CN113159154A - Time series characteristic reconstruction and dynamic identification method for crop classification - Google Patents
Time series characteristic reconstruction and dynamic identification method for crop classification Download PDFInfo
- Publication number
- CN113159154A CN113159154A CN202110399004.1A CN202110399004A CN113159154A CN 113159154 A CN113159154 A CN 113159154A CN 202110399004 A CN202110399004 A CN 202110399004A CN 113159154 A CN113159154 A CN 113159154A
- Authority
- CN
- China
- Prior art keywords
- classification
- dimension
- missing
- data
- value
- 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
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000006872 improvement Effects 0.000 claims abstract description 5
- 238000012795 verification Methods 0.000 claims description 18
- 238000012549 training Methods 0.000 claims description 15
- 238000011160 research Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 5
- 238000012217 deletion Methods 0.000 claims description 3
- 230000037430 deletion Effects 0.000 claims description 3
- 238000012804 iterative process Methods 0.000 claims description 2
- 238000012512 characterization method Methods 0.000 claims 1
- 230000007812 deficiency Effects 0.000 claims 1
- 238000005457 optimization Methods 0.000 abstract description 4
- 230000000875 corresponding effect Effects 0.000 description 13
- 238000003306 harvesting Methods 0.000 description 11
- 241000209094 Oryza Species 0.000 description 6
- 235000007164 Oryza sativa Nutrition 0.000 description 6
- 235000009566 rice Nutrition 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000003287 optical effect Effects 0.000 description 3
- 230000000306 recurrent effect Effects 0.000 description 3
- 244000068988 Glycine max Species 0.000 description 2
- 235000010469 Glycine max Nutrition 0.000 description 2
- 235000010716 Vigna mungo Nutrition 0.000 description 2
- 240000001417 Vigna umbellata Species 0.000 description 2
- 235000011453 Vigna umbellata Nutrition 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 241000209140 Triticum Species 0.000 description 1
- 235000021307 Triticum Nutrition 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000035558 fertility Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000002310 reflectometry Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000002689 soil Substances 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
A time series feature reconstruction and dynamic identification method for crop classification comprises the following steps: step 1: constructing an initial sample data set; step 2: predictive padding of missing data regions; and step 3: filling a value range of the missing data area; and 4, step 4: iterating the regressor and the classifier; and 5: dynamic classification; repeating the step 2 to the step 4, and constructing an R + 1-dimensional feature regressor and a classifier until the regressors and the classifiers of all R-dimensional features are established; and the reconstruction of the missing part of the data and the classification of crops are realized, and the dynamic identification of the crops is completed. The invention realizes the filling of the missing part of the time sequence remote sensing data by taking the LSTM network model as a frame and carrying out collaborative optimization on the regressor for filling the missing value and the classifier for classifying the crops, and realizes the gradual improvement of the precision of the classification result by classifying the crops according to the dynamically increased remote sensing data.
Description
Technical Field
The invention belongs to the field of image processing mode classification, and particularly relates to a crop classification-oriented time series feature reconstruction and dynamic identification method.
Background
Timely and accurate acquisition of spatial distribution information of crops is an important prerequisite for the work of growth monitoring, yield estimation, disaster assessment and the like of the crops, and has important significance for precision agriculture and related applications thereof. The remote sensing technology has the advantages of wide coverage area, short revisit period, relatively easy data acquisition, large information amount, low cost and the like, and is an efficient information acquisition method compared with the traditional field investigation mode. Thus, the crop type, planting area and spatial distribution information acquisition method based on remote sensing data is increasingly applied to the agricultural field. Since time series (hereinafter referred to as time series) Remote Sensing images have the capability of tracking the evolution of crop growth information, the method based on the time series Remote Sensing images is used for classifying crops to be an effective means for obtaining high-precision crop distribution mapping (Roy D P, Yan L. robust land-based crop time series modules [ J ]. Remote Sensing of Environment,2018: 110810).
A time series feature vector (such as a time series Normalized Difference Vegetation Index (NDVI)) describes the change of the land cover with time, and features of different time dimensions have strong correlation, namely, front and rear time phases are correlated. However, the traditional classification model (such as a deep convolutional neural network, a support vector machine, etc.) assumes that the dimensions of the features are independent from each other, so that the evolution of the time sequence feature expression cannot be effectively described, and the precision of the classification result is low. The Recurrent Neural Network (RNN) model enables the value h of the hidden layer to be hidden by introducing a ring into the Network structureiNot only depending on the current input value xiAlso dependent on the hidden layer value h obtained from the last inputi-1The structure enables the neural network to remember the previous network state, and realizes the modeling of the time correlation of the sequentially input dimensional features.
Among them, the Long Short-Term Memory network (LSTM) is a special recurrent neural network that effectively solves the problem of gradient surge or disappearance during training by introducing a special hidden layer called gate (or Memory cell) to remember the previous input. Currently, this network has applied crop classification based on time series images and achieved better precision than the traditional methods (Zhou Ya' nan, Luo Jiancheng, Feng Li, Yang Yingpin, Chen Yuehong, Wu Wei (2019): Long-short-term-memory-based crop classification using high-resolution optical images and multi-temporal SAR data [ J ], GIScience & Remote Sensing,56(8): 1170-1191).
However, the above methods have problems with both methods and data in applications such as crop type monitoring: in the method, the remote sensing technology is revisited according to a certain period, and is a streaming and dynamic data acquisition method. The traditional classification method requires all data to be processed simultaneously, i.e. all data need to be prepared at one time before classification; in the aspect of data, due to the existence of clouds and shadows thereof in remote sensing images, it is extremely difficult to directly acquire complete and non-missing crop time sequence information. This makes the features of the optical image dependent on the partially or totally missing data, resulting in inconsistent features in different dimensions, preventing accurate alignment of features (d.poiuliot and r.latifovic, "registration of land time series in the presence of accurate and sparse orientations: Development and assessment in not-efficient Alberta, Canada," Remote settings. environs., vol., 204.204, No. july 2017, pp.979-996,2018.).
Therefore, the dynamic classifier constructed by using the optical time series image needs to solve the problems of both the reconstruction of the missing region and the dynamic classification. RNNs represented by LSTM have achieved good results in a variety of applications, but all of these methods use the recurrent neural network as a feature extraction tool and a classifier with excellent classification performance. In contrast, the invention takes an LSTM network model as a framework, carries out collaborative optimization on missing value filling (called as a regressor) and crop classification (called as a classifier), realizes filling of missing parts of time sequence remote sensing data, classifies crops according to dynamically acquired remote sensing data, and realizes gradual improvement of classification precision and classification results.
Disclosure of Invention
The invention provides a time series characteristic reconstruction and dynamic identification method for crop classification, aiming at overcoming the defects in the prior art.
Firstly, constructing a regressor model by using an LSTM network, filling missing data and finishing time sequence characteristic reconstruction; then, constructing a classifier model based on the LSTM network, classifying the filled data, and then performing collaborative optimization on the regressor and the classifier according to the evaluation of the classification result; and finally, continuously generating a new regressor and a new classifier along with the dynamic increase of the time sequence crop data to realize the dynamic classification of the crops.
According to the principle, the time series feature reconstruction and dynamic identification method for crop classification comprises the following steps:
step 1: constructing an initial sample data set;
assuming that the research area has acquired R views x (R), sorting according to the acquisition date D yields:
X(R)=<X1,X2,...,XR> (1)
D(R)=<d1,d2,...,dR> (2)
wherein, XiE.X (R) denotes an image constituting a time series, diE.d represents the acquisition date of the corresponding image.
The time sequence feature x (r) obtained from the remote sensing image x (r) can be represented as:
x(R)=<x1,x2,...,xR> (3)
crop types are differentiated by the change in time by NDVI.
The n crops k (n) to be classified can be represented as:
K(n)=<k1,k2,...,kn> (4)
the labeled sample set S obtained by field data acquisition or image interpretation is expressed as:
S={s1,s2,...,sm}T (5)
where m is the number of samples obtained and T is the transposed symbol. Element s in the sample setiE.s marks the crop type and the corresponding set of labels is denoted L. To represent the characteristic dimension R available in the time series, the sample data set and the tag set are denoted S, respectivelyRAnd LR. There is one independent verification data set S 'at the same time'RAnd the corresponding tag set is represented as L'R。
The sample set is missing in part of the dimension due to the influence of clouds and cloud shadows.
Step 2: predictive padding of missing data regions;
initial data set S using regression model constructed based on LSTM networkRAnd verification dataset S'RThe missing data region of (a) is filled.
Assuming that R (1 < R ≦ R) represents the feature dimension to be filled, the missing region of the previous R-1 dimension feature is filled, and a suitable method can be used for filling.
The part of the initial data set without data loss on the r-dimension characteristic is represented as vrThe corresponding front r-1 dimensional partial data is denoted as Vr-1(ii) a The part of the initial data set with data missing on the r-dimension characteristic is represented as urThe corresponding front r-1 dimensional partial data is represented as Ur-1. Similarly, respective portions of the verification data set are each v'r,V'r-1,u'r,U'r-1And (4) showing.
Regression model M for constructing r-dimension characteristic based on LSTM networkreg(r) of (A). To represent the iterative process, we will represent it asThe superscript 0 indicates the number of iterations. The input during the model training is Vr-1Output is vr。
Will Ur-1Input deviceObtain a predicted result urWill urFilling corresponding positions of the data set to obtain a front r-dimensional characteristic non-missing data set Sr(P), and fill urThen obtaining the former r dimension partial data U of data deletion on the r dimension characteristicr. Obtaining a pre-r-dimensional feature-free dataset S of the validation dataset in the same mannerr'(P)。
And step 3: filling a value range of the missing data area;
first, using VrAs input, LrAs output, training classifier modelsThe U obtained in the step 2rInput deviceAnd (5) classifying, and comparing the classification result with the corresponding actual label. Data set UrComparing the label of each sample with the classification results of the previous r-1 dimension and the previous r dimension of the sample respectively, the following 4 cases are generated: (1) "T-T" case: classifying correctly (T) on the front r-1 dimension characteristic and the front r dimension characteristic, which shows that the actual characteristic and the predicted filling characteristic can be classified correctly; (2) "F-T" case: the front r-1 dimensional features are classified wrongly (F), but the r-th dimensional features are newly added to be classified correctly (T), which shows that the new features can provide valuable additional information to ensure that the classification result is correct; (3) "T-F" case: the classification of the front r-1 dimensional features is correct (T), but the classification is wrong after the new r-th dimensional features are added (F), which indicates that the classification of the model is guided by the error of the additional information in the new added features, and the sample needs to be adjusted to achieve correct classification; (4) "F-F" case: classification errors (F) are divided on the front r-1 dimension and the front r dimension, which shows that the existing features can not provide valuable additional information to enable the model to be classified correctly.
Obtaining a sample set Q with the classification result of the 3 rd T-F caser. The former r-1 dimensional features have been correctly classified, but the newly added features have not been correctly classified instead, mainly due to the newly added featuresMisleading information is added to the features, so that the classification result is wrong. Therefore, the invention carries out value taking in a certain value range and fills the missing value, and the specific method is as follows:
using vrCalculating crop type kiMean value of K-independent dataset on dimension r characteristicAnd standard deviation ofTo obtain crop kiReasonable range of values F (k) of the dataset on the r-th dimensioni):
Wherein, λ represents crop k in the r dimension characteristiciThe data is multiplied by the standard deviation of 1 or 2, since too large a will result in F (k)i) Outside the reasonable numerical range for NDVI.
Then, for QrFrom its corresponding label kiValue range F (k) ofi) First minimum valueAnd filling the r-dimension characteristic of the q, and then filling by taking the step size of 0.01 each time. Using each filled valueClassifying the sample q, recording the classification probability and corresponding filling value of using a new classification result as 'T-T' until the value exceeds F (k)i) The set of classification probabilities { q } of the case that the classification result is "T-T" after the sample q is filled is finally obtainedp(z), wherein z is a filling value corresponding to the classification probability. Taking the filling value corresponding to the maximum classification probability in the set as the optimal filling value Z of the sample q on the r-th dimensionm:
Zm=argmax{qp(z)} (7)
In the time series classification process, the probability of each class can be obtained through a Softmax function, and the maximum value of each class is used as an estimation result.
QrAfter all samples are filled with optimal values in a value range value filling mode, a new non-missing data set S is obtainedr(Q)。
And 4, step 4: iterating the regressor and the classifier;
using S obtained in step 3r(Q) training the regressor to obtain a regressor model iterated 1 timeReuse ofPredicting the r-dimensional missing part u 'of the verification dataset'rObtaining a verification data set S with the front r-dimension characteristic free of deficiencyr'。
Will Sr(Q) as input, LrAs output, training the classifier to obtain a 1-time iterated classifier modelWill Sr' inputThe classification is carried out to obtain a classification result.
The method is adopted to fill the missing area in the data set sample, so that the similarity of the crop samples of the same type can be improved, the prediction result of the regressor is gradually close to the actual type value, and the improvement of the classification precision is realized.
And 5: dynamic classification;
and (5) repeating the step (2) to the step (4), and constructing an R + 1-dimensional feature regressor and a classifier until the regressors and the classifiers of all the R-dimensional features are established. And the reconstruction of the missing part of the data and the classification of crops are realized, and the dynamic identification of the crops is completed.
The invention adopts the scheme has the advantages that: the invention realizes the filling of the missing part of the time sequence remote sensing data by taking the LSTM network model as a frame and carrying out collaborative optimization on the regressor for filling the missing value and the classifier for classifying the crops, and realizes the gradual improvement of the precision of the classification result by classifying the crops according to the dynamically increased remote sensing data.
Drawings
FIG. 1 is a flow chart of the process of the present invention.
FIGS. 2(a) to 2(b) show the positions of the regions of interest in the example of the present invention. FIG. 2(a) is a position of shou county in Anhui province; fig. 2(b) is a satellite image of sentry No. 2 in shou county, which is synthesized in a standard pseudo-color manner, i.e., near infrared, red, and green bands as red, green, and blue bands, respectively.
FIG. 3 is an image taken in accordance with an embodiment of the present invention. The dates of the images taken are indicated in the graph, and the percentage of each sub-image represents the percentage of the missing area in the current date image.
Fig. 4(a) to 4(h) show the results of local classification of the regions of interest according to the embodiment of the present invention. FIG. 4(a) is a local area of a summer crop in a 4-month 5-day image; FIG. 4(b) is a local area of a summer crop in a 5-month 20-day image; FIG. 4(c) local area of summer crop in the 6-month-4-day image; FIG. 4(e) is a partial area of autumn crop in 7 months and 14 days of image; FIG. 4(f) is a partial area of autumn crop in the image of day 22/9; FIG. 4(g) is a partial area of autumn crop in the image of day 29 of 10 months; fig. 4(d) and (h) are partial classification results of summer crops and autumn crops according to the embodiment of the present invention, respectively.
Detailed Description
In order to facilitate the understanding and practice of the present invention for those skilled in the art, the present invention will be described in further detail with reference to the following examples and the invention flow chart of fig. 1. It is to be understood that the specific embodiments described herein are for purposes of illustration only and are not to be construed as limiting the invention, and that the described embodiments are merely a subset of the embodiments of the invention, rather than a complete subset. Therefore, all other embodiments that can be obtained by a person skilled in the art without making any inventive step based on the embodiments of the present invention belong to the protection scope of the present invention.
In this embodiment, the research area is shou county in huai nan city, anhui (see fig. 2(a) to 2 (b)), and the whole year image of the sentinel 2 satellite 2019 is used in the research area, which has 145 scenes in total. Due to the influence of cloud and cloud shadow, the loss of the time sequence characteristics on partial dimensions is caused, the embodiment only reserves 26 scenes of images (as shown in fig. 3) which can provide effective information for classification, the more the number of the images is, the more the precision of the classification result can be improved, and the selection can be performed according to the requirements and the data acquisition condition. In this embodiment, the classification is performed in a parcel mode. The plot data is obtained by integrating 0.8m panchromatic data and 2m multispectral data obtained from China high-score No. 2 satellite. After that, 200 pieces of ten thousand land are extracted through a series of operations.
In the embodiment, a sentinel 2 satellite image of the whole year in 2019 collected in a research area is subjected to atmospheric correction by using a Sen2cor model to obtain a Level-2A-Level surface reflectivity product, and then a time sequence cloud-free image of the research area is obtained by combining a cloud mask algorithm based on deep learning with a manual visual interpretation mode. Selecting corresponding pixels on No. 2 cloudless images of sentinels in each period according to the phenological information of main crops in the research area, extracting NDVI data of the pixels on each scene image, and forming a high-dimensional feature vector based on the time sequence image to serve as a training sample data set.
This example divides the crops in the area of study into two broad categories: crops are harvested in summer (hereinafter, summer harvest, the growth cycle of the crops is 9 months, 10 months to 5 months of the second year) and autumn (hereinafter, autumn harvest, the growth cycle of the crops is 5 months to 10 months, 11 months). The summer crop comprises: wheat and rape, in order to maintain soil fertility, summer crops also contain more fallow areas; the autumn harvest of crops comprises: medium rice, late rice, and soybean.
The invention relates to a crop classification-oriented time series feature reconstruction and dynamic identification method, which aims at solving the problems that a traditional model cannot effectively describe the evolution process of time series feature expression and a traditional classification method cannot realize dynamic classification of flow-type acquired remote sensing crop data, and the specific implementation mode of the invention is described by combining the accompanying drawings, and the flow is as follows:
step 1: constructing an initial sample data set;
in this embodiment, the remaining 26 scene images are arranged according to the time sequence, and a set of pixels that meet the phenological characteristics of the crop type under study in this embodiment is selected from the images and given an interpreted crop type label. And then, calculating the NDVI of all the pixels in the pixel set on each scene image, and combining the labels to form a time sequence NDVI data set serving as an initial sample data set. Because images influenced by clouds and cloud shadows still exist in the reserved images, the time sequence characteristics are lost on partial dimensionality, and therefore the acquired data set contains lost data. In the embodiment, 2000 sample data sets are extracted, 1227 verification data sets are obtained, 500 summer crop samples and 727 autumn crop samples are obtained in the verification data sets.
Step 2: predictive padding of missing data regions;
and (3) dividing the data set obtained in the step (1) into a summer crop sample data set and an autumn crop sample data set according to the life cycles of the summer crops and the autumn crops. Since the autumn harvest crop lifecycle starts from month 5, the missing data area padding of the time sequence feature 5 months ago is required, and the autumn harvest crop feature contains all 26 days. In this embodiment, classification of autumn crops is described, R (R is more than 1 and less than or equal to R) is used to represent the feature dimension currently filled, and the missing region of the previous R-1-dimensional feature is filled completely, wherein the 1 st-dimensional feature is filled in other non-learning manners, and the other dimension features are filled in subsequent steps. The following steps are exemplified by the 14 th dimension feature (day 7, 24) of the autumn harvest crop data set, wherein the number of the missing samples on the dimension feature of the training data set is 342, and the number of the non-missing samples is 1658. The data missing samples of the verification data set on the dimensional features are 101, and the data non-missing samples are 626.
Construction of a regressor model of the 14 th dimensional features using an LSTM networkInputting a 14 th-dimension characteristic non-missing 1658 non-missing sample data set containing the first 13-dimension subjected to deletion filling into a regressorAnd (5) training. Then, 342 samples which are missing on the 14 th dimension characteristic and contain the first 13 dimensions and are subjected to missing filling are input into a trained regressorIn the above example, the data on the 14 th dimensional features of the 342 samples are predicted, and a training data set with no missing of the first 14 dimensional features is obtained. 101 missing data on the autumn crop verification data set are predicted in the same manner, and 727 verification data sets without missing front 14-dimensional features are obtained.
And step 3: filling a value range of the missing data area;
the procedure is to take autumn harvest crop medium rice as an example. And calculating the mean value and the standard deviation of the 14 th dimension characteristic of the sample of the type of the middle rice in the 1658 non-missing samples, and taking the parameter lambda as 2 to obtain the value range of the middle rice in the dimension.
Thereafter, the 1658 pieces of non-missing samples are used to train a classifier modelInputting the 342 samples filled by prediction in the step 2 into a classifierAnd comparing the obtained classification result with the actual label, and screening out 342 samples of which the type is middle rice and the classification on the 13-dimensional feature is correct but the classification on the 14-dimensional feature is wrong (namely, the classification result is T-F) as a sample set to be refilled.
And (3) filling each sample in the sample set to be refilled with the minimum value from the value range of the mid-season rice to the 14 th dimension characteristic of the sample, and then filling the minimum value with the step size of 0.01 each time. Each time a value is filled, the currently refilled sample is enteredAnd (4) classifying, and recording the classification probability and the current filling value when the classification is correct (namely the T-T condition) until the value exceeds the value range.
And selecting the filling value corresponding to the maximum probability from the recorded classification probabilities as the optimal filling value on the 14-dimensional feature of the current sample.
The sample sets to be refilled were filled with sample types late rice and soybean, respectively, in the same manner. After all the sample sets to be refilled obtain the optimal filling value, a new 2000 sample sets without missing first 14-dimensional features are obtained.
And 4, step 4: iterating the regressor and the classifier;
training a regressor model by using the new 2000 sample sets without missing first 14-dimensional features obtained in the step 3 to obtain an iterated regressor modelInputting 101 missing sample sets in the verification sample set into a regressorThe prediction filling is carried out to obtain 727 sample sets without loss of the first 14-dimensional characteristics.
Training a classifier model by using the new 2000 sample sets with the first 14-dimensional features obtained in the step 3 and without loss to obtain an iterative classifier modelIn thatMiddle input throughAnd predicting the filled 727 verification sample sets without missing front 14-dimensional features to obtain a classification result.
And 5: dynamic classification;
and (5) repeating the steps 2 to 4, and constructing the regressor and the classifier of the 15 th-dimensional feature until the regressors and the classifiers of all the 26-dimensional features are completely established. And the reconstruction of the missing part of the data and the classification of crops are realized, and the dynamic identification of the crops is completed.
The above-described process is primarily illustrative of the specific implementation of the method used in the present invention.
The embodiment includes two types of crops of summer harvest and autumn harvest, the operations of the steps are respectively executed on the two types of crops to obtain a classifier model and a regressor model of the crops of summer harvest and autumn harvest, the two models are used for classifying and predicting research areas, and predicted partial areas are selected to obtain classification results as shown in fig. 4(a) to 4 (h).
The embodiments described in this specification are merely illustrative of implementations of the inventive concept and the scope of the present invention should not be considered limited to the specific forms set forth in the embodiments but rather by the claims that follow the summary of the invention in equivalents thereof that would occur to those skilled in the art to which the inventive concept pertains.
Claims (3)
1. A time series feature reconstruction and dynamic identification method for crop classification comprises the following steps:
step 1: constructing an initial sample data set;
assuming that the research area has acquired R views x (R), sorting according to the acquisition date D yields:
X(R)=<X1,X2,...,XR〉 (1)
D(R)=<d1,d2,...,dR〉 (2)
wherein, XiE.X (R) denotes an image constituting a time series, diE is equal to D and represents the acquisition date of the corresponding image;
the time sequence feature x (r) obtained from the remote sensing image x (r) can be represented as:
x(R)=<x1,x2,...,xR> (3)
crop types are differentiated by changes in time through NDVI;
the n crops k (n) to be classified can be represented as:
K(n)=<k1,k2,...,kn> (4)
the labeled sample set S obtained by field data acquisition or image interpretation is expressed as:
S={s1,s2,...,sm}T (5)
wherein m is the number of samples obtained and T is a transposed symbol; element s in the sample setiThe e is marked with the crop type by S, and the corresponding label set is expressed as L; to represent the characteristic dimension R available in the time series, the sample data set and the tag set are denoted S, respectivelyRAnd LR(ii) a There is one independent verification data set S 'at the same time'RAnd the corresponding tag set is represented as L'R;
The sample set is missing in part of the dimensions due to the influence of clouds and cloud shadows;
step 2: predictive padding of missing data regions;
initial data set S using regression model constructed based on LSTM networkRAnd verification dataset S'RFilling the missing data area;
assuming that R (1 < R ≦ R) represents the feature dimension to be filled, and the missing region of the front R-1 dimension feature is filled completely, a proper method can be used for filling;
the part of the initial data set without data loss on the r-dimension characteristic is represented as vrThe corresponding front r-1 dimensional partial data is denoted as Vr-1(ii) a The part of the initial data set with data missing on the r-dimension characteristic is represented as urThe corresponding front r-1 dimensional partial data is represented as Ur-1(ii) a Similarly, respective portions of the verification data set are each v'r,V′r-1,u′r,U′r-1Represents;
regression model for constructing r-dimension characteristic based on LSTM networkMreg(r); to represent the iterative process, it is represented asThe superscript 0 indicates the number of iterations; the input during the model training is Vr-1Output is vr;
Will Ur-1Input deviceObtain a predicted result urWill urFilling corresponding positions of the data set to obtain a front r-dimensional characteristic non-missing data set Sr(P), and fill urThen obtaining the former r dimension partial data U of data deletion on the r dimension characteristicr(ii) a Obtaining a pre-r-dimensional feature-free dataset S 'of the verification dataset in the same manner'r(P);
And step 3: filling a value range of the missing data area;
first, using VrAs input, LrAs output, training classifier modelsThe U obtained in the step 2rInput deviceClassifying, and comparing a classification result with a corresponding actual label; data set UrComparing the label of each sample with the classification results of the previous r-1 dimension and the previous r dimension of the sample respectively, the following 4 cases are generated: (1) "T-T" case: classifying correctly (T) on the front r-1 dimension characteristic and the front r dimension characteristic, which shows that the actual characteristic and the predicted filling characteristic can be classified correctly; (2) "F-T" case: the front r-1 dimensional features are classified wrongly (F), but the r-th dimensional features are newly added to be classified correctly (T), which shows that the new features can provide valuable additional information to ensure that the classification result is correct; (3) "T-F" case: the front r-1 dimension feature is classified correctly (T), but the classification is wrong after the r dimension feature is added (F), which indicates that the new feature is addedThe classification of the model is guided by the error of the additional information in the characterization, and the sample needs to be adjusted to achieve the correct classification; (4) "F-F" case: classification errors (F) are divided on the front r-1 dimension characteristic and the front r dimension characteristic, which shows that the existing characteristic can not provide valuable additional information to enable the model to be classified correctly;
obtaining a sample set Q with the classification result of the 3 rd T-F caser(ii) a The former r-1 dimensional features can be correctly classified, but the newly added features cannot be correctly classified, mainly because misleading information is added into the newly added features, the classification result is wrong; therefore, the value is taken in the set value range, and the missing value is filled, and the specific method comprises the following steps:
using vrCalculating crop type kiMean value of K-independent dataset on dimension r characteristicAnd standard deviation ofTo obtain crop kiReasonable range of values F (k) of the dataset on the r-th dimensioni):
Wherein, λ represents crop k in the r dimension characteristiciλ times standard deviation of data;
then, for QrFrom its corresponding label kiValue range F (k) ofi) First minimum valueFilling the r-dimension characteristic of q, and then filling with a set step length value each time; using each filled valueClassifying the sample q, recording the classification probability and corresponding filling value of using a new classification result as 'T-T' until the value exceeds F (k)i) The set of classification probabilities { q } of the case that the classification result is "T-T" after the sample q is filled is finally obtainedp(z), wherein z is a filling value corresponding to the classification probability; taking the filling value corresponding to the maximum classification probability in the set as the optimal filling value Z of the sample q on the r-th dimensionm:
Zm=argmax{qp(z)} (7)
In the time series classification process, the probability of each category can be obtained through a Softmax function, and the maximum value of each category is used as an estimation result;
Qrafter all samples are filled with optimal values in a value range value filling mode, a new non-missing data set S is obtainedr(Q);
And 4, step 4: iterating the regressor and the classifier;
using S obtained in step 3r(Q) training the regressor to obtain a regressor model iterated 1 timeReuse ofPredicting the r-dimensional missing part u 'of the verification dataset'rObtaining a verification data set S 'with no deficiency of front r-dimension features'r;
Will Sr(Q) as input, LrAs output, training the classifier to obtain a 1-time iterated classifier modelIs prepared from S'rInput deviceClassifying to obtain a classification result;
the method is adopted to fill the missing area in the data set sample, so that the similarity of the crop samples of the same type can be improved, the prediction result of the regressor is gradually close to the actual type value, and the improvement of the classification precision is realized;
and 5: dynamic classification;
repeating the step 2 to the step 4, and constructing an R + 1-dimensional feature regressor and a classifier until the regressors and the classifiers of all R-dimensional features are established; and the reconstruction of the missing part of the data and the classification of crops are realized, and the dynamic identification of the crops is completed.
2. The method for reconstructing and dynamically identifying the time series characteristics of the crop classification according to claim 1, wherein: λ is 1 or 2.
3. The method for reconstructing and dynamically identifying the time series characteristics of the crop classification according to claim 1, wherein: to QrFor each sample q, the set step size value for each fill is 0.01.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110399004.1A CN113159154B (en) | 2021-04-12 | 2021-04-12 | Crop classification-oriented time sequence feature reconstruction and dynamic identification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110399004.1A CN113159154B (en) | 2021-04-12 | 2021-04-12 | Crop classification-oriented time sequence feature reconstruction and dynamic identification method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113159154A true CN113159154A (en) | 2021-07-23 |
CN113159154B CN113159154B (en) | 2024-03-29 |
Family
ID=76890639
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110399004.1A Active CN113159154B (en) | 2021-04-12 | 2021-04-12 | Crop classification-oriented time sequence feature reconstruction and dynamic identification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113159154B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116610662A (en) * | 2023-07-17 | 2023-08-18 | 金锐同创(北京)科技股份有限公司 | Filling method, filling device, computer equipment and medium for missing classification data |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110119717A (en) * | 2019-05-15 | 2019-08-13 | 中国科学院遥感与数字地球研究所 | A kind of Crop classification method based on multi-temporal NDVI and LST |
CN110287869A (en) * | 2019-06-25 | 2019-09-27 | 吉林大学 | High-resolution remote sensing image Crop classification method based on deep learning |
CN111523525A (en) * | 2020-07-02 | 2020-08-11 | 航天宏图信息技术股份有限公司 | Crop classification identification method and device and electronic equipment |
CN111598019A (en) * | 2020-05-19 | 2020-08-28 | 华中农业大学 | Crop type and planting mode identification method based on multi-source remote sensing data |
CN111798132A (en) * | 2020-07-06 | 2020-10-20 | 北京师范大学 | Dynamic farmland monitoring method and system based on multi-source time sequence remote sensing depth coordination |
-
2021
- 2021-04-12 CN CN202110399004.1A patent/CN113159154B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110119717A (en) * | 2019-05-15 | 2019-08-13 | 中国科学院遥感与数字地球研究所 | A kind of Crop classification method based on multi-temporal NDVI and LST |
CN110287869A (en) * | 2019-06-25 | 2019-09-27 | 吉林大学 | High-resolution remote sensing image Crop classification method based on deep learning |
CN111598019A (en) * | 2020-05-19 | 2020-08-28 | 华中农业大学 | Crop type and planting mode identification method based on multi-source remote sensing data |
CN111523525A (en) * | 2020-07-02 | 2020-08-11 | 航天宏图信息技术股份有限公司 | Crop classification identification method and device and electronic equipment |
CN111798132A (en) * | 2020-07-06 | 2020-10-20 | 北京师范大学 | Dynamic farmland monitoring method and system based on multi-source time sequence remote sensing depth coordination |
Non-Patent Citations (1)
Title |
---|
杜保佳;张晶;王宗明;毛德华;张淼;吴炳方;: "应用Sentinel-2A NDVI时间序列和面向对象决策树方法的农作物分类", 地球信息科学学报, no. 05, 5 June 2019 (2019-06-05) * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116610662A (en) * | 2023-07-17 | 2023-08-18 | 金锐同创(北京)科技股份有限公司 | Filling method, filling device, computer equipment and medium for missing classification data |
CN116610662B (en) * | 2023-07-17 | 2023-10-03 | 金锐同创(北京)科技股份有限公司 | Filling method, filling device, computer equipment and medium for missing classification data |
Also Published As
Publication number | Publication date |
---|---|
CN113159154B (en) | 2024-03-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103679675B (en) | Remote sensing image fusion method oriented to water quality quantitative remote sensing application | |
Xia et al. | Exploring the potential of Chinese GF-6 images for crop mapping in regions with complex agricultural landscapes | |
Solberg | Contextual data fusion applied to forest map revision | |
CN111368691B (en) | Unsupervised hyperspectral remote sensing image space spectrum feature extraction method | |
CN104463203B (en) | High-spectrum remote sensing semisupervised classification method based on the scoring of atural object classification degree of membership | |
CN111639587B (en) | Hyperspectral image classification method based on multi-scale spectrum space convolution neural network | |
Yang et al. | Semantic segmentation based on temporal features: Learning of temporal–spatial information from time-series SAR images for paddy rice mapping | |
CN113936214B (en) | Karst wetland vegetation community classification method based on fusion of aerospace remote sensing images | |
CN113159154B (en) | Crop classification-oriented time sequence feature reconstruction and dynamic identification method | |
Mirpulatov et al. | Pseudo-labeling approach for land cover classification through remote sensing observations with noisy labels | |
Bagan et al. | Sensitivity of the subspace method for land cover classification | |
Azizi et al. | Comprehensive wheat lodging detection after initial lodging using UAV RGB images | |
CN113139515A (en) | Hyperspectral image classification method based on conditional random field and depth feature learning | |
Baidar | Rice crop classification and yield estimation using multi-temporal sentinel-2 data: a case study of terai districts of Nepal | |
CN116863341A (en) | Crop classification and identification method and system based on time sequence satellite remote sensing image | |
Sun et al. | A highly efficient temporal-spatial probability synthesized model from multi-temporal remote sensing for paddy rice identification | |
Khobragade et al. | Optimization of statistical learning algorithm for crop discrimination using remote sensing data | |
Siok et al. | A simulation approach to the spectral quality of multispectral images enhancement | |
Yang et al. | 3D convolutional neural network for hyperspectral image classification using generative adversarial network | |
CN113420798A (en) | Hyperspectral image classification based on twin spectral attention consistency | |
CN114463642A (en) | Cultivated land plot extraction method based on deep learning | |
Guedes et al. | Image analysis for automatic measurement of crustose lichens | |
Luo et al. | Evaluation of machine learning-based urban surface mapping using a new moderate-resolution satellite imagery dataset | |
Abdalla et al. | Color Consistency of UAV Imagery using Multi-Channel CNN-based Image-to-Image Regression and Residual Learning | |
Bue et al. | Using spatial correspondences for hyperspectral knowledge transfer: Evaluation on synthetic data |
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 |