CN117788822A - Farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing image - Google Patents
Farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing image Download PDFInfo
- Publication number
- CN117788822A CN117788822A CN202311830023.0A CN202311830023A CN117788822A CN 117788822 A CN117788822 A CN 117788822A CN 202311830023 A CN202311830023 A CN 202311830023A CN 117788822 A CN117788822 A CN 117788822A
- Authority
- CN
- China
- Prior art keywords
- farmland
- image
- low
- pixel
- altitude
- 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
- 238000000605 extraction Methods 0.000 title claims abstract description 20
- 230000011218 segmentation Effects 0.000 claims abstract description 64
- 238000000034 method Methods 0.000 claims abstract description 41
- 238000012549 training Methods 0.000 claims abstract description 29
- 238000012545 processing Methods 0.000 claims abstract description 19
- 238000012795 verification Methods 0.000 claims abstract description 15
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims abstract description 10
- 238000006243 chemical reaction Methods 0.000 claims abstract description 8
- 238000011426 transformation method Methods 0.000 claims abstract description 3
- 239000003550 marker Substances 0.000 claims description 30
- 238000002372 labelling Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000009466 transformation Effects 0.000 claims description 8
- 238000005457 optimization Methods 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 5
- 238000003709 image segmentation Methods 0.000 claims description 4
- 230000000007 visual effect Effects 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 3
- 238000003708 edge detection Methods 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000005520 cutting process Methods 0.000 claims description 2
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 230000001131 transforming effect Effects 0.000 claims description 2
- 238000004422 calculation algorithm Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 241000196324 Embryophyta Species 0.000 description 2
- 230000004888 barrier function Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- 230000000877 morphologic effect Effects 0.000 description 2
- 230000001737 promoting effect Effects 0.000 description 2
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 238000012271 agricultural production Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a farmland boundary positioning information extraction method based on an unmanned aerial vehicle low-altitude remote sensing image, which comprises the following steps: constructing a low-altitude farmland image data set, and dividing the data set into a training set and a verification set; training an initial semantic segmentation network model based on the training set, and verifying the trained semantic segmentation network model through the verification set to obtain a semantic segmentation network model; processing the target low-altitude farmland image by utilizing a semantic segmentation network model to obtain a farmland mask image, and extracting farmland boundaries of the farmland mask image to obtain farmland boundary pixel coordinates; and constructing a conversion relation between the farmland boundary pixel coordinates and the boundary geographic coordinates based on an affine transformation method, and obtaining the farmland boundary geographic coordinates. The invention improves the segmentation efficiency and the interpretation accuracy to a certain extent; the intelligent degree is high, the universality is good, and the technical problems of low efficiency, long time consumption and poor accuracy of the conventional farmland segmentation method are solved.
Description
Technical Field
The invention relates to the technical field of intelligent agriculture, in particular to a farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing images.
Background
Smart agriculture is a brand new agricultural production mode for realizing high-quality, high-efficiency and low-cost development of agriculture, and unmanned farms are important ways for realizing smart agriculture. Accurate acquisition farmland plot boundary position information is the prerequisite and the basis of constructing high accuracy farmland electronic map, can provide accurate regional position information for intelligent agricultural machinery unmanned operation, also is the data foundation of construction unmanned farm simultaneously.
At present, farmland image data based on low-altitude remote sensing of high-resolution unmanned aerial vehicle are mostly extracted from large-area farmland in plain areas, and the conditions of great difficulty and relatively low interpretation precision exist in extracting farmland in hilly areas. Therefore, the invention provides a farmland boundary positioning information extraction method based on the unmanned aerial vehicle low-altitude remote sensing image, which realizes the acquisition of farmland boundary geographic coordinates in the southern hilly area.
Disclosure of Invention
The invention aims to help agricultural machinery to better understand farmland environment, realize intelligent agricultural machinery operation and provide a farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing images.
In order to achieve the above purpose, the invention provides a farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing images, comprising the following steps:
constructing a low-altitude farmland image data set, and dividing the data set into a training set and a verification set;
training an initial semantic segmentation network model based on the training set, and verifying the trained semantic segmentation network model through the verification set to obtain a semantic segmentation network model;
processing the target low-altitude farmland image by utilizing the semantic segmentation network model to obtain a farmland mask image, and extracting farmland boundaries of the farmland mask image to obtain farmland boundary pixel coordinates;
and constructing a conversion relation between the farmland boundary pixel coordinates and the boundary geographic coordinates based on an affine transformation method, and obtaining farmland boundary geographic coordinates.
Preferably, constructing the low-altitude farmland image dataset comprises:
and acquiring a low-altitude farmland image by using an unmanned aerial vehicle platform, carrying out distortion calibration on the low-altitude farmland image, constructing a two-dimensional orthophoto, adjusting the image to be of the same size specification, and constructing the low-altitude farmland image data set.
Preferably, the dividing the training set and the verification set includes cutting and labeling the low-altitude farmland image data set before the dividing includes:
and labeling the low-altitude farmland image by a manual visual dotting mode based on a labeling tool, generating a label graph with the same format, and performing data processing on the label graph to obtain a data format which can be input into the semantic segmentation network model.
Preferably, the semantic segmentation network model comprises:
feature extraction network: the method comprises the steps of extracting features from an image through a MobileNet V2 structure, and integrating low-resolution and high-semantic features of a bottom network with high-resolution and low-semantic features of a top network through a pyramid structure, wherein a basic unit of the MobileNet V2 structure is a depth separable convolution;
regional advice network: the method comprises the steps of generating anchor frames with different sizes on a feature map by adopting a sliding window method for the integrated features so as to identify regions of interest on different scales;
ROI alignment layer module: the region of interest generated by the region suggestion network is integrated, and the ROI alignment layer integrates region characteristics;
PointRend module: the method is used for carrying out self-adaptive feature pooling on the acquired fine granularity features by utilizing bilinear interpolation, combining the fine granularity features with local mask prediction of all sampling points, further combining the fine granularity features with global prediction, and carrying out target segmentation based on a multi-layer perceptron.
Preferably, training an initial semantic segmentation network model based on the training set includes:
setting training parameters of the initial semantic segmentation network model, setting a cross entropy loss function as a loss function of the initial semantic segmentation network model, and calculating a loss function between a feature map and a label map;
optimizing network parameters by using a Momentum optimization method, and setting the number of batch processing data and total iteration generation;
and performing parameter optimization by adopting a random gradient descent method, setting an initial learning rate and a weight attenuation rate, and training the initial semantic segmentation network model.
Preferably, the method for calculating the loss function between the feature map and the label map is as follows:
where N is the sum of the number of pixels in a batch size, m is the number of categories, y ij Representing the labeling of pixel i for category j,the probability that pixel i is of class j is represented.
Preferably, verifying the trained semantic segmentation network model through the verification set includes:
initializing a semantic segmentation network model by loading a MobileNet V2 weight pre-trained based on a Cityspaces public data set, performing scene analysis on an input farmland data set, transforming a sample space into a feature space, extracting pixel features by using a feature extraction network, inputting a combined feature map into a region suggestion network to obtain regions of interest with different sizes, gathering region features by an ROI alignment layer module, performing bilinear interpolation up-sampling on an output fine-grained feature map by a PointRend module, and performing subdivision prediction by using a multi-layer perceptron to output a feature map;
comparing the prediction probabilities of the same pixel point in the feature map under different categories, judging the pixel as a farmland pixel if the prediction probability of the farmland block is higher than a preset threshold value, judging the pixel as a background pixel if the prediction probability of the farmland block is lower than the preset threshold value, and outputting a segmentation mask;
and verifying the segmentation mask by adopting an image segmentation evaluation index mIoU average cross-correlation ratio to obtain the verification result.
Preferably, the method for verifying the segmentation mask is as follows:
wherein k+1 represents the number of categories, p ii Representing the number of pixels with a true value i and a predicted value i, p ij Representing the number of pixels, p, with a true value i and a predicted value j ji Representing the number of pixels for which the true value is j and the predicted value is i.
Preferably, obtaining the farmland boundary pixel coordinates includes:
performing binarization processing on the farmland mask image output by the semantic segmentation model to obtain a binarized image, and extracting a binarized image boundary by adopting Canny edge detection and calculation;
and extracting boundary pixel coordinates of the segmentation mask by an eight-neighborhood boundary tracking method to obtain the farmland boundary pixel coordinates.
Preferably, the method for obtaining the geographical coordinates of the farmland boundary comprises the following steps:
∑X＝A 0 n+A 1 ∑x-A 2 ∑y
∑xX＝A 0 ∑x+A 1 ∑x 2 -A 2 ∑xy
∑yX＝A 0 ∑y+A 1 ∑xy-A 2 ∑y 2
∑Y＝B 0 n+B 1 ∑x+B 2 ∑y
∑xY＝B 0 ∑x+B 1 ∑x 2 +B 2 ∑xy
∑yY＝B 0 ∑y+B 1 ∑xy+B 2 ∑y 2
wherein X represents n marker plane coordinates; a is that 0 、A 1 、A 2 、B 0 、B 1 、B 2 Is an affine transformation coefficient; n is the number of markers; y represents n marker plane coordinates; xX represents the product of the pixel abscissa X of the marker and its plane abscissa X; yX represents the product of the pixel ordinate y of the marker and its plane abscissa X; xY is the product of the pixel abscissa x of the marker and the plane ordinate Y thereof; yY represents the product of the pixel ordinate Y of the marker and its plane ordinate Y.
Compared with the prior art, the invention has the following advantages and technical effects:
(1) According to the invention, the feature extraction network ResNet 50 of the Mask R-CNN model is replaced by the MobileNet V2 with fewer calculation parameters, so that the model size is reduced; in a Mask prediction branch, replacing a Mask Head branch of a Mask R-CNN model with a PointRend model, solving the problem of fuzzy segmentation result, and improving the boundary segmentation precision;
(2) Aiming at the problem that the prediction result of the model in a narrow ridge area and a ridge breaking and connecting area caused by rolling of an agricultural machine is incomplete, a farmland boundary closing algorithm based on morphological processing is adopted, and farmland boundary closing is realized by carrying out morphological operations such as multiple expansion and corrosion on a segmented incomplete boundary, so that the method has important significance in acquiring accurate positioning information of the boundary and an agricultural machine positioning system based on machine vision;
(3) The deviation generated by the coordinate registration is within the centimeter level, the provided farmland area ground feature coordinate automatic acquisition mode can meet the requirement of unmanned agricultural machinery operation on map precision, can provide priori information for agricultural machinery operation path planning and obstacle perception, and reduces the requirement of unmanned application on single machine intellectualization.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application, illustrate and explain the application and are not to be construed as limiting the application. In the drawings:
fig. 1 is a flowchart of a farmland boundary positioning information extraction method based on a low-altitude remote sensing image of an unmanned aerial vehicle according to an embodiment of the present invention;
FIG. 2 is a diagram of a MobileNet V2-FPN-PointRend network according to an embodiment of the present invention;
FIG. 3 is a schematic view of farmland and segmentation effect according to an embodiment of the present invention;
FIG. 4 is a schematic view of farmland and boundary closure effect according to an embodiment of the present invention;
FIG. 5 is a schematic diagram showing the effect of farmland and boundary coordinates on mapping software according to an embodiment of the present invention;
fig. 6 is a schematic diagram of conversion between a planar coordinate system and an image pixel coordinate system according to an embodiment of the present invention.
Detailed Description
It should be noted that, in the case of no conflict, the embodiments and features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the accompanying drawings in conjunction with embodiments.
It should be noted that the steps illustrated in the flowcharts of the figures may be performed in a computer system such as a set of computer executable instructions, and that although a logical order is illustrated in the flowcharts, in some cases the steps illustrated or described may be performed in an order other than that illustrated herein.
The invention provides a farmland boundary positioning information extraction method based on an unmanned aerial vehicle low-altitude remote sensing image, as shown in fig. 1, comprising the following steps:
step one, collecting and marking low-altitude farmland picture data;
s11, data acquisition: in the embodiment, a large-scale Phantom 4RTK professional unmanned aerial vehicle is utilized to collect low-altitude farmland picture samples, and sample scenes are as many as possible;
in this embodiment, the unmanned aerial vehicle platform is used to collect a low-altitude farmland image, the collected image is in a JPEG format, distortion calibration is performed on the farmland image, a two-dimensional orthophoto is constructed by using professional mapping software, the regular grid cuts the image to a farmland image with the image data width of 2048 pixels and the height of 2048 pixels, because farmland boundary detection belongs to relatively fine image processing work. Such as some similar problems to ridge weeds, too low image pixels can be troublesome to label work, thereby affecting the final training accuracy.
S12, data marking: carrying out pixel-level labeling on the farmland ridges by using a Labelme labeling tool in a manual visual dotting mode, generating a file in a json format after labeling is completed, carrying out data processing on the file in the json format to generate a label graph, and finishing the label graph into a data format which can be input into a MobileNet V2-FPN-PointRend semantic segmentation network; because there are weeds, trees, illumination shadow, barriers such as wire pole in unmanned aerial vehicle low altitude collection farmland picture data, there is different degree to farmland field piece to shelter from, need walk around all barriers when annotating to do not extract the characteristic of the extraneous information outside the farmland field piece in the characteristic extraction process, and then influence the detection effect of algorithm to farmland region.
The process for processing the label graph comprises the following steps: and storing the marked low-altitude farmland pictures as a label file in a json format, and then converting the label file in the json format, wherein the specific operation is as follows: the json file is output to a computer, and a Python programming language script is written to process the json file which is input to the computer, wherein the processing process is as follows: creating a folder and naming the folder as field, wherein the folder named field is used for storing a tag file of field, and an image of the folder is training data; processing the json file by means of a labelme module package of Python, and analyzing a farmland image sample and a json tag graph corresponding to the farmland image sample; the analyzed farmland image sample is converted into a PNG format of the farmland image sample by using a PIL module of Python, the pixel width is adjusted to 512 and the pixel height is adjusted to 1024, the original image size is properly reduced under the condition of not sacrificing details, and the training speed of a later network can be accelerated.
The labelme labeling software is selected, and the labelme has the advantages that a labeling area can be freely circled according to the requirement, the label of the farmland field is named as field, and the labeled json file is output to a computer; selecting and writing Python as a programming language for processing json files input into a computer, wherein the processing steps are as follows:
1. processing the json format file by means of a labelme module package of Python, and analyzing an original image and a label image;
2. the analyzed original image is converted into a PNG format by using a PIL module of Python, pixels are changed into 512-width pixels and 1024-height pixels, the size of the original image is properly reduced under the condition of not sacrificing details, and the training speed of a later-stage deep neural network is accelerated;
3. the pictures of which 70% are taken as training set, 20% are taken as verification set, and the rest 10% are taken as test set. The picture names of the training set and the verification set are respectively stored in notepad files named as 'train. Txt', 'eval. Txt'.
Dividing a data set, and building a MobileNet V2-FPN-PointRend semantic segmentation network model;
s21, dividing a data set:
s22, building a MobileNet V2-FPN-PointRend semantic segmentation network model (as shown in figure 2): the MobileNet V2-FPN-PointRend semantic segmentation model includes: convolution layer, mobileNet V2 feature extraction network, feature pyramid structure (FPN), region suggestion network (RPN), ROI alignment layer module (ROIAlign), pointrand module.
Training a MobileNet V2-FPN-PointRend semantic segmentation network by using the obtained farmland semantic segmentation data set; FIG. 3 is a schematic view of farmland and segmentation effect.
S31, setting semantic segmentation network training parameters;
s32, setting a cross entropy loss function as a loss function of a model, and calculating the loss function between the feature map and the label map, wherein the calculation formula is as follows:
wherein N is the sum of the pixel numbers in a batch size, m is the category number, and in this embodiment, the farmland image is divided into farmland and background types, so m is 2; y is ij The pixel point i is marked on the category j, if the pixel point i is a farmland category, the pixel point i is 1, and if the pixel point i is a farmland category, the pixel point i is 0;the probability that pixel i is of class j is represented.
S34, optimizing network parameters by using a Momentum optimization algorithm, wherein the updating mode is poly, the input image size is 512 multiplied by 1024, the number of batch processing data is set to be 4, and the total iteration generation is 2000 times; parameter optimization is carried out by adopting a random gradient descent algorithm with a momentum factor of 0.9, the initial learning rate is set to be 0.001, and the weight attenuation rate is set to be 0.0001;
s35, training a MobileNet V2-FPN-PointRend semantic segmentation network.
The semantic segmentation network model comprises:
feature extraction network: the method comprises the steps of (1) a MobileNet V2 and a pyramid structure (Feature Pyramid Networks, FPN), wherein the MobileNet V2 basic unit is a depth separable convolution, the MobileNet V2 extracts features from an image, and the pyramid structure integrates low-resolution and high-semantic features of a bottom network and high-resolution and low-semantic features of a top network;
regional advice network: the integrated feature map is transmitted to a regional suggestion network, and the network adopts a sliding window method to generate anchor frames with various sizes on the feature map so as to identify regions of interest on different scales;
ROI alignment layer module: the integration region suggests a region of interest generated by the network, and the ROI aligns layer integration region characteristics;
PointRend module: and carrying out self-adaptive feature pooling on the acquired fine granularity features by utilizing bilinear interpolation, finally combining the fine granularity features with local mask prediction of all sampling points, further combining the fine granularity features with global prediction, and ensuring accurate target segmentation by applying a multi-layer perceptron. The region features integrated by the ROI alignment layer module are input into the PointRend module to be classified into fine granularity, and fine granularity features are generated.
Step four, verifying a training result based on the test data set and the trained model;
s41, firstly, a mobile Net V2-FPN-PointRend model adopts a migration learning method, a mobile Net V2 weight initialization network pre-trained based on a Cityspace public data set is loaded, then scene analysis is carried out on an input farmland data set, a sample space is transformed into a feature space, pixel features are extracted by using a mobile Net V2 backbone network and a feature pyramid FPN, the combined feature map is input into a regional suggestion network (Region Proposal Network, RPN) to obtain regions of interest with different sizes, bilinear interpolation upsampling is carried out on the output fine-grained feature map by an ROI alignment Layer (ROIAlign) gathering region features, a Multi-Layer Perceptron (MLP) subdivision prediction is adopted, a feature map with 2 channels is output, each channel represents different categories, channel 0 represents a background, and channel 1 represents farmland blocks;
s42, comparing the prediction probabilities of the same pixel point under two categories, if the prediction probability of the farmland block is higher than a preset threshold value of 0.5, judging the pixel as a farmland pixel, and if the prediction probability of the farmland block is lower than the preset threshold value of 0.5, otherwise, outputting a segmentation mask;
s43, evaluating the MobileNet V2-FPN-PointRend semantic segmentation model by adopting an image segmentation evaluation index mIoU (average cross-over ratio). Wherein, the calculation formula is as follows:
where k+1 represents the number of categories, the subject of this example includes a farm category, so k=1, p ii Representing the true value as i, the predicted value as i pixels, p ij Representing the number of pixels with a true value i and a predicted value j, and p ji Representing the number of pixels for which the true value is j and the predicted value is i.
Step five, extracting farmland boundaries from the verification results based on an edge extraction algorithm, and obtaining pixel coordinates of the farmland boundaries;
s51, binarizing the mask image output by the semantic segmentation model;
s52, extracting a binarized image boundary by adopting a Canny edge detection algorithm;
s53, adopting an eight-neighborhood boundary tracking algorithm to extract pixel coordinates of the segmentation mask.
And step six, constructing the conversion relation between the pixel coordinates and the geographic coordinates based on an affine transformation algorithm, and obtaining farmland geographic coordinates.
S61, data acquisition: the longitude and latitude coordinates of the marker were measured using a mobile station module of a span navigation N5 (inertial navigation version) GNSS satellite positioning system by placing the mobile station on a tensile centering bar, holding the bar directly above the marker by a tester and keeping the bubble level horizontal. Each point was measured for 5s and the measured value for each point was averaged.
S62, converting longitude and latitude coordinates acquired by a GNSS satellite positioning system into plane coordinates based on Gaussian-Kelvin projection transformation, wherein the calculation formula is as follows:
in the above formula: l=l-L 0 ；t＝tanB；When l=0, the meridian arc length from the equator is calculated as:
X＝C 0 B-cosB(C 1 sinB+C 2 sin 3 B+C 3 sin 5 B)
wherein: c (C) 0 、C 1 、C 2 、C 3 The values of (2) are related to the ellipsoidal parameters.
Wherein B represents latitude, L represents longitude, x represents abscissa, y represents ordinate, L 0 For the central meridian longitude, N is the radius of curvature of the circle of the mortise, e is the second eccentricity, l is the arc length of the longitude of the point and the equator of the central meridian distance, and X represents the meridian arc length from the equator; t=tan B (B represents latitude); a. c (C) 0 、C 1 、C 2 、C 3 The method comprises the following specific parameters:
| a | C 0 | C 1 | C 2 | C 3 |
| 6378137.0 | 6367499.15 | 32009.82 | 133.96 | 0.69755 |
s63, calculating a global coordinate conversion coefficient based on affine transformation, wherein the calculation formula is as follows:
∑X＝A 0 n+A 1 ∑x-A 2 ∑y
∑xX＝A 0 ∑x+A 1 ∑x 2 -A 2 ∑xy
∑yX＝A 0 ∑y+A 1 ∑xy-A 2 ∑y 2
∑Y＝B 0 n+B 1 ∑x+B 2 ∑y
∑xY＝B 0 ∑x+B 1 ∑x 2 +B 2 ∑xy
∑yY＝B 0 ∑y+B 1 ∑xy+B 2 ∑y 2
wherein X represents n marker plane coordinates; a is that 0 、A 1 、A 2 、B 0 、B 1 、B 2 Is an affine transformation coefficient; n is the number of markers; y represents n marker plane coordinates; xX represents the product of the pixel abscissa X of the marker and its plane abscissa X; yX represents the product of the pixel ordinate y of the marker and its plane abscissa X; xY is the product of the pixel abscissa x of the marker and the plane ordinate Y thereof; yY represents the product of the pixel ordinate Y of the marker and its plane ordinate Y.
A 0 、A 1 、A 2 、B 0 、B 1 、B 2 The method is solved for affine transformation coefficients (to be evaluated):
the method comprises the steps of (1) controlling a marker in a test area, measuring longitude and latitude coordinates of the marker by using a GNSS satellite positioning system, and converting the longitude and latitude coordinates of the marker into X and Y under a plane projection coordinate system by using Gaussian-gram projection;
from the planar coordinate system and image pixel coordinate system conversion schematic diagram (fig. 6), it can be deduced that:
X＝A 0 +A 1 x-A 2 y
Y＝B 0 +B 1 x+B 2 y
x is the abscissa of the marker plane, Y is the ordinate of the marker plane, X is the abscissa of the marker pixel, and Y is the ordinate of the marker pixel; a is that 0 、A 1 、A 2 ；B 0 、B 1 、B 2 For affine transformation coefficients (to be evaluated), the minimum sum of squares of the distance differences between the measured values and the calculated values of the formulas is used as a criterion for solving the optimal parameters, and can be deduced:
∑X＝A 0 n+A 1 ∑x-A 2 ∑y
∑xX＝A 0 ∑x+A 1 ∑x 2 -A 2 ∑xy
∑yX＝A 0 ∑y+A 1 ∑xy-A 2 ∑y 2
∑Y＝B 0 n+B 1 ∑x+B 2 ∑y
∑xY＝B 0 ∑x+B 1 ∑x 2 +B 2 ∑xy
∑yY＝B 0 ∑y+B 1 ∑xy+B 2 ∑y 2
obtaining geographic coordinates of a farmland boundary:
X geography ＝A 0 +A 1 u-A 2 v
Y Geography ＝B 0 +B 1 u+B 2 v
X Geography 、Y Geography Respectively representing the geographical coordinates (to be evaluated) of the boundary of the farmland, u and v respectively representThe horizontal and vertical coordinates of boundary pixels of farmland are shown, A 0 、A 1 、A 2 ；B 0 、B 1 、B 2 The conversion relationship between the plane coordinates and the pixel coordinates of the passing marker can be obtained.
S64, evaluating by adopting GB/T17160-2008 of 1:500 1:1 000 1:2 000 topography digital standard, determining the position and the size error of the marker, and verifying the precision. Wherein, the calculation formula is as follows:
P＝P′ i(x,y) -P″ i(x,y)
wherein P' i(x,y) The representing marker is obtained by a GNSS satellite positioning system and is projected and transformed by Gaussian-Kelvin to obtain plane coordinates, P i(x,y) Representing the planar coordinates of the markers acquired by the linear registration algorithm.
Fig. 4 is a schematic view of farmland and boundary closure effects, and fig. 5 is a schematic view of farmland and boundary coordinates effects on mapping software.
The traditional farmland segmentation method comprises the technical means of artificial visual dotting, random forest algorithm optimization texture characteristics and the like, and has the defects of long time consumption, low interpretation precision, resource waste, insufficient precision and the like in farmland type extraction; the invention provides improvement based on Mask R-CNN segmentation network model, and the segmentation efficiency and the interpretation accuracy are improved to a certain extent by applying farmland image segmentation; the intelligent degree is high, the universality is good, and the technical problems of low efficiency, long time consumption and poor accuracy of the conventional farmland segmentation method are solved;
the method is more efficient and accurate in farmland prediction result, provides an important basis for further obtaining high-precision farmland boundary positioning information and constructing high-precision maps of a plurality of farmland in a larger area, and plays a positive promoting role in promoting efficient and accurate farmland cultivation informatization management.
The foregoing is merely a preferred embodiment of the present application, but the scope of the present application is not limited thereto, and any changes or substitutions easily conceivable by those skilled in the art within the technical scope of the present application should be covered in the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.
Claims (10)
1. The farmland boundary positioning information extraction method based on the unmanned aerial vehicle low-altitude remote sensing image is characterized by comprising the following steps of:
constructing a low-altitude farmland image data set, and dividing the data set into a training set and a verification set;
training an initial semantic segmentation network model based on the training set, and verifying the trained semantic segmentation network model through the verification set to obtain a semantic segmentation network model;
processing the target low-altitude farmland image by utilizing the semantic segmentation network model to obtain a farmland mask image, and extracting farmland boundaries of the farmland mask image to obtain farmland boundary pixel coordinates;
and constructing a conversion relation between the farmland boundary pixel coordinates and the boundary geographic coordinates based on an affine transformation method, and obtaining farmland boundary geographic coordinates.
2. The method for extracting farmland boundary positioning information based on the unmanned aerial vehicle low-altitude remote sensing image according to claim 1, wherein constructing a low-altitude farmland image dataset comprises:
and acquiring a low-altitude farmland image by using an unmanned aerial vehicle platform, carrying out distortion calibration on the low-altitude farmland image, constructing a two-dimensional orthophoto, adjusting the image to be of the same size specification, and constructing the low-altitude farmland image data set.
3. The method for extracting farmland boundary positioning information based on unmanned aerial vehicle low-altitude remote sensing images according to claim 2, wherein the steps of cutting and labeling the low-altitude farmland image dataset before dividing the training set and the verification set include:
and labeling the low-altitude farmland image by a manual visual dotting mode based on a labeling tool, generating a label graph with the same format, and performing data processing on the label graph to obtain a data format which can be input into the semantic segmentation network model.
4. The method for extracting farmland boundary positioning information based on the unmanned aerial vehicle low-altitude remote sensing image according to claim 1, wherein the semantic segmentation network model comprises:
feature extraction network: the method comprises the steps of extracting features from an image through a MobileNet V2 structure, and integrating low-resolution and high-semantic features of a bottom network with high-resolution and low-semantic features of a top network through a pyramid structure, wherein a basic unit of the MobileNet V2 structure is a depth separable convolution;
regional advice network: the method comprises the steps of generating anchor frames with different sizes on a feature map by adopting a sliding window method for the integrated features so as to identify regions of interest on different scales;
ROI alignment layer module: the region of interest generated by the region suggestion network is integrated, and the ROI alignment layer integrates region characteristics;
PointRend module: the method is used for carrying out self-adaptive feature pooling on the acquired fine granularity features by utilizing bilinear interpolation, combining the fine granularity features with local mask prediction of all sampling points, further combining the fine granularity features with global prediction, and carrying out target segmentation based on a multi-layer perceptron.
5. The method for extracting farmland boundary positioning information based on unmanned aerial vehicle low-altitude remote sensing images according to claim 4, wherein training an initial semantic segmentation network model based on the training set comprises:
setting training parameters of the initial semantic segmentation network model, setting a cross entropy loss function as a loss function of the initial semantic segmentation network model, and calculating a loss function between a feature map and a label map;
optimizing network parameters by using a Momentum optimization method, and setting the number of batch processing data and total iteration generation;
and performing parameter optimization by adopting a random gradient descent method, setting an initial learning rate and a weight attenuation rate, and training the initial semantic segmentation network model.
6. The method for extracting farmland boundary positioning information based on the unmanned aerial vehicle low-altitude remote sensing image according to claim 5, wherein the method for calculating the loss function between the feature map and the tag map is as follows:
where N is the sum of the number of pixels in a batch size, m is the number of categories, y ij Representing the labeling of pixel i for category j,the probability that pixel i is of class j is represented.
7. The method for extracting farmland boundary positioning information based on the unmanned aerial vehicle low-altitude remote sensing image according to claim 1, wherein verifying the trained semantic segmentation network model through the verification set comprises:
initializing a semantic segmentation network model by loading a MobileNet V2 weight pre-trained based on a Cityspaces public data set, performing scene analysis on an input farmland data set, transforming a sample space into a feature space, extracting pixel features by using a feature extraction network, inputting a combined feature map into a region suggestion network to obtain regions of interest with different sizes, gathering region features by an ROI alignment layer module, performing bilinear interpolation up-sampling on an output fine-grained feature map by a PointRend module, and performing subdivision prediction by using a multi-layer perceptron to output a feature map;
comparing the prediction probabilities of the same pixel point in the feature map under different categories, judging the pixel as a farmland pixel if the prediction probability of the farmland block is higher than a preset threshold value, judging the pixel as a background pixel if the prediction probability of the farmland block is lower than the preset threshold value, and outputting a segmentation mask;
and verifying the segmentation mask by adopting an image segmentation evaluation index mIoU average cross-correlation ratio to obtain the verification result.
8. The method for extracting farmland boundary positioning information based on the unmanned aerial vehicle low-altitude remote sensing image according to claim 7, wherein the method for verifying the segmentation mask is as follows:
wherein k+1 represents the number of categories, p ii Representing the number of pixels with a true value i and a predicted value i, p ij Representing the number of pixels, p, with a true value i and a predicted value j ji Representing the number of pixels for which the true value is j and the predicted value is i.
9. The method for extracting farmland boundary positioning information based on the unmanned aerial vehicle low-altitude remote sensing image according to claim 8, wherein obtaining the farmland boundary pixel coordinates comprises:
performing binarization processing on the farmland mask image output by the semantic segmentation model to obtain a binarized image, and extracting a binarized image boundary by adopting Canny edge detection and calculation;
and extracting boundary pixel coordinates of the segmentation mask by an eight-neighborhood boundary tracking method to obtain the farmland boundary pixel coordinates.
10. The method for extracting farmland boundary positioning information based on the unmanned aerial vehicle low-altitude remote sensing image according to claim 1, wherein the method for obtaining the farmland boundary geographic coordinates is as follows:
∑X＝A 0 n+A 1 ∑x-A 2 ∑y
∑xX＝A 0 ∑x+A 1 ∑x 2 -A 2 ∑xy
∑yX＝A 0 ∑y+A 1 ∑xy-A 2 ∑y 2
∑Y＝B 0 n+B 1 ∑x+B 2 ∑y
∑xY＝B 0 ∑x+B 1 ∑x 2 +B 2 ∑xy
∑yY＝B 0 ∑y+B 1 ∑xy+B 2 ∑y 2
wherein X represents n marker plane coordinates; a is that 0 、A 1 、A 2 、B 0 、B 1 、B 2 Is an affine transformation coefficient; n is the number of markers; y represents n marker plane coordinates; xX represents the product of the pixel abscissa X of the marker and its plane abscissa X; yX represents the product of the pixel ordinate y of the marker and its plane abscissa X; xY is the product of the pixel abscissa x of the marker and the plane ordinate Y thereof; yY represents the product of the pixel ordinate Y of the marker and its plane ordinate Y.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311830023.0A CN117788822A (en) | 2023-12-28 | 2023-12-28 | Farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311830023.0A CN117788822A (en) | 2023-12-28 | 2023-12-28 | Farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117788822A true CN117788822A (en) | 2024-03-29 |
Family
ID=90385136
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311830023.0A Pending CN117788822A (en) | 2023-12-28 | 2023-12-28 | Farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117788822A (en) |
-
2023
- 2023-12-28 CN CN202311830023.0A patent/CN117788822A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11403846B2 (en) | Crop boundary detection in images | |
| CN110263717B (en) | Method for determining land utilization category of street view image | |
| Kwong et al. | Tree height mapping and crown delineation using LiDAR, large format aerial photographs, and unmanned aerial vehicle photogrammetry in subtropical urban forest | |
| Liu et al. | Establishing a citywide street tree inventory with street view images and computer vision techniques | |
| Hecht et al. | Automatic derivation of urban structure types from topographic maps by means of image analysis and machine learning | |
| Veljanovski et al. | Object-based image analysis of VHR satellite imagery for population estimation in informal settlement Kibera-Nairobi, Kenya | |
| CN112560623B (en) | Unmanned aerial vehicle-based rapid mangrove plant species identification method | |
| CN113343858B (en) | Road network geographic position identification method and device, electronic equipment and storage medium | |
| CN109657728B (en) | Sample production method and model training method | |
| Ming et al. | Cropland extraction based on OBIA and adaptive scale pre-estimation | |
| CN115527027A (en) | Remote sensing image ground object segmentation method based on multi-feature fusion mechanism | |
| Helmholz et al. | Semi-automatic quality control of topographic data sets | |
| CN116822980A (en) | Base vector guided typical earth surface element intelligent monitoring method and system | |
| CN117788822A (en) | Farmland boundary positioning information extraction method based on unmanned aerial vehicle low-altitude remote sensing image | |
| CN109376638A (en) | A kind of literary ground rate calculation method based on remote sensing images and GIS-Geographic Information System | |
| CN115205704A (en) | High-resolution remote sensing image small sample high-precision building segmentation and extraction method and device | |
| Guo et al. | Population estimation in Singapore based on remote sensing and open data | |
| CN117115566B (en) | Urban functional area identification method and system by utilizing full-season remote sensing images | |
| Carpanese et al. | Automated estimation of building height through image processing | |
| Zhang et al. | Assessing Urban Park Open Space by Semantic Segmentation of Geo-tagged Panoramic Images | |
| Francis et al. | Estimating Chicago's tree cover and canopy height using multi-spectral satellite imagery | |
| Zhou | Areal uncertainty in traffic island polygons extracted from airborne laser point clouds | |
| Lin et al. | MLS-assisted validation of WorldView-2 panchromatic image for estimating Pinus sylvestris crown height | |
| Ghiasvand Nanji | Detecting urban tree canopy using convolutional neural networks with aerial images and LiDAR data | |
| Gustin et al. | Mapping the Suitability of Land Use to the Regional Spatial Plans in Batam City |
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 |