CN111814563A - A kind of classification method and device of planting structure - Google Patents

Abstract

The invention relates to a method and a device for classifying a planting structure, which mainly preprocess remote sensing data, namely extract a planting structure index of the remote sensing data, add the determined planting structure index into original remote sensing image data, fuse images of a plurality of wave bands, use the fused images as a training set after labeling, train a constructed network model, preprocess the remote sensing image to be classified and input the preprocessed remote sensing image into the trained network model to classify the planting structure. According to the invention, the spatial resolution and the spectrum richness of the images for training are increased, so that the parameters in the subsequent network model training are determined, and the remote sensing images to be tested are accurately classified.

Description

Method and device for classifying planting structures

technical field

The invention relates to a planting structure classification method and device, belonging to the technical field of agricultural information.

Background technique

The planting structure of crops is an important part of the spatial pattern of crops, and its spatiotemporal change information is an important expression of the crop planting method of a region or production unit. Regional crop yield estimation, agricultural planting structure adjustment and optimization depend on accurate and timely determination of agricultural planting structure change information. The use of remote sensing combined with high and new technology to extract crop planting structure has good application prospects and huge development potential.

At the same time, with the development of machine learning, algorithms such as neural network (NNS) and support vector machine (SVM) have been applied to the classification of high-resolution remote sensing images, but they are all shallow learning algorithms and cannot express complex functions well . Therefore, these models cannot adapt to the semantic segmentation problem with large sample size and high complexity.

Subsequently, according to the advantages of automatic feature extraction and automatic classification of convolutional neural network (CNN), the researchers combined it with the field of remote sensing to realize the classification of planting structures. However, it should be noted that the data source of the current method is too single, the classification quality of the land use in the village is not high, and the information of the land use in the village cannot be accurately obtained.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a classification method and device for planting structures, so as to solve the problem in the prior art that the information of the land use situation in the village cannot be accurately obtained.

For achieving the above object, the technical scheme of the classification method of the planting structure of the present invention comprises the following steps:

1) Obtain the original remote sensing image data;

2) Preprocessing the original remote sensing image data:

According to the original remote sensing image data, extract at least one planting structure index among NDVI, AWEI, and SAVI; wherein, NDVI is the normalized difference vegetation index, AWEI is the automatic water extraction index, and SAVI is the soil adjustment vegetation index;

Add the extracted planting structure index to the original remote sensing image data to obtain an image;

Fusion of images of multiple bands;

Label the fused images to obtain a training set;

3) Build a network model;

4) Input the training set into the constructed network model for training to obtain a trained network model;

5) Input the remote sensing images to be classified into the trained network model after preprocessing to classify the planting structure.

The beneficial effects of the present invention are:

In the present invention, by preprocessing the remote sensing data, that is, by extracting the planting structure index from the remote sensing data, the determined planting structure index is added to the original remote sensing image data, and images of multiple bands are fused, that is By increasing the spatial resolution and spectral richness of the training images, the parameters in the subsequent network model training can be determined, and the remote sensing images to be tested can be accurately classified.

Further, the fusion adopts the G-S image fusion method.

Further, the formula of the normalized difference vegetation index is:

NDVI=(NIR-RED)/(NIR+RED)

where NIR is the reflectance in the near-infrared band, and RED is the reflectance in the red band.

Further, the formula of the automatic water extraction index is:

AWEI=BLUE+2.5*GREEN-1.5*(NIR+SWIR1)-0.25*SWIR2

where BLUE is the blue light band, GREEN is the green band, and SWIR1 and SWIR2 are the short-wave infrared band.

Further, the formula for the soil adjustment vegetation index is:

In the formula, L is the soil adjustment factor corresponding to the vegetation coverage, and L is taken as 0.5.

Further, the constructed network model is a U-Net network model, and the U-Net network model includes a depthwise separable convolution module and an activation function.

The present invention also provides a technical solution of a planting structure classification device, the device includes a processor and a memory, and the processor executes the technical solution of the above-mentioned classification method of the planting structure stored in the memory.

Description of drawings

Fig. 1 is the flow chart of the classification method embodiment of planting structure of the present invention;

Fig. 2 is the structural representation of the U-Net network model of the present invention;

Figure 3-a is the conventional U-Net training accuracy curve;

Fig. 3-b is the Separable-UNet training accuracy rate curve of the present invention;

Fig. 3-c is the Mish-Separable-UNet training accuracy rate curve of the present invention;

Figure 4-a is the seven-band classification result of the original remote sensing impact data;

Figure 4-b is the classification result of adding NDVI to the seven bands of the original remote sensing influence data according to the present invention;

Figure 4-c is the classification result of adding AWEI to the seven bands of the original remote sensing impact data of the present invention;

Fig. 4-d is the classification result that adds SAVI in the seven bands of original remote sensing influence data of the present invention;

Figure 4-e is the classification result of adding NDVI and AWEI to the seven bands of the original remote sensing influence data of the present invention;

Fig. 4-f is the classification result of adding NDVI and SAVI in the seven bands of original remote sensing influence data of the present invention;

Fig. 4-g is the classification result of adding AWEI and SAVI in the seven bands of original remote sensing influence data of the present invention;

Fig. 4-h is the classification result of adding three kinds of planting result indices in the seven bands of original remote sensing influence data of the present invention;

FIG. 5 is a schematic structural diagram of an embodiment of a device for classifying planting structures according to the present invention.

Detailed ways

The solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

Example of planting structure classification method:

Taking an irrigation area of a city in a certain province as an example, using the Landsat8 remote sensing image data of the area on a certain date in May, which has been stored in the database, the planting structure classification method of the present invention is specifically introduced.

The planting structure classification method of the present invention, as shown in Figure 1, specifically comprises the following steps:

1) Obtain the original remote sensing image data;

In this embodiment, the original remote sensing image data is the acquired Landsat8 commercial multispectral satellite data set.

2) Preprocessing the original remote sensing image data:

According to the original remote sensing image data, extract at least one planting structure index among NDVI, AWEI, and SAVI; wherein, NDVI is the normalized difference vegetation index, AWEI is the automatic water extraction index, and SAVI is the soil adjustment vegetation index;

Add the extracted planting structure index to the original remote sensing image data to obtain an image;

Fusion of images of multiple bands;

Label the fused images to obtain a training set;

Among them, the Normalized Difference Vegetation Index (NDVI) is based on the fact that the chlorophyll in the vegetation has a relatively low reflectance in the red band of the visible light band and relatively high reflectance in the green band, and the reflectivity in the near-infrared band to the mid-infrared band gradually decreases. The most representative normalized difference vegetation index is constructed, and its formula is as follows:

NDVI=(NIR-RED)/(NIR+RED)

where NIR is the reflectance in the near-infrared band, and RED is the reflectance in the red band.

The Automated Water Extraction Index (AWEI) improves the factors such as low classification accuracy and relatively unfixed threshold selection in water extraction. Its formula is as follows:

AWEI _sh =BLUE+2.5*GREEN-1.5*(NIR+SWIR1)-0.25*SWIR2

where BLUE is the blue light band, GREEN is the green band, and SWIR1 and SWIR2 are the short-wave infrared band.

The Soil Adjusted Vegetation Index (SAVI) introduces a soil adjustment factor to improve the identification ability of the normalized index under different soil backgrounds. The formula is as follows:

In the formula, L is the vegetation coverage corresponding to the soil adjustment factor, 0 is no vegetation, and 1 is complete vegetation coverage. In this paper, according to the actual situation in the study area, L=0.5 is selected as the optimal factor value.

It should be noted that in this embodiment, three planting indices, NDVI, AWEI and SAVI, are selected based on different index characteristics and added to the 1-7 band remote sensing image to improve the spectral complexity, wherein NDVI and SAVI can effectively distinguish the spectral difference between vegetation and vegetation and soil. Differences, AWEI has the ability to efficiently identify water bodies, and the selection of these three planting structure indices at the same time can clarify the land use situation in the region, which is of great significance for the rational planning of subsequent land.

As another embodiment, the selection of the planting structure index in the present invention may also select one or both of the above to obtain an index image map. However, it should be noted that when one or two of them are selected, such as AWEI, it has the ability to efficiently identify water bodies and can efficiently distinguish water bodies, but the distinction between vegetation is not obvious.

The G-S image fusion method is used to fuse images of multiple bands.

It should be pointed out that the synthesis process of the index and the remote sensing image map in this embodiment is as follows: increasing the dimension of the image matrix data to 3 dimensions, superimposing multiple band images and index image data matrices to the second dimension in the 3-dimensional matrix, and generating 3D multi-channel data, realize the synthesis of index and remote sensing image map.

Then, on the basis of the above synthesis, the synthesized data is fused with the panchromatic band image using the G-S image fusion method to achieve image fusion that improves the spatial resolution and spectral complexity of the data. The fusion data realized by the above process is sliced into the network and trained with the label data.

In this embodiment, the labeling of the image is performed according to known data or after field investigation.

In this embodiment, 80% of them are randomly selected as training data to make a training set for training the network, and the other 20% are used as test set data for result evaluation; 128*128 pixel slices are made for training images and validation set images and correspond one-to-one .

3) Build a network model;

The network model in this embodiment adopts the U-Net network model, which includes a depthwise separable convolution module and an activation function, as shown in Figure 2;

Among them, the depthwise separable convolution is a combination of depthwise (DW) and pointwise (PW) to extract features. Compared with conventional convolution operations, the number of parameters and operation cost are relatively low. .

In this embodiment, the depthwise separable convolution is adopted, which effectively reduces the amount of parameter calculation. The depthwise separable convolution considers both channel and region changes, while the conventional convolution only considers the region first, and then considers the channel, realizing the separation of the channel and the region.

Tanh x值域为(-1，1)，该函数为过原点且穿越I、III象限的严格单调递增曲线。Among them, the Mish activation function, its formula is: Mish=x*tanh(ln(1+e ^x )), where the formula of the Tanh function is:

The Tanh x value range is (-1, 1), and the function is a strictly monotonically increasing curve passing through the origin and passing through the I and III quadrants.

The Mish activation function in this embodiment avoids saturation due to capping, and the smooth activation function allows better information to penetrate deep into the neural network, thereby obtaining better accuracy and generalization, and further improving operational efficiency and accuracy. .

4) Input the training set into the constructed network model for training to obtain a trained network model;

5) Input the remote sensing images to be classified into the trained network model after preprocessing to classify the planting structure.

Input the training data set into the network model for training, apply the trained weight file to predict the test set, and obtain the classification result.

Further, in order to evaluate the accuracy of the method of the present application, the step of evaluating the test results is also included. Specifically: in this embodiment, the precision, the recall rate, the harmonic mean of the two, and the kappa coefficient are used as the indicators of the evaluation method. These metrics are calculated from a confusion matrix, where the formula for accuracy is:

where C _ii indicates the number of correctly classified samples, and C _ij indicates that class I samples are mistaken for class J.

Recall represents the average correct proportion of pixels classified into a class, and the formula for Recall is:

In addition, the model can be further evaluated by calculating the harmonic mean F1 of F precision and recall, the formula for the F1 value is:

F1=2*(Precision*Recall)/(Precision+Recall) (3)

The kappa coefficient measures the agreement of predicted classes with human labels. The formula for the kappa coefficient is:

In order to further verify the effectiveness of the method of the present invention, the present invention only compares and analyzes U-Net and its improved model. In the present invention, the conventional U-Net model, the U-Net model with the depth separable convolution module, and the modified activation function network model are obtained by training 200 batches of data in the same test set, as shown in Figure 3-a to Figure 3-c curve shown.

It can be seen that, compared with the traditional U-Net model, after adding the depth separable convolution module, the model achieves overfitting batches from 100 to 30, which greatly reduces the amount of calculation and improves the fitting rate. After using the Mish activation function, the accuracy and calculation rate are slightly improved due to its smoothness. The improved network model achieves lightweight parameters, and the weight file size is reduced from 303MB to 53MB, a reduction of 82.5%, which greatly saves production costs.

At the same time, the present invention synthesizes the seven bands 1-7 in the original remote sensing image with 1, 2, and 3 indices respectively, and then uses the G-S method to fuse the panchromatic bands to verify the prediction results. The prediction results are shown in the figure. 4-a to 4-h.

Table 1 below compares the prediction results of each method with the accuracy of the real label data. After investigation, most of the forest land area is economic crops such as jujube trees, so wheat, forest land and cotton are classified into crop categories.

Table 1

As shown in Table 1, the overall accuracy rate of the 7-band fusion image without index addition reaches 90.3%, and the accuracy rate of the image fusion with the addition of the index image is higher than that without the index. Structural classification had a positive effect.

In the case of adding a single index, all three indexes show an improvement in the overall accuracy, among which AWEI has the greatest impact on the classification of planting structure, and the accuracy of crop identification is increased by 2.4%. The recognition is better than the other two indexes by 81.4% and 93.9%, and the overall accuracy rate is 92.7%, which is the best among the three indexes.

Under the condition of two index 9-band fusion images, the AWEI index still plays an important role, and the classification effect of training images containing AWEI is better than that of NDVI-SAVI 9-band fusion images. The NDVI index further improved AWEI's crop identification ability and improved the mutual identification ability between crops. The accuracy rates of wheat, cotton, and woodland were increased to 85.2%, 74.4%, and 87.5%, respectively. At the same time, it also guarantees a high water body and village identification accuracy of 81.8% and 93.5%, respectively. Only the water body identification accuracy rate is lower than the AWEI-SAVI method, with a difference of 0.3%, and the overall accuracy rate reaches 93.7%.

Compared with the NDVI-AWEI 9-band fusion image method, the overall accuracy of the 10-band fusion image is improved by only 0.1%, and the recognition degree of crop classification is reduced. The improvement of the fusion image by SAVI is limited, and the priority is lower considering the production cost. To sum up, the NDVI-AWEI 9-band fusion image method combined with the improved U-Net network model has superiority in classifying the planting structure.

The present invention evaluates the reliability and accuracy of the results of planting structure classification by each index method combined with the improved U-Net model by calculating the kappa coefficient, Precision, Recall and the harmonic mean value F1 of the two, as shown in Table 2 below, Compared with the images without index, the classification accuracy of the fusion image with index improved significantly, and the kappa coefficient increased from 0.868 to more than 0.874 with index, which further improved the consistency between predicted classes and manual labels. The precision, recall and the harmonic mean of the two all improved by more than 0.02. In the new method, the NDVI-AWEI fusion method and the three-exponential method have the same accuracy of 0.873, and the Kappa coefficient and mean F1 are better than the three-exponential method, 0.886 and 0.872, respectively, which can provide accurate and stable crop planting structure classification results.

Table 2

Example of planting structure classification device:

The planting structure classification device proposed in this embodiment, as shown in FIG. 5 , includes a processor and a memory. The memory stores a computer program that can be run on the processor. When the processor executes the computer program, the above-mentioned planting structure classification method is implemented. example method.

That is to say, it should be understood that the methods of the above embodiments of the planting structure classification method can be implemented by computer program instructions to implement the flow of the planting structure classification method. The computer program instructions may be provided to a processor such that execution by the processor of the instructions results in the implementation of the functions specified by the above-described method flows.

The processor referred to in this embodiment refers to a processing device such as a microprocessor MCU or a programmable logic device FPGA.

The memory referred to in this embodiment includes a physical device for storing information, and usually the information is digitized and then stored in an electrical, magnetic, or optical medium. For example: all kinds of memories that use electrical energy to store information, RAM, ROM, etc.; all kinds of memories that use magnetic energy to store information, hard disks, floppy disks, magnetic tapes, magnetic core memories, magnetic bubble memories, U disks; use optical methods to store information of all kinds of memory, CD or DVD. Of course, there are other ways of memory, such as quantum memory, graphene memory, and so on.

The device constituted by the above-mentioned memory, processor and computer program is realized by the processor executing corresponding program instructions in the computer, and the processor can be equipped with various operating systems, such as windows operating system, linux system, android, and iOS system.

As other implementation manners, the device may further include a display, and the display is used for displaying the test results for the reference of the staff.

The above are only preferred embodiments of the present invention, and the present invention has been described in detail by general description and specific embodiments, but is not intended to limit the present invention. For those skilled in the art, the present invention may have Various modifications or improvements. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims (7)

1. A method of classifying a plant structure, comprising the steps of:

1) acquiring original remote sensing image data;

2) preprocessing original remote sensing image data:

extracting at least one planting structure index of NDVI, AWEI and SAVI according to the original remote sensing image data; the NDVI is a normalized difference vegetation index, the AWEI is an automatic water body extraction index, and the SAVI is a soil regulation vegetation index;

adding the extracted planting structure index into original remote sensing image data to obtain an image;

fusing images of a plurality of wave bands;

labeling the fused images to obtain a training set;

3) constructing a network model;

4) inputting the training set into the constructed network model for training to obtain a trained network model;

5) and inputting the remote sensing images to be classified into the trained network model after preprocessing to classify the planting structure.

2. The method for classifying a plant structure according to claim 1, wherein said fusing employs a G-S image fusion method.

3. The method for classifying a planting structure according to claim 1,

the normalized difference vegetation index is formulated as:

NDVI＝(NIR-RED)/(NIR+RED)

in the formula, NIR is the reflectivity of a near infrared band, and RED is the reflectivity of a RED band.

4. The method for classifying a planting structure of claim 1, wherein the formula of the automatic water body extraction index is:

AWEI＝BLUE+2.5*GREEN-1.5*(NIR+SWIR1)-0.25*SWIR2

in the formula, BLUE is a BLUE light waveband, GREEN is a GREEN waveband, and SWIR1 and SWIR2 are short-wave infrared wavebands.

5. The method for classifying a planting structure according to claim 1, wherein the soil adjusted vegetation index is formulated as:

in the formula, L is the vegetation coverage condition corresponding to the soil conditioning factor, and L is 0.5.

6. The method for classifying a plant structure according to claim 1, wherein said constructed network model is a U-Net network model comprising a depth separable convolution module and an activation function.

7. An apparatus for classifying a plant structure comprising a processor and a memory, wherein the processor executes a program of the method for classifying a plant structure according to any one of claims 1-6 stored in the memory.

Images