CN115937698A - Self-adaptive tailing pond remote sensing deep learning detection method - Google Patents
Self-adaptive tailing pond remote sensing deep learning detection method Download PDFInfo
- Publication number
- CN115937698A CN115937698A CN202211212391.4A CN202211212391A CN115937698A CN 115937698 A CN115937698 A CN 115937698A CN 202211212391 A CN202211212391 A CN 202211212391A CN 115937698 A CN115937698 A CN 115937698A
- Authority
- CN
- China
- Prior art keywords
- tailing pond
- features
- feature
- positioning
- remote sensing
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The self-adaptive remote sensing deep learning detection method for the tailing pond is used for identifying and extracting the tailing pond in a large range. The invention provides a brand-new remote sensing extraction framework of a tailing pond, which comprises the following steps: firstly, performing depth feature learning of tailing ponds with different scales by adopting a feature pyramid network based on a tailing pond remote sensing image data set; secondly, positioning the tailing pond according to the characteristic values by adopting a cascade positioning optimization module and optimizing positioning coordinates by using a multi-stage detector; thirdly, acquiring a candidate detection object according to the positioning optimization value, and extracting and fusing multi-dimensional features based on a deep learning algorithm and a traditional feature extraction algorithm to generate a feature vector; and finally, performing ensemble learning and classification by adopting a self-adaptive feature learning classifier based on the feature vector, removing the false alarm target and extracting the tailing pond. The invention has the advantages of high positioning precision, low false alarm rate, wide application range and the like, and can provide important basis for the dynamic update and the safety management of the distribution information of the tailing pond.
Description
Technical Field
The invention belongs to the technical field of mine safety monitoring and remote sensing image processing, and particularly relates to a remote sensing deep learning detection method for a tailing pond by coupling cascade optimization and characteristic self-adaption. In the safety control of the tailing pond, the accurate position information of the tailing pond in each area is extracted, and the method can be used for scientifically carrying out safety monitoring and risk assessment work of the tailing pond.
Background
The tailings pond is a tailings resource place for storing industrial production of mines, is usually constructed by artificial dams, is a debris flow danger source with high potential energy, and has serious harm to the surrounding environment due to the fact that the instability probability of the tailings pond rises year by year along with the year accumulation of tailings and accumulated water. Due to the requirement of mineral resources, tailings ponds of different grades are increased from place to place year by year, the large tailings pond has the requirement of tailings dam safety assessment all the year round due to high potential energy, the small tailings pond is small in scale and low in threshold, the problems of low safety level and tailings leakage accidents are easy to happen, and therefore accurate acquisition of spatial distribution of the tailings pond and information updating are very important for scientific management of the tailings pond.
In the existing research, various feature analysis and tailing pond detection methods are provided in the technical field of remote sensing image processing, and the technical barrier of tailing pond extraction is successfully broken through. The existing methods are mainly divided into two branches: the method comprises the steps of constructing a tailing pond feature extraction model based on a traditional machine learning method and performing a target extraction method based on a deep learning framework. A large amount of researches are mainly carried out on scene composition characteristics of a tailing pond based on a traditional machine learning method, wanhua and the like use DT-tailings rules to carry out extraction of a single tailing pond in a high-resolution one-number remote sensing image based on a decision tree mode, B.D.Ma and the like use principal component analysis and a rapid atmospheric correction method to carry out processing on remote sensing images and extraction of spectral characteristics of the tailing pond, bear literature and the like carry out characteristic analysis on a manganese ore tailing pond and qualitatively analyze scattering characteristics of the manganese ore tailing pond and differences between a Synthetic Aperture Radar (SAR) and an optical image, and an object-oriented characteristic learning method is explored by utilizing a space and structure mode. The Wangchang et al analyzes the spectral, texture and spatial geometrical characteristics of the tailings and solid waste in the Quickbird data set, and then determines the optimal segmentation threshold of the tailings and the solid waste. The research obtains great achievement in the aspect of analyzing the characteristics of the tailing pond, but the extraction methods can only be realized by manually adjusting the threshold and segmenting the image, and have no autonomy and flexibility. With the development of a deep learning framework in the field of artificial intelligence and strong remote sensing image processing capacity thereof, part of target detection frameworks are primarily tried to be applied to remote sensing detection of a tailing pond, single-stage models such as a YOLO v4 and an SSD detection method, double-stage models such as Faster R-CNN and feature pyramid and attention mechanism-based improvement models are used for learning and extracting features of the tailing pond in a remote sensing image, and in addition, in order to solve the difficulty of site selection and boundary extraction of the tailing pond, a multi-task branch network (MTBunet) is proposed to optimize coordinates of a detection candidate frame, so that the precision of the tailing pond detection model is further improved.
By the analysis of the method and the evaluation of the existing research method, the high-precision large-scale remote sensing detection of the tailing pond is still a difficult task. Different from a conventional homogeneous target detection object, tailing ponds in different areas and different levels are different in size and irregular in shape boundaries, which provides challenges for feature learning of remote sensing images, and the problem of inaccurate positioning often exists due to object scale diversity in the traditional remote sensing image detection of directly transferring a well-represented frame for detecting fixed-shape objects (such as tanks, ships and airplanes) to a large-scale tailing pond; in addition, due to the fact that scene composition of tailing ponds at different stages is different, characteristics cannot be effectively learned, and the detection results are low in identification accuracy due to false detection of more similar targets. By combining the above methods, two problems in the existing tailing pond detection are still not effectively solved: 1) The positioning accuracy is low, the position and category information of an object can be output by the existing detection method in a mode of a region suggestion frame, and when the positioning accuracy of a boundary frame is low, misjudgment is generally carried out to lose a plurality of real detection targets; 2) The false alarm rate is high, the scene of the tailing pond in remote sensing images is complex, and in the detection result of a deep learning framework, the false alarm rate is usually accompanied by a false alarm target.
In recent years, the application of artificial intelligence technology brings new development opportunities to various industries. The rapid development of artificial intelligence technology, especially deep learning technology, has also become an important driving force for the development of surveying and mapping remote sensing subject. With the development of satellite commercialization at home and abroad and the popularization of unmanned aerial vehicles, the remote sensing data volume is rapidly increased, the spatial resolution, the temporal resolution and the spectral resolution are continuously improved, basic data guarantee is provided for remote sensing monitoring and other multi-field application, new direction is brought for rapid, accurate and automatic identification of remote sensing targets, and the development trend in the future is also provided. The extraction of target ground feature information in remote sensing data, such as attribute information, position range, change information and the like, is realized, but the reliability of the information is influenced by the limitations of cognitive level and technology. The traditional remote sensing target identification method is developed into an intelligent remote sensing target identification method, the method has the advantages of high data updating frequency, time saving and labor saving, further method improvement aiming at the characteristics of the tailing pond is worth reproducing the existing method, and the identification precision and the robustness of the method are improved.
Disclosure of Invention
The technical problems to be solved by the invention are as follows: the remote sensing deep learning detection method for the tailing pond coupled with cascade optimization and feature self-adaption is provided, so that extraction of the tailing pond in a large range can be accurately realized with high precision, distribution conditions of the tailing pond can be accurately reflected, and scientific reference basis is provided for dynamic updating and safety management of regional information of the tailing pond.
The invention adopts the following technical scheme for solving the technical problems:
the invention provides a self-adaptive remote sensing deep learning detection method for a tailing pond, which is mainly realized by a multi-scale feature extraction network, a cascade positioning optimization module, a multi-dimensional feature extraction and fusion module and a feature self-adaptive learning classifier, wherein: the multi-scale feature extraction network firstly realizes extraction and calculation of scene features of tailings ponds of different sizes in remote sensing images, then a cascade positioning optimization module continuously optimizes coordinates of a detection frame with low precision based on a feature map obtained by calculation of the multi-scale feature extraction network through a regression detector to realize coordinate positioning of the tailings ponds, then a multi-dimensional feature extraction and fusion module obtains typical features of the tailings ponds from multiple dimensions and fuses the typical features into feature vectors for distinguishing false alarm targets, and finally self-adaptive learning and binding of the tailings ponds features are carried out through a self-adaptive integrated learning classifier based on the multi-dimensional features to realize extraction of the tailings ponds and elimination of the false alarm targets.
In the method, a multi-stage cascade optimization remote sensing image tailing pond deep learning detection method can be adopted, and the method specifically comprises the following steps: based on the existing high-resolution remote sensing image data, the positioning optimization and classification optimization are carried out on the identification of the tailings ponds with different scales through multi-stage cascading positioning and classification, the high-precision extraction of the tailings ponds in the large-range remote sensing image is realized, and the technical support can be provided for the safety management and information update of the tailings ponds in the large-range.
The invention also provides a self-adaptive remote sensing deep learning detection method for the tailing pond, which is a tailing pond remote sensing detection method for coupling deep learning cascade optimization positioning and feature self-adaptive learning, and specifically comprises the following steps: the method comprises the steps of firstly, carrying out multi-level scene deep learning feature extraction on a remote sensing image, then using a cascade classifier to carry out positioning of the position of a tailing pond and coordinate optimization of a detection target to obtain a tailing pond detection candidate area, taking six dimensions into consideration, extracting and fusing to obtain detection object features, using a self-adaptive integrated classifier to carry out feature learning and cascade classification of the tailing pond, and obtaining a high-precision tailing pond detection result.
In the method, the position of the detection of the tailing pond is optimized based on multi-scale features and a cascade detector, and the classification of the tailing pond is optimized based on multi-dimensional feature extraction and self-adaptive feature learning, so that the tailing ponds with different grades can be detected, and false alarm targets in the detection result can be removed through the multi-dimensional features, and the method specifically comprises the following steps:
(1) Position optimization based on multi-scale feature extraction and cascade positioning:
a. inputting an original remote sensing image to be detected, performing up-sampling on the image, and acquiring a three-level image;
b. downsampling the extracted features to enlarge the visual field;
c. combining the features on two sides through linking, fusing the multi-scale features and supplementing information content;
d. inputting a feature layer to position a candidate region of the feature;
e. designing a cascaded detector to carry out positioning optimization of a tailing pond;
f. outputting the coordinates of the detection candidate frame;
(2) And (3) identification optimization based on self-adaptive multidimensional feature integrated learning classification:
a. cutting the recognition result according to the tailing pond provided by the first-step detection framework, wherein the cut object is used for further classification;
b. based on RGB three-channel remote sensing images, respectively extracting robust features of candidate set objects: geometric, probabilistic, scene, topological, spectral, textural features;
d. normalizing the features;
e. single-side classification and sampling;
f. and (4) self-adaptive feature learning strong and weak feature binding is carried out to reduce the dimension of the features, and classification rules are learned.
In the method, the following steps can be adopted:
firstly, inputting an original high-resolution remote sensing image, respectively using sliding windows with the scales of {340, 170, 84 and 43} to perform up-sampling and extracting multi-scale features, combining an FPN network and a Resnet152 to perform residual error network, and respectively using convolutional layers { C2, C3, C4 and C5} to perform feature extraction, wherein the step lengths are respectively set to {4,8, 16 and 32};
then, respectively retaining the semantic features of the scene by using a down-sampling mode;
finally, the transverse connection mode is used for fusing the features to enhance the detail information.
In the method, the candidate target can be positioned according to the extraction of the multi-scale features, and then the extraction of the positioning coordinates of the extraction of the tailing pond is continued, wherein b = (b) x ，b y ，b w ，b h ) And then, optimizing the positioning of the regression frame reclassification by using different thresholds, so that four positioning coordinates of the tailing pond are closer to the real position of the tailing pond, wherein x is the detected candidate object of the tailing pond, g is the position coordinate set of the tailing pond after the regression positioning is carried out, and each detected sample i is input (g) i ，b i ) Training an optimization model, and minimizing a loss function of the detection frame in a training process; the distance vector is also optimized in the regression process to ensure the stability of regression positioning, and the vector is specifically defined as follows:
wherein (g) x ，g y ，g w ，g h ) The four-vertex coordinates of the training sample g are indicated, (b) x ，b y ，b w ，b h ) Is the coordinate of the target tailing pond extracted by the model in the image, (delta) x ，δ y ，δ w ，δ h ) The values of the parameters are optimized at the coordinates of the four corner points.
In the method, the position can be optimized by using the cascade detector comprising the positioning optimization module, so that the coordinates of the tailings pond positioned in the first detection process can be positioned again according to the characteristics of the tailings pond.
For effective positioning optimization, a single regression function is iteratively applied in the cascade detection process of the tailing pond, and is specifically defined as follows:
where x denotes a candidate, b denotes a suggestion box,
representing the connections between each iteration, in the proposed framework a regression structure is designed based on these cascaded detectors to perform the optimization:
in the formula: t denotes the number of connected layers, the regressor f at each stage T Is optimized according to the sample distribution;
wherein for each detection frame, there is an IoU value u to evaluate and evaluate the detection frame, and for each candidate frame tailings pond, the label is seized according to u:
wherein, g y Representing the label of sample g, the threshold of IoU is defined at each stage to evaluate the effect of each stage of detector, it is difficult to realize high-precision positioning based on single detector in the detection process of tailings pond, in the proposed model, different loss functions are defined to optimize positioning, wherein the value of u is {0.5,0.6,0.7} in the original frame of R-CNN, and the classifier h is t And the regressor in each stage t, u t Is set to optimize according to the corresponding IoU threshold when u t >u t-1 The loss function is defined as follows:
L(x t ，g)＝L cls (h t (x t )，y t )+λ[y t ≥1|L loc (f t (x t ，b t )′g) (5)
wherein, b t ＝f t-1 (x t-1 ，b t-1 ) G denotes the real detection box x t λ =1 balance parameter, y t Is represented by x t Is labeled with t In the case of (c).
The method can adopt a method comprising the following steps, so that multi-dimensional geographic feature extraction including geometry, texture, spectrum, topology and the like can be carried out on the detected target for screening the correctly identified tailing pond target, and the method comprises the following steps:
firstly, cutting a candidate area according to optimized positioning coordinates to obtain a tailing pond candidate area RGB sample set,
then, extracting geometric, probability, scene, topology, spectrum and texture features of candidate set objects based on the three-channel images respectively, calculating the obtained feature values To form feature vectors correspondingly, for the ith sample, the corresponding feature vectors are Fi = { Ge1, ge2, ge3, pr1, pr2, se1, se2, se3, se4, to1, to2, to3, sp1, sp2, sp3, te1, te2, te3, te4, te5, te6, te7, te8}, wherein each element represents the height, width and height-width ratio of the sample respectively, and the probability of the image being a tailing pond and the area ratio of the tailing pond based on VGG16 is defined by image classification, the probability of the image being a tailing pond is obtained by scene segmentation based on U-Net, the ratio of the area of the whole tailing pond, the water body, the sand body and the corresponding tailing pond is obtained by segmentation, the probability ratio of the water body, the sand body and the dam body, calculating the corresponding topological relation among the R, G and B gray level values of the samples respectively, the spectral and the maximum spectral difference and the entropy of the corresponding feature matrix of the maximum correlation of the probability of the water body and the dam body of the sample, and the dam body, and the maximum correlation of the probability of the tailings of the probability of the dam are calculated by using the spectrum.
The method can adopt a method comprising the following steps, so that the extracted characteristics can be applied to the extraction of the tailing pond, and the method comprises the following steps:
firstly, sampling data of a tailing pond by using a unilateral sampling mode for an extracted data set of characteristics of the tailing pond to reduce the calculated amount of the tailing pond in a sample data set, only reserving data with a larger gradient during data sampling, and introducing a constant for a sample with a small gradient during gain calculation to balance in order to avoid the change of sample distribution caused by discarding data with a small gradient; the GOSS algorithm firstly sorts all values of the characteristics to be split in a descending order according to absolute values, and a% of data with the largest absolute values are selected; then randomly selecting b% data from the rest of the smaller gradient data; multiplying the b% of data by a constant 1-a/b, paying more attention to samples with insufficient training without changing the distribution of an original data set too much, and finally calculating the gain of information by using (a + b)%;
then, splitting the tailing pond samples by using the estimated variance gain based on the data set during classification, wherein the specific expression mode is as follows:
wherein the data set { x ] is trained for the tailings pond 1 ，…，x n Sample in (v) }, χ i Representing features χ in a multi-dimensional feature vector space S s Each instance based on an estimated variance gain
The splitting of the instance is performed at feature j and node d, based on the dataset a £ b where O represents the training dataset for the tailings reservoir at the decision tree fixed node,
coefficient->
Is used to perform a gradient sum normalization A on the data set B c The size of (d);
finally, for extracting the features representing the tailing pond, in order to reduce the calculation complexity, the original values and the offsets are added, the special features are stored in different data sets, and in order to avoid reducing the loss of precision of the recognition result caused by the number of the features, the features to be bound are selected by using a self-adaptive binding strategy; firstly, constructing a feature relation graph by referring to conflicts among the proposed features, and taking weight calculation as an edge; secondly, classifying the features according to the degree descending order of the features in the graph; finally, each feature is examined in the ordered list and evaluated to determine whether to assign to an existing bundle with small conflicts or create as a new bundle.
The method can adopt a method comprising the following steps:
firstly, inputting a remote sensing image to be detected, carrying out target detection on the remote sensing image to be detected, cutting a tailing pond sample according to a model candidate area coordinate, and generating a candidate set of a model;
then, the multi-dimensional feature calculation is carried out on the candidate set, the feature vector and the label of the candidate object are made,
and finally, carrying out sample reclassification on the candidate set by using an adaptive integrated classifier based on the characteristic values, screening and extracting the detected tailings pond target, and carrying out drawing.
Compared with the prior art, the invention has the following main advantages:
(1) The invention makes a remote sensing identification data set of the tailing pond based on Google 16-grade remote sensing images, firstly, a multi-scale feature extraction network realizes extraction and calculation of scene features of tailing ponds with different sizes in the remote sensing images, and a cascading positioning optimization module realizes coordinate positioning of the tailing pond based on feature maps obtained by calculation of the multi-scale feature extraction network through continuously optimizing coordinates of detection frames with lower precision by different regression detectors, thereby completing a high-precision positioning task of the tailing pond.
(2) According to the invention, six types of multi-dimensional features including geometry, probability, scene, topology, space and texture are carried out on a tailing pond candidate set generated by target monitoring to obtain typical features of a tailing pond and are fused into feature vectors for distinguishing false alarm targets, and finally, the analysis and binding of the tailing pond features are carried out on the basis of a multi-dimensional feature self-adaptive integrated learning classifier to realize the reclassification of the tailing pond, so that the false alarm targets in the detection result are eliminated, and the category precision in the detection result is improved.
(3) The invention can realize the training of a tailing pond detection model and a classification model under the support of a high-resolution remote sensing image tailing pond sample, can realize the high-precision detection of a regional tailing pond based on the original data of a large-range high-resolution remote sensing image, has higher regional applicability and robustness in the migration application of the model due to the wide distribution of the sample, has higher accuracy and precision of a recognition result, and can be applied to the detection and drawing of the large-range tailing pond.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a schematic diagram of different IoU threshold modifications.
Fig. 3 is a schematic diagram of sample distribution in a tailings pond area.
Fig. 4 is a schematic diagram of the positioning result of the tailings pond.
FIG. 5 is a schematic representation of feature importance.
FIG. 6 is a schematic diagram of an exemplary zone application.
Detailed Description
The invention provides a remote sensing deep learning detection method for a tailing pond by coupling cascade optimization and characteristic self-adaptation, which is used for identifying and extracting a large-range tailing pond. The invention provides a brand-new remote sensing extraction framework of a tailing pond, which comprises the following steps: firstly, performing depth feature learning of tailing ponds with different scales by adopting a feature pyramid network based on a tailing pond remote sensing image data set; secondly, positioning the tailing pond according to the characteristic values by adopting a cascade positioning optimization module and optimizing positioning coordinates by using a multi-stage detector; thirdly, acquiring a candidate detection object according to the positioning optimization value, and extracting and fusing multi-dimensional features based on a deep learning algorithm and a traditional feature extraction algorithm to generate a feature vector; and finally, performing ensemble learning and classification by adopting a self-adaptive feature learning classifier based on the feature vector, removing the false alarm target and extracting the tailing pond. The invention has the advantages of high positioning precision, low false alarm rate, wide application range and the like, and can provide important basis for the dynamic update and the safety management of the distribution information of the tailing pond.
The present invention will be further described with reference to the following examples and drawings, but is not limited to the following.
The remote sensing deep learning detection method of the tailing pond based on the coupled cascade optimization and the characteristic self-adaptation is used for identifying and extracting the tailing pond in a large range. As shown in fig. 1: the method is mainly realized by a multi-scale feature extraction network, a cascade positioning optimization module, a multi-dimensional feature extraction and fusion module and a feature self-adaptive learning classifier. The invention has the advantages of high positioning precision, low false alarm rate, wide application range and the like, and can provide important basis for the dynamic update and the safety management of the regional information of the tailings reservoir.
The multi-scale feature extraction network can adopt the feature pyramid to carry out down-sampling on remote sensing images of the tailing pond to obtain tailing pond features of different scales, and obtains feature information with high resolution and strong semantics in a mode of up-sampling and image fusion.
The cascade positioning optimization module can adopt three cascade IoU threshold detectors which are arranged in a cascade mode and are increased layer by layer for optimizing the positioning of the detection frame.
The multi-dimensional feature extraction and fusion module can adopt an image processing algorithm to obtain the geometric features, the probability features, the scene features, the topological features, the spatial features and the textural features of the detected target image, and is used for distinguishing and characterizing different types of tailing ponds.
The feature adaptive learning classifier can sample data of the tailing pond in a unilateral sampling mode to reduce the calculated amount of the tailing pond in a sample data set; for extracting the features representing the tailing pond, the features to be bound are selected by using a self-adaptive binding strategy, the calculation complexity is reduced, the original values and the offset are added, the proprietary features are stored in different sets, and the accuracy improvement of the identification results brought by various features is ensured.
The invention provides a remote sensing deep learning detection method of a tailing pond by coupling cascade optimization and characteristic self-adaption, which provides a brand new remote sensing extraction framework of the tailing pond, and the framework comprises the following steps: firstly, inputting a remote sensing image data set to be detected into a multi-scale feature extraction network to calculate deep learning features of tailings ponds with different scales; secondly, inputting the characteristic layer into a cascade positioning optimization module, positioning the tailing pond according to the characteristic value and optimizing a positioning coordinate by using a regression model; thirdly, outputting candidate detection objects according to the positioning optimization value, and extracting and fusing multi-dimensional features based on a deep learning algorithm and a traditional feature extraction algorithm to generate feature vectors; and finally, inputting the feature vectors into a feature self-adaptive learning classifier to perform integrated learning and classification, removing false alarm targets and extracting a tailing pond.
In the method, the following method can be adopted to realize the deep learning detection of the remote sensing image tailing pond by multistage cascade optimization, and specifically comprises the following steps:
(1) Position optimization based on multi-scale feature extraction and cascade positioning:
a. inputting an original remote sensing image to be detected, performing up-sampling on the image, and acquiring a three-level image;
b. downsampling the extracted features to enlarge the visual field;
c. combining the characteristics of two sides through links, fusing the characteristics of multiple scales and supplementing information content;
d. inputting a characteristic layer to position a candidate region of the characteristic;
e. designing a cascaded detector to carry out positioning optimization of a tailing pond;
f. and outputting the coordinates of the detection candidate frame.
(2) And (3) identification optimization based on self-adaptive multi-dimensional feature ensemble learning classification:
a. and (4) cutting the recognition result according to the tailings pond provided by the first-step detection framework, wherein the cut object is used for further classification.
b. Based on the RGB three-channel remote sensing image, respectively extracting robust features of candidate set objects: geometric, probabilistic, scene, topological, spectral, textural features.
d. Normalizing the features;
e. single-side classification and sampling;
f. the self-adaptive features realize feature dimension reduction by learning strong and weak features and binding, and learning classification rules;
the method can be used for acquiring the multilevel scene characteristics of the tailing pond in the remote sensing image by the following steps:
firstly, inputting an original high-resolution remote sensing image, performing up-sampling and extracting multi-scale features by respectively using sliding windows with the scales of {340, 170, 84 and 43}, combining an FPN network and a Resnet152 to perform residual error network, and performing feature extraction by respectively using convolutional layers { C2, C3, C4 and C5}, wherein the step lengths are respectively set to {4,8, 16 and 32};
then, respectively keeping the semantic features of the scene by using a down-sampling mode;
finally, the characteristics are fused in a transverse connection mode to enhance the detail information.
The invention can also adopt the following method to obtain the candidate positioning area of the tailing pond in the remote sensing image and carry out positioning optimization:
firstly, according to the extraction of the multi-scale features, the candidate target is positioned, and then the extraction of the positioning coordinates of the extraction of the tailing pond is continued, wherein b = (b) x ，b y ，b w ，b h ) And then optimizing the positioning of the regression frame reclassification by using different thresholds, wherein x is the detected tailings pond candidate object, g is the position coordinate set of the tailings pond subjected to regression positioning, and each detected sample i is input (g) i ，b i ) Training an optimization model, and minimizing a loss function of the detection frame in the training process; the distance vector is also optimized in the regression process to ensure the stability of regression positioning, and the vector is specifically defined as follows:
wherein (g) x ，g y ，g w ，g h ) The four-vertex coordinates of the training sample g are shown, (b) x ，b y ，b w ，b h ) Is the coordinate of the target tailing pond extracted by the model in the image, (delta) x ，δ y ，δ w ，δ h ) The values of the parameters are optimized at the coordinates of the four corner points.
Then, in order to perform effective positioning optimization, a single regression function is used to be iterated in the process of cascade detection of the tailing pond, which is specifically defined as follows:
where x denotes a candidate, b denotes a suggestion box,
representing the connections between each iteration, in the proposed framework the regression structure is designed based on these cascaded detectors to perform the optimization:
wherein: t denotes the number of connected layers, the regressors at each stage, f T Is optimized according to the sample distribution.
Wherein for each detection frame, there is an IoU value u to evaluate and evaluate the detection frame, and for each candidate frame tailings pond, the label is seized according to u:
wherein, g y The label of the sample g is indicated and the threshold for IoU is defined at each level to evaluate the effect of each level of detector.
In the detection process of the tailings pond, high-precision positioning is difficult to realize only based on a single detector, in the proposed model, different loss functions are defined for optimizing the positioning, wherein the value of u is {0.5,0.6 and 0.7} in the original frame of R-CNN, and a classifier h is used for optimizing the positioning t And a regressor in eachIn stage t, u t Is set to optimize according to the corresponding IoU threshold when u is t >u t-1 The loss function is defined as follows:
L(x t ，g)＝L cls (h t (x t )，y t )+λ[y t ≥1]L loc (f t (x t ，b t )，g) (5)
wherein, b t ＝f t-1 (x t-1 ，b t-1 ) And g represents the real detection box x t λ =1 balance parameter, y t Is represented by x t In u t In the case of (1).
The invention can also adopt the following method to obtain the multidimensional characteristics of the candidate positioning area of the tailing pond:
firstly, cutting a candidate area according to optimized positioning coordinates to obtain a tailings pond candidate area RGB sample set,
then, extracting geometric, probability, scene, topology, spectrum and texture features of a candidate set object based on a three-channel image, calculating To obtain feature values, correspondingly forming feature vectors respectively, regarding an ith sample, correspondingly defining the feature vectors as Fi = { Ge1, ge2, ge3, pr1, pr2, se1, se2, se3, se4, to1, to2, to3, sp1, sp2, sp3, te1, te2, te3, te4, te5, te6, te7 and Te8}, wherein each element respectively represents the height, width and aspect ratio of the sample, and defining the probability of the image being a tailing pond and the area ratio of the tailing pond based on VGG16 image classification, carrying out scene segmentation based on U-Net To obtain the ratio of the area of the whole tailing pond, the water body, the sand body and the tailing pond corresponding To the area, calculating the mean value of R, G and B values among the samples, respectively calculating the average value, the spectral value, the maximum ratio of the spectral value and the inverse energy ratio of the three-component matrixes of the relative energy.
The invention can also adopt the following method to eliminate false alarms according to the multidimensional characteristics to obtain the target of the accurate tailing pond:
firstly, sampling data of a tailing pond by using a unilateral sampling mode for an extracted data set of characteristics of the tailing pond so as to reduce the calculated amount of the tailing pond in a sample data set, only reserving data with a larger gradient during data sampling, and introducing a constant for a sample with a small gradient during gain calculation so as to avoid change of sample distribution caused by discarding data with a small gradient. The GOSS algorithm firstly sorts all values of the characteristics to be split in a descending order according to absolute values, and a% of data with the largest absolute values are selected. Then b% of the data were randomly selected among the remaining smaller gradient data. Then multiply the b% data by a constant 1-a/b, focus more on the samples with insufficient training without changing the distribution of the original data set too much, and finally use (a + b)% to calculate the gain of the information. Thus, the algorithm will focus more on the under-trained samples without changing the distribution of the original data set too much.
Then, splitting the tailing pond samples by using the estimated variance gain based on the data set during classification, wherein the specific expression mode is as follows:
wherein the data set { x ] is trained for the tailings pond 1 ，…，x n Sample in (b) }, x i Representing features χ in a multi-dimensional feature vector space S s Each instance based on an estimated variance gain
The splitting of the instance is performed at feature j and node d, based on the dataset a £ b where O represents the training dataset for the tailings reservoir at the decision tree fixed node,
coefficient>
Is used to perform a gradient sum normalization A on the data set B c Of (c) is used.
Finally, for extracting the features representing the tailing pond, in order to reduce the calculation complexity, the original values and the offset are added, the special features are stored in different data sets, and in order to avoid the loss of precision of the identification result caused by the reduction of the number of the features, the features to be bound are selected by using an adaptive binding strategy. Firstly, constructing a feature relation graph by referring to conflicts among the proposed features, and calculating the weight as an edge; secondly, classifying the features according to the degree descending order of the features in the graph; finally, each feature is examined in the ordered list and evaluated to determine whether to assign to an existing bundle with small conflicts or create as a new bundle.
Application example:
the invention is applied to three main typical tailing pond distribution areas in Hebei province: in Chicheng, zhangbei and Chongli, a high-resolution remote sensing image download platform diagram new earth (http:// www.locarpace.cn /) is used for downloading high-resolution remote sensing images of a sample acquisition area and a demonstration application area, nine parts of tailing reservoirs in Hebei province, shaanxi province, sichuan province, guizhou province, guangxi province, hunan province, hubei province, jiangxi province and Anhui province are searched in a manual visual interpretation mode as shown in FIG. 2, 16-level high-resolution remote sensing images of Google are cut, and LabelImg software is used for marking tailing reservoir detection samples to manufacture target detection samples, wherein 996 images are marked in total and contain 1260 samples of the tailing reservoirs, and the invention is further explained by combining with the attached drawings.
Firstly, inputting a training set sample original high-resolution remote sensing image, respectively using sliding windows with the scales of {340, 170, 84 and 43} to perform up-sampling and extract multi-scale features, combining an FPN network and a Resnet152 to perform residual error network, and respectively using convolutional layers { C2, C3, C4 and C5} to perform feature extraction, wherein the step length is respectively set to {4,8, 16 and 32}; then, respectively retaining the semantic features of the scene by using a down-sampling mode; and finally, in order to fully reserve the scene characteristics reserved in the upsampling process, performing characteristic fusion in a transverse connection mode to enhance the detail information.
Secondly, according to the extraction of the multi-scale features, the candidate target is positioned, then the extraction of the extracted positioning coordinates of the tailings pond is continuously carried out for optimization of positioning, wherein x is the detected tailings pond candidate set, and g is the tailings pond position coordinate set subjected to regression positioning, and the comparison is shown in fig. 3. For each detection frame, the value of the IoU is set to be 0.5 by the first-stage detector, the value of the second-stage detector is set to be 0.6, the value of the third-stage detector is set to be 0.7, the detection candidate frames and the truth frames are respectively compared, and the high-precision detection frames are gradually reserved for positioning and optimizing the areas with lower identification precision by using the detection frames with different confidence degrees.
And fourthly, cutting the candidate area according To the optimized positioning coordinates To obtain RGB sample sets of the candidate area of the tailing pond, extracting object geometric, probability, scene characteristics, topology, spectrum and texture characteristics of the candidate set respectively based on the three-channel image, calculating characteristic values To respectively correspond To the characteristic vectors To form characteristic vectors, for the ith sample, the corresponding characteristic vectors are Fi = { Ge1, ge2, ge3, pr1, pr2, se1, se2, se3, se4, to1, to2, to3, sp1, sp2, sp3, te1, te2, te3, te4, te5, te6, te7 and Te8}, wherein each element respectively represents the height, width and height-width ratio of the sample, and image classification based on VGG16 defines that the image is the probability of the tailing pond and the non-tailing pond, carrying out scene segmentation based on U-Net To obtain the ratio of the areas of the tailing pond, the area corresponding To the sand body, the ratio of the area corresponding To the dam body, the corresponding probability value of the dam body, the entropy of the water body, the corresponding To the water body, the spectrum ratio, the entropy, the three-To-degree ratio of the integral tailing pond, the integral database, and the entropy, and the corresponding To the characteristic values, and the correlation of the dam body are respectively obtained by calculating the difference. Candidate features (parts) are detected for the training set as shown in table 1, and candidate features (parts) are detected for the test set as shown in table 2.
And fifthly, sampling the data of the tailing pond by using a unilateral sampling mode for the extracted data set of the characteristics of the tailing pond so as to reduce the calculated amount of the tailing pond in a sample data set, only retaining data with a larger gradient when sampling the data, but introducing a constant for the sample with a small gradient for balancing when calculating gain in order to avoid the change of sample distribution caused by discarding the data with a small gradient. The GOSS algorithm firstly sorts all values of the characteristics to be split in a descending order according to absolute values, and selects 20% of data with the largest absolute values. Then 30% of the remaining smaller gradient data were randomly selected. Then multiply the 10% data by a constant 1-2/3, focus more on the under-trained samples without changing the distribution of the original data set too much, and finally use 50% to calculate the gain of the information. Thus, the algorithm will focus more on the under-trained samples without changing the distribution of the original data set too much.
Finally, for extracting the features representing the tailing pond, in order to reduce the calculation complexity, the original values and the offset are added, the proprietary features are stored in different sets, and in order to avoid reducing the loss of precision of the number of the features on the identification result, an adaptive binding strategy is used for selecting the features to be bound. Firstly, constructing a feature relation graph by referring to conflicts among the proposed features, and calculating the weight as an edge; secondly, classifying the features according to the degree descending order of the features in the graph; finally, each feature is examined in the ordered list and evaluated to determine whether to assign to an existing bundle with small conflicts or create as a new bundle. As shown in fig. 4, the processed features can balance the contribution of the processed features in the classification process, and the registration module represents the contribution ratio of the features in the method, which is compared with the conventional methods RF, GBDT, catBoost and AdaBoost.
And sixthly, calling the trained target detection positioning optimization model and classification model, inputting the remote sensing image to be extracted in the demonstration area of the tailing pond, as shown in fig. 5, wherein fig. 5 (a) shows the positioning result of the large-scale tailing pond, fig. 5 (b) shows the positioning result of the medium-sized tailing pond, fig. 5 (c) shows the positioning result of the small-sized tailing pond, and fig. 5 (d) shows the detection result after reclassification is carried out on the target with the false alarm in the detection process. Fig. 6 is a graph of the results of tailings pond detection in the akang county, the kangbei district, and the chongli district of the north river province and the demonstration district.
TABLE 1 training set detection candidate features
TABLE 1 training set detection candidate features (continuation)
TABLE 2 test set detection candidate features
TABLE 2 test set detection candidate features (continuation)
Claims (10)
1. A self-adaptive remote sensing deep learning detection method for a tailing pond is characterized by being mainly realized through a multi-scale feature extraction network, a cascade positioning optimization module, a multi-dimensional feature extraction and fusion module and a feature self-adaptive learning classifier, wherein: the multi-scale feature extraction network firstly realizes extraction and calculation of scene features of tailings ponds of different sizes in remote sensing images, then a cascade positioning optimization module continuously optimizes coordinates of a detection frame with low precision based on a feature map obtained by calculation of the multi-scale feature extraction network through a regression detector to realize coordinate positioning of the tailings ponds, then a multi-dimensional feature extraction and fusion module obtains typical features of the tailings ponds from multiple dimensions and fuses the typical features into feature vectors for distinguishing false alarm targets, and finally self-adaptive learning and binding of the tailings ponds features are carried out through a self-adaptive integrated learning classifier based on the multi-dimensional features to realize extraction of the tailings ponds and elimination of the false alarm targets.
2. The adaptive remote sensing deep learning detection method for the tailings pond according to claim 1, which is characterized in that the remote sensing image tailings pond deep learning detection method adopting multi-stage cascade optimization specifically comprises the following steps: based on the existing high-resolution remote sensing image data, the positioning optimization and classification optimization are carried out on the identification of the tailings ponds with different scales through multi-stage cascade positioning and classification, the high-precision extraction of the tailings ponds in the large-range remote sensing image is realized, and the technical support can be provided for the safety management and information updating of the tailings ponds in the large-range.
3. A self-adaptive tailing pond remote sensing deep learning detection method is characterized in that the tailing pond remote sensing detection method is a tailing pond remote sensing detection method which is coupled with deep learning cascade optimization positioning and characteristic self-adaptive learning: the method comprises the steps of firstly, carrying out multi-level scene deep learning feature extraction on a remote sensing image, then using a cascade classifier to carry out positioning of the position of a tailing pond and coordinate optimization of a detection target to obtain a tailing pond detection candidate area, taking six dimensions into consideration, extracting and fusing to obtain detection object features, using a self-adaptive integrated classifier to carry out feature learning and cascade classification of the tailing pond, and obtaining a high-precision tailing pond detection result.
4. The method as claimed in claim 3, wherein the position optimization of tailing pond detection based on multi-scale features and cascade detectors and the classification optimization of tailing ponds based on multi-dimensional feature extraction and adaptive feature learning are specifically as follows:
(1) Position optimization based on multi-scale feature extraction and cascade positioning:
a. inputting an original remote sensing image to be detected, performing up-sampling on the image, and acquiring a three-level image;
b. downsampling the extracted features to enlarge the visual field;
c. combining the characteristics of two sides through links, fusing the characteristics of multiple scales and supplementing information content;
d. inputting a feature layer to position a candidate region of the feature;
e. designing a cascaded detector to carry out positioning optimization of a tailing pond;
f. outputting the coordinates of the detection candidate frame;
(2) And (3) identification optimization based on self-adaptive multi-dimensional feature ensemble learning classification:
a. cutting the recognition result according to the tailing pond provided by the first-step detection framework, wherein the cut object is used for further classification;
b. based on the RGB three-channel remote sensing image, respectively extracting robust features of candidate set objects: geometric, probabilistic, scene, topological, spectral, textural features;
d. normalizing the features;
e. single-side classification and sampling;
f. and (4) self-adaptive feature learning strong and weak feature binding is carried out to reduce the dimension of the features, and classification rules are learned.
5. The method of claim 4, wherein:
firstly, inputting an original high-resolution remote sensing image, respectively using sliding windows with the scales of {340, 170, 84 and 43} to perform up-sampling and extracting multi-scale features, combining an FPN network and a Resnet152 to perform residual error network, and respectively using convolutional layers { C2, C3, C4 and C5} to perform feature extraction, wherein the step lengths are respectively set to {4,8, 16 and 32};
then, respectively retaining the semantic features of the scene by using a down-sampling mode;
finally, the transverse connection mode is used for fusing the features to enhance the detail information.
6. The method of claim 4, wherein:
according to the extraction of the multi-scale features, the candidate target is positioned, and then the extraction of the positioning coordinates of the extraction of the tailing pond is continued, wherein b = (b) x ，b y ，b w ，b h ) And then, optimizing the positioning of the regression frame reclassification by using different thresholds, wherein x is the detected candidate object of the tailings pond, and g is the target after the regression positioningFor each detected sample i, input (g) into the tailings pond location coordinate set of (2) i ，b i ) Training an optimization model, and minimizing a loss function of the detection frame in the training process; the distance vector is also optimized in the regression process to ensure the stability of regression positioning, and the vector is specifically defined as follows:
wherein (g) x ，g y ，g w ，g h ) The four-vertex coordinates of the training sample g are shown, (b) x ，b y ，b w ，b h ) Is the coordinate of the target tailing pond extracted by the model in the image (delta) x ，δ y ，δ w ，δ h ) The coordinates representing the four coordinate corner points are used to optimize the parameter values.
7. The method of claim 4, wherein: optimizing the position by using a cascade detector containing a positioning optimization module;
in order to perform effective positioning optimization, a single regression function is iteratively applied in the process of cascade detection of the tailing pond, which is specifically defined as follows:
where x denotes a candidate, b denotes a suggestion box,
representing the connections between each iteration, in the proposed framework the regression structure is designed based on these cascaded detectors to perform the optimization:
in the formula: t denotes the number of connected layers, the regressor f at each stage T Is optimized according to the sample distribution;
wherein for each detection frame, there is an IoU value u to evaluate and evaluate the detection frame, and for each candidate frame tailings pond, the label is seized according to u:
wherein, g y Representing the label of sample g, the threshold of IoU is defined at each stage to evaluate the effect of each stage of detector, it is difficult to realize high-precision positioning based on single detector in the detection process of tailings pond, in the proposed model, different loss functions are defined to optimize the positioning, wherein the value of u is {0.5,0.6,0.7} in the original frame of R-CNN, and the classifier h is t And the regressor in each stage t, u t Is set to optimize according to the corresponding IoU threshold when u t ＞u t-1 The loss function is defined as follows:
L(x t ，g)＝L cls (h t (x t )，y t )+λ[y t ≥1]L loc (f t (x t ，b t )，g) (5)
wherein, b t ＝f t-1 (x t-1 ，b t-1 ) And g represents the real detection box x t λ =1 balance parameter, y t Is represented by x t Is labeled with t In the case of (1).
8. The method of claim 4, wherein:
firstly, cutting a candidate area according to optimized positioning coordinates to obtain a tailing pond candidate area RGB sample set,
then, extracting geometric, probability, scene, topology, spectrum and texture features of a candidate set object based on a three-channel image, calculating To obtain feature values, correspondingly forming feature vectors respectively, regarding an ith sample, correspondingly defining the feature vectors as Fi = { Ge1, ge2, ge3, pr1, pr2, se1, se2, se3, se4, to1, to2, to3, sp1, sp2, sp3, te1, te2, te3, te4, te5, te6, te7 and Te8}, wherein each element respectively represents the height, width and aspect ratio of the sample, and defining the probability of the image being a tailing pond and the area ratio of the tailing pond based on VGG16 image classification, carrying out scene segmentation based on U-Net To obtain the ratio of the area of the whole tailing pond, the water body, the sand body and the tailing pond corresponding To the area, calculating the mean value of R, G and B values among the samples, respectively calculating the average value, the spectral value, the maximum ratio of the spectral value and the inverse energy ratio of the three-component matrixes of the relative energy.
9. The method of claim 4, wherein:
firstly, sampling data of a tailing pond by using a unilateral sampling mode for an extracted data set of characteristics of the tailing pond to reduce the calculated amount of the tailing pond in a sample data set, only reserving data with a larger gradient during data sampling, and introducing a constant for a sample with a small gradient during gain calculation to balance in order to avoid the change of sample distribution caused by discarding data with a small gradient; the GOSS algorithm firstly sorts all values of the characteristics to be split in a descending order according to absolute values, and a% of data with the largest absolute values are selected; then randomly selecting b% data from the rest of the smaller gradient data; then multiplying the b% data by a constant l-a/b, paying more attention to the samples with insufficient training without changing the distribution of the original data set too much, and finally calculating the gain of the information by using (a + b)%;
then, splitting the tailing pond samples by using the estimated variance gain based on the data set during classification, wherein the specific expression mode is as follows:
wherein, the data set { x is trained for the tailings pond 1 ，…，x n Sample in (1) }, x i Representing features x in a multi-dimensional feature vector space S s Each instance based on an estimated variance gain
Splitting of the incoming instance at feature j and node d, based on the dataset a $ b, where O represents the training dataset of the tailings reservoir at the decision tree fixed node, n = ∑ I [ x = i ∈O]，/>
A l ＝{x i ∈A：x ij ≤d}，A r ＝{x i ∈A：x ij ＞d}，B l ＝{x i ∈B：x ij ≤d}，B l ＝{x i ∈B：x ij > d }, coefficient ≥ r>
Is used to perform a gradient sum normalization A on the data set B c The size of (d);
finally, for extracting the features representing the tailing pond, in order to reduce the calculation complexity, the original values and the offsets are added, the special features are stored in different data sets, and in order to avoid reducing the loss of precision of the recognition result caused by the number of the features, the features to be bound are selected by using a self-adaptive binding strategy; firstly, constructing a feature relation graph by referring to conflicts among the proposed features, and calculating the weight as an edge; secondly, classifying the features according to the degree descending order of the features in the graph; finally, each feature is examined in the ordered list and evaluated to determine whether to assign to an existing bundle with small conflicts or create as a new bundle.
10. The method of claim 4, wherein:
firstly, inputting a remote sensing image to be detected, carrying out target detection on the remote sensing image to be detected, cutting a tailing pond sample according to a model candidate area coordinate, and generating a candidate set of a model;
then, the multi-dimensional feature calculation is carried out on the candidate set, the feature vector and the label of the candidate object are made,
and finally, carrying out sample reclassification on the candidate set by using an adaptive integrated classifier based on the characteristic values, screening and extracting the detected tailings pond target, and carrying out drawing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211212391.4A CN115937698A (en) | 2022-09-29 | 2022-09-29 | Self-adaptive tailing pond remote sensing deep learning detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211212391.4A CN115937698A (en) | 2022-09-29 | 2022-09-29 | Self-adaptive tailing pond remote sensing deep learning detection method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115937698A true CN115937698A (en) | 2023-04-07 |
Family
ID=86554687
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211212391.4A Pending CN115937698A (en) | 2022-09-29 | 2022-09-29 | Self-adaptive tailing pond remote sensing deep learning detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115937698A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116304932A (en) * | 2023-05-19 | 2023-06-23 | 湖南工商大学 | Sample generation method, device, terminal equipment and medium |
| CN116665067A (en) * | 2023-08-01 | 2023-08-29 | 吉林大学 | Ore finding target area optimization system and method based on graph neural network |
-
2022
- 2022-09-29 CN CN202211212391.4A patent/CN115937698A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116304932A (en) * | 2023-05-19 | 2023-06-23 | 湖南工商大学 | Sample generation method, device, terminal equipment and medium |
| CN116304932B (en) * | 2023-05-19 | 2023-09-05 | 湖南工商大学 | Sample generation method, device, terminal equipment and medium |
| CN116665067A (en) * | 2023-08-01 | 2023-08-29 | 吉林大学 | Ore finding target area optimization system and method based on graph neural network |
| CN116665067B (en) * | 2023-08-01 | 2023-10-03 | 吉林大学 | Ore finding target area optimization system and method based on graph neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112132006B (en) | Intelligent forest land and building extraction method for cultivated land protection | |
| CN111310861B (en) | License plate recognition and positioning method based on deep neural network | |
| Movaghati et al. | Road extraction from satellite images using particle filtering and extended Kalman filtering | |
| CN115937698A (en) | Self-adaptive tailing pond remote sensing deep learning detection method | |
| EP4091109A1 (en) | Systems for multiclass object detection and alerting and methods therefor | |
| CN103020605B (en) | Bridge identification method based on decision-making layer fusion | |
| CN110222767B (en) | Three-dimensional point cloud classification method based on nested neural network and grid map | |
| CN111931505A (en) | Cross-language entity alignment method based on subgraph embedding | |
| CN111400572A (en) | Content safety monitoring system and method for realizing image feature recognition based on convolutional neural network | |
| Bhagwat et al. | A novel multilabel classification of remote sensing images using XGBoost | |
| CN115393631A (en) | Hyperspectral image classification method based on Bayesian layer graph convolution neural network | |
| CN115984632A (en) | Hyperspectral plastic waste material rapid classification method and device and storage medium | |
| CN111860359A (en) | Point cloud classification method based on improved random forest algorithm | |
| Sutha | Object based classification of high resolution remote sensing image using HRSVM-CNN classifier | |
| CN115019163A (en) | City factor identification method based on multi-source big data | |
| CN108960005B (en) | Method and system for establishing and displaying object visual label in intelligent visual Internet of things | |
| Taha et al. | Assessment of Approaches for the Extraction of Building Footprints from Plé iades Images | |
| Aksoy et al. | Image mining using directional spatial constraints | |
| Joshi et al. | Automatic rooftop detection using a two-stage classification | |
| Bhanu et al. | Analysis of terrain using multispectral images | |
| Chandra et al. | A cognitive perspective on road network extraction from high resolution satellite images | |
| García et al. | Improving texture pattern recognition by integration of multiple texture feature extraction methods | |
| CN117593890B (en) | Detection method and device for road spilled objects, electronic equipment and storage medium | |
| CN116630814B (en) | Quick positioning and evaluating method for building disasters based on machine learning | |
| Hormese et al. | A Comparison of Convolutional Neural Network Architectures for Road Classification from Satellite 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 |