CN115690591A - Remote sensing image farmland non-agricultural change detection method based on deep learning - Google Patents
Remote sensing image farmland non-agricultural change detection method based on deep learning Download PDFInfo
- Publication number
- CN115690591A CN115690591A CN202310010258.9A CN202310010258A CN115690591A CN 115690591 A CN115690591 A CN 115690591A CN 202310010258 A CN202310010258 A CN 202310010258A CN 115690591 A CN115690591 A CN 115690591A
- Authority
- CN
- China
- Prior art keywords
- agricultural
- land
- remote sensing
- change
- images
- 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.)
- Granted
Links
- 230000008859 change Effects 0.000 title claims abstract description 109
- 238000001514 detection method Methods 0.000 title claims abstract description 63
- 238000013135 deep learning Methods 0.000 title claims abstract description 30
- 238000000034 method Methods 0.000 claims abstract description 58
- 239000013598 vector Substances 0.000 claims abstract description 47
- 238000010276 construction Methods 0.000 claims abstract description 16
- 238000012545 processing Methods 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 13
- 238000001228 spectrum Methods 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 13
- 238000001914 filtration Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000003595 spectral effect Effects 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 238000012216 screening Methods 0.000 claims description 5
- 238000005520 cutting process Methods 0.000 claims description 4
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 4
- 235000002566 Capsicum Nutrition 0.000 claims description 3
- 239000006002 Pepper Substances 0.000 claims description 3
- 235000016761 Piper aduncum Nutrition 0.000 claims description 3
- 235000017804 Piper guineense Nutrition 0.000 claims description 3
- 235000008184 Piper nigrum Nutrition 0.000 claims description 3
- 239000000654 additive Substances 0.000 claims description 3
- 230000000996 additive effect Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 150000003839 salts Chemical class 0.000 claims description 3
- 244000203593 Piper nigrum Species 0.000 claims 1
- 230000008569 process Effects 0.000 abstract description 11
- 230000008901 benefit Effects 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 8
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000000605 extraction Methods 0.000 description 4
- 238000003709 image segmentation Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 238000012937 correction Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000012805 post-processing Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 241000722363 Piper Species 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000009313 farming Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000000513 principal component analysis Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 239000003905 agrochemical Substances 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 238000009412 basement excavation Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007123 defense Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 239000002420 orchard Substances 0.000 description 1
- 238000004886 process control Methods 0.000 description 1
- 230000005855 radiation Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 239000002689 soil Substances 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a method for detecting non-agricultural change of cultivated land by remote sensing images based on deep learning, which comprises the following steps: s1: constructing a classification sample library of agricultural land and non-agricultural land; s2: training a non-agricultural land detection model by using the classification sample library constructed in the step S1; s3: and processing the detected non-agricultural land, correcting the result and outputting the result in a vector mode. The method for detecting the non-agricultural change of the cultivated land by using the remote sensing image based on the deep learning has the advantages that the training of the model for detecting the cultivated land change and the output of the cultivated land change information can be completed by using the multi-period high-resolution remote sensing satellite image and the conventional land feature marking information through the unsupervised model training without additional remote sensing change detection marking, so that the manual marking cost of the change pattern spot is effectively saved; the output boundary information of the non-agricultural land is accurate and reliable; the method is used for high-resolution remote sensing satellite images and unmanned aerial vehicle images, the construction process is simple, and the sample library can be combined and reused pertinently.
Description
Technical Field
The invention relates to the technical field of deep learning, in particular to a method for detecting non-agricultural change of cultivated land by using remote sensing images based on deep learning.
Background
The key point of protecting the safety of the grains lies in the strategy of storing grains in the ground and storing grains in the technology. The cultivated land is the root of life of grain production, and in order to guarantee grain supply, firstly, the safety of cultivated land is ensured, the red line of 18 hundred million mu cultivated land of the country is kept, and a strict cultivated land supervision system needs to be established to deal with the severe situation of cultivated land protection, and no matter whether a management and control system is established or process control and responsibility are studied, the fact is used as a guideline to allow data to speak. At present, each supervision department mainly adopts a visual interpretation method to check the non-agricultural illegal behaviors of the cultivated land such as the number of the randomly occupied cultivated land building houses in the countryside, namely, suspected pattern spots are manually extracted from images, and after a third national survey result is utilized to determine the land parcel range, field check is carried out. For the conditions of afforestation, construction of pond water surface, road construction, land surface excavation and the like which destroy cultivated land, the problem of non-farming of cultivated land can be basically solved by using an interior interpretation method through the feature and texture information of the ground features and the texture information of the remote sensing satellite images, but the problem that economic crops, orchards and crops are difficult to distinguish on the image texture is difficult to distinguish, and meanwhile, along with the continuous increase of the resolution ratio and the data volume of the satellite remote sensing, the traditional manual interpretation method is difficult to support the task and the requirement and has the defects of easy loss, low efficiency and the like.
The research aiming at the 'non-agricultural chemical' aspect of the farmland is mainly theoretical research on the aspects of time and space distribution, driving factors, relation analysis with economic growth and the like, the essence of the research is the change of land utilization types, namely, the farmland is changed into other land types, but the precision requirement is higher in practical application, all change pattern spots are required firstly in practical application, and then farmland change information is screened out in a targeted manner. Due to the rapid development of the deep learning convolutional neural network, currently, the mainstream change detection adopts a method of directly comparing two-phase images, namely, the change characteristics of the two-phase images are obtained through the convolutional neural network, and then a change pattern spot is directly output end to end. For example, a deep learning-based farmland change detection technology is proposed for peaks and the like, wu Gujie and the like utilize a transfer learning mechanism to extract illegal land patterns and the like; zhang Cuijun, etc., an improved UNet network model is provided, which converts the problem of change detection into the problem of pixel-level classification, and then inputs the superposed front and rear 2 remote sensing images into the UNet network model, thus realizing the change detection of the end-to-end remote sensing images, but the model network framework is large and the training process is complex; the twin DeepLabv3+ binary change detection network is provided by the Furfu, so that the detection accuracy is effectively improved, but the change detection effect of the detection network on targets with different scales is poor; wang Minshui is also based on Deep-Labv3+ network, and 2 remote sensing images in different time phases are superposed by using a random patch optimization strategy through matrix operation and then are used as network input, so that compared with UNet, the accuracy is higher, the memory consumption is low, and the detection of the boundary is not smooth; yang Zhan provides a twin network model, which extracts features from the original image directly through 2 networks sharing weights, and then calculates the distance between 2 feature tensors to obtain the change detection result. The detection of the boundary by the model is smooth, but the process is complicated. In order to realize an end-to-end twin network change detection model, daudt proposes 2 classical twin network models FC-Sim-conc and FC-Sim-diff, and based on a U-shaped jump connection structure, images of front and back 2 time phases are respectively sent to a decoder to realize change detection in a splicing and absolute value difference calculation mode, but the precision is low. Chen proposes a DASNet network, which improves the change detection performance of the model through a double-attention mechanism, but has limited change detection capability for small-scale targets.
However, the research and analysis find that the existing deep learning farmland non-agricultural change detection method based on the two-stage remote sensing image has the following problems: (1) Due to the fact that spectral characteristics of the ground objects are comprehensively influenced by physical characteristics of the ground objects, imaging time, environment and other external factors, the different-deep learning method is limited in detection capability aiming at different scale changes, and the detection result edge is not smooth generally; (2) The existing deep learning image segmentation model change method comprises a large number of parameters, and the training process and the output loss calculation process are often complex; (3) The methods all need a large amount of image change sample labels to learn, and the sample acquisition cost is time-consuming and labor-consuming.
Chinese patent document CN114155440a discloses an automatic detection method and system for farmland non-farming, the method comprising: obtaining a historical farmland vector diagram of a first region to be detected; acquiring real-time remote sensing image data of the first detection area to be detected; constructing a multi-level image segmentation model; inputting the historical farmland vector diagram and the real-time remote sensing image data into the multistage image segmentation model to obtain small farmland image spots; constructing a land screening model; inputting the small image spots of the cultivated land output by the multistage image segmentation model into the land use screening model for traversal detection to obtain multiple types of land use image spots for construction; and generating a first occupation detection result by combining the multiple types of construction land image spots. But the detection precision of the technical scheme is still not high.
Chinese patent document CN114913432a discloses a remote sensing detection method for urban construction land change, which comprises: inputting two Landsat5-TM remote sensing images with different time phases and subjected to geometric correction and radiation correction; carrying out geometric accurate registration on the two remote sensing images; carrying out principal component analysis and dimensionality reduction processing on 7 wave bands of the two remote sensing images to respectively obtain corresponding first components; calculating texture characteristics of TM7 wave bands of the two remote sensing images by utilizing the gray level co-occurrence matrix; calculating the normalized vegetation index NDVI of the two remote sensing images; calculating the normalized building index NDBI of the two remote sensing images; performing image difference operation on the obtained 4 pairs of change detection factors; performing principal component analysis on the obtained four difference result gray level images, and taking a first component of the four difference result gray level images as a city construction land change detection result; and (4) segmenting the grey-scale image of the urban construction land change detection result by using a threshold value, and extracting to obtain a change detection binary image. But the detection precision of the technical scheme is still not high.
At present, by analyzing the collected non-agricultural change information of typical cultivated land, the change of the cultivated land is known to follow the rule that the similarity attenuates along with the distance on the spatial distribution, namely, the change of the cultivated land is reduced in time, and particularly, the change of the cultivated land in half a year in winter is less. Therefore, the emphasis of farmland protection should be placed on residential areas and surrounding areas of traffic lines, how to combine the artificial intelligence technology with remote sensing image data and space-time characteristics of non-agricultural land, and how to realize more accurate and automatic farmland non-agricultural change information extraction is still a problem commonly explored in the remote sensing application fields at home and abroad.
Disclosure of Invention
The invention aims to solve the problem that the non-agricultural change information of the cultivated land is difficult to find by utilizing a multi-period remote sensing image, provides a detection method for the non-agricultural change of the cultivated land based on a deep learning remote sensing image, and realizes more accurate and automatic extraction of the non-agricultural change information of the cultivated land.
In order to solve the technical problems, the technical scheme of the invention is as follows: the method for detecting the non-agricultural change of the remote sensing image cultivated land based on the deep learning specifically comprises the following steps:
s1: collecting remote sensing images and obtaining a plurality of time phase images to be detected; constructing a classification sample library of agricultural land and non-agricultural land through a plurality of time phase images;
s2: training a non-agricultural land detection model by using the classification sample library constructed in the step S1;
s3: and (3) detecting a plurality of time phase images by using the non-agricultural land detection model in the step (S2), processing the detected non-agricultural land, correcting the result, outputting the result in a vector mode, and obtaining non-agricultural land boundary information.
By adopting the technical scheme, the classification sample library is constructed by using the high-resolution remote sensing satellite image and the unmanned aerial vehicle image, the construction process is simple, the sample library can be combined and reused in a targeted manner, the non-agricultural land detection model is trained through the unsupervised model, additional remote sensing change detection marking is not needed, the training of the model for detecting the change of the cultivated land and the output of the change information of the cultivated land can be completed by using the multi-period high-resolution remote sensing image and the conventional land feature marking information, and the manual marking cost for changing the pattern spots is effectively saved; and the output non-agricultural land boundary information is accurate and reliable.
As a preferred technical solution of the present invention, the step S1 of constructing a classification sample library of agricultural land and non-agricultural land comprises the specific steps of:
s11: extracting a color characteristic matrix and a gray level co-occurrence matrix of the remote sensing satellite images of a plurality of time phases to be detected;
s12: selecting an image from remote sensing satellite images of a plurality of time phases, setting the ground feature color style of the image as a target style, replacing other images with the target style by using a method of exchanging low-frequency components through Fourier transform, and obtaining the images with the same style;
s13: cutting the images with the same style processed in the step S12 and the corresponding land utilization data and the change survey data to obtain sample blocks with the same size;
s14: and screening the cut sample blocks with the same size to obtain positive samples and negative samples.
As a preferred embodiment of the present invention, the step S12 of replacing the image with the target style by using the method of exchanging low frequency components by fourier transform comprises the specific steps of:
s121: respectively reading an image of a target style and an image to be processed, segmenting color channels to obtain color channels R, G, B and color channels Ra, bb and Cc, and initializing a low-frequency filter and a high-frequency filter according to the image size regulation;
s122: calculating the forward Fourier transform and the frequency spectrum centralization of the color channels Ra, bb and Cc by using a color channel R, G, B, respectively, then performing filtering processing on 6 color channels (a color channel R, G, B and color channels Ra, bb and Cc) by using a low-frequency filter and a high-frequency filter in the step S121, merging the frequency spectrums of the corresponding color channels after the processing is completed, and calculating the frequency spectrum centralization again;
s123: and (4) performing inverse Fourier transform on the frequency spectrum decentralization result obtained in the step (S122), calculating gray values of three color channels, and performing normalized superposition to obtain an image with a changed style, namely the image with the target style.
As a preferred technical solution of the present invention, the farmland non-agricultural change detection model in step S2 includes a change detection module and a non-agricultural land classification module, wherein the change detection module is composed of 1 spectral change feature extractor, 1 feature vector space solver, and a change category filter.
As a preferred technical solution of the present invention, the spectral change feature extractor is composed of a 5-layer depth separable convolution network, and after each convolution operation, a batch normalization and a ReLU function activation operation are added, so as to perform a filtering operation and a difference between each band of an image to be detected, obtain a variation of each pixel value in each band, and then compose a variation vector from the variations of each band. Wherein, the depth Separable Convolution network (Depthwise Separable Convolition) is: convolution kernel W convolution kernel H (picture W-convolution kernel W + 1) (picture H-convolution kernel H + 1) input channel number, and the size of the 5-layer depth separable convolution is 1 × 1, 3 × 3, 1 × 1, respectively.
As a preferred technical scheme of the present invention, the eigenvector space solver is composed of a group of convolution blocks, and is configured to extract N non-overlapping feature blocks with the same size from an integrated feature data block, then solve an eigenvector space on the non-overlapping feature blocks by using a PCA algorithm, and finally project overlapping data blocks around each pixel with the same size as the non-overlapping feature blocks to the eigenvector space, thereby creating an eigenvector space on the entire integrated data block; the comprehensive characteristic data block is formed by superposing color characteristics, gray level co-occurrence matrixes and variation vectors. Wherein the size of a group of rolling blocks is 2 x 2, and the step length is 2; the size of the non-overlapping feature block is h multiplied by h (h is more than or equal to 2), and N is the width and height of the input comprehensive feature data divided by h to be rounded. PCA is an existing common data dimensionality reduction method.
As a preferred technical solution of the present invention, the change category filter comprises a deep clustering self-encoder and an additive angle interval LOSS (ARC _ LOSS) calculation unit, and is configured to cluster a feature vector space into two different clusters, and calculate a normalized euclidean distance between the feature vector and an average feature vector of the clusters, where the calculation formula is:
(ii) a Wherein,x i 、y i representing two n-dimensional feature vectors, namely a mathematical description of a normalized Euclidean distance summation operation on the selected n-dimensional feature vectors; n refers to the number of spatial dimensions of the feature vector processed at a time; v is the variance between the two divided by n-1, and each pixel on the differential data block is assigned to one of the clusters according to the distance to generate the variation graph. Wherein N has the same meaning as the characteristics of the N non-overlapping feature blocks with the same size, and can freely select integral multiple of 2 in actual operation.
The spectrum change feature extractor and the change category filter are adopted to realize an unsupervised solving process, namely, the change pattern spots are obtained by calculating the distance difference of deep features of the processed images in two stages instead of learning by manual change labeling, so that the automation degree and the extraction precision are further improved.
As a preferred technical scheme of the invention, the non-agricultural land classification module consists of a 6-layer full convolution network and 1 dense connection layer, and batch standardization and ReLU function activation operation are added after each convolution operation.
As a preferred embodiment of the present invention, the step S3 of removing salt and pepper noise and pseudo-variation processing and correcting the boundary of the detected non-agricultural land comprises the following specific steps:
s31: solving a closed operation connection fracture part of the farmland non-agricultural change pattern spot, and removing a change edge burr by using a median filtering algorithm;
s32: replacing and removing the isolated pixel clusters by the pixel classes on the left side by using area and length threshold values;
s33: and correcting the boundary, namely calculating a pixel point at the center of the non-agricultural category, and optimizing the area energy item and the edge energy item by taking the pixel point as the category vertex of the image to obtain regular non-agricultural land grid or vector data. The median filtering algorithm is an existing conventional post-processing method, and can reduce edge obstacles for subsequent processing.
As a preferred embodiment of the present invention, the positive type sample in step S14 includes: ploughing and plastic greenhouses in different seasons; the negative class samples include: construction land, newly-opened construction site, mound, gardens, land for transportation, water body.
Compared with the prior art, the invention has the beneficial effects that:
(1) The method has the advantages that the training process of the model is simple, the characteristic space is solved through the convolutional neural network and then clustering is carried out to obtain the change pattern spots, the method not only has the image characteristic extraction capability of the convolutional neural network, and the characteristic space clustering is solved through overlapping of the color characteristic, the gray level co-occurrence matrix and the change vector, so that the accuracy of the change result is ensured, but also the problem of complexity and time consumption of the end-to-end output training process is greatly saved;
(2) According to the method, additional remote sensing change detection labeling is not needed, training of a change detection model and output of farmland change information can be completed by utilizing multi-stage high-resolution remote sensing images and conventional ground feature labeling information, and the manual labeling cost of change pattern spots is effectively saved;
(3) The method has the advantages that the method relates to farmland information, the accuracy of the output farmland non-agricultural boundary information has high requirements, edge fracture and burrs of a general deep learning method can be effectively eliminated by pertinently combining a plurality of different post-processing information, and energy correction is carried out on independent change pixel groups to obtain accurate farmland non-agricultural change vector data;
(4) The method has high robustness on image resolution, can be used for high-resolution remote sensing satellite images or unmanned aerial vehicle images, is simple in construction process of a sample library, can continuously expand images with different scales according to task requirements, and can be combined and reused in a targeted manner, so that non-agricultural change cases of farmlands with different scales can be well identified, the method can be used for multiple farmland monitoring works such as national soil change investigation, geographic national condition monitoring, defense law enforcement, greenhouse illegal violation illegal building and the like, and technical support is provided for farmland resource protection.
Drawings
FIG. 1 is a flow chart of farmland non-agricultural change detection of the method for detecting farmland non-agricultural change based on deep learning remote sensing images of the invention;
FIGS. 2 a-2 b are sample diagrams of non-agricultural change detection results of cultivated land in an embodiment of the detection method for non-agricultural change of cultivated land based on deep learning of remote sensing images according to the present invention; fig. 2a is a sample diagram of 2016 and fig. 2b is a sample diagram of 2020.
Detailed Description
The technical solution in the embodiment of the present invention will be described clearly and completely with reference to the accompanying drawings in the drawings of the embodiment of the present invention;
example (b): as shown in FIG. 1, the intelligent detection method for farmland non-agricultural change for deep learning based on the detection method for farmland non-agricultural change of remote sensing images for deep learning comprises the following steps:
s1: collecting remote sensing images and obtaining a plurality of time phase images to be detected; constructing a classification sample library of agricultural land and non-agricultural land through a plurality of time phase images;
the step S1 of establishing the classification sample library of the agricultural land and the non-agricultural land comprises the following specific steps:
s11: extracting a color characteristic matrix and a gray level co-occurrence matrix of the remote sensing satellite images of a plurality of time phases to be detected;
s12: selecting a proper image from the remote sensing satellite images of a plurality of time phases, setting the ground feature color style of the image as a target style, and replacing other images with the target style by using a method of exchanging low-frequency components through Fourier transform to obtain images with the same style;
the specific step of replacing the image with the target style in the step S12 by using the method of exchanging low frequency components by fourier transform is as follows:
s121: respectively reading an image of a target style and an image to be processed, segmenting color channels to obtain color channels R, G, B and color channels Ra, bb and Cc, and initializing a low-frequency filter and a high-frequency filter according to the image size regulation;
s122: calculating the forward Fourier transform and the frequency spectrum centralization of the color channels Ra, bb and Cc by using a color channel R, G, B, respectively, then performing filtering processing on 6 color channels (a color channel R, G, B and color channels Ra, bb and Cc) by using a low-frequency filter and a high-frequency filter in the step S121, merging the frequency spectrums of the corresponding color channels after the processing is completed, and calculating the frequency spectrum centralization again;
s123: performing inverse Fourier transform on the frequency spectrum decentralization result obtained in the step S122, calculating gray values of three color channels, and performing normalized superposition to obtain an image with a changed style, namely an image with a target style;
s13: cutting the high-resolution images with the same style processed in the step S12 and the corresponding land utilization data and the change survey data to obtain sample blocks with the same size; cutting a vector or a grid of the high-resolution satellite image into sample blocks with the size of 128 × 128;
s14: screening the cut sample blocks with the same size to obtain positive samples and negative samples; as a preferred embodiment of the present invention, the positive type sample in step S14 includes: ploughing and plastic greenhouses in different seasons; the negative class samples include: construction land, newly-opened construction site, mound, garden, transportation land and water body;
s2: training a non-agricultural land detection model by using the classification sample library constructed in the step S1;
the farmland non-agricultural change detection model in the step S2 comprises a change detection module and a non-agricultural land classification module, wherein the change detection module consists of 1 spectral change feature extractor, 1 feature vector space solver and a change category filter; the spectrum change feature extractor consists of 5 layers of depth Separable convolutional networks, and the depth Separable convolutional networks (Depthwise Separable Convolition) are as follows: convolution kernel W × convolution kernel H (picture W-convolution kernel W + 1) × (picture H-convolution kernel H + 1) × input channel number; the sizes of the 5-layer depth separable convolution are respectively 1 × 1, 3 × 3 and 1 × 1, batch standardization and ReLU function activation operation are added after each convolution operation, filtering operation is carried out on each wave band of an image to be detected and difference is carried out on each wave band, the variation of each pixel value in each wave band is obtained, and then variation vectors are formed by the variation of each wave band; the feature vector space solver consists of a group of convolution blocks, the size of each convolution block is 2 multiplied by 2, and the step length is 2; the method comprises the steps of extracting N (N is the width and height of input comprehensive characteristic data divided by h to be integer) non-overlapped characteristic blocks with the size h multiplied by h (h is more than or equal to 2) on a comprehensive characteristic data block, then solving a characteristic vector space on the h multiplied by h non-overlapped characteristic blocks by adopting a PCA algorithm, and finally projecting overlapped h multiplied by h data blocks around each pixel with the same size as the non-overlapped characteristic blocks to the characteristic vector space so as to create a characteristic vector space on the whole comprehensive characteristic data block; the comprehensive characteristic data block is formed by overlapping color characteristics, a gray level co-occurrence matrix and a variation vector; the change category filter comprises a deep clustering self-encoder and an additive angle interval LOSS (ARC _ LOSS) calculation unit, is used for clustering the feature vector space into two different clusters, and simultaneously calculates the standardized Euclidean distance between the feature vector and the average feature vector of the clusters, and has the calculation formula as follows:
The non-agricultural land classification module consists of a full convolution network with 6 layers and 1 dense connection layer, and batch standardization and ReLU function activation operation are added after each convolution operation;
s3: detecting a plurality of time phase images by using the non-agricultural land detection model in the step S2, processing the detected non-agricultural land, correcting the result and outputting the result in a vector mode to obtain boundary information of the non-agricultural land;
the specific steps of removing salt and pepper noise and pseudo-change processing and correcting the boundary of the detected non-agricultural land in the step S3 are as follows:
s31: solving a closed operation connection fracture part of the farmland non-agricultural change pattern spot, and removing a change edge burr by using a median filtering algorithm; the median filtering algorithm is an existing conventional post-processing method and can reduce edge barriers for subsequent processing;
s32: replacing and removing the isolated pixel clusters by the pixel classes on the left side by using area and length threshold values;
s33: and correcting the boundary, namely calculating a pixel point at the center of the non-agricultural category, and optimizing the area energy item and the edge energy item by taking the pixel point as the category vertex of the image to obtain regular non-agricultural land grid or vector data.
The specific embodiment is as follows: in order to verify and explain the effectiveness of the method, the method for detecting the non-agricultural change of the cultivated land based on the deep learning remote sensing image is adopted to test the non-agricultural change data (the two-stage image time is 2016 and 2020 respectively) of the cultivated land at the portion of hong Kong in Jiangsu, and the cultivated land change pattern spots detected by the method for detecting the non-agricultural change of the cultivated land based on the deep learning remote sensing image are compared with the visual interpretation results one by one to evaluate the change detection precision, and the table 1 is a matrix for confusion of the cultivated land change detection pattern spots of the image Gao Bian Gao.
TABLE 1 dot matrix for detecting farmland change by using image of Gao-fen-two of family and harbor
The total area of the test area is about 260km 2 Co-detection of1064 change pattern spot samples are obtained, the total change pattern spot detection rate is about 95%, the total time from the training of the model to the result output is about 40min by using the method, and the total time for manual visual comparison and inspection of the result is about 36h; as can be seen from the table 1, about 198 pattern spots belonging to non-agricultural change of the cultivated land in the changed pattern spots are correctly detected 186 by the method, the non-agricultural correct detection rate of the cultivated land is about 95.4%, the missing score is 12, and the missing score is about 6%; the sample diagrams of the detected results are shown in fig. 2 a-2 b, wherein fig. 2a is the result sample diagram of an image of 2016, fig. 2b is the result sample diagram of an image of 2020, and as can be seen from fig. 2 a-2 b, the range drawn by the dotted line is the range of non-agricultural change of the cultivated land into the factory and the attached facilities thereof, the boundary of the whole result is smooth and regular, the patented method can comprehensively process the change growth and the regression of different land features before the two-stage images, can effectively remove the part of abandoned land above the cultivated land, and can obtain the precise change range with pertinence. The image interval time of the two periods is about 4 years, the color difference of vegetation samples in the cultivated land is large, once a large building such as a newly-added factory building appears in a cultivated land block, the spectral standard deviation of the block, the standard deviation of the sub-object mean value and the contrast based on the gray level co-occurrence matrix are obviously increased.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or scope of the invention as defined in the appended claims.
Claims (10)
1. A remote sensing image cultivated land non-agricultural change detection method based on deep learning is characterized by comprising the following steps:
s1: collecting remote sensing images and obtaining a plurality of time phase images to be detected; constructing a classification sample library of agricultural land and non-agricultural land through a plurality of time phase images;
s2: training a non-agricultural land detection model by using the classification sample library constructed in the step S1;
s3: and (3) detecting a plurality of time phase images by using the non-agricultural land detection model in the step (S2), processing the detected non-agricultural land, correcting the result, outputting the result in a vector mode, and obtaining non-agricultural land boundary information.
2. The method for detecting non-agricultural changes in cultivated land based on remote sensing images of deep learning according to claim 1, wherein the specific steps of constructing the classification sample library of agricultural land and non-agricultural land in step S1 are as follows:
s11: extracting a color characteristic matrix and a gray level co-occurrence matrix of the remote sensing satellite images of a plurality of time phases to be detected;
s12: selecting an image from the remote sensing satellite images of a plurality of time phases, setting the ground feature color style of the image as a target style, replacing the other images with the target style by using a method of exchanging low-frequency components through Fourier transform, and obtaining the images with the same style;
s13: cutting the images with the same style processed in the step S12 and the corresponding land utilization data and the change survey data to obtain sample blocks with the same size;
s14: and screening the cut sample blocks with the same size to obtain positive samples and negative samples.
3. The method for detecting non-agricultural changes in farmland based on remote sensing images of deep learning according to claim 2, characterized in that the specific steps of replacing the images of the method for exchanging low frequency components by Fourier transform in step S12 with a target style are as follows:
s121: respectively reading an image of a target style and an image to be processed, segmenting color channels to obtain color channels R, G, B and color channels Ra, bb and Cc, and initializing a low-frequency filter and a high-frequency filter according to the image size regulation;
s122: calculating the forward Fourier transform and the frequency spectrum centralization of the color channels Ra, bb and Cc by using a color channel R, G, B, then respectively performing filtering processing on the 6 color channels by using the low-frequency filter and the high-frequency filter in the step S121, merging the frequency spectrums of the corresponding color channels after the processing is finished, and calculating the frequency spectrum decentration again;
s123: and (4) performing inverse Fourier transform on the frequency spectrum decentralization result obtained in the step (S122), calculating gray values of three color channels, performing normalization and superposition, and obtaining an image with a changed style, namely the image with the target style.
4. The method for detecting non-agricultural changes in cultivated land based on remote sensing images of deep learning of claim 2, wherein the model for detecting non-agricultural changes in cultivated land in step S2 comprises a change detection module and a non-agricultural land classification module, wherein the change detection module comprises 1 spectral change feature extractor, 1 feature vector space solver and a change class screener.
5. The method for detecting non-agricultural changes in cultivated land by remote sensing images based on deep learning of claim 4, wherein the spectral change feature extractor is composed of 5 layers of depth separable convolution networks, batch normalization and ReLU function activation operations are added after each convolution operation, and are used for performing filtering operation and difference between each wave band of an image to be detected, obtaining the variation of each pixel value in each wave band, and then forming a variation vector by the variation of each wave band.
6. The method for detecting non-agricultural changes in cultivated land based on remote sensing images of deep learning according to claim 5, characterized in that the feature vector space solver is composed of a set of convolution blocks and is used for extracting N non-overlapping feature blocks with the same size from a comprehensive feature data block, then solving a feature vector space on the non-overlapping feature blocks by adopting a PCA algorithm, and finally projecting overlapping data blocks with the same size as the non-overlapping feature blocks around each pixel to the feature vector space, thereby creating a feature vector space on the whole comprehensive data block; the comprehensive characteristic data block is formed by superposing color characteristics, gray level co-occurrence matrixes and variation vectors.
7. The method for detecting non-agricultural changes in cultivated land by remote sensing images based on deep learning of claim 6, wherein the change category filter comprises a deep clustering self-encoder and an additive angle interval loss calculation unit, and is used for clustering the feature vector space into two different clusters, and calculating the normalized Euclidean distance between the feature vector and the average feature vector of the clusters, and the calculation formula is as follows:
(ii) a Wherein,x i 、y i representing two n-dimensional feature vectors, namely a mathematical description of a normalized Euclidean distance summation operation on the selected n-dimensional feature vectors; n refers to the number of spatial dimensions of the feature vector processed at a time; v is the variance between the two divided by n-1, and each pixel on the differential data block is assigned to one of the clusters according to the distance to generate the variation graph.
8. The method for detecting non-agricultural changes in the cultivated land based on the remote sensing image of the deep learning of claim 4, wherein the non-agricultural land classification module is composed of 6 layers of full convolution networks and 1 dense connection layer, and batch standardization and ReLU function activation operation are added after each convolution operation.
9. The method for detecting non-agricultural changes in cultivated land based on remote sensing image of deep learning according to claim 7, wherein the step S3 of removing salt and pepper noise and pseudo-change processing and correcting the boundary of the detected non-agricultural land comprises the following specific steps:
s31: solving a closed operation connection fracture part of the farmland non-agricultural change pattern spot, and removing a change edge burr by using a median filtering algorithm;
s32: replacing and removing the isolated pixel clusters by the pixel classes on the left side by using area and length threshold values;
s33: and correcting the boundary, namely calculating a pixel point at the center of the non-agricultural category, and optimizing the area energy item and the edge energy item by taking the pixel point as the category vertex of the image to obtain regular non-agricultural land grid or vector data.
10. The method for detecting non-agricultural changes in farmland based on remote sensing images of deep learning according to claim 8, wherein the positive type samples in step S14 comprise: ploughing and plastic greenhouses in different seasons; the negative class samples include: construction land, newly-opened construction site, mound, gardens, land for transportation, water body.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310010258.9A CN115690591B (en) | 2023-01-05 | 2023-01-05 | Remote sensing image farmland non-agricultural change detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310010258.9A CN115690591B (en) | 2023-01-05 | 2023-01-05 | Remote sensing image farmland non-agricultural change detection method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115690591A true CN115690591A (en) | 2023-02-03 |
| CN115690591B CN115690591B (en) | 2023-03-31 |
Family
ID=85057259
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310010258.9A Active CN115690591B (en) | 2023-01-05 | 2023-01-05 | Remote sensing image farmland non-agricultural change detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115690591B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117575979A (en) * | 2023-08-01 | 2024-02-20 | 广东省国土资源测绘院 | Remote sensing image change detection method and device |
| WO2024159812A1 (en) * | 2023-08-09 | 2024-08-08 | 广东省国土资源测绘院 | Cultivated land non-agricultural patch extraction method with multi-source sar data fused therein, and device |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110929592A (en) * | 2019-11-06 | 2020-03-27 | 北京恒达时讯科技股份有限公司 | Extraction method and system for outer boundary of mariculture area |
| CN112149719A (en) * | 2020-09-04 | 2020-12-29 | 国家卫星气象中心(国家空间天气监测预警中心) | Multipurpose one-stop remote sensing classification method and device and computer equipment |
| CN113392748A (en) * | 2021-06-07 | 2021-09-14 | 中国煤炭地质总局勘查研究总院 | Remote sensing image farmland information extraction method based on convolutional neural network |
| CN113569815A (en) * | 2021-09-22 | 2021-10-29 | 速度时空信息科技股份有限公司 | Method for detecting remote sensing image change based on image segmentation and twin neural network |
| CN113989657A (en) * | 2021-10-11 | 2022-01-28 | 中国测绘科学研究院 | Method and device for detecting farmland range change based on invariant information sample screening |
| CN114155440A (en) * | 2021-12-08 | 2022-03-08 | 二十一世纪空间技术应用股份有限公司 | Automatic detection method and system for farmland non-farming |
| CN114972191A (en) * | 2022-04-25 | 2022-08-30 | 航天宏图信息技术股份有限公司 | Method and device for detecting farmland change |
-
2023
- 2023-01-05 CN CN202310010258.9A patent/CN115690591B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110929592A (en) * | 2019-11-06 | 2020-03-27 | 北京恒达时讯科技股份有限公司 | Extraction method and system for outer boundary of mariculture area |
| CN112149719A (en) * | 2020-09-04 | 2020-12-29 | 国家卫星气象中心(国家空间天气监测预警中心) | Multipurpose one-stop remote sensing classification method and device and computer equipment |
| CN113392748A (en) * | 2021-06-07 | 2021-09-14 | 中国煤炭地质总局勘查研究总院 | Remote sensing image farmland information extraction method based on convolutional neural network |
| CN113569815A (en) * | 2021-09-22 | 2021-10-29 | 速度时空信息科技股份有限公司 | Method for detecting remote sensing image change based on image segmentation and twin neural network |
| CN113989657A (en) * | 2021-10-11 | 2022-01-28 | 中国测绘科学研究院 | Method and device for detecting farmland range change based on invariant information sample screening |
| CN114155440A (en) * | 2021-12-08 | 2022-03-08 | 二十一世纪空间技术应用股份有限公司 | Automatic detection method and system for farmland non-farming |
| CN114972191A (en) * | 2022-04-25 | 2022-08-30 | 航天宏图信息技术股份有限公司 | Method and device for detecting farmland change |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117575979A (en) * | 2023-08-01 | 2024-02-20 | 广东省国土资源测绘院 | Remote sensing image change detection method and device |
| WO2024159812A1 (en) * | 2023-08-09 | 2024-08-08 | 广东省国土资源测绘院 | Cultivated land non-agricultural patch extraction method with multi-source sar data fused therein, and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115690591B (en) | 2023-03-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115690591B (en) | Remote sensing image farmland non-agricultural change detection method based on deep learning | |
| Kamilaris et al. | Deep learning in agriculture: A survey | |
| CN108830870B (en) | Satellite image high-precision farmland boundary extraction method based on multi-scale structure learning | |
| CN111626269B (en) | Practical large-space-range landslide extraction method | |
| Zhang et al. | Deep learning-based automatic recognition network of agricultural machinery images | |
| Ünsalan et al. | A system to detect houses and residential street networks in multispectral satellite images | |
| CN103971115A (en) | Automatic extraction method for newly-increased construction land image spots in high-resolution remote sensing images based on NDVI and PanTex index | |
| Bhagat et al. | Eff-UNet++: A novel architecture for plant leaf segmentation and counting | |
| CN107832797B (en) | Multispectral image classification method based on depth fusion residual error network | |
| CN105303184A (en) | Method for accurately identifying ground features in satellite remote-sensing image | |
| Rad et al. | Developing an automatic phenology-based algorithm for rice detection using sentinel-2 time-series data | |
| Liang et al. | Low-cost weed identification system using drones | |
| CN103198480A (en) | Remote sensing image change detection method based on area and Kmeans clustering | |
| Li et al. | Temporal Sequence Object-based CNN (TS-OCNN) for crop classification from fine resolution remote sensing image time-series | |
| Tan et al. | Rapid recognition of field-grown wheat spikes based on a superpixel segmentation algorithm using digital images | |
| Zhan et al. | Vegetation land use/land cover extraction from high-resolution satellite images based on adaptive context inference | |
| CN103971377A (en) | Building extraction method based on prior shape level set segmentation | |
| Gebrehiwot et al. | Automated indunation mapping: Comparison of methods | |
| Waleed et al. | An automated method for detection and enumeration of olive trees through remote sensing | |
| Liu et al. | Contour-oriented cropland extraction from high resolution remote sensing imagery using richer convolution features network | |
| Sharma et al. | Land cover classification: a comparative analysis of clustering techniques using Sentinel-2 data | |
| Dehkordi et al. | Performance Evaluation of Temporal and Spatial-Temporal Convolutional Neural Networks for Land-Cover Classification (A Case Study in Shahrekord, Iran) | |
| Dimov et al. | Feature extraction and machine learning for the classification of active cropland in the Aral Sea Basin | |
| CN116310864A (en) | Automatic identification method, system, electronic equipment and medium for crop lodging | |
| Chang et al. | Rice field mapping using sentinel-1a time series data and deep learning model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP03 | Change of name, title or address | ||
| CP03 | Change of name, title or address |
Address after: 210042 8 Blocks 699-22 Xuanwu Avenue, Xuanwu District, Nanjing City, Jiangsu Province Patentee after: Speed Technology Co.,Ltd. Address before: 210000 8 -22, 699 Xuanwu Road, Xuanwu District, Nanjing, Jiangsu. Patentee before: SPEED TIME AND SPACE INFORMATION TECHNOLOGY Co.,Ltd. |