Soil type and soil diagnosis layer identification method based on deep learning

Abstract

The invention discloses a soil type and soil diagnosis layer identification method based on deep learning. The method comprises the following steps: obtaining a soil profile image data set; establishing a soil outline data set, a subline data set and a soil level data set based on the soil profile image data set; establishing a soil outline classification model and a subclass classification model based on an image classification algorithm, establishing a hierarchical segmentation model based on a semantic segmentation algorithm, and training by adopting a soil outline data set, a subclass data set and a soil hierarchical data set respectively; and carrying out actual soil type and soil diagnosis layer identification after training. According to the invention, soil type and soil diagnosis layer identification is carried out on the soil profile image through the image classification algorithm and the semantic segmentation algorithm, and the efficiency and convenience of soil type and soil diagnosis layer identification are improved.

Description

Technical Field

[0001] The present application relates to the technical field of soil image recognition, and in particular to a method for identifying soil types and soil diagnostic layers based on deep learning. Background Art

[0002] An important direction of soil research is soil classification. As scientific plans for responding to food security such as precision agriculture, arable land quality evaluation, and rural land planning are adopted by the Ministry of Agriculture and Rural Affairs, the requirements for the precision of digital soil mapping and the accuracy of soil classification are becoming increasingly higher. Therefore, it is necessary to study soil types and soil diagnostic layer identification methods.

[0003] Traditional soil classification methods are mainly based on the theoretical knowledge system of soil genesis. They are qualitative classification methods determined by experts. The classification results are prone to certain differences due to different subjective experiences and lack the support of quantitative indicator systems. In view of the shortcomings of traditional soil genesis classification methods, the quantitative soil system classification that has gradually developed in recent years can more accurately express the differences in soil properties and is the basis for the quantitative expression and application of soil information. Soil system classification uses quantitative occurrence environment and soil characteristics to reflect soil formation processes of different intensities and then distinguish soils. Because of the advantage of quantitative indicators, soil system classification is being praised and studied by more and more scholars. In the specific identification process of soil system classification, the diagnostic layer, as the basic element of soil system classification, is a specific soil layer with quantitative properties, which can determine the final determination of soil type. In addition, the accurate division and discrimination of soil layers are also important for the study and evaluation of soil carbon storage, bulk density, water movement and soil microbial distribution. At the same time, accurate division of soil layers is also an important part of digital soil mapping.

[0004] However, in the actual process of identifying soil types and diagnostic layers, a lot of manpower and material resources are needed to collect samples, measure soil physical and chemical data, and then judge the soil diagnostic layer and diagnostic characteristics. Only after comprehensively considering environmental covariates can the final result be obtained. The efficiency and convenience of identifying soil types and soil diagnostic layers are poor. Summary of the invention

[0005] The present application provides a soil type and soil diagnostic layer identification method based on deep learning to solve the problem of poor efficiency and convenience in soil type and soil diagnostic layer identification in the prior art.

[0006] On the one hand, the present application provides a method for identifying soil types and soil diagnostic layers based on deep learning, comprising the following steps:

[0007] Step 1: Obtain soil profile image dataset.

[0008] Step 2: Copy the soil profile image dataset into three copies, and perform soil order labeling, sub-class labeling and level labeling on the three soil profile image datasets respectively to obtain a soil order dataset, a sub-class dataset and a soil level dataset.

[0009] Step three, establish soil order classification model and subclass classification model based on image classification algorithm, and establish hierarchical segmentation model based on semantic segmentation algorithm.

[0010] Step 4: Use the soil order dataset, the sub-class dataset and the soil level dataset to train the soil order classification model, the sub-class classification model and the level segmentation model respectively to obtain a soil order soil category recognition model, a sub-class soil category recognition model and a soil diagnostic layer recognition model.

[0011] Step five: using the soil order soil category recognition model, the suborder soil category recognition model and the soil diagnostic layer recognition model to identify the actual soil type and soil diagnostic layer.

[0012] In a possible implementation, in step 1, after acquiring the soil profile image dataset, the image data in the soil profile image dataset is preprocessed.

[0013] The preprocessing includes: removing isolated samples, cropping, compression and data enhancement.

[0014] In a possible implementation, in step three, the image classification algorithm includes: a ResNet network, an AlexNet network, a VGGNet network, and a DenseNet network.

[0015] In a possible implementation, in step three, the image classification algorithm uses a ResNet-18 network in a ResNet network, including a starting convolutional layer, eight residual blocks and a final fully connected layer, each residual block contains two convolutional layers, for a total of 18 layers of the network.

[0016] In a possible implementation, in step three, the semantic segmentation algorithm includes: UNet++ network, UNet network, SegNet network and DeepLab network.

[0017] In a possible implementation, in step three, the semantic segmentation algorithm uses a UNet++ network, including an encoder path, a decoder path, a skip path, and a convolution operation.

[0018] The encoder path is used for downsampling, the decoder path is used for upsampling, the skip path is used to connect the encoder and the decoder, and the convolution operation is used to extract image features and generate a feature map.

[0019] In one possible implementation, in step four, accuracy and recall are used as performance evaluation indicators of the soil order soil category recognition model and the suborder soil category recognition model, and average intersection-over-union ratio and pixel accuracy are used as performance evaluation indicators of the soil diagnostic layer recognition model.

[0020] In a possible implementation, step five includes:

[0021] Collect the soil profile image to be identified.

[0022] The soil profile image to be identified is respectively input into the soil order soil category identification model, the sub-order soil category identification model and the soil diagnostic layer identification model to obtain the soil order identification result, the sub-order identification result and the diagnostic layer identification result.

[0023] The soil type and soil diagnostic layer identification method based on deep learning in this application has the following advantages:

[0024] Soil types and soil diagnostic layers of soil profile images are identified through image classification algorithm and semantic segmentation algorithm, which improves the efficiency and convenience of soil type and soil diagnostic layer identification.

[0025] By establishing soil order dataset, subclass dataset and soil level dataset based on the soil profile image dataset, and using the soil order dataset, subclass dataset and soil level dataset to train the soil order classification model and subclass classification model established based on the image classification algorithm, and the hierarchical segmentation model established based on the semantic segmentation algorithm, the correlation between the various recognition models and the recognition accuracy of the same soil profile image are improved.

[0026] After obtaining the soil profile image dataset, the image data in the soil profile image dataset is preprocessed, including: eliminating isolated samples, cropping, compression and data enhancement, which improves the reliability and diversity of the dataset, thereby improving the generalization ability and accuracy of the recognition model.

[0027] The proposed image classification algorithm adopts the ResNet-18 network in the ResNet network and the specific network structure setting to improve the recognition accuracy of the soil order soil category recognition model and the soil subclass soil category recognition model.

[0028] The proposed semantic segmentation algorithm adopts the UNet++ network and the specific network structure setting to improve the recognition accuracy of the soil diagnostic layer recognition model.

[0029] The proposed accuracy and recall rate are used as performance evaluation indicators for soil order and subclass soil category recognition models, and the average intersection-over-union ratio and pixel accuracy are used as performance evaluation indicators for soil diagnostic layer recognition models, which improves the reliability and applicability of the models. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] 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.

[0031] Figure 1 A schematic diagram of a process for identifying soil types and soil diagnostic layers based on deep learning provided in an embodiment of the present application;

[0032] Figure 2 A schematic diagram of the structure of the ResNet-18 network provided in an embodiment of the present application;

[0033] Figure 3 Provided in the embodiments of this application Figure 2 Schematic diagram of the structure of the residual block in ;

[0034] Figure 4 A schematic diagram of the structure of the UNet++ network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0035] 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.

[0036] like Figure 1 As shown, the embodiment of the present application provides a method for identifying soil types and soil diagnostic layers based on deep learning, comprising the following steps:

[0037] Step 1: Obtain soil profile image dataset.

[0038] Step 2: Copy the soil profile image dataset into three copies, and perform soil order labeling, sub-class labeling and level labeling on the three soil profile image datasets respectively to obtain a soil order dataset, a sub-class dataset and a soil level dataset.

[0039] Step three, establish soil order classification model and subclass classification model based on image classification algorithm, and establish hierarchical segmentation model based on semantic segmentation algorithm.

[0040] Step 4: Use the soil order dataset, the sub-class dataset and the soil level dataset to train the soil order classification model, the sub-class classification model and the level segmentation model respectively to obtain a soil order soil category recognition model, a sub-class soil category recognition model and a soil diagnostic layer recognition model.

[0041] Step five: using the soil order soil category recognition model, the suborder soil category recognition model and the soil diagnostic layer recognition model to identify the actual soil type and soil diagnostic layer.

[0042] Specifically, in this embodiment, in step 1, the image data in the soil profile image data set is an RGB soil profile image collected by a SLR camera.

[0043] Exemplarily, in step one, after acquiring the soil profile image dataset, the image data in the soil profile image dataset is also preprocessed.

[0044] The preprocessing includes: removing isolated samples, cropping, compression and data enhancement.

[0045] Specifically, in this embodiment, the removal of isolated samples is to remove image data below a preset clarity threshold in the soil profile image data set to avoid interference of blurred images on the training process.

[0046] In this embodiment, cropping is to perform random region cropping on the image data in the soil profile image data set to improve the robustness of the model to changes in the image region.

[0047] In this embodiment, the image data in the soil profile image data set is compressed using a lossless compression method, thereby reducing the size of the image data without losing information, thereby reducing the demand for storage and computing resources.

[0048] In this embodiment, data enhancement includes flipping, rotating, scaling, color conversion, and adding noise to the image data in the soil profile image dataset, thereby enhancing the diversity of the image data in the soil profile image dataset and thereby improving the generalization ability of the subsequent recognition model.

[0049] Specifically, in this embodiment, in step 2, soil order labeling includes labeling an unlabeled soil profile image dataset with the soil type of the soil order level obtained in advance by the soil system classification method to obtain a soil order dataset. Specifically, labeling the image data in the soil profile image dataset with the type of the soil order level.

[0050] Subclass annotation includes: annotating an unannotated soil profile image dataset with the soil type at the subclass level obtained in advance by the soil system classification method to obtain a subclass dataset. The soil type at the subclass level is a refinement of the soil type at the soil class level. Specifically, the image data in the soil profile image dataset is annotated with the type at the subclass level.

[0051] The hierarchical annotation includes: annotating an unlabeled soil profile image dataset with the soil diagnostic layer type obtained in advance by the soil system classification method to obtain a soil hierarchical dataset. Specifically, the image data in the soil profile image dataset is segmented and annotated with the soil diagnostic layer region.

[0052] Exemplarily, in step three, the image classification algorithm includes: ResNet network, AlexNet network, VGGNet network and DenseNet network.

[0053] Exemplarily, in step three, the image classification algorithm adopts the ResNet-18 network in the ResNet network, including a starting convolutional layer, eight residual blocks and a final fully connected layer, each residual block contains two convolutional layers, a total of 18 layers of network.

[0054] Specifically, in this embodiment, the soil order classification model and the subclass classification model both use the ResNet-18 network with the same network architecture.

[0055] like Figure 2 As shown, in this embodiment, the convolution at the beginning of the ResNet-18 network is a 7×7 convolution layer, and the pooling after the convolution is a maximum pooling layer, which is to quickly reduce the amount of input data and adjust it to the size and number of channels required by the subsequent residual block. The convolution kernel weights and averages the pixels in a small area of ​​the image to become each corresponding pixel in the output image. The pooling at the end of the ResNet-18 network is an average pooling layer, and the full connection is a fully connected layer using the activation function Softmax, which is to convert the final output into a column vector and reduce it to the specified number of soil classes or subclasses. Figure 2 Chestnut soil, black loam soil, brown soil, yellow-brown soil and brown soil are examples of soil orders.

[0056] like Figure 3As shown, each residual block of the ResNet-18 network has two branch paths, one is the identity mapping (identity) of x→x, and the other is the nonlinear mapping of F(x), namely the residual mapping. The residual mapping in the residual block is composed of a convolution layer (Conv), a batch normalization layer (BN) and a nonlinear activation function (ReLU). In this embodiment, specifically: x passes through a 3×3 convolution (Conv), and is processed by a batch normalization layer (BN) and a nonlinear activation function (ReLU), and then passes through a 3×3 convolution (Conv) and a batch normalization layer (BN) to obtain F(x). The batch normalization layer (BN) adjusts the output of the convolution (Conv) to satisfy the normal distribution, so that the activation input values ​​are distributed in the area where the nonlinear activation function (ReLU) is more sensitive. The nonlinear activation function (ReLU) is to add nonlinear factors to deal with the problem of linear inseparability. Figure 3 The “+” in means that F(x) is added to x by matrix values, and then output through a nonlinear activation function (ReLU).

[0057] Exemplarily, in step three, the semantic segmentation algorithm includes: UNet++ network, UNet network, SegNet network and DeepLab network.

[0058] Exemplarily, in step three, the semantic segmentation algorithm adopts a UNet++ network, including an encoder path, a decoder path, a skip path and a convolution operation.

[0059] The encoder path is used for downsampling, the decoder path is used for upsampling, the skip path is used to connect the encoder and the decoder, and the convolution operation is used to extract image features and generate a feature map.

[0060] like Figure 4 As shown, specifically, in this embodiment, the left side of the UNet++ network is the encoder path (a), and the right side is the decoder path (b). The original image is downsampled four times in the encoder path, and convolution operations are performed on each layer, with a convolution kernel size of 3×3×3. The resolution of each layer of feature maps gradually becomes lower, the image gradually becomes blurred and abstract, the number of channels gradually increases, and detail information such as spatial positioning features is gradually lost, but more image features are extracted. The decoder path uses the feature map to perform corresponding four-layer upsampling, restore the compressed image until it is restored, and output the image of the segmented soil diagnosis layer. The jump path merges the channels of the feature maps on the symmetrical encoder path and decoder path, and directly passes the features extracted during downsampling to the upsampling, integrating more shallow feature information. The feature maps on the encoder and decoder paths are further connected through more densely nested jump paths, that is, Figure 4The green part (c) in . Multiple cascaded jump paths, upsampling and restoring the generated original image (shallow output, middle output, deep output) from multiple decoder branches, that is Figure 4 Finally, all output images from multiple decoder branches are averaged to obtain the diagnosis layer hierarchical results.

[0061] Exemplarily, in step four, accuracy and recall are used as performance evaluation indicators of the soil order soil category recognition model and the suborder soil category recognition model, and average intersection-over-union ratio and pixel accuracy are used as performance evaluation indicators of the soil diagnostic layer recognition model.

[0062] Specifically, in this embodiment, in step four, the soil order data set is randomly divided into a soil order training set and a soil order test set in a ratio of 7:3, the soil order training set is used to train the soil order classification model to obtain the soil order soil category recognition model, and the soil order test set is used to perform performance testing on the soil order soil category recognition model.

[0063] In this embodiment, the subclass data set is randomly divided into a subclass training set and a subclass test set in a ratio of 7:3. The subclass training set is used to train the subclass classification model to obtain the subclass soil category recognition model. The subclass test set is used to perform performance testing on the subclass soil category recognition model.

[0064] In this embodiment, the soil layer data set is randomly divided into a soil layer training set and a soil layer test set in a ratio of 7:3. The soil layer training set is used to train the hierarchical segmentation model to obtain a soil diagnostic layer recognition model. The soil layer test set is used to perform performance testing on the soil diagnostic layer recognition model.

[0065] Specifically, in this embodiment, accuracy and recall are used as performance evaluation indicators of the soil order and sub-order soil category recognition models.

[0066] Accuracy ACC is an evaluation index in image classification. It refers to the ratio of the total number of correct samples to the total number of samples when the trained model is applied to classify the image data set. It is an evaluation index for judging the accuracy of the model as a whole. The calculation method is shown in the following formula:

[0067]

[0068] Among them, TP (True Positive) refers to the number of positive examples predicted by the model but actually are positive examples, so the prediction is correct. FN (False Negative) refers to the number of negative examples predicted by the model but actually are positive examples, so the prediction is wrong. Based on this rule, it can be deduced that TN represents the number of negative examples predicted and correctly judged as negative examples, and FP represents the number of positive examples predicted but wrongly judged as negative examples.

[0069] Recall ratio R is an evaluation index in image classification. It refers to the ratio of the number of correctly predicted samples in a class to the total number of samples in this class. The calculation method is shown in the following formula:

[0070]

[0071] Specifically, in this embodiment, average intersection-over-union ratio and pixel accuracy are used as performance evaluation indicators of the soil diagnostic layer recognition model.

[0072] In semantic segmentation, in order to evaluate the model, the manually circled layers are considered to be true values ​​to verify the performance of the model. Intersection over Union (IoU) and Mean Intersection over Union (MIoU) are evaluation indicators in semantic segmentation. The IoU is the ratio between the intersection of the true value and the predicted value of the pixel in the image and the union of the two. The calculation method is as follows:

[0073]

[0074] Among them, P represents the pixel prediction value and G represents the pixel true value.

[0075] The mean intersection over union (MIoU) is the mean of the IoUs in a category, and is calculated as follows:

[0076]

[0077] Where i represents the category index, that is, the i-th category, and n represents the total number of categories.

[0078] Pixel Accuracy PA (Pixel Accuracy) is an evaluation indicator in semantic segmentation. It is the ratio between correctly segmented pixels and total image pixels, reflecting the accuracy of the model. The calculation method is as follows:

[0079]

[0080] Among them, TP` represents the number of pixels that the model segmented as positive examples are actually positive examples; FN` represents the number of pixels that the model segmented as negative examples but are actually positive examples; TN` represents the number of pixels that the model segmented as negative examples but are actually negative examples; FP` represents the number of pixels that the model segmented as positive examples but are actually negative examples.

[0081] Exemplarily, step five includes:

[0082] Collect the soil profile image to be identified.

[0083] The soil profile image to be identified is respectively input into the soil order soil category identification model, the sub-order soil category identification model and the soil diagnostic layer identification model to obtain the soil order identification result, the sub-order identification result and the diagnostic layer identification result.

[0084] The embodiment of the present application uses an image classification algorithm and a semantic segmentation algorithm to identify soil types and soil diagnostic layers of a soil profile image, thereby improving the efficiency and convenience of identifying soil types and soil diagnostic layers.

[0085] By establishing soil order dataset, subclass dataset and soil level dataset based on the soil profile image dataset, and using the soil order dataset, subclass dataset and soil level dataset to train the soil order classification model and subclass classification model established based on the image classification algorithm, and the hierarchical segmentation model established based on the semantic segmentation algorithm, the correlation between the various recognition models and the recognition accuracy of the same soil profile image are improved.

[0086] After obtaining the soil profile image dataset, the image data in the soil profile image dataset is preprocessed, including: eliminating isolated samples, cropping, compression and data enhancement, which improves the reliability and diversity of the dataset, thereby improving the generalization ability and accuracy of the recognition model.

[0087] The proposed image classification algorithm adopts the ResNet-18 network in the ResNet network and the specific network structure setting to improve the recognition accuracy of the soil order soil category recognition model and the soil subclass soil category recognition model.

[0088] The proposed semantic segmentation algorithm adopts the UNet++ network and the specific network structure setting to improve the recognition accuracy of the soil diagnostic layer recognition model.

[0089] The proposed accuracy and recall rate are used as performance evaluation indicators for soil order and subclass soil category recognition models, and the average intersection-over-union ratio and pixel accuracy are used as performance evaluation indicators for soil diagnostic layer recognition models, which improves the reliability and applicability of the models.

[0090] Although the preferred embodiments of the present application have been described, those skilled in the art may make other 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.

[0091] 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 soil type and soil diagnostic layer identification method based on deep learning, characterized in that: The following steps are involved: Step 1: Obtain soil profile image dataset; Step 2, copying the soil profile image dataset into three copies, and respectively performing soil order labeling, sub-class labeling and level labeling on the three soil profile image datasets to obtain a soil order dataset, a sub-class dataset and a soil level dataset; Step 3: Establish soil order classification model and sub-order classification model based on image classification algorithm, and establish hierarchical segmentation model based on semantic segmentation algorithm; Step 4, respectively using the soil order dataset, the sub-order dataset and the soil level dataset to train the soil order classification model, the sub-order classification model and the level segmentation model to obtain a soil order soil category recognition model, a sub-order soil category recognition model and a soil diagnostic layer recognition model; Step five: using the soil order soil category recognition model, the suborder soil category recognition model and the soil diagnostic layer recognition model to identify the actual soil type and soil diagnostic layer.

2. The method for identifying soil types and soil diagnostic layers based on deep learning according to claim 1, characterized in that: In step 1, after acquiring the soil profile image dataset, the image data in the soil profile image dataset is also preprocessed; The preprocessing includes: removing isolated samples, cropping, compression and data enhancement.

3. The method for identifying soil types and soil diagnostic layers based on deep learning according to claim 1, characterized in that: In step three, the image classification algorithms include: ResNet network, AlexNet network, VGGNet network and DenseNet network.

4. The method for identifying soil types and soil diagnostic layers based on deep learning according to claim 3, characterized in that: In step three, the image classification algorithm uses the ResNet-18 network in the ResNet network, including a starting convolutional layer, eight residual blocks and a final fully connected layer. Each residual block contains two convolutional layers, for a total of 18 layers of the network.

5. The method for identifying soil types and soil diagnostic layers based on deep learning according to claim 1, characterized in that: In step three, the semantic segmentation algorithms include: UNet++ network, UNet network, SegNet network and DeepLab network.

6. The method for identifying soil types and soil diagnostic layers based on deep learning according to claim 5, characterized in that: In step 3, the semantic segmentation algorithm adopts the UNet++ network, including an encoder path, a decoder path, a skip path and a convolution operation; The encoder path is used for downsampling, the decoder path is used for upsampling, the skip path is used to connect the encoder and the decoder, and the convolution operation is used to extract image features and generate a feature map.

7. The method for identifying soil types and soil diagnostic layers based on deep learning according to claim 1, characterized in that: In step 4, accuracy and recall are used as performance evaluation indicators of the soil order soil category recognition model and the suborder soil category recognition model, and average intersection-over-union ratio and pixel accuracy are used as performance evaluation indicators of the soil diagnostic layer recognition model.

8. The method for identifying soil types and soil diagnostic layers based on deep learning according to claim 1, characterized in that: Step five includes: Collecting soil profile images to be identified; The soil profile image to be identified is respectively input into the soil order soil category identification model, the sub-order soil category identification model and the soil diagnostic layer identification model to obtain the soil order identification result, the sub-order identification result and the diagnostic layer identification result.

Images