Field tobacco plant growth state classification method based on deep learning

Abstract

The invention provides a field tobacco plant growth state classification method based on deep learning, and the method comprises the steps: collecting tobacco information image data in real time through employing a tobacco field camera device, building a set of tobacco field data set under a complex background, and discriminating the growth and development stage and nutritional condition of tobacco through employing a deep learning algorithm. Due to the fact that environmental factors such as soil, fertilizing amount and weather in the same land parcel are the same, the tobacco plants basically belong to the same growth and development stage and nutritional status, the growth and development stage and nutritional status of the tobacco plants in the tobacco field can be known in real time in the tobacco planting stage, and the quality and yield of the tobacco plants are increased by adopting effective measures such as adjusting the fertilizing amount. The method solves the problems of low accuracy, high human subjectivity, long time consumption, loss of manpower and material resources and the like in the current tobacco field for judging the state of the stage to which the tobacco plants belong, and can accurately judge the growth state of the tobacco plants in the field. A set of private data set is established for the tobacco field, and data support is provided for later establishment of an intelligent agriculture algorithm model and scientific planting in the field.

Description

Technical Field

[0001] The present invention relates to the field of intelligent identification of tobacco fields, and provides a method for capturing and collecting images using a tobacco field camera device and a method for distinguishing and classifying tobacco leaf growth states based on deep learning. Background Art

[0002] Tobacco is an important cash crop in my country. my country's tobacco production accounts for an average of about 35% of the world's total. The tobacco industry has become an important pillar of my country's social economy. According to statistics, in 2023, the tobacco industry achieved a total industrial and commercial tax and profit of 1,521.7 billion yuan, a year-on-year increase of 5.6%, and a total fiscal revenue of 1,502.8 billion yuan, a year-on-year increase of 4.3%, achieving a record high in total tax and profit and total fiscal revenue, making positive contributions to stabilizing the economy, ensuring the increase of national and local fiscal revenue, and promoting economic and social development. At the same time, tobacco planting has brought huge economic profits to tobacco farmers and effectively changed the economic outlook of the planting areas.

[0003] Tobacco leaves are the main raw material for Chinese cigarette production. Scientific and reasonable classification of tobacco leaves according to their quality is of great significance for the procurement of raw tobacco leaves and the design of cigarette formulas. During the tobacco leaf growth period, observing the growth and development stage and nutritional status of tobacco plants can better reflect and predict the quality of tobacco leaves. The results of observation are conducive to timely setting up plans to adjust the amount of fertilizer and watering in tobacco fields, strictly controlling the various nutritional conditions required for each stage of tobacco leaf growth, and better improving the yield and quality of tobacco fields. At present, the growth and development stage and nutritional status of tobacco leaves mainly rely on manual experience, and manual grading is easily affected by subjective experience and environmental conditions, and there are problems such as low grading accuracy and poor consistency.

[0004] Based on this, the present invention provides a classification method for discriminating the growth status of tobacco plants based on deep learning. Use tobacco field camera equipment to collect tobacco leaf information image data in real time, establish a tobacco field data set under a complex background, and use a deep learning algorithm to discriminate the growth and development stage and nutritional status of tobacco leaves. Since the soil, fertilizer application, weather and other environmental factors are the same in the same plot, the tobacco plants basically belong to the same growth and development stage and nutritional status. This method can accurately discriminate the overall growth and development stage and nutritional status of the tobacco field. Summary of the invention

[0005] The purpose of the present invention is to address the problems of low accuracy, strong subjectivity, long time consumption and large loss of human resources in the current manual identification of the growth and development stages and nutritional status of tobacco plants, and propose a field tobacco plant growth status classification method based on deep learning. The method establishes a tobacco field data set of a complete tobacco plant development cycle under a complex background by using tobacco field camera equipment to capture images, and proposes an improved neural network that can accurately identify the growth and development stages and nutritional status of tobacco plants in the field.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for classifying the growth status of field tobacco plants based on deep learning, comprising the following steps:

[0008] Step 1: Establish a tobacco plant growth status dataset, create corresponding labels, and pre-process the collected dataset to obtain image data;

[0009] Step 2: Establish a classification model and train the image data generated in step 1. The classification model is based on multi-layer convolution in the ResNet34 network to extract features and combines the multi-scale context aggregator (MSCA) for deep feature learning.

[0010] Step 3: Use the trained classification model to input the test set images and identify the growth status of tobacco plants.

[0011] Preferably, step 1 comprises the following steps:

[0012] Step 1.1, the dataset images cover the complete growth stage of tobacco plants. The tobacco field camera used is a special camera for the industry, model DS-2SK8CSP-ZZYC network high-definition high-speed intelligent full-color gun-ball integrated machine, and the camera is fixed-focus, fixed-point, and timed to take image data. The focal length is divided into three distances: long-range, medium-range, and close-range. According to the industry, the area in the long-range image contains the entire tobacco field, the area in the medium-range image contains 15 complete ridges, and the area in the close-range image is roughly 4-8 ridges, with about 10 tobacco plants in each ridge. Use the same focal length to rotate the camera on the same plot, and capture 8 or more pictures at each depth of field. During the capture process, ensure that the camera rotates one circle on the same horizontal line to obtain a complete field image. The time is fixed at 10:00, 14:30, and 16:30 every day to capture images.

[0013] Step 1.2, ask experts to label the pictures captured by the camera, marking the growth and development stage and nutritional status of the tobacco fields in the pictures.

[0014] Step 1.3, pre-process the image size by compressing, cropping, rotating, setting different grayscale values, etc.

[0015] Step 1.4, use the HSV color threshold segmenter to segment the hue, saturation and brightness of the tobacco plant, which can intuitively express the brightness, hue and vividness of the tobacco plant color. The HSV color space can be described by a cone space model. The image is binary processed to eliminate the complex background interference information in the image, solving the problem of low image recognition accuracy under complex background.

[0016] Step 1.5, divide the data set into training set, validation set, and test set in a ratio of 6:2:2. The training set is used to evaluate the model, learn the sample data set, and establish the classifier by matching parameters. The validation set is used to determine the parameters of the network structure or control the complexity of the model, and adjust the parameters of the classifier for the learned model. The test set is used to test the performance of the optimal model finally selected, mainly to test the resolution ability of the trained model.

[0017] Preferably, step 2 comprises the following steps:

[0018] Step 2.1, input the processed image in step 1 into the classification model according to H*W*3. At this time, the image is a two-dimensional image with H*W pixels and 3 channels. Use ResNet34 as the backbone network, delete the initial fully connected layer, and the convolution operation of the first layer is a convolution layer with a convolution kernel size of 7×7 and a stride of 2. The size of the input image is reduced to H / 2*W / 2, and the number of channels is increased from 3 to 64. The first layer is used to extract half-sized shallow features. After that, a 3×3 maximum pooling layer with a stride of 2 is further used to capture features of one-quarter the size of the original image, in order to filter important features and reduce the number of parameters. Followed by a 3*3 maximum pooling layer with a stride of 2. Then use 3 modules composed of residual blocks, each module uses several residual blocks with the same number of output channels. The number of channels of the first module is the same as the number of input channels. Since the maximum pooling layer with a stride of 2 has been used before, there is no need to reduce the height and width. Each subsequent module doubles the number of channels of the previous module in the first residual block and halves the height and width. The output channels of each ResBlock are 64, 128, 256, and 512 respectively. The sequence is flattened and transposed at the final output.

[0019] Step 2.2, the model first extracts features through a multi-layer convolutional network based on ResNet34. ResNet34 effectively alleviates the gradient vanishing problem in deep networks through its residual structure, and can reduce the number of parameters and computational complexity while maintaining high accuracy. On this basis, the multi-scale context aggregator (MSCA) is combined to further improve the diversity and expression ability of features. MSCA aggregates and fuses context information of feature maps at different scales, enhancing the model's sensitivity to objects of different scales. In particular, when dealing with complex backgrounds or multi-scale targets, it can improve the learning effect of deep features. In this way, the model can not only capture local detail information, but also obtain global context information through the MSCA module, achieving more accurate feature expression and classification.

[0020] The specific method is as follows: The generated features are used to capture and aggregate multi-scale contextual information from three features of different sizes using three TokenEncoders and three TokenDecoders built on the transformer architecture. First, the TokenEncoder uses a spatial attention module to convert the input features into 3D tokens embeddings of the target size for subsequent transformer modules while considering the limitations of computation and storage. Given an input feature, called F∈R b ×c×h×w , the spatial attention module uses a 1×1 convolutional layer to obtain intermediate features, called F′∈R b×l×h×w Then, both F and F′ will be reshaped into 3D tokens, called f∈R b×c×(h×w) and f′∈R b×l×(h×w) Through the einsum operation, f and f′ are converted into token embeddings t∈Rb×l×c, which can be expressed as t blc =f' bl(hw) f bc(hw) Where b, c, h, and w represent the batch size, number of channels, height, and width of the input feature F, respectively; Subsequently, the Transformer encoder is used to model the contextual information in the token, where a set of trainable parameters are first added element-by-element to the token t for position embedding. The Transformer encoder adopts the standard Transformer structure, including a multi-head attention block and a feedforward neural network block, and a layer normalization layer is applied before each block. The token decoder receives two inputs, the convolutional features F from the feature extractor and the token embedding T from the token encoder. In the token decoder, a trainable parameter is first added to F for position embedding, and then F and T are input to the transformer decoder, the purpose of which is to re-project the token embedding into the pixel space and enhance the contextual information in F. In the Transformer decoder, a layer normalization layer with shared weights is first applied to F and T, and then a multi-head attention module is used. In the process of encoding and decoding tokens at different scales, context aggregation connections are introduced to aggregate higher-level Transformer features into lower-level convolutional features.

[0021] In step 2.3, global average pooling is used to downsample the overall feature information by one layer. The output feature map of the last convolutional layer is converted into a fixed-size feature vector of size H×W×C, and a vector of length C is output, where each element is the average of all elements on the corresponding channel.

[0022] Step 2.4, input the final feature vector into the fully connected layer, perform a linear transformation through a weight matrix, then add a bias term, and finally perform a nonlinear transformation through the activation function (Relu) to complete the image classification task, which is expressed by the formula: y = б(w*x+b). Among them, y is the output vector, б is the activation function, w is the weight matrix, and b is the bias vector.

[0023] Preferably, the specific method of step 3 is to input the images of the test set into the trained classification model, determine the growth and development stage and nutritional status category of the tobacco leaves, compare them with the real category (referring to the status given by field experts), and detect the accuracy of the classification model.

[0024] The advantages of the present invention over the prior art are: the use of tobacco field camera equipment to capture and collect images and the method of distinguishing the growth and development stage and nutritional status of tobacco leaves based on deep learning can achieve real-time understanding of the growth and development stage and nutritional status of tobacco plants in the tobacco field during the tobacco planting stage, and accordingly adjust the amount of fertilizer and take effective measures to increase the quality and yield of tobacco plants. Provide a method to solve the current problems of low accuracy in judging the stage of tobacco plants in tobacco fields, strong human subjectivity, and long time consumption and loss of manpower and material resources. And establish a set of private data sets for tobacco fields to provide more data support for the later establishment of smart agricultural algorithm models and scientific planting in the field. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a tobacco field image acquisition and identification method in an embodiment of the present invention;

[0026] Figure 2 Schematic diagram of the improved neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0028] The present invention proposes a tobacco leaf maturity image classification method based on a multi-scale multi-path structure. RGB images of tobacco leaves at different growth and development stages and nutritional status are made into a data set, a label corresponding to each image is made, and the image data is preprocessed; the processed image is input into a classification model for training, and after the training is completed, the tobacco leaf RGB image can be input for classification, and the output is the tobacco leaf growth and development stage and nutritional status category determined by the model.

[0029] like Figure 1 As shown, a classification method for the growth status of field tobacco plants based on deep learning is characterized by comprising the following steps:

[0030] Step 1: Establish a tobacco plant growth status dataset, create corresponding labels, and pre-process the collected dataset to obtain image data. Specifically, it includes:

[0031] Step 1.1, the dataset images cover the complete growth stage of tobacco plants. The tobacco field camera used is a special camera for the industry, model DS-2SK8CSP-ZZYC network high-definition high-speed intelligent full-color gun-ball integrated machine. The camera is fixed-focus, fixed-point, and timed to take image data. The focal length is divided into three distances: distant view, medium view, and close view. According to industry specifications, the area in the distant view image includes the entire tobacco field, the area in the medium view image includes 15 complete ridges, and the area in the close view image is roughly 4-8 ridges, with about 10 tobacco plants in each ridge. The camera is rotated at the same focal length on the same plot, and 8 or more images are captured at each depth of field. During the capture process, the camera is ensured to rotate one circle on the same horizontal line to obtain a complete field image. The time is fixed at 10:00, 14:30, and 16:30 every day to capture images. These three time periods are all times with sufficient light every day, which can ensure the quality of the captured images. The images captured by the camera are collected and sorted using the Hikvision Cloud Eye platform to provide support for the subsequent establishment of a dataset.

[0032] Step 1.2, ask experts to label the pictures captured by the camera, marking the growth and development stage and nutritional status of the tobacco fields in the pictures.

[0033] Step 1.3, pre-process the image size by compressing, cropping, rotating, setting different grayscale values, etc.

[0034] Step 1.4, use the HSV color threshold segmenter to segment the hue, saturation and brightness of the tobacco plant, which can intuitively express the brightness, hue and vividness of the tobacco plant color. The HSV color space can be described by a cone space model. The image is binary processed to eliminate the complex background interference information in the image, solving the problem of low image recognition accuracy under complex background.

[0035] Step 1.5, divide the data set into training set, validation set, and test set in a ratio of 6:2:2.

[0036] Step 2, such as Figure 2 As shown, a classification model is established to train the image data generated in step 1. The classification model is based on the multi-layer convolution feature extraction in the ResNet34 network, adding a multi-scale context aggregator (MSCA), and modifying the network output part to a multi-branch prediction head (MBPH) to enhance deep feature learning. Specifically, it includes:

[0037] Step 2.1, input the processed image in step 1 into the classification model according to H*W*3. At this time, the image is a two-dimensional image with H*W pixels and 3 channels. Use ResNet34 as the backbone network, delete the initial fully connected layer, and the convolution operation of the first layer is a convolution layer with a convolution kernel size of 7×7 and a stride of 2. The size of the input image is reduced to H / 2*W / 2, and the number of channels is increased from 3 to 64. The first layer is used to extract shallow features of half size. After that, a 3×3 maximum pooling layer with a stride of 2 is further used to capture features of one-quarter the size of the original image, in order to filter important features and reduce the number of parameters. It is followed by a 3*3 maximum pooling layer with a stride of 2. Then, three modules composed of residual blocks are used, and each module uses several residual blocks with the same number of output channels. The number of channels of the first module is the same as the number of input channels. Each subsequent module doubles the number of channels of the previous module in the first residual block and halves the height and width. The output channels of each ResBlock are 64, 128, 256, and 512. The sequences are flattened and transposed at the final output.

[0038] Step 2.2, the generated features are used to capture and aggregate multi-scale contextual information from three features of different sizes using three Token Encoders and three Token Decoders built on the transformer architecture. First, the Token Encoder uses a spatial attention module to convert the input features into 3D tokens embeddings of the target size for subsequent transformer modules while considering the limitations of computation and storage. Given an input feature, called F∈R b×c×h×w , the spatial attention module uses a 1×1 convolutional layer to obtain intermediate features, called F′∈R b×l×h×w Then, both F and F′ will be reshaped into 3D tokens, called f∈R b×c×(h×w) and f′∈R b×l×(h×w) . Through the einsum operation, f and f′ are converted into token embedding t∈Rb×l×c, which can be expressed as:

[0039] t blc =f' bl(hw) f bc(hw)

[0040] Where b, c, h, and w represent the batch size, number of channels, height, and width of the input feature F, respectively; Subsequently, the Transformer encoder is used to model the contextual information in the token, where a set of trainable parameters are first added element-by-element to the token t for position embedding. The Transformer encoder adopts the standard Transformer structure, including a multi-head attention block and a feedforward neural network block, and a layer normalization layer is applied before each block. The token decoder receives two inputs, the convolutional features F from the feature extractor and the token embedding T from the token encoder. In the token decoder, a trainable parameter is first added to F for position embedding, and then F and T are input to the transformer decoder, the purpose of which is to re-project the token embedding into the pixel space and enhance the contextual information in F. In the Transformer decoder, a layer normalization layer with shared weights is first applied to F and T, and then a multi-head attention module is used. In the process of encoding and decoding tokens at different scales, context aggregation connections are introduced to aggregate higher-level Transformer features into lower-level convolutional features.

[0041] In step 2.3, global average pooling is used to downsample the overall feature information by one layer. The output feature map of the last convolutional layer is converted into a fixed-size feature vector of size H×W×C, and a vector of length C is output, where each element is the average of all elements on the corresponding channel.

[0042] In step 2.4, the final feature vector is input into the fully connected layer, linearly transformed through a weight matrix, then a bias term is added, and finally a nonlinear transformation is performed through the activation function (Relu) to complete the image classification task, which is expressed by the formula:

[0043] y=б(w*x+b).

[0044] Among them, y is the output vector, б is the activation function, w is the weight matrix, and b is the bias vector.

[0045] Step 3: Use the trained classification model to input the test set images and identify the growth status of tobacco plants. Input the test set images into the trained classification model, determine the tobacco leaf growth and development stage and nutritional status categories, compare them with the real categories, and test the accuracy of the classification model.

[0046] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention are equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A classification method for the growth status of field tobacco plants based on deep learning, characterized in that: The following steps are involved: Step 1: Establish a tobacco plant growth status dataset, create corresponding labels, and pre-process the collected dataset to obtain image data; Step 2: Establish a classification model to train the image data generated in step 1. The classification model is based on multi-layer convolution in the ResNet34 network to extract features and combines a multi-scale context aggregator (MSCA) for deep feature learning. Step 3: Use the trained classification model to input the test set images and identify the growth status of tobacco plants.

2. The method for classifying the growth status of field tobacco plants based on deep learning according to claim 1, characterized in that: Step 1 contains the following steps: Step 1.1, the dataset images cover the complete growth stages of tobacco plants. The tobacco field camera used is a special camera for the industry, model DS-2SK8CSP-ZZYC network high-definition high-speed intelligent full-color gun-ball integrated camera. The camera is fixed-focus, fixed-point, and timed to capture image data; Step 1.2, ask experts to annotate the pictures captured by the camera, marking the growth and development stage and nutritional status of the tobacco fields in the pictures; Step 1.3, pre-process the image size by compressing, cropping, rotating, setting different grayscale values, etc. Step 1.4, use the HSV color threshold segmenter to segment the tobacco plant pixels, and perform binary processing on the image to remove the complex background interference information in the image, and solve the problem of low image recognition accuracy under complex background; Step 1.5, divide the data set into training set, validation set, and test set in a ratio of 6:2:2; The training set is used to evaluate the model, learn the sample data set, and establish the classifier by matching parameters; the validation set is used to determine the parameters of the network structure or control the complexity of the model, and adjust the parameters of the classifier for the learned model; the test set is used to test the performance of the optimal model finally selected, mainly to test the resolution ability of the trained model.

3. The method for classifying the growth status of field tobacco plants based on deep learning according to claim 2, characterized in that: The camera model used in step 1.1 is DS-2SK8CSP-ZZYC network high-definition high-speed intelligent full-color gun-ball integrated camera.

4. The method for classifying the growth status of field tobacco plants based on deep learning according to claim 2, characterized in that: In step 1.1, the focal length is divided into three distances: distant view, mid-view, and close-up. According to the industry, the distant view image area includes the entire tobacco field, the mid-view image area includes 15 complete ridges, and the close-up image area is approximately 4-8 ridges, with about 10 tobacco plants in each ridge. The camera is rotated at the same focal length on the same plot, and 8 or more images are captured at each depth of field. During the capture process, the camera is ensured to rotate one circle on the same horizontal line to obtain a complete field image. The time is fixed at 10:00, 14:30, and 16:30 every day to capture images.

5. The method for classifying the growth status of field tobacco plants based on deep learning according to claim 1, characterized in that: Step 2 contains the following steps: Step 2.1, input the processed image in step 1 into the classification model according to H*W*3. At this time, the image is a two-dimensional image with H*W pixels and 3 channels. Use ResNet34 as the backbone network, delete the initial fully connected layer, and the convolution operation of the first layer is a convolution layer with a convolution kernel size of 7×7 and a stride of 2. The size of the input image is reduced to H / 2*W / 2, and the number of channels is increased from 3 to 64. The first layer is used to extract shallow features of half size. After that, a 3×3 maximum pooling layer with a stride of 2 is further used to capture features of one-quarter the size of the original image, in order to filter important features and reduce the number of parameters. It is followed by a 3*3 maximum pooling layer with a stride of 2. Then, use 3 modules composed of residual blocks, each module uses several residual blocks with the same number of output channels; the number of channels of the first module is the same as the number of input channels. Since the maximum pooling layer with a stride of 2 has been used before, there is no need to reduce the height and width. Each subsequent module doubles the number of channels of the previous module in the first residual block and halves the height and width; the output channels of each ResBlock are 64, 128, 256, and 512 respectively; the sequence is flattened and transposed at the final output; Step 2.2, the model first extracts features through a multi-layer convolutional network based on ResNet34. ResNet34 effectively alleviates the gradient vanishing problem in deep networks through its residual structure, and can reduce the number of parameters and computational complexity while maintaining high accuracy. On this basis, the multi-scale context aggregator (MSCA) is combined to further improve the diversity and expression ability of features. MSCA aggregates and fuses context information of feature maps at different scales, enhancing the model's sensitivity to objects of different scales. In particular, when dealing with complex backgrounds or multi-scale targets, it can improve the learning effect of deep features. In this way, the model can not only capture local detail information, but also obtain global context information through the MSCA module, achieving more accurate feature expression and classification. Step 2.3, use global average pooling to downsample the overall feature information by one layer, and the output feature map of the last convolutional layer is converted into a fixed-size feature vector of size H×W×C, and outputs a vector of length C, where each element is the average of all elements on the corresponding channel; In step 2.4, the final feature vector is input into the fully connected layer, linearly transformed through a weight matrix, and then a bias term is added. Finally, a nonlinear transformation is performed through the activation function (Relu) to complete the image classification task, which is expressed by the formula: y = б(w*x+b), where y is the output vector, б is the activation function, w is the weight matrix, and b is the bias vector.

6. The method for classifying the growth status of field tobacco plants based on deep learning according to claim 5, characterized in that: The specific method of step 2.2 is as follows: the generated features are used to capture and aggregate multi-scale contextual information from three features of different sizes using three Token Encoders and three Token Decoders built on the transformer architecture; first, the Token Encoder uses a spatial attention module to convert the input features into 3D tokens embedding of the target size for subsequent transformer modules while considering the limitations of computation and storage; given an input feature, called F∈R b×c×h×w , the spatial attention module uses a 1×1 convolutional layer to obtain intermediate features, called F′∈R b×l×h×w Then, both F and F′ will be reshaped into 3D tokens, called f∈R b×c×(h×w) and f′∈R b×l×(h×w) ; Through the einsum operation, f and f′ are converted into token embedding t∈Rb×l×c, which can be expressed as t blc =f' bl(hw) f bc(hw) Where b, c, h, and w represent the batch size, number of channels, height, and width of the input feature F, respectively. Subsequently, the Transformer encoder is used to model the contextual information in the token, where a set of trainable parameters are first added element-by-element to the token t for position embedding. The Transformer encoder adopts the standard Transformer structure, including a multi-head attention block and a feedforward neural network block, and a layer normalization layer is applied before each block. The token decoder receives two inputs, the convolutional features F from the feature extractor and the token embedding T from the token encoder. In the token decoder, a trainable parameter is first added to F for position embedding, and then F and T are input into the transformer decoder, the purpose of which is to reproject the token embedding into the pixel space and enhance the contextual information in F. In the Transformer decoder, a layer normalization layer with shared weights is first applied to F and T, and then a multi-head attention module is used. In the encoding and decoding process of tokens at different scales, context aggregation connections are introduced to aggregate higher-level Transformer features into lower-level convolutional features.

7. The method for classifying the growth status of field tobacco plants based on deep learning according to claim 1, characterized in that: Step 3 specifically includes: inputting the tobacco leaf image to be predicted into the classification model, determining the tobacco leaf growth and development stage and nutritional status category and comparing it with the real category, and detecting the accuracy of the classification model.

Images