Channel cultivated land non-grain identification method and system based on machine learning, and storage medium

Abstract

The invention belongs to the technical field of satellite remote sensing area application, and particularly relates to a channel cultivated land non-grain identification method and system based on machine learning and a storage medium. The identification method comprises the following steps: S1, analyzing a multi-collinearity problem between non-grain formation and natural formation factors of the trench cultivated land; s2, establishing a channel cultivated land non-grain training data set, and completing the extraction of multi-channel layer information; and S3, constructing a DeepLabV3 + machine learning model suitable for the research area, and realizing accurate extraction of non-grain of the trench cultivated land. Wherein the DeepLabV3 + machine learning model can perform classification based on an object block in a machine learning network of a convolutional structure and output a classification probability, and a convolutional neural network model is utilized to learn internal and overall spatial information of the object block to express context spatial semantic information of a category and express a probability of the category at an object block level, so that the classification efficiency is improved. And the classification accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of satellite remote sensing regional application technology, and in particular to a method, system and storage medium for identifying non-grain cultivated land in ditches based on machine learning. Background Art

[0002] Non-agriculturalized plots of cultivated land will show unique morphological and texture features in remote sensing images. These features are in sharp contrast with the surrounding environment and can be identified from remote sensing images. The water and heat conditions in the gullies of the Loess Plateau are good, and the soil moisture content in the gullies is nearly three times that of the plateau slopes. It contains very rich reserve resources of cultivated land and potential for increasing grain production. However, with the adjustment of agricultural structure and the lack of labor in some areas, cultivated land has been abandoned, and some farmers have illegally planted trees and dug ponds on basic farmland, resulting in serious "non-grainization". The non-grainization of cultivated land affects food security, farmers' welfare and rural stability. The terrain conditions in the Loess Hilly Region are changeable, and gully cultivated land accounts for 80% of the total cultivated land. Therefore, the identification of non-grainization is crucial to the intensive and economical use of land resources, which can improve the efficiency of cultivated land resource utilization and ensure regional food security. Machine learning networks observe large amounts of data and train, use algorithms that automatically "learn" and analyze data to discover patterns, identify data trends and patterns that humans may miss, acquire the ability to analyze and solve problems, and then use the patterns to make predictions for new samples. They are mainly used in common land features such as remote sensing image classification, building extraction, and road extraction. Research on the extraction and identification of non-grain features is still in its infancy. Summary of the invention

[0003] In order to solve the technical problems mentioned above, the present application provides a method, system and storage medium for identifying non-grain cultivated land in ditches based on machine learning.

[0004] To achieve the purpose of this application, the technical solutions provided are as follows:

[0005] A method for identifying non-grain cultivated land in ditch areas based on machine learning, the method comprising the following steps:

[0006] S1: Analyze the multicollinearity problem between the non-grainization of ditch cultivated land and natural formation factors;

[0007] S2: Establish a training dataset of non-grain cultivated land in ditch areas and complete the extraction of multi-channel layer information;

[0008] S3: Construct a DeepLabV3+ machine learning model suitable for the study area to extract non-grain cultivated land in ditch areas.

[0009] Preferably, the DeepLabV3+ machine learning model is based on the DeepLabV3 model that uses pooling features with multiple expansion rates to obtain contextual information, and incorporates an Encoder Decoder structure; the Encoder Decoder structure is used to fuse multi-scale information and obtain image segmentation results, and balances the resolution accuracy and running time of the output coding features by controlling the hole convolution. The Encoder uses hole convolution to extract features, replacing the downsampling process, and the Decoder is used to restore the target boundary details.

[0010] Preferably, the DeepLabV3+ machine learning model sets the internal feature extraction model to a Resnet18 model.

[0011] Preferably, the input data of the Resnet18 model is a total of 226 ditch farmland non-grain training samples under the 04th, 17th, and 28th combination bands, and the original thumbnails of the training samples are linked to the label thumbnails.

[0012] Preferably, the step S2 uses ENVI to complete the preprocessing of the remote sensing image of the study area; the step S2 uses Arcmap to extract 267 non-grain patches of ditch cultivated land and non-ditch cultivated land as positive samples and negative samples, respectively, containing a total of 5580 pixels.

[0013] Another technical solution adopted in this application is:

[0014] A non-grain identification system for ditch cultivated land based on machine learning, the system comprises:

[0015] The collection and analysis module is used to analyze the multicollinearity problem between the non-grainization of ditch cultivated land and natural formation factors;

[0016] Remote sensing processing module, used to establish a training data set for non-grain cultivated land in ditch areas and complete the extraction of multi-channel layer information;

[0017] The information database module is used to build a DeepLabV3+ machine learning model suitable for the study area and realize the extraction of non-grain cultivated land in ditch areas.

[0018] Another technical solution adopted in this application is:

[0019] A computer-readable storage medium includes a computer program. When the computer program is run on a computer, any of the above-mentioned identification methods is executed.

[0020] The beneficial effects of this application are:

[0021] (1) The DeepLabV3+ machine learning model can classify object blocks based on the convolutional machine learning network and output the classification probability. It uses the convolutional neural network model to learn the internal and overall spatial information of the object block to represent the contextual spatial semantic information of the category.

[0022] (2) The DeepLabV3+ machine learning model represents the probability of categories at the object block level, which can avoid pixel-level mismatches and improve classification accuracy.

[0023] (3) The DeepLabV3+ machine learning model can directly read remote sensing images without destroying the data structure, and can obtain richer multi-source data sets, making the trained model robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 This is a flowchart of the non-grain identification method of ditch farmland based on machine learning in this application.

[0026] Figure 2 This is the architecture diagram of the DeepLabV3+ machine learning model in this application.

[0027] Figure 3 This is a flowchart of the preprocessing of remote sensing images in this application.

[0028] Figure 4 This is an example map of the non-grain extraction of ditch cultivated land in the study area for this application. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0030] refer to Figure 1 , a method for identifying non-grain cultivated land in ditch based on machine learning, comprising the following steps:

[0031] S1: Analyze the multicollinearity problem between the non-grainization of ditch cultivated land and natural formation factors;

[0032] S2: Establish a training dataset of non-grain cultivated land in ditch areas and extract multi-channel layer information;

[0033] S3: Construct a DeepLabV3+ machine learning model suitable for the study area to extract non-grain cultivated land in ditch areas.

[0034] Specifically, the DeepLabV3+ machine learning model is a semantic segmentation method that combines deep convolutional neural networks and probabilistic graph models.

[0035] like Figure 2 As shown in the figure, the deep convolutional neural network mainly includes convolutional layers, activation functions, pooling layers and fully connected layers.

[0036] The pooling layer is the process of reducing statistical information and extracting feature information. The pooling operation can reduce the size of the feature map, effectively reduce the amount of calculation and suppress the occurrence of overfitting. In addition, since the maximum pooling operation can always select the maximum value element in a certain area, it can eliminate the error caused by some small displacements.

[0037] The fully connected layer integrates the local information with categories extracted by the convolutional layer and the pooling layer.

[0038] Furthermore, the DeepLabV3+ machine learning model incorporates the Encoder Decoder structure based on the DeepLabV3 model's use of pooling features with multiple expansion rates to obtain rich contextual information.

[0039] The Encoder Decoder structure is used to fuse multi-scale information and obtain image segmentation results. The resolution accuracy and running time of the output coding features are balanced by controlling the dilated convolution. The Encoder uses dilated convolution to extract features, replacing the downsampling process. The Decoder is used to restore the target boundary details.

[0040] Furthermore, the Resnet18 model is used to extract image features in the DCNN (deep convolutional neural network) semantic segmentation task in the DeepLabV3+ machine learning model. The Resnet18 model contains 17 convolutional layers with weights and 1 fully connected layer, and the remaining pooling layers do not contain weights. Its learning results are greatly affected by network weights and data fluctuations.

[0041] Exemplarily, the DeepLabV3+ machine learning model adopts two model training strategies, one is a fixed learning rate throughout the training process, and the other is that after a certain number of training rounds, the learning rate will gradually decrease according to a fixed multiple until the end of training. When the initial learning rate is 0.001 and the reduction multiple is 0.63 times per round, the learning rate is close to 0.0001 and 0.00001 learning rates at the 25th and 50th rounds of training, respectively. Under the control of the remaining training parameters, according to the learning rate can be gradually reduced to the original 1 / 10 experience every certain number of rounds, the initial learning rate is 0.001, the variable learning rate is reduced to 0.63 times every five rounds, and the DeepLabV3+ machine learning model with three fixed learning rates of 0.001, 0.0001 and 0.00001 is trained respectively.

[0042] The original images and label images in the training set were used to test the overall classification accuracy of the DeepLabV3+ machine learning model under the four learning rates, and the results were compared and analyzed: the 0.0001 learning rate model had the highest accuracy, followed by the 0.001 learning rate and variable learning rate models, and the worst was the 0.00001 learning rate model. The recall rate, precision, and IOU (overlap) values ​​of the four models differed by 27.27%, 8.32%, and 0.1712. After comprehensive model accuracy evaluation, the learning rate parameter of the DeepLadV3+ machine learning model in this embodiment is set to 0.0001.

[0043] Specifically, the detection methods used in step S1 to analyze the multicollinearity problem include Pearson correlation coefficient method, variance inflation factor and information growth rate.

[0044] Furthermore, step S1 retains all influencing factors that do not have multicollinearity problems in the training samples based on the calculation results of Pearson correlation analysis, variance inflation factor and information growth rate.

[0045] refer to Figure 3 In step S2, ENVI is used to pre-process the remote sensing images of the study area. The modules mainly used by ENVI include radiation calibration, atmospheric correction and orthorectification modules.

[0046] The radiation calibration module is mainly used to convert the digital quantization value of remote sensing images into physically meaningful atmospheric apparent radiation brightness values.

[0047] The atmospheric correction module is mainly used to eliminate the radiation error caused by atmospheric images and invert the true reflectivity of various ground objects.

[0048] The orthorectification module is mainly used to correct image tilt errors.

[0049] Furthermore, in step S2, Arcmap is used to extract 267 non-grain patches of ditch cultivated land and non-ditch cultivated land as positive samples and negative samples, which contain 5580 pixels in total. Among the 5580 pixels, 3906 pixels are selected as training sets and 1674 pixels are selected as test sets in a ratio of 7:3.

[0050] The present application also provides a method system for identifying non-grain-producing ditch cultivated land based on machine learning, including:

[0051] The collection and analysis module is used to analyze the multicollinearity problem between the non-grainization of ditch cultivated land and natural formation factors;

[0052] Remote sensing processing module, used to establish a training data set for non-grain cultivated land in ditch areas and complete the extraction of multi-channel layer information;

[0053] The information database module is used to build a DeepLabV3+ machine learning model suitable for the study area to achieve accurate extraction of non-grain cultivated land in ditch areas.

[0054] A computer-readable storage medium includes a computer program. When the computer program is run on a computer, any of the above-mentioned identification methods is executed.

[0055] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0056] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for identifying non-grain cultivated land in ditch areas based on machine learning, characterized in that: The identification method comprises the following steps: S1: Analyze the multicollinearity problem between the non-grainization of ditch cultivated land and natural formation factors; S2: Establish a training dataset of non-grain cultivated land in ditch areas and extract multi-channel layer information; S3: Construct a DeepLabV3+ machine learning model suitable for the study area to extract non-grain cultivated land in ditch areas.

2. The method for identifying non-grain cultivated land in ditch areas based on machine learning according to claim 1 is characterized in that: The DeepLabV3+ machine learning model is based on the DeepLabV3 model that uses pooling features with multiple expansion rates to obtain context information and incorporates the Encoder Decoder structure; The Encoder-Decoder structure is used to fuse multi-scale information and obtain image segmentation results. The resolution accuracy and running time of the output coding features are balanced by controlling the dilated convolution. The Encoder uses dilated convolution to extract features instead of the downsampling process, and the Decoder is used to restore the target boundary details.

3. The method for identifying non-grain cultivated land in ditch areas based on machine learning according to claim 1, characterized in that: The DeepLabV3+ machine learning model sets the internal feature extraction model to the Resnet18 model.

4. The method for identifying non-grain cultivated land in ditch areas based on machine learning according to claim 3 is characterized in that: The input data of the Resnet18 model are 226 non-grain training samples of ditch cultivated land in the 04th, 17th and 28th combined bands, and the original thumbnails of the training samples are linked with the label thumbnails.

5. The method for identifying non-grain cultivated land in ditch areas based on machine learning according to claim 1, characterized in that: The step S2 also includes completing the preprocessing of the remote sensing image of the study area using ENVI; In step S2, Arcmap is used to extract 267 non-grain grain patches of ditch cultivated land and non-ditch cultivated land as positive samples and negative samples respectively.

6. A non-grain identification system for ditch cultivated land based on machine learning, characterized by: The system includes: The collection and analysis module is used to analyze the multicollinearity problem between the non-grainization of ditch cultivated land and natural formation factors; Remote sensing processing module, used to establish a training data set for non-grain cultivated land in ditch areas and complete the extraction of multi-channel layer information; The information database module is used to build a DeepLabV3+ machine learning model suitable for the study area and realize the extraction of non-grain cultivated land in the ditch.

7. A computer-readable storage medium, characterized in that: The invention comprises a computer program, which enables the identification method according to any one of claims 1 to 5 to be executed when the computer program is run on a computer.

Images