A Crop Recognition Method Based on UAV Remote Sensing Images

Abstract

The present application discloses a method for crop recognition based on unmanned aerial vehicle (UAV) remote sensing images, which relates to the technical field of crop recognition. The designed trained crop recognition model adopts an improved DeepLab V3+ model. The improved DeepLab V3+ model is obtained by replacing the backbone structure in the DeepLab V3+ model with a MobileNet V2 model with an attention mechanism to improve the DeepLab V3+ model. After obtaining the UAV remote sensing images taken of the crops, the trained crop recognition model is directly used to recognize the UAV remote sensing images to obtain the crop recognition results. Compared with the DeepLab V3+ model, using the improved DeepLab V3+ model to recognize the crops has higher recognition accuracy.

Description

Technical Field

[0001] This application relates to the technical field of crop recognition, and particularly to a method for crop recognition based on unmanned aerial vehicle (UAV) remote sensing images. Background Art

[0002] The spatial distribution information of crops is important basic data for agricultural production management, crop yield estimation, and adjustment of agricultural land structure. Traditional on-site investigation methods have a long cycle and high cost, resulting in high uncertainty and time lag in the obtained crop spatial distribution information. Timely and accurately grasping the crop spatial distribution information is of great significance for adjusting the crop planting structure, formulating agricultural production layout plans, promoting the efficient use of water and soil resources, and ensuring food security.

[0003] With the rapid development of UAVs, they have been widely used in fields such as agriculture, surveying engineering, urban and rural construction, and ecological environment. As an emerging technology, UAV remote sensing technology has gradually become one of the mainstream ways to obtain low-altitude remote sensing images, effectively making up for the shortcomings of aerospace remote sensing. Currently, UAV remote sensing technology is often used to obtain UAV remote sensing images of crops and identify the crops from the images. In recent years, in the research field of UAV remote sensing image classification and recognition, the application of convolutional neural network (CNN) based on deep learning has gradually increased, and the obtained research results also show good performance. The convolutional neural network uses convolutional layers and fully connected layers as key structural modules. Due to the partial connection and shared parameter structure of the convolutional layer, it can automatically extract features and allows for deeper feature extraction, having better robustness advantages. It can effectively avoid the problems of insufficient feature extraction and difficulty in reconstructing diverse data in the work of manual feature extraction. The fully connected layer, as a module for completing the classification task, realizes image classification based on the features extracted by the convolutional layer and usually has good results. Many researchers in the remote sensing field have found that while the convolutional neural network can automatically extract features, it can also maintain high image classification and recognition accuracy, and have thus carried out research on the application of convolutional neural network in remote sensing image classification. Nowadays, deep learning has received extensive attention from experts because of its low requirements for feature extraction, no need for expert participation, reduction of human intervention, and ability to extract features more comprehensively. For the recognition of crops, the recognition accuracy of the DeepLab V3+ model still needs to be improved, and it is not sufficient to accurately grasp the crop structure in a certain area, and there are still biases in local information extraction. Summary of the Invention

[0004] The purpose of this application is to provide a method for crop recognition based on UAV remote sensing images, which uses an improved DeepLab V3+ model to recognize crops with high recognition accuracy.

[0005] To achieve the above purpose, this application provides the following solutions:

[0006] This application provides a method for identifying crops based on drone remote sensing images. The method for identifying crops based on drone remote sensing images includes:

[0007] Obtain the drone remote sensing images obtained by photographing the crops;

[0008] Using the drone remote sensing images as input, identify them using a trained crop identification model. The trained crop identification model uses an improved DeepLab V3+ model. The improved DeepLab V3+ model is obtained by replacing the backbone structure in the DeepLab V3+ model with a MobileNetV2 model with an attention mechanism to improve the DeepLabV3+ model. The crop identification results include the crop types at various positions in the drone remote sensing images.

[0009] Optionally, the MobileNetV2 model with an attention mechanism includes: a MobileNet V2 model and an attention mechanism connected in sequence, and the attention mechanism is a CA attention mechanism.

[0010] Optionally, the MobileNetV2 model includes an input layer, a residual block layer, and a classification layer connected in sequence;

[0011] The input layer includes a first convolutional block;

[0012] The residual block layer includes a second convolutional block, several third convolutional blocks, and a fourth convolutional block connected in sequence;

[0013] The classification layer includes a fifth convolutional block.

[0014] Optionally, the first convolutional block includes a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence;

[0015] The second convolutional block includes a convolutional layer, a batch normalization layer, an activation function layer, a convolutional layer, and a batch normalization layer connected in sequence;

[0016] The third convolutional block includes a convolutional layer, a batch normalization layer, an activation function layer, a convolutional layer, a batch normalization layer, an activation function layer, a convolutional layer, and a batch normalization layer connected in sequence;

[0017] The fourth convolutional block includes a batch normalization layer, an activation function layer, a convolutional layer, a batch normalization layer, an activation function layer, a convolutional layer, and a batch normalization layer connected in sequence;

[0018] The fifth convolutional block includes a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence.

[0019] Optionally, before using the UAV remote sensing image as input and using the trained crop recognition model for recognition to obtain the crop recognition result, the crop recognition method based on UAV remote sensing image further includes:

[0020] Obtain a data set; the data set includes a plurality of sample UAV remote sensing images and a label image corresponding to each sample UAV remote sensing image, and the label image is obtained by annotating the crop types at various positions in the sample UAV remote sensing image;

[0021] Construct an initial crop recognition model; the initial crop recognition model uses the improved DeepLab V3+ model;

[0022] Use the data set to train the initial crop recognition model to obtain the trained crop recognition model.

[0023] Optionally, obtaining the data set specifically includes:

[0024] Obtain the original UAV remote sensing image;

[0025] Crop the original UAV remote sensing image multiple times to obtain a plurality of sample UAV remote sensing images;

[0026] For each sample UAV remote sensing image, annotate the sample UAV remote sensing image to obtain the label image corresponding to the sample UAV remote sensing image; form a data set by combining all the sample UAV remote sensing images and the label image corresponding to each sample UAV remote sensing image.

[0027] Optionally, before using the data set to train the initial crop recognition model, the crop recognition method based on UAV remote sensing image further includes: performing data augmentation on the data set to obtain an augmented data set, and using the augmented data set as a new data set;

[0028] Among them, performing data augmentation on the data set specifically includes:

[0029] Perform image augmentation on each sample UAV remote sensing image and each label image in the data set; the image augmentation includes image translation and image rotation.

[0030] Optionally, when using the data set to train the initial crop recognition model, the optimizer uses the stochastic gradient descent optimizer.

[0031] According to the specific embodiments provided in this application, the following technical effects are disclosed in this application:

[0032] The present application provides a method for crop recognition based on UAV remote sensing images. The trained crop recognition model uses the improved DeepLabV3+ model. The improved DeepLabV3+ model is obtained by replacing the backbone structure in the DeepLab V3+ model with the MobileNetV2 model with an attention mechanism to improve the DeepLab V3+ model. After obtaining the UAV remote sensing image taken of the crops, directly using the UAV remote sensing image as the input, and using the trained crop recognition model for recognition to obtain the crop recognition result. The crop recognition result includes the crop types at various positions in the UAV remote sensing image. Simulation proves that compared with the DeepLab V3+ model, using the improved DeepLab V3+ model to recognize crops has higher recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0034] Figure 1 It is an application environment diagram of a method for crop recognition based on UAV remote sensing images provided in Embodiment 1 of the present application.

[0035] Figure 2 It is a schematic flowchart of a method for crop recognition based on UAV remote sensing images provided in Embodiment 1 of the present application.

[0036] Figure 3 It is a schematic technical route diagram of a method for crop recognition based on UAV remote sensing images provided in Embodiment 1 of the present application.

[0037] Figure 4 It is a schematic flowchart of format conversion provided in Embodiment 1 of the present application.

[0038] Figure 5 It is a comparison schematic diagram of UAV remote sensing image A provided in Embodiment 1 of the present application; Figure 5 (a) is UAV remote sensing image A; Figure 5 (b) is the crop recognition result obtained by using the improved DeepLab V3+ model to recognize UAV remote sensing image A; Figure 5 (c) is the crop recognition result obtained by using the DeepLab V3+ model to recognize UAV remote sensing image A.

[0039] Figure 6Schematic comparison diagram of the UAV remote sensing image B provided in Embodiment 1 of the present application; Figure 6 (a) is the UAV remote sensing image B; Figure 6 (b) is the crop recognition result obtained by using the improved DeepLab V3+ model to recognize the UAV remote sensing image B; Figure 6 (c) is the crop recognition result obtained by using the DeepLab V3+ model to recognize the UAV remote sensing image B.

[0040] Figure 7 Schematic comparison diagram of the UAV remote sensing image C provided in Embodiment 1 of the present application; Figure 7 (a) is the UAV remote sensing image C; Figure 7 (b) is the crop recognition result obtained by using the improved DeepLab V3+ model to recognize the UAV remote sensing image C; Figure 7 (c) is the crop recognition result obtained by using the DeepLab V3+ model to recognize the UAV remote sensing image C. Detailed implementation manners

[0041] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0042] Embodiment 1

[0043] The crop recognition method based on UAV remote sensing images provided in the embodiments of the present application can be applied to the application environment as Figure 1 shown. Among them, the terminal communicates with the server through the network. The data storage system can store the data that the server needs to process. The data storage system can be set up separately, integrated on the server, or placed on the cloud or other servers. The terminal can send the UAV remote sensing image to be processed to the server. After receiving the UAV remote sensing image to be processed, for the UAV remote sensing image to be processed, the server uses the trained crop recognition model to recognize it with the UAV remote sensing image as the input, and obtains the crop recognition result. The server can feedback the obtained crop recognition result for the UAV remote sensing image to the terminal.

[0044] In addition, in some embodiments, the method for identifying crops based on UAV remote sensing images can also be implemented by a server or a terminal alone. For example, the terminal can directly identify the to-be-processed UAV remote sensing images, or the server can obtain the to-be-processed UAV remote sensing images from the data storage system and identify the to-be-processed UAV remote sensing images.

[0045] Among them, the terminal can be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0046] As Figure 2 shown, this embodiment provides a method for identifying crops based on UAV remote sensing images. This method is executed by a computer device, and specifically can be executed alone by a computer device such as a terminal or a server, or can be jointly executed by a terminal and a server. In the embodiments of the present application, taking this method applied to Figure 1 the server in as an example for illustration, the method for identifying crops based on UAV remote sensing images includes:

[0047] S1: Obtain the UAV remote sensing images obtained by photographing crops.

[0048] S2: Using the UAV remote sensing images as inputs, use the trained crop recognition model for recognition to obtain the crop recognition results; the trained crop recognition model adopts the improved DeepLab V3+ model. The improved DeepLab V3+ model is obtained by replacing the backbone structure in the DeepLabV3+ model with the MobileNetV2 model with an attention mechanism to improve the DeepLab V3+ model; the crop recognition results include the crop types at various positions in the UAV remote sensing images.

[0049] Implementing the above steps S1 to S2, this embodiment improves the existing DeepLab V3+ model. Specifically, the backbone structure of the DeepLab V3+ model is replaced, and at the same time, an attention mechanism is added to achieve the expected effect. Using the improved DeepLabV3+ model to identify UAV remote sensing images can improve the recognition accuracy, more accurately extract the complex structures of crops, provide strong support for future agricultural structure adjustment work, and at the same time provide an important basis for agricultural development.

[0050] The DeepLab V3+ model includes an encoder and a decoder.

[0051] The encoder includes a backbone structure and an Atrous Spatial Pyramid Pooling (ASPP) structure. The backbone structure can be Xception. The features output by the intermediate convolutional layer of the backbone structure are fed into the decoder, and the features output by the last convolutional layer of the backbone structure are fed into the ASPP structure. The ASPP structure includes a 1×1 convolutional block connected in parallel, three 3×3 atrous convolutional blocks with different dilation rates (the dilation rates can be 6, 12, and 18 respectively), and a global average pooling block. The five modules connected in parallel are then connected to a concat layer and a 1×1 convolutional block.

[0052] The decoder includes a 1×1 convolutional block, a concat layer, a 3×3 convolutional block, and an upsampling layer connected in sequence. Another input end of the concat layer is also connected to another upsampling layer. The features output by the intermediate convolutional layer of the backbone structure are dimension-reduced in channels through the 1×1 convolutional block, and the features output by the ASPP structure are interpolated and upsampled through the upsampling layer. The features obtained by dimension reduction in channels and the features obtained by interpolated upsampling are concatenated through the concat layer, and then convolved through the 3×3 convolutional block in sequence and interpolated and upsampled through the upsampling layer.

[0053] In this embodiment, the backbone structure (backbone) in the DeepLab V3+ model is replaced with a MobileNetV2 model with an attention mechanism to improve the DeepLab V3+ model, resulting in an improved DeepLab V3+ model. Among them, the MobileNetV2 model with an attention mechanism includes: a MobileNet V2 model and an attention mechanism connected in sequence. The attention mechanism can be a CA attention mechanism (Coordinate attention, coordinate attention mechanism).

[0054] The MobileNetV2 model includes an input layer, a residual block layer, and a classification layer connected in sequence.

[0055] The input layer includes a first convolutional block, and the first convolutional block includes a 3×3 two-dimensional convolutional layer, a two-dimensional batch normalization layer, and a ReLU6 activation function layer connected in sequence.

[0056] The residual block layer includes a second convolutional block, a plurality of third convolutional blocks, and a fourth convolutional block connected in sequence. The second convolutional block includes a 3×3 two-dimensional convolutional layer, a two-dimensional batch normalization layer, a ReLU6 activation function layer, a 1×1 two-dimensional convolutional layer, and a two-dimensional batch normalization layer connected in sequence. The third convolutional block includes a 1×1 two-dimensional convolutional layer, a two-dimensional batch normalization layer, a ReLU6 activation function layer, a 3×3 two-dimensional convolutional layer, a two-dimensional batch normalization layer, a ReLU6 activation function layer, a 1×1 two-dimensional convolutional layer, and a two-dimensional batch normalization layer connected in sequence, and the number of the third convolutional blocks can be 15. The fourth convolutional block includes a two-dimensional batch normalization layer, a ReLU6 activation function layer, a 3×3 two-dimensional convolutional layer, a two-dimensional batch normalization layer, a ReLU6 activation function layer, a 1×1 two-dimensional convolutional layer, and a two-dimensional batch normalization layer connected in sequence.

[0057] The classification layer is used to map the number of channels to a higher dimension through 1×1 convolution operation, and then adjust the feature representation through batch normalization and ReLU6 activation function to better adapt to subsequent model learning and classification tasks. The classification layer includes a fifth convolutional block, and the fifth convolutional block includes a 1×1 two-dimensional convolutional layer, a two-dimensional batch normalization layer, and a ReLU6 activation function layer connected in sequence.

[0058] In this embodiment, the MobileNetV2 model introduces a new network structure, which flips the standard residual structure. This structure uses lightweight 1×1 convolution to expand the input, then performs 3×3 depthwise separable convolution, and finally compresses it with 1×1 convolution. This design helps to reduce the computational cost while maintaining the feature expression ability. The CA attention mechanism encodes precise position information into the neural network to model channel relationships and long-term dependencies. It consists of two steps: coordinate information embedding and coordinate attention generation. In the coordinate information embedding step, the position information is embedded into the input feature map. In the coordinate attention generation step, the attention map is generated using the position information and applied to the input feature map to emphasize the representations of interest.

[0059] In this embodiment, the features output by the last two-dimensional convolutional layer of the last third convolutional block in the residual block layer are sent as the features output by the intermediate convolutional layer of the backbone structure into the decoder, and the features output by the two-dimensional convolutional layer of the fifth convolutional layer in the classification layer are sent as the features output by the last convolutional layer of the backbone structure into the dilated convolution structure. Finally, two features are obtained from the MobileNetV2 model with the attention mechanism. The shallow features (i.e., the features output by the intermediate convolutional layer of the backbone structure) directly enter the decoder, and the deep features (i.e., the features output by the last convolutional layer of the backbone structure) enter the dilated convolution structure.

[0060] Before using the trained crop recognition model to recognize using drone remote sensing images as input and obtaining the crop recognition result, as Figure 3 shown, the crop recognition method based on drone remote sensing images in this embodiment further includes:

[0061] (1) Obtain a dataset, where the dataset includes multiple sample drone remote sensing images and the label image corresponding to each sample drone remote sensing image. The label image is obtained after annotating the crop types at various positions in the sample drone remote sensing image.

[0062] Obtaining the dataset specifically includes:

[0063] (1.1) Obtain the original drone remote sensing image.

[0064] Use drone remote sensing technology to photograph the crops to obtain the original drone remote sensing image.

[0065] (1.2) Crop the original drone remote sensing image multiple times to obtain multiple sample drone remote sensing images.

[0066] For the original drone remote sensing image, in this embodiment, a cropping tool is used to crop it multiple times. The cropping size is set to 512*512, and a total of 370 pictures (i.e., sample drone remote sensing images) are cropped. The size of each sample drone remote sensing image is 512*512. In this embodiment, the PIL function is used to crop the original drone remote sensing image. First, the original drone remote sensing image needs to be referenced. Specifically, using the open(file) method of Image can return the opened original drone remote sensing image, and then using the crop((x0, y0, x1, y1)) method can crop the original drone remote sensing image. x0, y0, x1, y1 are used to define the cropping range. Among them, x0 is the x coordinate of the upper left corner position point of the original drone remote sensing image, y0 is the y coordinate of the upper left corner position point of the original drone remote sensing image, x1 is the x coordinate after translating 512 pixel points to the right from the upper left corner position point, and y1 is the y coordinate after translating 512 pixel points downward from the upper left corner position point.

[0067] (1.3) For each sample drone remote sensing image, annotate the sample drone remote sensing image to obtain the label image corresponding to the sample drone remote sensing image, and form a dataset with all the sample drone remote sensing images and the label images corresponding to each sample drone remote sensing image.

[0068] Label the cropped sample UAV remote sensing images. Specifically, the labelme software can be used for labeling. Select representative crop types according to needs, such as rice, corn, peanuts, soybeans, etc. Label the position points of the crops (the types are the above-selected representative crop types) in the sample UAV remote sensing images, and label the crop types at these position points. Label the position points without crops in the sample UAV remote sensing images as the background. After labeling the sample UAV remote sensing images, the obtained labeled file is a json file. At this time, it is necessary to further convert the labeled json file into a png image format to obtain a label image.

[0069] As Figure 4 shown, the process of converting the json file into a png image format includes: first, extract the json file information, then extract the connected components, then fill the connected components, and finally write to png until the end.

[0070] (2) Construct an initial crop recognition model, and the initial crop recognition model adopts the improved DeepLab V3+ model.

[0071] The structures of the initial crop recognition model and the trained crop recognition model are exactly the same, only the network parameters are different.

[0072] (3) Use the dataset to train the initial crop recognition model to obtain a trained crop recognition model.

[0073] Before using the dataset to train the initial crop recognition model, the crop recognition method based on UAV remote sensing images in this embodiment further includes: performing data augmentation on the dataset to obtain an augmented dataset, and using the augmented dataset as a new dataset.

[0074] Among them, performing data augmentation on the dataset specifically includes: performing image augmentation on each sample UAV remote sensing image and each label image in the dataset, and the image augmentation includes image translation and image rotation.

[0075] When performing image augmentation on the sample UAV remote sensing images and the labeled label images, the used image augmentation methods include:

[0076] (1) Image translation

[0077] Translate the images (sample UAV remote sensing images and label images), and the translation formula is as follows:

[0078]

[0079] In Equation (1), x′ and y′ are the x - coordinate and y - coordinate of pixel point A in the translated image respectively; Δx and Δy are the translation amounts in the custom x - direction and y - direction respectively; x and y are the x - coordinate and y - coordinate of pixel point A in the original image respectively.

[0080] The above Equation (1) can also be expressed as:

[0081]

[0082] (2) Image rotation

[0083] Rotate the images (sample UAV remote sensing images and label images), and the rotation formula is as follows:

[0084]

[0085] In Equation (3), x′ and y′ are the x - coordinate and y - coordinate of pixel point A in the rotated image respectively; θ is the custom rotation angle, cosθ and sinθ are the rotation amounts in the custom x - direction and y - direction respectively; x and y are the x - coordinate and y - coordinate of pixel point A in the original image respectively.

[0086] The above Equation (3) can also be expressed as:

[0087]

[0088] Load the enhanced dataset (including sample UAV remote sensing images and label images) into the improved DeepLabV3+ model and train it. During the training process, the GPU type used in this embodiment is NVIDIA GeForce RTX3090, the CUDA (a parallel computing framework for completing complex parallel computations) version is 11.3, the code adopted is based on the torch framework in Python, and the training parameters are set as follows: Select the SGD (Stochastic Gradient Descent) optimizer as the optimizer, which can adaptively adjust the learning rate during the training process. The initial learning rate is set to 0.001, the batch size (the number of samples selected for one training) is 10, the downsample factor is 8, the number of iteration cycles is 80, and the num_workers (the number of threads for loading data) is 4. Then in this embodiment, when training the initial crop recognition model using the dataset, the optimizer adopts the stochastic gradient descent optimizer.

[0089] This embodiment is directed to the field of agricultural remote sensing, and proposes a set of crop recognition methods. By adopting the technical process of this complete set of crop recognition methods, practical problems in scientific research and production are solved. The specific process includes: preprocessing the original UAV remote sensing images, specifically cropping the original UAV remote sensing images and setting the size to 512*512 to obtain sample UAV remote sensing images. Further, the labelme software is used to annotate the sample UAV remote sensing images. During the annotation process, representative crop types are selected to obtain a data set. Subsequently, data augmentation operations are performed on the data set. The data augmentation operations include basic operations such as translation and rotation to obtain an augmented data set. The augmented data set is divided into a training set and a validation set according to a ratio of 7:3 for training and validating the model to obtain a trained crop recognition model. Subsequently, the trained crop recognition model can be used to recognize UAV remote sensing images to obtain crop recognition results.

[0090] In terms of the model, this embodiment improves the DeepLab V3+ model and uses the improved DeepLab V3+ model to recognize crops. Further, a comparative analysis is performed on the recognition results of the improved DeepLabV3+ model and the original DeepLab V3+ model to evaluate the improvement in the accuracy of the improved DeepLabV3+ model. As Figures 5 - 7 shown, Figure 5 (a), Figure 6 (a) and Figure 7 (a) are all UAV remote sensing images. Figure 5 (b), Figure 6 (b) and Figure 7 (b) are all crop recognition results obtained by using the improved DeepLabV3+ model to recognize UAV remote sensing images. Figure 5 (c), Figure 6 (c) and Figure 7 (c) are all crop recognition results obtained by using the DeepLab V3+ model to recognize UAV remote sensing images. It can be seen that compared with the existing DeepLab V3+ model, the recognition accuracy of this embodiment is higher, which can complete the accurate recognition of UAV remote sensing images and has a lower occupancy of computer resources. The advantages are as follows:

[0091] (1) High recognition accuracy: All aspects of the improved DeepLab V3+ model, which combines the MobileNetV2 model and the CA attention mechanism as the backbone structure, are superior to the original DeepLab V3+ model. The simulation results show that when the backbone structure is the MobileNetV2 model with the added CA attention mechanism, compared with the original DeepLab V3+ model, the MIoU (mean intersection over union) of the improved DeepLab V3+ model increases by 1.04%, the MPA (mean pixel accuracy) increases by 2.37%, the Recall increases by 2.37%, the Precision increases by 0.05%, the F1-Score increases by 1%, the classification accuracy of corn increases by 0.09%, the classification accuracy of walnuts increases by 0.51%, the classification accuracy of soybeans increases by 1.75%, and the classification accuracy of ginger increases by 3.50%.

[0092] (2) Accurate recognition of UAV remote sensing images: The core of this embodiment is to accurately extract and recognize the planting structure of crops. Obviously, compared with the original DeepLab V3+ model, the improved DeepLab V3+ model has clearer boundaries, a wider and more accurate extraction range in crop extraction.

[0093] (3) Low computer resource occupancy: Similar to the MobileNet V1 model, the MobileNet V2 model also uses two hyperparameters, the width multiplier and the resolution multiplier, to adjust the size and computational complexity of the model, which allows users to make trade-offs according to device performance and application scenarios, and balance the accuracy and speed of the model. Compared with the MobileNet V1 model, the MobileNet V2 model further improves performance while maintaining lightweight, low latency, and low computational resource occupancy.

[0094] The embodiment of the present application also provides an application scenario that applies the above-mentioned crop recognition method based on UAV remote sensing images. Specifically, the crop recognition method based on UAV remote sensing images provided in this embodiment can be applied in a crop recognition scenario. The crop recognition scenario includes an image acquisition link and an image recognition link. The UAV remote sensing image is acquired through the image acquisition link, and the UAV remote sensing image is recognized through the image recognition link to obtain a crop recognition result. The crop recognition method based on UAV remote sensing images provided in this embodiment belongs to the image recognition link.

[0095] In this article, specific examples are used to elaborate on the principles and implementation manners of this application. The descriptions of the above embodiments are only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manners and application scopes. To sum up, the content of this specification should not be construed as a limitation to this application.

Claims

1. A method for crop recognition based on UAV remote sensing images, characterized in that, The crop recognition method based on UAV remote sensing images includes: Obtaining UAV remote sensing images obtained by photographing crops; Using the trained crop recognition model to perform recognition with the UAV remote sensing images as the input to obtain crop recognition results; the trained crop recognition model uses the improved DeepLab V3+ model, and the improved DeepLab V3+ model is obtained by replacing the backbone structure in the DeepLab V3+ model with the MobileNet V2 model with an attention mechanism to improve the DeepLab V3+ model; the crop recognition results include the crop types at various positions in the UAV remote sensing images, and the crop types include corn, walnut, soybean, and ginger; The MobileNet V2 model with an attention mechanism includes: the MobileNet V2 model and the attention mechanism connected in sequence, and the attention mechanism is the CA attention mechanism; The MobileNet V2 model includes an input layer, a residual block layer, and a classification layer connected in sequence; The input layer includes a first convolutional block; The residual block layer includes a second convolutional block, a plurality of third convolutional blocks, and a fourth convolutional block connected in sequence; The classification layer includes a fifth convolutional block; The first convolutional block includes a 3*3 convolutional layer, a batch normalization layer, and an activation function layer connected in sequence; The second convolutional block includes a 3*3 convolutional layer, a batch normalization layer, an activation function layer, a 1*1 convolutional layer, and a batch normalization layer connected in sequence; The third convolutional block includes a 1*1 convolutional layer, a batch normalization layer, an activation function layer, a 3*3 convolutional layer, a batch normalization layer, an activation function layer, a 1*1 convolutional layer, and a batch normalization layer connected in sequence; The fourth convolutional block includes a batch normalization layer, an activation function layer, a 3*3 convolutional layer, a batch normalization layer, an activation function layer, a 1*1 convolutional layer, and a batch normalization layer connected in sequence; The fifth convolutional block includes a 1*1 convolutional layer, a batch normalization layer, and an activation function layer connected in sequence.

2. The crop recognition method based on UAV remote sensing images according to claim 1, wherein, Before using the trained crop recognition model to perform recognition with the UAV remote sensing images as the input to obtain crop recognition results, the crop recognition method based on UAV remote sensing images further includes: Obtaining a data set; the data set includes a plurality of sample UAV remote sensing images and the label images corresponding to each sample UAV remote sensing image, and the label images are obtained after annotating the crop types at various positions in the sample UAV remote sensing images; Constructing an initial crop recognition model; the initial crop recognition model uses the improved DeepLab V3+ model; Training the initial crop recognition model with the data set to obtain the trained crop recognition model.

3. The crop recognition method based on drone remote sensing images according to claim 2, wherein Obtaining the data set specifically includes: Obtaining the original UAV remote sensing images; Performing multiple croppings on the original UAV remote sensing images to obtain a plurality of sample UAV remote sensing images; For each of the sample UAV remote sensing images, annotate the sample UAV remote sensing image to obtain a label image corresponding to the sample UAV remote sensing image; form a data set by combining all the sample UAV remote sensing images and the label image corresponding to each sample UAV remote sensing image.

4. The crop recognition method based on UAV remote sensing images according to claim 2, wherein Before training the initial crop recognition model using the data set, the crop recognition method based on UAV remote sensing images further includes: performing data augmentation on the data set to obtain an augmented data set, and using the augmented data set as a new data set; Among them, performing data augmentation on the data set specifically includes: Performing image augmentation on each sample UAV remote sensing image and each label image in the data set respectively; the image augmentation includes image translation and image rotation.

5. The crop recognition method based on drone remote sensing images according to claim 2, characterized in that, When training the initial crop recognition model using the data set, the optimizer uses a stochastic gradient descent optimizer.

Images