensing image is cut and blocked, and index information of each image block is stored at the same time; secondly, training the culture pond detection semantic segmentation model in advance until convergence to obtain a culture pond semantic segmentation model; then inputting different remote sensing image blocks into the semantic segmentation model of the culture pond to obtain a remote sensing image block culture pond binary mask image; and finally, fusing basic index information of the remote sensing image blocks obtained after cutting to complete the combination of the binary mask images and generate a mask grid image of the culture pond. The method utilizes the deep learning technology to train the culture pond detection model, and has accurate extraction result, good robustness and high adaptability; and the remote sensing image culture pond detection integrated process is designed, the manual participation degree is reduced, and the method has a certain practical value.
Description
Technical Field
The invention relates to a deep learning technology, mainly relates to pixel-level semantic segmentation, and discloses an end-to-end remote sensing image culture pond detection method.
Background
Aquaculture refers to commercial activities for raising aquatic organisms (including fish, mollusks, crustaceans, and aquatic plants), and can be classified into 3 categories, land, water, and beach, depending on the nature of the basal plane of operation. Land-based systems mainly include ponds, rice fields and other facilities constructed on land; the development system based on the water surface comprises a bay, a fence, a net cage and raft culture, and is usually positioned in coastal or inland water areas provided with a fence; the mudflat-based culture system comprises pond culture and high-level pond culture. China is the first aquaculture country in the world and is the only country in the world with aquaculture yield exceeding fishing yield, and the aquaculture scale is still continuously and rapidly increasing at present. While making a great contribution to meeting the demand of aquatic products in the world, the aquaculture in China is under pressure in various aspects such as the increasingly worsening of the water environment condition, the supervision of social public opinion, the monitoring of policies and regulations, the increasingly higher demand on the quality of the aquatic products and the like, and the aquaculture is increasingly becoming one of the current research hotspots.
Due to the convenience and the economy of aquaculture in the coastal zone land area, a large number of aquaculture ponds are built in ecological protection areas of the coastal zone, wastewater with excessive nitrogen and phosphorus in the aquaculture ponds is directly discharged into the sea through concealed pipes, the sea defense forest land is illegally occupied, and the ecological environment is greatly damaged. Remote sensing has the advantages of wide detection range, high data acquisition speed, short period, high timeliness, low cost, high economic benefit and the like. In recent years, with the development of remote sensing technology, multi-platform, multi-type and multi-resolution remote sensing data provide opportunity for aquiculture information acquisition, and a culture pond can be identified and positioned quickly by using remote sensing images.
The method for extracting the remote sensing image from the culture pond mainly comprises visual interpretation, namely, according to interpretation marks (such as positions, shapes, sizes, tones, shadows and textures) and interpretation experiences, the requirement on the professional level of an interpreter is high, and meanwhile, the method has the advantages of high workload, labor and time consumption and low timeliness of quantitative analysis on massive spatial information; secondly, the method based on the space structure analysis is not suitable for extracting the culture pond with isolated distribution and small plaque; in addition, the object-oriented information extraction method firstly segments an image into image objects with certain meanings, then comprehensively utilizes information such as spectral characteristics, textures, shapes, proximity relations and the like of ground features to determine the types of the segmented objects, is greatly influenced by characteristic selection, and needs to be combined with a visual interpretation method.
A semantic segmentation algorithm based on a Deep convolutional neural network has a good effect in the field of computer vision, and a semantic segmentation algorithm designed based on a network structure classified by labels such as AlexNet, VGGNet and GoogleNet is used as a basis in the early stage, L ong and the like replace a full connection layer of the convolutional neural network with a convolutional layer, the convolutional layer is adopted and gradually recovered through an interpolation method, and a semantic segmentation mask map with the same size as the original input size is generated.
In order to effectively and quickly extract the information of the culture pond and enable the extraction of the culture pond to be more intelligent and automatic, a remote sensing image culture pond detection method based on semantic segmentation is provided.
Disclosure of Invention
The invention aims to provide a remote sensing image culture pond detection method based on semantic segmentation, which can effectively and quickly identify culture pond information in a remote sensing image, avoids the characteristics of large image, time and labor waste caused by artificial intervention in the traditional method, and provides technical support and basis for culture pond supervision and protection of coastal zone ecological environment.
The invention provides a semantic segmentation-based remote sensing image culture pond retrieval method, which is realized by adopting the following technical scheme:
the method comprises the following steps that firstly, a sub-meter-level remote sensing image to be extracted is cut and processed in a blocking mode, remote sensing image blocks are obtained after cutting, and index information of each image block is stored;
step two, when information extraction is carried out on the remote sensing image block, a culture pond detection semantic segmentation model needs to be trained in advance, the sub-meter remote sensing image is used as training basic data, a visual interpretation mode is adopted, culture pond information in the remote sensing image is obtained, a training sample is made, data are preprocessed and input into a designed semantic segmentation model for training, and after cost errors are converged, the culture pond semantic segmentation model is obtained;
extracting culture pond information of the remote sensing image block, preprocessing the culture pond information, and inputting the preprocessed culture pond information into a trained semantic segmentation model to obtain a binary mask map of the culture pond of the remote sensing image block;
combining the stored basic index information of the remote sensing image blocks and the real parameter size of the merging area, merging the extracted binary mask images of the remote sensing image blocks to obtain the mask raster image of the culture pond with the geographic information.
Further, the remote sensing image in the first step is cut and blocked, and the specific cutting standard is determined by the size of a preset block and the size of an overlapping area;
further, the remote sensing image cutting and blocking in the first step comprises the following steps:
a) setting cutting parameters, wherein the size of the pixels of the cut image blocks is set to be 1024 x 1024, and the size of the pixels of the length of the overlapped area is 144 and is represented by o;
b) the method comprises the steps of cutting an original sub-meter-level remote sensing image in a block-by-block mode according to a line sequence, wherein stored index information is (x, y, w, h), x is the column number of the remote sensing image block in the original sub-meter-level image, y is the line number of the remote sensing image block in the original sub-meter-level image, w and h are the width and the height of the remote sensing image block, and due to the fact that a GPU training model is used, in order to take account of the size of GPU video memory, the stored index information is uniformly set to be 1024.
Further, the training of the semantic segmentation model of the culture pond in the second step comprises the following steps:
a) manufacturing a training sample by using a sub-meter satellite image, and extracting culture pond information in the image by means of visual interpretation by means of remote sensing professional software;
b) processing the extracted culture pond information data and the original sub-meter satellite image to generate a training sample and a test sample containing culture pond mask information;
c) carrying out preprocessing such as random mirror image, random rotation, random blur, random affine transformation, random color stretching transformation and the like on training data input into the model;
d) setting training parameters, and taking a Dice coefficient of a set similarity measurement function and a binary cross entropy BCE as an error cost function, wherein the error cost function specifically comprises the following steps:
Loss＝DiceLoss+BCELoss (3)
wherein p isnIs the predicted value of the nth pixel point, rnThe actual label value of the nth pixel point is a smoothing coefficient, N is the number of the pixel points, and the finally obtained error cost function is shown as (3) and is the result of adding the error costs of the Dice and the BCE;
training the model by using an Adam optimizer, wherein a specific training strategy is shown as a formula (4):
wherein m istAnd vtβ being first and second order momentum terms, respectively1And β2The power value is usually 0.9 and 0.999 respectively;
and
respectively are respective correction values; wtRepresenting the parameter of the t time, namely the t-th iteration model; gt＝ΔJ(Wt) Representing the gradient magnitude of the t times of iterative cost function relative to W; taking a very small number (generally 1e-8) to prevent the denominator from being 0;
e) parallel training is carried out by adopting 4 Titan xp GPUs until the model converges, and model testing is carried out in a test set, wherein the testing IoU is 0.75.
Further, the extraction of the information of the culture pond for the remote sensing image block in the third step comprises the following steps
The method comprises the following steps:
a) the method comprises the steps that a model structure is divided into two parts, namely encoding and decoding, after input remote sensing block data is subjected to standardization processing, encoding of the model is firstly carried out, namely a characteristic extraction stage, an encoding layer adopts a ResNet34 network and is formed by connecting FirstConv-BN-Relu, FirstMaxPool, Encoder1, Encoder2, Encoder3 and Encoder4 in sequence, wherein Conv represents convolution, BN represents batch normalization, Relu is an activation function, and Relu (x) is max (0, x); inputting R, G, B three-channel images with the size of 1024 × 1024, namely three-dimensional data, wherein the size of a convolution filter of the FirstConv-BN-Relu is 7 × 7, the convolution step size is 2, and the convolution filter contains batch normalization BN and an activation function Relu; encoder is a residual block of a ResNet34 network under different scales, the output of each Encoder part is reserved as bottom layer characteristics, and the output bottom layer characteristics are down-sampled by 2 times;
b) the output of the coding layer is connected with a cavity convolution module DilationBlock which is formed by combining and superposing cavities with different expansion rates, so that the receptive field of the model is increased; the modules are sequentially connected by the void convolution layers with the expansion rates of 1, 2, 4 and 8, and finally output as output superposition values of void convolution with different expansion rates;
c) the output of the hole convolution module is used as the input of a decoding layer which is formed by sequentially connecting DecoderBlock4, DecoderBlock3, DecoderBlock2, DecoderBlock1, Final-Deconv-Relu-1, Final-Conv-Relu-2 and Final-Conv-Sigmoid-3, the output of the hole convolution module is connected with DecoderBlock4, the bottom layer features stored in the coding layer are connected with the decoding layer after being connected through L inkBlock connection processing, Encoder3 is connected with L inkBlock3, Encoder2 is connected with L inkBlock2, Encoder1 is connected with 7 inkBlock1, the output of the DecoderBlock4 is combined with the output of 4 as the input of the DecoderBlock 72, the output of the DecoderBlock4 is combined with the output of the lamination layer 4 as the input of a DecoderBlock deconvolution function, the DecoderBlock4, the convolution function of each of the DecoderBlock 72 and the DecoderBlock4 and the activation function of the DecoderBlock4, and the convolution function of the DecoderBlock4, wherein the convolution function of the DecoderBlock 72 comprises the volume-4, the volume-363, the volume-4, the volume-3
d) And obtaining a predicted value of each pixel point through a decoding layer, and generating a remote sensing image culture pond binary mask image according to the predicted value, wherein the pixel value of a culture pond area in the image is 1, and the pixel values of other areas are 0.
Further, the combining of the two-value mask maps of the culture pond in the fourth step specifically comprises:
and the two-value mask map of the culture pond obtained by the semantic segmentation model does not contain geographic information, each generated culture pond block corresponds to the stored index information of the remote sensing image block, the two-value mask maps of the culture pond are merged through coordinate mapping, the geographic information is added, and a grid map of the culture pond corresponding to the original input remote sensing image is generated.
Compared with the prior art, the invention has the following beneficial effects:
compared with the traditional method, the method for detecting the remote sensing image culture pond based on semantic segmentation improves the universality of the model, and can effectively extract culture pond information in the remote sensing image with relative blur and color deviation;
the network of the invention is an end-to-end structure, namely, a remote sensing image is input, a mask image containing culture pond information is directly output, no human intervention is needed, and the degree of automation is higher.
Brief description of the drawings
FIG. 1 is an overall flow chart of the remote sensing image culture pond detection method.
FIG. 2 is a schematic diagram of a network structure of the remote sensing image culture pond detection method of the invention.
FIG. 3 is a schematic diagram of a coding layer structure in the remote sensing image culture pond extraction model.
FIG. 4 is a schematic structural diagram of a cavity convolution module DialationBlock in the remote sensing image culture pond extraction model.
FIG. 5 is a schematic structural diagram of a DecoderBlock module of a decoding layer in the remote sensing image culture pond extraction model.
FIG. 6 is a schematic structural diagram of an L inkBlock module in the remote sensing image culture pond extraction model.
FIG. 7 is a schematic diagram showing changes in training error cost values of the remote sensing image culture pond extraction model.
Fig. 8 is a jilin No. one sub-meter-level remote sensing image in the embodiment of the present invention.
FIG. 9 is a drawing of the overall effect of extraction of the Jilin No. one sub-meter-level remote sensing image culture pond.
Detailed Description
The technical scheme of the invention is explained in detail below with reference to the attached drawings:
FIG. 1 is an overall flow chart of the remote sensing image culture pond detection method, which comprises the following specific steps:
the first step, the sub-meter remote sensing grid image size is bigger under general conditions, needs to be cut and blocked for prediction:
1.1, firstly setting cutting parameters, setting the size of pixels of a cutting image block to be 1024 x 1024 during cutting, setting the length o of a superposition area to be 144, wherein an superposition area refers to that a common area exists between every two adjacent cutting blocks during cutting, and because a prediction result of each remote sensing block possibly has a fault at a joint in the final splicing process, the superposition area is set, and influences are eliminated by taking the intersection of the prediction results of the superposition area;
1.2, cutting the original sub-meter-level remote sensing image in blocks according to the line sequence, and storing index information.
And a second step, which is a model training part and actually needs to be trained in advance, wherein the specific training process is as follows:
2.1, manufacturing a training sample by using the Jilin I sub-meter-level remote sensing image, and extracting culture pond information in the image by means of visual interpretation by means of remote sensing professional software;
2.2, processing the extracted culture pond information data and the original sub-meter-level remote sensing image to generate 2000 training samples containing culture pond mask information and 300 test samples;
2.3, preprocessing such as random mirror image, random rotation, random fuzzy, random affine transformation, random color stretching transformation and the like is carried out on the training data input into the model, and the input data are preprocessed, so that the data are expanded, the sample diversity is increased, and the robustness of the model is improved to a certain extent;
and 2.4, setting training parameters, adopting a set similarity measurement function Dice coefficient and a binary cross entropy BCE as error cost functions, and improving the performance of the model by using different loss functions.
The model is trained by using the Adam optimizer, the Adam optimizer is simple to implement, efficient in calculation, low in memory requirement, free of influence of gradient expansion transformation on parameters, capable of automatically adjusting learning rate, suitable for large-scale data and parameter scenes, and excellent in performance.
Fig. 7 is a schematic diagram of the decrease of the error cost value along with the number of iterations, and it can be seen through graphical representation that the model convergence is fast, and the model training is ended when the training error cost value fluctuates and tends to be stable.
2.5, performing parallel training by adopting 4 Titan xp GPUs, and accelerating the convergence speed of the model through parallel calculation during training; during testing and actual application, the detection efficiency of the model can be greatly improved through multi-GPU parallel computing. The concrete framework of the model is shown in fig. 2, until the model converges, the model is tested in the test set, where test IoU is 0.75.
Inputting the remote sensing image blocks cut and partitioned into blocks into a trained model, and acquiring binary mask information of the culture pond:
3.1 after carrying out standardization processing on the input remote sensing block data, firstly entering a model coding stage, namely a feature extraction stage, wherein a coding layer adopts a ResNet34 network and is formed by connecting FirstConv-BN-Relu, FirstMaxPool, Encoder1, Encoder2, Encoder3 and Encoder4 in sequence, as shown in figure 3, the output of each Encoder part is reserved as a bottom layer feature, and the sampling rate of the output bottom layer feature is reduced by 2 times;
3.2 the output of the coding layer is connected to a cavity convolution module, which is formed by combining and superposing cavities with different expansion rates, as shown in FIG. 4, and the receptive field of the model is increased. The modules are sequentially connected by the void convolution layers with the expansion rates of 1, 2, 4 and 8, and finally output as output superposition values of void convolution with different expansion rates.
And 3.3, the output of the cavity convolution module is used as the input of the decoding layer, the cavity convolution increases the scope of the receptive field of the convolution, the cavity convolution with different expansion rates is used, the cavity convolution is carried out on different scales, and the context semantic information of the characteristics is enriched.
The decoding layer is formed by sequentially connecting DecoderBlock4, DecoderBlock3, DecoderBlock2, DecoderBlock1, Final-Deconv-Relu-1, Final-Conv-Relu-2 and Final-Conv-Sigmoid-3. the output of the hole convolution module is connected with DecoderBlock4, the bottom layer features stored in the coding layer are connected with the decoding layer after L inkBlock connection processing, such as the DecoderBlock module shown in FIG. 5, the Encoder3 is connected with L inkBlock3, such as the L inkBlock module shown in FIG. 6, the Encoder2 is connected with L inkBlock2, the Encoder1 is connected with L inkBlock1, the output of the DecoderBlock1 is combined with the output of 1, the DecoderBlock1 is used as the input of the DecoderBlock1, the output of the DecoderBlock1 is combined with the output of the DecoderBlock1, and the DecoderBlock1 is used as the input and output of the DecoderBlock1, and the DecoderBlock 1.
And fourthly, merging the two-value mask images of the culture ponds, obtaining a plurality of two-value mask image blocks of the culture ponds through a semantic segmentation model, enabling each generated culture pond block to correspond to the stored index information of the remote sensing image block, merging the two-value mask images of the culture ponds through coordinate mapping, adding geographic information, and generating a grid image of the culture ponds corresponding to the original input remote sensing image, wherein input sub-meter-level remote sensing image images and corresponding two-value mask images of the culture ponds are shown in the images 8 and 9 respectively.
The above embodiments are only for illustrating the invention and are not to be construed as limiting the invention, and those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention, therefore, all equivalent technical solutions also belong to the scope of the invention, and the scope of the invention is defined by the claims.
Claims (6)
1. A remote sensing image culture pond detection method based on semantic segmentation is characterized by comprising the following steps:
the method comprises the following steps that firstly, a sub-meter-level remote sensing image to be extracted is cut and processed in a blocking mode, remote sensing image blocks are obtained after cutting, and index information of each image block is stored;
secondly, when information extraction is carried out on the remote sensing image block, a model structure needs to be designed, the model is trained, the sub-meter remote sensing image is used as training basic data, a visual interpretation mode is adopted, culture pond information in the remote sensing image is obtained, a training sample is made, data are preprocessed and input into a designed semantic segmentation model for training, and after cost errors are converged, the culture pond semantic segmentation model is obtained;
extracting culture pond information of the remote sensing image block, preprocessing the culture pond information, and inputting the preprocessed culture pond information into a trained semantic segmentation model to obtain a binary mask image of the culture pond of the remote sensing image block, wherein the method comprises the following steps of:
a) the model structure is divided into two major parts of coding and decoding, after the input remote sensing block data is subjected to standardization processing, the coding of the model is firstly carried out, namely a characteristic extraction stage, a ResNet34 network is adopted in a coding layer, and the coding layer is formed by sequentially connecting FirstConv-BN-Relu, FirstMaxPool, Encoder1, Encoder2, Encoder3 and Encoder4, wherein the size of a convolution filter of the FirstConv-BN-Relu is 7, the convolution step length is 2, and the convolution filter contains batch normalization BN and an activation function Relu; encoder is a residual block of a ResNet34 network under different scales, the output of each Encoder part is reserved as bottom layer characteristics, and the output bottom layer characteristics are down-sampled by 2 times;
b) the output of the coding layer is connected with a cavity convolution module DilationBlock which is formed by combining and superposing cavities with different expansion rates, so that the receptive field of the model is increased; the modules are sequentially connected by the void convolution layers with the expansion rates of 1, 2, 4 and 8, and finally output as output superposition values of void convolution with different expansion rates;
c) the output of the hole convolution module is used as the input of a decoding layer consisting of DecoderBlock4, DecoderBlock3, DecoderBlock2, DecoderBlock1 and Final-DeconThe method comprises the steps of generating a code layer, generating a tail layer
x represents the input characteristic value of a neuron in the forward transmission process of the convolutional neural network;
d) obtaining a predicted value of each pixel point through a decoding layer, and generating a remote sensing image culture pond binary mask image according to the predicted value, wherein the pixel value of a culture pond area in the image is 1, and the pixel values of other areas are 0;
combining the stored basic index information of the remote sensing image blocks and the real parameter size of the merging area, merging the extracted binary mask images of the remote sensing image blocks to obtain the mask raster image of the culture pond with the geographic information.
2. The method for detecting the remote sensing image culture pond based on the semantic segmentation as claimed in claim 1, wherein the remote sensing image in the first step is cut into blocks, and the specific cutting standard is determined by a preset block size and an overlapping area size.
3. The method for detecting the remote sensing image culture pond based on the semantic segmentation as claimed in claim 1 or 2, wherein the remote sensing image cutting and blocking in the first step comprises the following steps:
a) setting cutting parameters, wherein the size of the pixels of the cut image blocks is set to be 1024 x 1024, and the size of the pixels of the length of the overlapped area is 144 and is represented by o;
b) the original sub-meter-level remote sensing image is cut in a block-by-block mode according to the line sequence, stored index information is (x, y, w, h), wherein x is the column number of the remote sensing image block in the original sub-meter-level image, y is the line number of the remote sensing image block in the original sub-meter-level image, and w and h are the width and the height of the remote sensing image block and are uniformly set to be 1024.
4. The remote-sensing image culture pond detection method based on semantic segmentation as claimed in claim 3, wherein the culture pond semantic segmentation model training in the second step comprises the following steps:
a) manufacturing a training sample by using a sub-meter satellite image, and extracting culture pond information in the image by means of visual interpretation by means of remote sensing professional software;
b) processing the extracted culture pond information data and the original sub-meter satellite image to generate a training sample and a test sample containing culture pond mask information;
c) carrying out 5 combination modes of random mirror image, random rotation, random blur, random affine transformation and random color stretching transformation on training data input into the model for pretreatment;
d) setting training parameters, and taking a Dice coefficient of a set similarity measurement function and a binary cross entropy BCE as an error cost function, wherein the error cost function specifically comprises the following steps:
Loss＝DiceLoss+BCELoss (3)
wherein p isnIs the predicted value of the nth pixel point, rnSetting the actual label value of the nth pixel point as a smooth coefficient and setting the value as 1e-8, setting N as the number of the pixel points, and finally obtaining an error cost function which is shown as (3) and is the result of adding the error costs of Dice and BCE, wherein w isnRepresenting the weight size of the nth iteration batch;
training the model by using an Adam optimizer, wherein a specific training strategy is shown as a formula (4):
wherein m istAnd vtβ being first and second order momentum terms, respectively1And β2The power value is 0.9 and 0.999 respectively;
and
respectively are respective correction values; wtRepresenting the parameter of the t time, namely the t-th iteration model; gt＝ΔJ(Wt) Representing the gradient magnitude of the cost function with respect to W for t iterations, J representing the cost function, i.e., the L oss value above, in order to prevent the denominator from being 0, the smoothing coefficient is set to 1 e-8;
e) and (3) performing parallel training by adopting 4 Titan xp GPUs until the model is converged, and performing model test on a test set, wherein the test IoU is 0.75, and IoU refers to the proportion of intersection and union of a real culture pond area and a model prediction culture pond area, and represents the performance index of the semantic segmentation model.
5. The remote sensing image culture pond detection method based on semantic segmentation as claimed in claim 1 or 4, wherein the culture pond binary mask map merging of step four specifically is:
and merging the two-value mask images of the culture pond through coordinate mapping, adding geographic information, and generating a grid image of the culture pond corresponding to the original input remote sensing image.
6. The remote sensing image culture pond detection method based on semantic segmentation as claimed in claim 4, characterized in that the generated training samples are 2000 and the test samples are 300.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910333358.9A CN110059758B (en) | 2019-04-24 | 2019-04-24 | Remote sensing image culture pond detection method based on semantic segmentation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910333358.9A CN110059758B (en) | 2019-04-24 | 2019-04-24 | Remote sensing image culture pond detection method based on semantic segmentation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110059758A CN110059758A (en) | 2019-07-26 |
| CN110059758B true CN110059758B (en) | 2020-07-10 |
Family
ID=67320418
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910333358.9A Active CN110059758B (en) | 2019-04-24 | 2019-04-24 | Remote sensing image culture pond detection method based on semantic segmentation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110059758B (en) |
Families Citing this family (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110599452B (en) * | 2019-08-07 | 2022-02-22 | 全球能源互联网研究院有限公司 | Rust detection system, method, computer device and readable storage medium |
| CN110781775B (en) * | 2019-10-10 | 2022-06-14 | 武汉大学 | Remote sensing image water body information accurate segmentation method supported by multi-scale features |
| CN110751075A (en) * | 2019-10-12 | 2020-02-04 | 海南长光卫星信息技术有限公司 | Remote sensing image culture pond detection method based on example segmentation |
| CN111080602B (en) * | 2019-12-12 | 2020-10-09 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting foreign matters in water leakage hole of railway wagon |
| CN111080650B (en) * | 2019-12-12 | 2020-10-09 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting looseness and loss faults of small part bearing blocking key nut of railway wagon |
| CN111259797A (en) * | 2020-01-16 | 2020-06-09 | 南开大学 | Iterative remote sensing image road extraction method based on points |
| CN111339237A (en) * | 2020-02-28 | 2020-06-26 | 深圳前海微众银行股份有限公司 | Farm risk prediction method, device, equipment and storage medium |
| CN113112446A (en) * | 2020-03-05 | 2021-07-13 | 成都理工大学 | Tunnel surrounding rock level intelligent judgment method based on residual convolutional neural network |
| CN111368843B (en) * | 2020-03-06 | 2022-06-10 | 电子科技大学 | Method for extracting lake on ice based on semantic segmentation |
| CN111461130B (en) * | 2020-04-10 | 2021-02-09 | 视研智能科技(广州)有限公司 | High-precision image semantic segmentation algorithm model and segmentation method |
| CN111539435A (en) * | 2020-04-15 | 2020-08-14 | 创新奇智(合肥)科技有限公司 | Semantic segmentation model construction method, image segmentation equipment and storage medium |
| CN111753647B (en) * | 2020-05-11 | 2021-01-29 | 广东无限阵列科技有限公司 | Automatic identification method and device for livestock breeding shed and storage medium |
| CN112836727B (en) * | 2020-07-27 | 2021-08-10 | 盐城郅联空间科技有限公司 | Image interpretation optimization method based on space-time big data mining analysis technology |
| CN112733745A (en) * | 2021-01-14 | 2021-04-30 | 北京师范大学 | Cultivated land image extraction method and system |
| CN112766144A (en) * | 2021-01-15 | 2021-05-07 | 雾实(福建)科技有限公司 | Method, device, equipment and medium for distinguishing routing inspection detail difference of island infrastructure |
| CN112883900B (en) * | 2021-03-12 | 2022-03-04 | 中科三清科技有限公司 | Method and device for bare-ground inversion of visible images of remote sensing images |
| CN113743383B (en) * | 2021-11-05 | 2022-06-07 | 航天宏图信息技术股份有限公司 | SAR image water body extraction method and device, electronic equipment and storage medium |
| CN113763384B (en) * | 2021-11-10 | 2022-03-15 | 常州微亿智造科技有限公司 | Defect detection method and defect detection device in industrial quality inspection |
| CN115019192B (en) * | 2022-05-30 | 2023-06-06 | 杭州电子科技大学 | Flood variation detection method and system based on dual-channel backbone network and joint loss function |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107193877A (en) * | 2017-04-24 | 2017-09-22 | 中国科学院遥感与数字地球研究所 | Land cover classification system and method |
| US10860919B2 (en) * | 2017-09-27 | 2020-12-08 | Google Llc | End to end network model for high resolution image segmentation |
| CN108376232A (en) * | 2018-01-04 | 2018-08-07 | 北京星衡科技有限公司 | A kind of method and apparatus of automatic interpretation for remote sensing image |
| CN108875659B (en) * | 2018-06-26 | 2022-04-22 | 上海海事大学 | Sea chart cultivation area identification method based on multispectral remote sensing image |
| CN109145730B (en) * | 2018-07-13 | 2021-08-13 | 安徽大学 | Automatic semantic segmentation method for mining area in remote sensing image |
-
2019
- 2019-04-24 CN CN201910333358.9A patent/CN110059758B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN110059758A (en) | 2019-07-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110059758B (en) | Remote sensing image culture pond detection method based on semantic segmentation | |
| CN110751075A (en) | Remote sensing image culture pond detection method based on example segmentation | |
| CN111986099B (en) | Tillage monitoring method and system based on convolutional neural network with residual error correction fused | |
| CN111259898B (en) | Crop segmentation method based on unmanned aerial vehicle aerial image | |
| CN112561876B (en) | Image-based water quality detection method and system for ponds and reservoirs | |
| CN110414509B (en) | Port docking ship detection method based on sea-land segmentation and characteristic pyramid network | |
| CN111178304B (en) | High-resolution remote sensing image pixel level interpretation method based on full convolution neural network | |
| CN114898212B (en) | Method for extracting multi-feature change information of high-resolution remote sensing image | |
| CN114022408A (en) | Remote sensing image cloud detection method based on multi-scale convolution neural network | |
| US20230306730A1 (en) | Information extraction method of offshore raft culture based on multi-temporal optical remote sensing images | |
| CN112990085A (en) | Method and device for detecting change of culture pond and computer readable storage medium | |
| CN112001293A (en) | Remote sensing image ground object classification method combining multi-scale information and coding and decoding network | |
| CN116935043A (en) | Typical object remote sensing image generation method based on multitasking countermeasure network | |
| CN115359029A (en) | Semi-supervised medical image segmentation method based on heterogeneous cross pseudo-supervised network | |
| CN115810149A (en) | High-resolution remote sensing image building extraction method based on superpixel and image convolution | |
| CN113469097B (en) | Multi-camera real-time detection method for water surface floaters based on SSD network | |
| CN114387446A (en) | Automatic water body extraction method for high-resolution remote sensing image | |
| CN111862134B (en) | Offshore aquaculture pond extraction method based on Sentinel-1 and GEE | |
| CN117058367A (en) | Semantic segmentation method and device for high-resolution remote sensing image building | |
| CN116543165A (en) | Remote sensing image fruit tree segmentation method based on dual-channel composite depth network | |
| CN114494910B (en) | Multi-category identification and classification method for facility agricultural land based on remote sensing image | |
| CN116012707A (en) | Remote sensing identification method for fish light complementary photovoltaic cultivation mode | |
| CN115578645A (en) | SAR image buoyant raft culture information extraction method for generating confrontation network through semi-supervised cyclic consistency | |
| CN114998587A (en) | Remote sensing image building semantic segmentation method and system | |
| Qin et al. | U_EFF_NET: High-precision segmentation of offshore farms from high-resolution SAR remote sensing images |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |