CN109409261B - Crop classification method and system - Google Patents
Crop classification method and system Download PDFInfo
- Publication number
- CN109409261B CN109409261B CN201811183874.XA CN201811183874A CN109409261B CN 109409261 B CN109409261 B CN 109409261B CN 201811183874 A CN201811183874 A CN 201811183874A CN 109409261 B CN109409261 B CN 109409261B
- Authority
- CN
- China
- Prior art keywords
- model
- rsnet
- remote sensing
- crop
- sensing network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/188—Vegetation
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a crop classification method and a crop classification system. The classification method comprises the following steps: constructing a historical crop marking sample according to the historical multi-temporal segmentation image and the historical crop weather history data; constructing a high-precision small-area current crop marker sample according to the current multi-temporal high-resolution image and the current crop phenological calendar data; establishing a remote sensing network RSNet model according to the ResNet model and the PSPNet model; pre-training the remote sensing network RSNet model according to the historical training sample, and establishing a pre-trained remote sensing network RSNet model; fine tuning the pre-trained remote sensing network RSNet model according to the current training sample, and establishing a fine tuning remote sensing network RSNet model; establishing an optimal pre-training remote sensing network RSNet model and an optimal fine-tuning remote sensing network RSNet model; and classifying the crops according to an optimal pre-training remote sensing network RSNet model and the optimal fine-tuning remote sensing network RSNet model. By adopting the classification method and the classification system provided by the invention, the workload of training sample selection can be reduced, and the classification efficiency can be improved.
Description
Technical Field
The invention relates to the field of crop classification, in particular to a crop classification method and a crop classification system.
Background
The method has the advantages of accurately acquiring and mastering the crop planting area and the spatial distribution condition thereof in real time, accurately estimating and predicting the crop yield, monitoring agricultural meteorological disasters and ensuring national grain safety, and has very important significance.
The remote sensing technology is a main technical means for monitoring the area and spatial distribution of crops at present, and with the development of the remote sensing technology, more and more remote sensing images with high time and high spatial resolution can be freely obtained, so that the data sources for remote sensing identification of crops are enriched, and the difficulty of the remote sensing image information extraction method is increased. The remote sensing classification of crops mainly comprises: the method comprises the steps of ground investigation, classification system determination, remote sensing data selection and pretreatment, classification feature selection, training sample selection, classification algorithm selection, precision evaluation and the like. The classification algorithm, the classification features and the training samples are three important aspects of remote sensing classification of crops, wherein the classification features and the training samples are key factors for judging whether the relation classification is successful or not. The traditional crop remote sensing classification method is mature and has achieved huge achievement under specific conditions. However, the selection of traditional classification features and training samples restricts the efficient application of remote sensing classification of crops: firstly, the feature of the ground feature in the medium-high resolution image is complex, the traditional feature selection method is difficult to extract the multi-scale and high-level features of the target ground feature, and the features extracted in a specific region can only be applied to the classification of the region, and the trans-region feature migration cannot be realized; secondly, under the influence of the landscape characteristics of crops, in the traditional classification algorithm, training samples in specific areas and specific time can only be applied to the classification of the specific areas, and the workload of selecting the training samples is large, so that the classification efficiency is reduced.
Disclosure of Invention
The invention aims to provide a crop classification method and a crop classification system, which aim to solve the problems of large workload of training sample selection and low classification efficiency.
In order to achieve the purpose, the invention provides the following scheme:
a method of classifying a crop, comprising:
acquiring historical multi-temporal time-sharing images and historical crop weather calendar data in a large-scale region of historical years; the large range is the area of province and above province;
constructing a historical crop marking sample according to the historical multi-temporal segmentation image and the historical crop phenological calendar data; the historical crop marking samples comprise historical training samples and historical verification samples, wherein the historical training samples are 85%, and the historical verification samples are 15%;
acquiring present multi-temporal high-resolution images of small areas in the current year and current crop climate calendar data; the small region is the area of a plurality of counties and regions;
constructing a high-precision small-area current agricultural crop marking sample according to the current multi-temporal high-resolution image and the current crop weather calendar data; the high-precision small-area present crop marker sample comprises a present training sample and a present verification sample, wherein the present training sample is 85%, and the present verification sample is 15%;
establishing a remote sensing network RSNet model according to the residual error network ResNet model and the pyramid pooling network PSPNet model; the remote sensing network RSNet model is a multispectral remote sensing image crop end-to-end classification model;
pre-training the remote sensing network RSNet model according to the historical training sample, and establishing a pre-training remote sensing network RSNet model;
performing precision evaluation on the pre-training remote sensing network RSNet model, and establishing an optimal pre-training remote sensing network RSNet model;
fine tuning the pre-trained remote sensing network RSNet model according to the present training sample, and establishing a fine tuning remote sensing network RSNet model;
performing precision evaluation on the fine tuning remote sensing network RSNet model, and establishing an optimal fine tuning remote sensing network RSNet model;
and classifying the crops according to the optimal pre-training remote sensing network RSNet model and the optimal fine-tuning remote sensing network RSNet model.
Optionally, the constructing a historical crop marking sample according to the historical multi-temporal segmentation image and the historical crop calendar data specifically includes:
and classifying the crops by using a method of combining target change detection and a support vector machine according to the historical multi-time phase segmentation images and the historical crop weather calendar data to construct a historical crop marking sample.
Optionally, the constructing a high-precision small-area current crop marker sample according to the current crop phenological calendar data and the current multi-temporal high-resolution image specifically includes:
according to the present multi-temporal high-resolution image and the current crop calendar data, performing image object segmentation by using an object-oriented method to construct different spectral features and texture features;
performing feature extraction and optimization on the different spectral features and textural features by using a random forest algorithm, performing high-precision crop classification by using a support vector machine classifier, and determining a high-resolution crop classification result;
and resampling the high-resolution crop classification result into a resolution which is the same as that of the medium-resolution crop marking sample by using a resampling method based on a mode principle, and constructing a high-precision small-area current crop marking sample.
Optionally, the precision evaluation is performed on the pre-trained remote sensing network RSNet model, and an optimal pre-trained remote sensing network RSNet model is established, which specifically includes:
according to the historical verification sample, performing precision evaluation on the pre-training remote sensing network RSNet model, and establishing an optimal pre-training remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
Optionally, the precision evaluation is performed on the fine tuning remote sensing network RSNet model, and the establishment of the optimal fine tuning remote sensing network RSNet model specifically includes:
according to the situation verification sample, performing precision evaluation on the fine tuning remote sensing network RSNet model, and establishing an optimal fine tuning remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
A crop classification system comprising:
the historical parameter acquisition module is used for acquiring historical multi-temporal time-sharing images and historical crop weather calendar data in a large-scale region of historical years; the large range is the area of province and above province;
the historical crop marking sample construction module is used for constructing a historical crop marking sample according to the historical multi-time-phase segmentation image and the historical crop weather data; the historical crop marking samples comprise historical training samples and historical verification samples, wherein the historical training samples are 85%, and the historical verification samples are 15%;
the current parameter acquisition module is used for acquiring the present multi-temporal high-resolution images of the small area of the current year and the current crop phenological calendar data; the small region is the area of a plurality of counties and regions;
the present crop marking sample construction module is used for constructing a high-precision small-area present crop marking sample according to the present multi-temporal high-resolution image and the current crop climate calendar data; the high-precision small-area present crop marker sample comprises a present training sample and a present verification sample, wherein the present training sample is 85%, and the present verification sample is 15%;
the remote sensing network RSNet model establishing module is used for establishing a remote sensing network RSNet model according to a residual error network ResNet model and a pyramid pooling network PSPNet model; the remote sensing network RSNet model is a multispectral remote sensing image crop end-to-end classification model;
the pre-training remote sensing network RSNet model establishing module is used for pre-training the remote sensing network RSNet model according to the historical training sample and establishing a pre-training remote sensing network RSNet model;
the optimal pre-training remote sensing network RSNet model establishing module is used for carrying out precision evaluation on the pre-training remote sensing network RSNet model and establishing an optimal pre-training remote sensing network RSNet model;
the fine-tuning remote sensing network RSNet model establishing module is used for fine-tuning the pre-trained remote sensing network RSNet model according to the situation training sample and establishing a fine-tuning remote sensing network RSNet model;
the system comprises an optimal fine-tuning remote sensing network RSNet model establishing module, a fine-tuning remote sensing network RSNet model establishing module and a fine-tuning remote sensing network RSNet model establishing module, wherein the optimal fine-tuning remote sensing network RSNet model establishing module is used for performing precision evaluation on the fine-tuning remote sensing network RSNet model and establishing an optimal fine-tuning remote sensing network RSNet;
and the classification module is used for classifying the crops according to the optimal pre-training remote sensing network RSNet model and the optimal fine-tuning remote sensing network RSNet model.
Optionally, the historical crop marking sample construction module specifically includes:
and the historical crop marking sample construction unit is used for classifying crops by using a method of combining target change detection and a support vector machine according to the historical multi-time phase segmentation image and the historical crop weather calendar data to construct a historical crop marking sample.
Optionally, the present pesticide marker sample construction module specifically includes:
the segmentation unit is used for performing image object segmentation by using an object-oriented method according to the present multi-temporal high-resolution image and the current crop climate calendar data to construct different spectral features and texture features;
the high-resolution crop classification result determining unit is used for extracting and optimizing the characteristics of the different spectral characteristics and the different texture characteristics by using a random forest algorithm, classifying high-precision crops by using a support vector machine classifier and determining a high-resolution crop classification result;
and the present crop marking sample construction unit is used for resampling the high-resolution crop classification result into a resolution which is the same as that of the medium-resolution crop marking sample by using a resampling method based on a mode principle, and constructing the high-precision small-area present crop marking sample.
Optionally, the optimal pre-training remote sensing network RSNet model establishing module specifically includes:
the optimal pre-training remote sensing network RSNet model establishing unit is used for carrying out precision evaluation on the pre-training remote sensing network RSNet model according to the historical verification sample and establishing an optimal pre-training remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
Optionally, the optimal fine-tuning remote sensing network RSNet model establishing module specifically includes:
the optimal fine-tuning remote sensing network RSNet model establishing unit is used for carrying out precision evaluation on the fine-tuning remote sensing network RSNet model according to the situation verification sample and establishing an optimal fine-tuning remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects: the invention provides a crop classification method and a crop classification system.A Residual Network (ResNet) model is utilized, and a new Remote Sensing Network RSNet (Remote Sensing Network) model suitable for multispectral Remote Sensing image end-to-end classification is generated by a ResNet-50 model and a Pyramid pooling Network (PSPNet) model; since the original ResNet model is only applicable to 3-band and input image processing with fixed size, the invention improves the RSNet model with 8-band input and combines with the PSPNet model with a multi-scale pooling layer and a deconvolution layer, so that the input image of the RSNet model can be of any size.
The invention fully utilizes the historical marked samples to achieve the aim of reducing the workload of sample selection, and obtains better classification results in the crop classification results of the regional situation images based on the pre-training remote sensing network RSNet model of the historical crop marked samples, thereby proving the reusability of the historical samples; the method has the advantages that the remote sensing network RSNet model is finely adjusted and pre-trained on the basis of the small quantity of high-precision small-area present crop marking samples, high classification precision is obtained in the classification of the area present image crops, the generalization capability of training sample time and space scales is realized, the problem of data starvation of a large quantity of marking samples in the remote sensing image classification is solved, the workload of manually selecting the training samples is greatly reduced, and the classification efficiency is improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
FIG. 1 is a flow chart of a method for classifying crops according to the present invention;
FIG. 2 is a diagram of a RSNet model network framework provided by the present invention;
FIG. 3 is a diagram of the accuracy evaluation result of the pre-trained RSNet model provided by the present invention;
FIG. 4 is a diagram showing the result of the precision evaluation of the post-fine-tuned RSNet model provided by the present invention;
FIG. 5 is a diagram of the classification results provided by the present invention;
FIG. 6 shows RSNet-based technology provided by the present invention2017A first classification result graph of the region partition crops of the model;
FIG. 7 shows RSNet-based technology provided by the present invention2017A second classification result graph of the region division crops of the model;
FIG. 8 shows RSNet-based technology provided by the present invention2017A third classification result graph of the region partition crops of the model;
fig. 9 is a view showing a structure of a crop classification system according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a crop classification method and a crop classification system, which can reduce the workload of training sample selection and improve the classification efficiency.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
Fig. 1 is a flowchart of a crop classification method according to the present invention, and as shown in fig. 1, a crop classification method includes:
step 101: acquiring historical multi-temporal time-sharing images and historical crop weather calendar data in a large-scale region of historical years; the large area is provincial area and above provincial area.
Step 102: constructing a historical crop marking sample according to the historical multi-temporal segmentation image and the historical crop phenological calendar data; the historical crop marking samples comprise historical training samples and historical verification samples, wherein the historical training samples are 85%, and the historical verification samples are 15%.
Based on historical multi-temporal segmentation images (such as a high-segmentation first-number 16-meter spatial resolution image) in a large range (such as a province) of historical years and crop climate data, crop classification is carried out by using a method of combining target change detection and a Support Vector Machine (SVM), historical crop marking samples are constructed, wherein 85% of the historical crop marking samples are used as training samples, and 15% of the historical crop marking samples are used as verification samples.
Step 103: acquiring present multi-temporal high-resolution images of small areas in the current year and current crop climate calendar data; the small region is the county area of a plurality of counties.
Step 104: constructing a high-precision small-area current agricultural crop marking sample according to the current multi-temporal high-resolution image and the current crop weather calendar data; the high-precision small-area spot crop marker sample comprises a spot training sample and a spot verification sample, wherein the spot training sample is 85%, and the spot verification sample is 15%.
Based on the present multi-temporal high-resolution images (such as high-resolution one-number 2 m/8 m spatial resolution fusion images) and crop climate calendar data in small areas (such as two counties) in the current year, carrying out image object segmentation by using an object-oriented method, constructing different spectral features and texture features based on image objects, carrying out feature extraction and optimization by using a Random Forest algorithm (RF), and carrying out high-precision crop classification by using an SVM classifier; and then resampling the high-resolution crop classification result into a resolution which is the same as that of the medium-resolution crop marking sample by adopting a resampling method of a mode principle, and constructing a high-precision small-area current crop marking sample, wherein 85% is used as a training sample, and 15% is used as a verification sample.
Step 105: establishing a remote sensing network RSNet model according to the residual error network ResNet model and the pyramid pooling network PSPNet model; the remote sensing network RSNet model is a multispectral remote sensing image crop end-to-end classification model.
The method improves a ResNet-50 model of a Residual Network (ResNet), combines the ResNet-50 model with a Pyramid pooling Network model (PSPNet) to generate a model suitable for multispectral Remote sensing image crop end-to-end classification, namely a Remote sensing Network model (RSNet), and is characterized in that FIG. 2 is a RSNet model Network framework diagram provided by the invention, and Table 1 is a model RSNet parameter table provided by the invention.
TABLE 1
Step 106: and pre-training the remote sensing network RSNet model according to the historical training samples to establish a pre-trained remote sensing network RSNet model.
And (3) based on a large-range historical crop mark training sample, migrating a ResNet-50 model, and pre-training the RSNet model to enable the model to learn and extract the characteristics related to the remote sensing image crops, so that the pre-training model can be suitable for remote sensing image crop classification.
Step 107: and carrying out precision evaluation on the pre-training remote sensing network RSNet model, and establishing an optimal pre-training remote sensing network RSNet model.
Step 108: and fine-tuning the pre-trained remote sensing network RSNet model according to the present training sample, and establishing a fine-tuned remote sensing network RSNet model.
Based on the small-area high-precision current crop marker training sample, the fully pre-trained RSNet model is subjected to fine adjustment, so that the RSNet can further learn the characteristics of the crops in the current remote sensing image, and the classification precision of the crops in the current remote sensing image is improved.
Step 109: and carrying out precision evaluation on the fine tuning remote sensing network RSNet model, and establishing an optimal fine tuning remote sensing network RSNet model.
The data for model precision evaluation in the invention are respectively 15% of verification samples in a large-range historical marking sample and 15% of verification samples in a small-area present marking sample; the indexes of the model precision evaluation are Overall classification precision (OA), drawing Precision (PA), user precision (UA), Kappa Coefficient (K) and Mean intersection ratio (MIoU) of the confusion matrix. Table 2 is a model accuracy evaluation index table, and the specific meaning and formula of each evaluation index are shown in table 2:
TABLE 2
Historical pre-training model result analysis
Fig. 3 is a diagram of the accuracy evaluation result of the pre-trained RSNet model provided by the present invention, as shown in fig. 3, it can be seen from fig. 3 that the Loss function value Loss rapidly decreases from 10 ten thousand to about 2 ten thousand, then slowly becomes flat, and finally remains at about 1.5 ten thousand; when the number of model iterations is less than 100 ten thousand, the precision fluctuation of each evaluation index is large, when the number of model iterations is more than 100 ten thousand, the model gradually tends to be stable, and the overall classification precision, the user precision of rice and corn and the drawing precision all reach more than 80%; through comparison of all accuracy evaluation indexes, a model with the highest overall classification accuracy and the highest Kappa coefficient is finally selected as an optimal pre-training model, namely a model with 1332000 iteration times (as shown in table 3) and named as RSNet2016。
TABLE 3
Fine-tuning model result analysis
Optimal pre-training RSNet-1332000 based on small-area present crop marker training sample pair2016The model is fine-tuned to obtain the precision evaluation result of the RSNet model after fine tuning, and FIG. 4 is the fine-tuned model provided by the present inventionAs shown in fig. 4, fig. 4 shows that the Loss rapidly decreases from 2.5 ten thousand to about 1.2 ten thousand, then slowly decreases, and finally remains at about 1.0 ten thousand; when the iteration number of the model is more than 6 ten thousand, the overall precision, the Kappa coefficient and the MIoU all have a descending trend, and the Loss value is still in the descending trend; this indicates that when the number of model iterations exceeds 6 ten thousand, the model begins to produce an overfit. Therefore, the lower the Loss value, the higher the model accuracy. By comparing the accuracy evaluation indexes, the overall accuracy and the Kappa coefficient of the model are the highest and are respectively 91.66% and 86.96% when the iteration number is 52000, and at the moment, the drawing accuracy and the user accuracy of the rice and the corn are both larger than 87.75% (as shown in table 4), so that the model is selected as the optimal RSNet model after fine adjustment and is named as the RSNet model2017。
TABLE 4
Step 110: and classifying the crops according to the optimal pre-training remote sensing network RSNet model and the optimal fine-tuning remote sensing network RSNet model.
The invention respectively pre-trains the history (RSNet) models2016) Fine tuning model (RSNet)2017) The method is applied to regional crop classification, precision evaluation is carried out by using high-precision small-region current crop marking samples, precision evaluation indexes are respectively total classification precision (OA), drawing Precision (PA), user precision (UA), kappa coefficient (K), and F1 score representing weighted average of drawing precision and user precision, F1 is [2 × (PA × UA) ]]/[(PA+UA)]。
Regional crop classification result analysis based on RSNet model
Table 5 is a table of the regional crop identification accuracy evaluation results of the 2 RSNet models provided by the present invention, as shown in table 5, and fig. 5 is a graph of the classification results provided by the present invention, as shown in fig. 5. As can be seen from table 5, the classification result of the model after the trimming of a small number of labeled samples in 2017 is good, the overall classification precision reaches 91.01%, the Kappa coefficient is 0.84, and the result shows that the trimming of the historical model based on a small number of current labeled samples is feasible; the precision of classifying the crops in the 2017 region by directly using the model pre-trained by the 2016 label sample is lower than that of a fine-tuning model, but the overall classification precision also reaches 75.06%, and the F1 scores of rice and corn are above 72.94%; it can be seen from the classification result graph that the historical model prediction result is better in the region with flat and regular terrain, but the corn and the rice have more serious misappearance in the low mountain and hilly areas in the middle and south and the relatively broken region in the northern plot. The reasons for this may be: firstly, the training samples in the low hilly area are less, so that the model in the hilly area is not fully trained; secondly, when model training is carried out, the small-class samples of the land parcel crushing area are convoluted into a large class (background value). As a whole, the classification results of the 2 RSNet models prove that the marked samples can be recycled and migrated.
TABLE 5
Based on RSNet2017Region-partitioned crop classification result analysis of models
Training samples participating in model fine adjustment exist in the overall region classification result, and also participate in model prediction, so that the false impression that the model accuracy is improved due to prediction of the training samples is possibly caused. To avoid the above problems, the present study is based solely on RSNet2017The model classifies crops of 2017 region verification samples, performs partition analysis, and the precision evaluation result is shown in table 6, and RSNet-based results are shown in fig. 6 to 82017And comparing the area crop identification result of the model with the truth value.
TABLE 6
As can be seen from Table 6, the overall classification accuracy of 14 scenes of images is greater than 88.38%, wherein the drawing accuracy, the user accuracy and the F1 score of corn are allThe number is above 83.01%. For rice, the precision is lower than that of corn overall, the drawing precision and the user precision of most images are more than 80%, wherein the drawing precision, the user precision and the F1 score of 2 images rice numbered as 9 and 48 are all lower than 80%, as can be seen from fig. 6 and 7, the rice area in the images is smaller. The drawing precision of the rice in the image with the number 29 is less than 80%, but the user precision is more than 90%, the place where the rice is wrong can be obviously seen from the classification result diagram, and the user precision of the rice in the image with the number 69 is less than 80%, but the drawing precision is 87.90%, which shows that the rice has a more obvious wrong division phenomenon, the place where the rice is wrong can also be obviously seen from the classification result diagram, and the rice area ratio in the 2 images is also smaller. According to the high-precision crop identification result in 2017, the areas of rice, corn and other types in the research area are 292.44km respectively2、2108.94km2、1807.48km2From this, it can be seen that rice belongs to the subclass, while corn and others belong to the major class, with the problem of sample imbalance. This shows that when the pre-training model is fine-tuned, if the training samples are unbalanced, the trained model is biased to the large class, i.e. the large class has good performance and the small class has poor performance. Thus, rice is less accurate overall than corn. However, as a whole, RSNet based2017The spatial distribution of regional crops identified by the model has strong consistency with the true value of the sample, the result enhances the reliability of predicting the situation image based on a small number of sample fine-tuning historical models, the generalization capability of the RSNet model is proved, the migration of the sample in the spatial scale is realized, and the defect that the training sample in the traditional classification method is limited to the specific image and the specific region is overcome.
Fig. 9 is a structural view of a crop classification system according to the present invention, and as shown in fig. 9, a crop classification system includes:
a historical parameter obtaining module 901, configured to obtain historical multi-temporal segmentation images and historical crop weather calendar data in a large-scale area of a historical year; the large area is provincial area and above provincial area.
A historical crop marking sample construction module 902, configured to construct a historical crop marking sample according to the historical multi-temporal segmentation image and the historical crop phenological calendar data; the historical crop marking samples comprise historical training samples and historical verification samples, wherein the historical training samples are 85%, and the historical verification samples are 15%.
The historical crop marking sample construction module 902 specifically includes: and the historical crop marking sample construction unit is used for classifying crops by using a method of combining target change detection and a support vector machine according to the historical multi-time phase segmentation image and the historical crop weather calendar data to construct a historical crop marking sample.
A current parameter obtaining module 903, configured to obtain a current multi-temporal high-resolution image of a small area in the current year and current crop climate calendar data; the small region is the county area of a plurality of counties.
A present crop marking sample construction module 904, configured to construct a high-precision small-area present crop marking sample according to the present multi-temporal high-resolution image and the current crop climate calendar data; the high-precision small-area spot crop marker sample comprises a spot training sample and a spot verification sample, wherein the spot training sample is 85%, and the spot verification sample is 15%.
The present pesticide marking sample construction module 904 specifically comprises:
the segmentation unit is used for performing image object segmentation by using an object-oriented method according to the present multi-temporal high-resolution image and the current crop climate calendar data to construct different spectral features and texture features;
the high-resolution crop classification result determining unit is used for extracting and optimizing the characteristics of the different spectral characteristics and the different texture characteristics by using a random forest algorithm, classifying high-precision crops by using a support vector machine classifier and determining a high-resolution crop classification result;
and the present crop marking sample construction unit is used for resampling the high-resolution crop classification result into a resolution which is the same as that of the medium-resolution crop marking sample by using a resampling method based on a mode principle, and constructing the high-precision small-area present crop marking sample.
The remote sensing network RSNet model establishing module 905 is used for establishing a remote sensing network RSNet model according to a residual error network ResNet model and a pyramid pooling network PSPNet model; the remote sensing network RSNet model is a multispectral remote sensing image crop end-to-end classification model.
And the pre-training remote sensing network RSNet model establishing module 906 is used for pre-training the remote sensing network RSNet model according to the historical training samples and establishing the pre-training remote sensing network RSNet model.
And the optimal pre-training remote sensing network RSNet model establishing module 907 is used for performing precision evaluation on the pre-training remote sensing network RSNet model and establishing an optimal pre-training remote sensing network RSNet model.
The optimal pre-training remote sensing network RSNet model establishing module 907 specifically includes:
the optimal pre-training remote sensing network RSNet model establishing unit is used for carrying out precision evaluation on the pre-training remote sensing network RSNet model according to the historical verification sample and establishing an optimal pre-training remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
And the fine-tuning remote sensing network RSNet model establishing module 908 is used for fine-tuning the pre-trained remote sensing network RSNet model according to the situation training sample and establishing the fine-tuning remote sensing network RSNet model.
And an optimal fine-tuning remote sensing network RSNet model establishing module 909, configured to perform precision evaluation on the fine-tuning remote sensing network RSNet model, and establish an optimal fine-tuning remote sensing network RSNet model.
The optimal fine-tuning remote sensing network RSNet model building module 909 specifically includes:
the optimal fine-tuning remote sensing network RSNet model establishing unit is used for carrying out precision evaluation on the fine-tuning remote sensing network RSNet model according to the situation verification sample and establishing an optimal fine-tuning remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
And the classification module 910 is configured to classify the crop according to the optimal pre-trained remote sensing network RSNet model and the optimal fine-tuning remote sensing network RSNet model.
The research classifies the high-resolution images by using an object-oriented classification method and constructs high-precision marking sample data; the method comprises the steps of taking a medium-resolution image as a core data source, migrating a model pre-trained by using historical marking samples based on a Remote Sensing Network model (RSNet), constructing a deep learning model suitable for a present task, carrying out efficient and accurate Remote Sensing automatic classification on crops, aiming at realizing generalization of time scale and space scale of the marking samples, reducing workload of marking sample selection, improving Remote Sensing identification efficiency of crops with medium-high resolution images, and providing powerful support for crop Remote Sensing estimation, agricultural meteorological disaster monitoring and the like.
Thus, the present invention has at least the following advantages:
1) ResNet and PSPNet are innovatively developed, and an RSNet model suitable for remote sensing image classification is constructed, wherein the pre-training model can be suitable for remote sensing image crop classification, and the classification precision of the remote sensing image crops can be improved by finely adjusting the remote sensing network model;
2) the historical labeled sample information is fully utilized, the problem that a large number of training samples are lost in the prior classification technology is solved, and generalization and migration of the labeled samples on the time-space scale are realized;
3) the workload of marking sample selection is reduced, and the efficiency of remote sensing identification of medium and high resolution image crops is improved.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.
Claims (10)
1. A method of classifying a crop, comprising:
acquiring historical multi-temporal time-sharing images and historical crop weather calendar data in a large-scale region of historical years; the large range is the area of the area above province;
constructing a historical crop marking sample according to the historical multi-temporal segmentation image and the historical crop phenological calendar data; the historical crop marking samples comprise historical training samples and historical verification samples, wherein the historical training samples are 85%, and the historical verification samples are 15%;
acquiring present multi-temporal high-resolution images of small areas in the current year and current crop climate calendar data; the small region is the area of a plurality of counties and regions;
constructing a high-precision small-area current agricultural crop marking sample according to the current multi-temporal high-resolution image and the current crop weather calendar data; the high-precision small-area present crop marker sample comprises a present training sample and a present verification sample, wherein the present training sample is 85%, and the present verification sample is 15%;
establishing a remote sensing network RSNet model according to the residual error network ResNet model and the pyramid pooling network PSPNet model; the remote sensing network RSNet model is a multispectral remote sensing image crop end-to-end classification model;
pre-training the remote sensing network RSNet model according to the historical training sample, and establishing a pre-training remote sensing network RSNet model;
performing precision evaluation on the pre-training remote sensing network RSNet model, and establishing an optimal pre-training remote sensing network RSNet model;
fine tuning the pre-trained remote sensing network RSNet model according to the present training sample, and establishing a fine tuning remote sensing network RSNet model;
performing precision evaluation on the fine tuning remote sensing network RSNet model, and establishing an optimal fine tuning remote sensing network RSNet model;
and classifying the crops according to the optimal pre-training remote sensing network RSNet model and the optimal fine-tuning remote sensing network RSNet model.
2. The classification method according to claim 1, wherein the constructing a historical crop marker sample from the historical multi-temporal segmentation image and the historical crop phenological calendar data comprises:
and classifying the crops by using a method of combining target change detection and a support vector machine according to the historical multi-time phase segmentation images and the historical crop weather calendar data to construct a historical crop marking sample.
3. The classification method according to claim 1, wherein the constructing of the high-precision small-area spot crop marker sample from the spot multi-temporal high-resolution image and the current crop phenology data specifically comprises:
according to the present multi-temporal high-resolution image and the current crop calendar data, performing image object segmentation by using an object-oriented method to construct different spectral features and texture features;
performing feature extraction and optimization on the different spectral features and textural features by using a random forest algorithm, performing high-precision crop classification by using a support vector machine classifier, and determining a high-resolution crop classification result;
and resampling the high-resolution crop classification result into a resolution which is the same as that of the medium-resolution crop marking sample by using a resampling method based on a mode principle, and constructing a high-precision small-area current crop marking sample.
4. The classification method according to claim 1, wherein the precision evaluation is performed on the pre-trained remote sensing network RSNet model to establish an optimal pre-trained remote sensing network RSNet model, and the method specifically comprises the following steps:
according to the historical verification sample, performing precision evaluation on the pre-training remote sensing network RSNet model, and establishing an optimal pre-training remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
5. The classification method according to claim 1, wherein the precision evaluation of the fine-tuning remote sensing network RSNet model to establish an optimal fine-tuning remote sensing network RSNet model specifically comprises:
according to the situation verification sample, performing precision evaluation on the fine tuning remote sensing network RSNet model, and establishing an optimal fine tuning remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
6. A crop classification system, comprising:
the historical parameter acquisition module is used for acquiring historical multi-temporal time-sharing images and historical crop weather calendar data in a large-scale region of historical years; the large range is the area of the area above province;
the historical crop marking sample construction module is used for constructing a historical crop marking sample according to the historical multi-time-phase segmentation image and the historical crop weather data; the historical crop marking samples comprise historical training samples and historical verification samples, wherein the historical training samples are 85%, and the historical verification samples are 15%;
the current parameter acquisition module is used for acquiring the present multi-temporal high-resolution images of the small area of the current year and the current crop phenological calendar data; the small region is the area of a plurality of counties and regions;
the present crop marking sample construction module is used for constructing a high-precision small-area present crop marking sample according to the present multi-temporal high-resolution image and the current crop climate calendar data; the high-precision small-area present crop marker sample comprises a present training sample and a present verification sample, wherein the present training sample is 85%, and the present verification sample is 15%;
the remote sensing network RSNet model establishing module is used for establishing a remote sensing network RSNet model according to a residual error network ResNet model and a pyramid pooling network PSPNet model; the remote sensing network RSNet model is a multispectral remote sensing image crop end-to-end classification model;
the pre-training remote sensing network RSNet model establishing module is used for pre-training the remote sensing network RSNet model according to the historical training sample and establishing a pre-training remote sensing network RSNet model;
the optimal pre-training remote sensing network RSNet model establishing module is used for carrying out precision evaluation on the pre-training remote sensing network RSNet model and establishing an optimal pre-training remote sensing network RSNet model;
the fine-tuning remote sensing network RSNet model establishing module is used for fine-tuning the pre-trained remote sensing network RSNet model according to the situation training sample and establishing a fine-tuning remote sensing network RSNet model;
the system comprises an optimal fine-tuning remote sensing network RSNet model establishing module, a fine-tuning remote sensing network RSNet model establishing module and a fine-tuning remote sensing network RSNet model establishing module, wherein the optimal fine-tuning remote sensing network RSNet model establishing module is used for performing precision evaluation on the fine-tuning remote sensing network RSNet model and establishing an optimal fine-tuning remote sensing network RSNet;
and the classification module is used for classifying the crops according to the optimal pre-training remote sensing network RSNet model and the optimal fine-tuning remote sensing network RSNet model.
7. The classification system according to claim 6, wherein the historical crop marker sample construction module specifically comprises:
and the historical crop marking sample construction unit is used for classifying crops by using a method of combining target change detection and a support vector machine according to the historical multi-time phase segmentation image and the historical crop weather calendar data to construct a historical crop marking sample.
8. The classification system according to claim 6, wherein the present crop marker sample construction module specifically comprises:
the segmentation unit is used for performing image object segmentation by using an object-oriented method according to the present multi-temporal high-resolution image and the current crop climate calendar data to construct different spectral features and texture features;
the high-resolution crop classification result determining unit is used for extracting and optimizing the characteristics of the different spectral characteristics and the different texture characteristics by using a random forest algorithm, classifying high-precision crops by using a support vector machine classifier and determining a high-resolution crop classification result;
and the present crop marking sample construction unit is used for resampling the high-resolution crop classification result into a resolution which is the same as that of the medium-resolution crop marking sample by using a resampling method based on a mode principle, and constructing the high-precision small-area present crop marking sample.
9. The classification system according to claim 6, wherein the optimal pre-trained remote sensing network RSNet model building module specifically comprises:
the optimal pre-training remote sensing network RSNet model establishing unit is used for carrying out precision evaluation on the pre-training remote sensing network RSNet model according to the historical verification sample and establishing an optimal pre-training remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
10. The classification system according to claim 6, wherein the optimal fine-tuning remote sensing network RSNet model building module specifically comprises:
the optimal fine-tuning remote sensing network RSNet model establishing unit is used for carrying out precision evaluation on the fine-tuning remote sensing network RSNet model according to the situation verification sample and establishing an optimal fine-tuning remote sensing network RSNet model; the indexes of the precision evaluation are the overall classification precision, drawing precision, user precision, Kappa coefficient and average cross-over ratio of the confusion matrix.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811183874.XA CN109409261B (en) | 2018-10-11 | 2018-10-11 | Crop classification method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811183874.XA CN109409261B (en) | 2018-10-11 | 2018-10-11 | Crop classification method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109409261A CN109409261A (en) | 2019-03-01 |
| CN109409261B true CN109409261B (en) | 2020-09-15 |
Family
ID=65466994
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811183874.XA Active CN109409261B (en) | 2018-10-11 | 2018-10-11 | Crop classification method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109409261B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111753834B (en) * | 2019-03-29 | 2024-03-26 | 中国水利水电科学研究院 | Planting land block structure semantic segmentation method and device based on deep neural network |
| CN110287869B (en) * | 2019-06-25 | 2022-03-18 | 吉林大学 | High-resolution remote sensing image crop classification method based on deep learning |
| WO2021068176A1 (en) * | 2019-10-11 | 2021-04-15 | 安徽中科智能感知产业技术研究院有限责任公司 | Crop planting distribution prediction method based on time series remote sensing data and convolutional neural network |
| CN113033279A (en) * | 2020-12-10 | 2021-06-25 | 四川航天神坤科技有限公司 | Crop fine classification method and system based on multi-source remote sensing image |
| CN113033453B (en) * | 2021-04-06 | 2023-09-15 | 北京艾尔思时代科技有限公司 | Method and system suitable for remote sensing identification of crop types in landscape crushing area |
| CN113269794A (en) * | 2021-05-27 | 2021-08-17 | 中山大学孙逸仙纪念医院 | Image area segmentation method and device, terminal equipment and storage medium |
| CN114283335B (en) * | 2021-12-27 | 2022-11-22 | 河南大学 | Historical period remote sensing identification precision verification preparation method |
| CN114332570B (en) * | 2022-03-17 | 2022-06-03 | 北京艾尔思时代科技有限公司 | Method and system for carrying out crop migration classification based on CDL deep learning |
| CN116740468A (en) * | 2023-07-18 | 2023-09-12 | 阳光农业相互保险公司 | Crop classification method based on 3DUNet |
| CN116664959B (en) * | 2023-07-28 | 2023-11-03 | 航天宏图信息技术股份有限公司 | Sample-free crop classification method and device based on multi-source satellite remote sensing |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103679131B (en) * | 2013-01-23 | 2016-11-16 | 福州大学 | Many seasons of based on sequential remote sensing image crop automatic recognition method |
| CN104899897B (en) * | 2015-05-27 | 2018-02-27 | 中国科学院地理科学与资源研究所 | A kind of high score remote sensing image land cover pattern change detecting method excavated based on historical data |
| CN108363949B (en) * | 2017-12-27 | 2020-08-28 | 二十一世纪空间技术应用股份有限公司 | Cotton remote sensing monitoring method based on phenological analysis |
| CN108304791A (en) * | 2018-01-23 | 2018-07-20 | 山东农业大学 | Seeds multispectral remote sensing recognition methods is easily obscured in a kind of mountain area based on cloud model |
-
2018
- 2018-10-11 CN CN201811183874.XA patent/CN109409261B/en active Active
Non-Patent Citations (1)
| Title |
|---|
| "FULLY CONVOLUTIONAL NEURAL NETWORKS FOR LARGE SCALE CROPLAND MAPPING WITH HISTORICAL LABEL DATASET";Dujuan Zhang 等;《IGARSS 2018》;20180727;第4659-4662页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109409261A (en) | 2019-03-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109409261B (en) | Crop classification method and system | |
| Jin et al. | Land-cover mapping using Random Forest classification and incorporating NDVI time-series and texture: A case study of central Shandong | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN109034224B (en) | Hyperspectral classification method based on double branch network | |
| CN110458048A (en) | Take population distribution Spatio-temporal Evolution and the cognition of town pattern feature into account | |
| CN110110642A (en) | A kind of pedestrian's recognition methods again based on multichannel attention feature | |
| CN107392901A (en) | A kind of method for transmission line part intelligence automatic identification | |
| CN108984803B (en) | Method and system for spatializing crop yield | |
| CN108052966A (en) | Remote sensing images scene based on convolutional neural networks automatically extracts and sorting technique | |
| CN111639787A (en) | Spatio-temporal data prediction method based on graph convolution network | |
| CN109145885B (en) | Remote sensing classification method and system for large-scale crops | |
| CN112347970B (en) | Remote sensing image ground object identification method based on graph convolution neural network | |
| CN109740483A (en) | A kind of rice growing season detection method based on deep-neural-network | |
| CN112749627A (en) | Method and device for dynamically monitoring tobacco based on multi-source remote sensing image | |
| CN111507296A (en) | Intelligent illegal building extraction method based on unmanned aerial vehicle remote sensing and deep learning | |
| CN113435254A (en) | Sentinel second image-based farmland deep learning extraction method | |
| CN113469122A (en) | Deep learning based crop space-time generalization classification method and system | |
| CN113887517A (en) | Crop remote sensing image semantic segmentation method based on parallel attention mechanism | |
| CN115860269A (en) | Crop yield prediction method based on triple attention mechanism | |
| CN114187532A (en) | Method and system for generating space-time sample of remote sensing image and intelligently iteratively classifying space-time sample | |
| CN114707785A (en) | Rural residential point multi-scale spatial feature analysis method based on deep learning | |
| CN113963040A (en) | Point cloud registration method and system based on depth feature consistency | |
| Sun et al. | Automatic building age prediction from street view images | |
| Shifaw et al. | Farmland dynamics in Pingtan, China: understanding its transition, landscape structure and driving factors | |
| Wang | [Retracted] Landscape Classification Method Using Improved U‐Net Model in Remote Sensing Image Ecological Environment Monitoring System |
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 |