Crop growth environment perception and chlorophyll content detection device and detection method based on atlas fusion

Abstract

The application provides a crop growth environment perception and chlorophyll content detection device based on atlas fusion and a use method thereof. The device is constructed with an aluminum waterproof shell to build a protective main body, is powered by a power supply, and is internally integrated with a Raspberry Pi development board and a driving board assembly to work cooperatively. A miniature camera and a miniature spectrometer are respectively responsible for collecting crop color images and spectral data, cooperate with a light supplementing system, and guarantee data acquisition quality. Data is integrated through a connecting assembly, images are classified based on a mobile lightweight crop background classification model MLCBC-net, clear, turbid and green algae covered backgrounds are identified, corresponding chlorophyll content detection models are then called, and background classification and chlorophyll content detection results are presented on a touch screen. The application solves the problem of complex background interference, realizes high-precision detection of crop chlorophyll content, provides strong support for precise crop cultivation and management, and promotes the progress of intelligent agriculture.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of agricultural spectrum detection, and particularly relates to a crop growth environment perception and chlorophyll content detection device and method based on atlas fusion. BACKGROUND

[0002] In modern agricultural production, the chlorophyll content of crops is a key indicator reflecting their nutritional status and photosynthetic capacity. Precise detection is of great significance for scientific regulation of cultivation measures and guarantee of high yield and quality. Traditional chlorophyll detection relies on chemical analysis, which is highly accurate but destructive, time-consuming and costly in the detection process, and cannot meet the needs of large-scale, rapid and non-destructive detection in agriculture. In recent years, deep learning and spectral analysis technology have become a research hotspot for chlorophyll content detection due to their real-time, efficient and non-destructive advantages, opening up new avenues for non-destructive detection of chlorophyll. Many studies have attempted to integrate deep learning and spectral technology into detection devices for crop chlorophyll content detection. However, in practical applications, complex field environments (such as clear and turbid backgrounds formed by irrigation water of different qualities, or green algae coverage) can interfere with spectral signals, and their optical characteristics differ, causing different interference to crop spectra, resulting in detection errors. Moreover, existing detection devices mostly use a unified model to process data under different backgrounds, pay less attention to the application of background classification in detection, lack technology for constructing specific crop chlorophyll content detection models for different backgrounds and integrating them into devices, which limits the detection accuracy. Therefore, it is of great practical significance and application value to develop a crop background environment perception and chlorophyll content detection method and device based on atlas combination. SUMMARY

[0003] To solve the problems in the prior art, the application provides a crop growth environment perception and chlorophyll content detection device and method based on atlas fusion, which solves the problems of interference of existing devices by complex backgrounds and lack of specificity of chlorophyll content detection models. By constructing a background precise classification model and a chlorophyll content detection exclusive model and integrating them into a portable device, the chlorophyll content detection accuracy of crops is improved.

[0004] The application achieves the above technical purposes through the following technical means.

[0005] The application discloses a crop growth environment perception and chlorophyll content detection device based on atlas fusion, which comprises a touch screen fixed in a touch screen fixed waterproof shell, wherein the touch screen is used for displaying detection results and an operation interface; the touch screen fixed waterproof shell, an aluminum waterproof shell and a handheld aluminum waterproof shell jointly form a detection device shell; and the handheld aluminum waterproof shell is provided with a miniature camera, a miniature spectrometer, a driving plate assembly, a Raspberry Pi development board, a connecting assembly, a power supply and a light supplementing system; the Raspberry Pi development board is provided with a mobile lightweight crop background classification model MLCBC-net and a crop chlorophyll content detection model for different backgrounds, which are used for classifying the collected images and selecting corresponding detection models to calculate the chlorophyll content according to the classification results.

[0006] The application further discloses a crop growth environment perception and chlorophyll content detection method based on atlas fusion.

[0007] Step 1: in the touch screen operation interface, set the integral time, the light intensity and the average number of times;

[0008] Step 2: hold the handheld aluminum waterproof shell, collect crop color image data and spectral image data by using the miniature camera and the miniature spectrometer respectively, and transmit the data to the Raspberry Pi development board;

[0009] Step 3: the Raspberry Pi development board classifies the collected crop color image by using the MLCBC-net, obtains the crop background type, and processes the spectral image by using a crop chlorophyll self-adaptive spatial reduction grouping query attention module, namely a CCA-SR-GQA module, in the MLCBC-net;

[0010] Step 4: the Raspberry Pi development board selects a corresponding chlorophyll content detection model stored in the internal storage according to the obtained crop background type to detect the chlorophyll content;

[0011] Step 5: the Raspberry Pi development board processes the spectral image collected by the miniature spectrometer by using the selected chlorophyll content detection model, and calculates the crop chlorophyll content;

[0012] Step 6: the Raspberry Pi development board sends result information to the touch screen to realize result visualization display.

[0013] Further, the structure of the MLCBC-net comprises four parts, namely four stages, each stage comprises different modules, and the crop color image is processed by the four stages of the MLCBC-net in sequence.

[0014] The MLCBC-net is arranged as follows:

[0015] Stage 1: the input is 4482 Conv2D module with 3x3 kernel, input is 224 2 FusedIB module with 32x32 kernel, input is 112 2 FusedIB module with 48x48 kernel;

[0016] Stage 2: input is 56 2 ExtraDW module with 80x80 kernel, input is 28 2 Inverted Bottleneck module with 160x160 kernel, input is 28 2 CCA-SR-GQA module with 160x160 kernel, input is 28 2 Inverted Bottleneck module with 160x160 kernel, input is 28 2 CCA-SR-GQA module with 160x160 kernel, input is 28 2 Inverted Bottleneck module with 160x160 kernel, input is 28 2 CCA-SR-GQA module with 160x160 kernel, input is 28 2 Inverted Bottleneck module with 160x160 kernel, input is 28 2 CCA-SR-GQA module with 160x160 kernel, input is 28 2 ConvNext module with 160x160 kernel;

[0017] Stage 3: input is 28 2 ExtraDW module with 160x160 kernel, input is 14 2 ExtraDW module with 256x256 kernel, input is 14 2 CCA-SR-GQA module with 256x256 kernel, input is 14 2 Inverted Bottleneck module with 256x256 kernel, input is 14 2 CCA-SR-GQA module with 256x256 kernel, input is 14 2 Inverted Bottleneck module with 256x256 kernel, input is 14 2 CCA-SR-GQA module with 256x256 kernel, input is 14 2 Inverted Bottleneck module with 256x256 kernel, input is 14 2 CCA-SR-GQA module with 256x256 kernel, input is 14 2 Inverted Bottleneck module with 256x256 kernel;

[0018] Stage 4: input is 14 2 Conv2D module with 256x256 kernel, input is 142 AvgPool module with a capacity of 256 and an input of 1. 2 Conv2D module of ×960, input is 1 2 Conv2D module with a resolution of ×1280;

[0019] Based on MLCBC-net, crop backgrounds are categorized into clear backgrounds, murky backgrounds, and green algae-covered backgrounds.

[0020] Furthermore, in step 3, the formula for the CCA-SR-GQA module is:

[0021]

[0022] in, To attract more attention Feature splicing, ; Encode the background type vector; , The weight matrix is ​​a learnable weight matrix;

[0023]

[0024] in, , , For image feature queries, keys, and values; Spectral characteristics; Adjust parameters for the dimension; This represents the attention mechanism; , , , , This is the learnable weight matrix for each part; This is a matrix transpose operation; To reduce attention in space, it is defined as follows:

[0025]

[0026]

[0027] in, Standardize for layers; Dimension adjustment operation; , These are the reduction ratios for image features and spectral features, respectively. , , The weight matrix is ​​a learnable weight matrix; This is element-wise multiplication.

[0028] Further, the step 4, the chlorophyll content detection model includes a first chlorophyll content detection model corresponding to a clear background, a second chlorophyll content detection model corresponding to a turbid background and a third chlorophyll content detection model corresponding to a green algae coverage background; Specifically as follows:

[0029] The first chlorophyll content detection model includes a standardization scaling and information variable elimination method spectral data preprocessing method, and a partial least squares regression chlorophyll content detection model;

[0030] The second chlorophyll content detection model includes a multi-source scattering correction and genetic algorithm spectral data preprocessing method, and a support vector regression chlorophyll content detection model;

[0031] The third chlorophyll content detection model includes a detrend correction and genetic algorithm spectral data preprocessing method, and a convolutional neural network chlorophyll content detection model.

[0032] The present application has the following beneficial effects:

[0033] (1) The mobile lightweight crop background classification model (MLCBC-net) is used for accurate background classification of the crop color image, and the clear, turbid and green algae coverage backgrounds are clearly distinguished, so that the interference of the complex background on the spectral signal is effectively excluded. The specific chlorophyll content detection model is called for different backgrounds, and the detection accuracy is improved.

[0034] (2) The miniature camera is matched with the miniature spectrometer and the light supplement system, so that high-quality crop image and spectral data can be quickly and stably obtained, and the integrity and accuracy of the data are ensured.

[0035] (3) The Raspberry Pi development board and the driving board assembly work cooperatively, the background classification and the corresponding crop chlorophyll content detection model calling are realized based on the MLCBC-net, and the detection result is output in real time through the touch screen, so that the operation is simple, the operation speed is fast, the processing capacity is strong, and the needs of agricultural production are met. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 It is a schematic diagram of the overall appearance of the detection device;

[0037] Figure 2 It is a schematic diagram of the internal power supply arrangement of the detection device;

[0038] Figure 3 It is a three-dimensional schematic diagram of the arrangement of various components on the handheld aluminum waterproof shell;

[0039] Figure 4 It is a schematic diagram of the internal structure of the detection device;

[0040] Figure 5 It is a schematic diagram of the light supplement lamp arrangement;

[0041] Figure 6 A schematic diagram showing the arrangement of the supplementary lighting, miniature camera, and miniature spectrometer;

[0042] Figure 7 MLCBC-net model diagram;

[0043] Figure 8 A diagram showing the chlorophyll content detection model under different backgrounds.

[0044] In the diagram: 1. Touch screen; 2. Waterproof housing for the touch screen; 3. Waterproof aluminum housing; 4. Handheld waterproof aluminum housing; 5. Miniature camera; 6. Miniature spectrometer; 7. Driver board assembly; 8. Raspberry Pi development board; 9. Connector assembly; 10. Power supply; 101. Power supply bracket; 102. Power supply mounting bracket; 111. First fill light; 112. Second fill light. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0046] like Figures 1 to 6 As shown, the crop growth environment sensing and chlorophyll content detection device based on map fusion of the present invention includes a touch screen 1, a fixed waterproof shell for the touch screen 2, an aluminum waterproof shell 3, a handheld aluminum waterproof shell 4, a miniature camera 5, a miniature spectrometer 6, a driver board assembly 7, a Raspberry Pi development board 8, a connection assembly 9, a power supply 10, and a supplementary lighting system.

[0047] like Figure 1 , 2 As shown, the touch screen 1 is fixed inside the touch screen fixing waterproof shell 2 and is used to display the test results and operation interface; the touch screen fixing waterproof shell 2, the aluminum waterproof shell 3, and the handheld aluminum waterproof shell 4 together constitute the outer shell of the testing device, which is used to house the main components of the testing device; the two ends of the handheld aluminum waterproof shell 4 that extend out of the outer shell of the testing device are all designed as ring handles for easy hand operation by the user.

[0048] like Figure 3 , 4 As shown, the miniature camera 5 and the miniature spectrometer 6 are both mounted on the handheld aluminum waterproof shell 4, and are used to acquire crop color images and crop spectral images, respectively; the drive board assembly 7 is also mounted on the handheld aluminum waterproof shell 4, and is used to drive the miniature camera 5, the miniature spectrometer 6 and the supplementary lighting system.

[0049] like Figure 3As shown, the Raspberry Pi development board 8 is installed on the handheld aluminum waterproof shell 4, which carries a mobile lightweight crop background classification model (MLCBC-net) and a crop chlorophyll content detection model for different backgrounds, and is used for background classification of the collected images and selection of the corresponding detection model to calculate the chlorophyll content according to the classification result.

[0050] As shown in Figure 3 , 4 The connecting assembly 9 is installed on the handheld aluminum waterproof shell 4, which is used to connect each component and realize data transmission; the power supply 10 is fixed above the handheld aluminum waterproof shell 4 by screws, screw sleeves, power supply supports 101, and power supply fixing frames 102, which supplies power to each component of the detection device.

[0051] As shown in Figure 5 , 6 The light supplementing system includes a first light supplementing lamp 111 and a second light supplementing lamp 112, both of which are installed below the handheld aluminum waterproof shell 4, and are used to provide auxiliary lighting when collecting images and spectra.

[0052] The crop growth environment perception and chlorophyll content detection method of the crop growth environment perception and chlorophyll content detection device based on atlas fusion comprises the following processes:

[0053] Step 1: In the detection device touch screen 1 operation interface, set the integration time, light intensity, and average number; wherein the integration time setting is to adjust the exposure time when the miniature spectrometer 6 collects spectral data; the light intensity setting is because the light intensity changes at any time when collecting spectral data outdoors, and the change of light intensity will affect the quality of spectral data collection, and if the light intensity can be adjusted, the spectral data can be stably collected; the average number setting is to reduce the error of spectral data collection, for example, when collecting the first data, the data quality may be poor, or other reasons may cause poor data quality, so multiple collections are needed to average, which can reduce the time error of data collection as much as possible.

[0054] Step 2: Hold the handheld aluminum waterproof shell 4, and use the miniature camera 5 and the miniature spectrometer 6 thereon to collect crop color images and spectral images, respectively;

[0055] Step 3: The Raspberry Pi development board 8 classifies the collected crop color image by using a mobile lightweight crop background classification (MLCBC-net) model, obtains the crop background type, and processes the spectral image by using a crop-chlorophyll adaptive spatial reduction grouped-query attention (CCA-SR-GQA) module in the MLCBC-net;

[0056] The structure of the MLCBC-net includes four parts, i.e., four stages, each stage includes different modules, and the crop color image is processed by the four stages of the MLCBC-net in turn, as shown in Figure 7 The MLCBC-net is arranged as follows:

[0057] Stage 1: sequentially through a Conv2D module with an input feature map size (hereinafter referred to as input) of 448 2 × 3, a FusedIB module with an input of 224 2 × 32, and a FusedIB module with an input of 112 2 × 48;

[0058] Stage 2: sequentially through an ExtraDW module with an input of 56 2 × 80, an InvertedBottleneck module with an input of 28 2 × 160, a CCA-SR-GQA module with an input of 28 2 × 160, an InvertedBottleneck module with an input of 28 2 × 160, a CCA-SR-GQA module with an input of 28 2 × 160, an InvertedBottleneck module with an input of 28 2 × 160, a CCA-SR-GQA module with an input of 28 2 × 160, an InvertedBottleneck module with an input of 28 2 × 160, a CCA-SR-GQA module with an input of 28 2 × 160, and a ConvNext module with an input of 28 2 × 160;

[0059] The formula of the CCA-SR-GQA is as follows:

[0060]

[0061] in, To attract more attention Feature splicing, ; Encode the background type vector; , The weight matrix is ​​a learnable weight matrix;

[0062]

[0063] in, , , For image features, the query, key, and value are defined. Spectral characteristics; Adjust parameters for the dimension; This represents the attention mechanism; , , , , This is the learnable weight matrix for each part; This is a matrix transpose operation; Spatial reduction attention, used to adjust the feature dimensions of the variables within parentheses, is defined as follows:

[0064]

[0065]

[0066] in, Standardize for layers; This is a dimensionality adjustment operation to match the dimensions of image and spectral features; , These are the reduction ratios for image features and spectral features, respectively. , , The weight matrix is ​​a learnable weight matrix; For element-wise multiplication, background encoding is used. Weight modulation is applied to the fusion features of images and spectra to achieve adaptive spatial reduction under different backgrounds.

[0067] Phase 3: Inputting 28 sequentially 2 The ExtraDW module has a size of ×160 and an input of 14. 2 The ExtraDW module has a capacity of 256 and an input of 14. 2 The CCA-SR-GQA module has a capacity of 256 and an input of 14.2 Inverted Bottleneck module with input of 14 2 CCA-SR-GQA module with input of 14 2 Inverted Bottleneck module with input of 14 2 CCA-SR-GQA module with input of 14 2 Inverted Bottleneck module with input of 14 2 CCA-SR-GQA module with input of 14 2 Inverted Bottleneck module with input of 14

[0068] Stage 4: sequentially through the input of 14 2 Conv2D module with input of 14 2 AvgPool module with input of 1 2 Conv2D module with input of 1 2 Conv2D module with input of 1280

[0069] The MLBCB-net can divide the crop background into clear background, turbid background and green algae covered background.

[0070] Step 4: as shown in Figure 8 The raspberry development board 8 selects the corresponding chlorophyll content detection model stored in the internal storage according to the obtained crop background type to detect the chlorophyll content;

[0071] The chlorophyll content detection model includes a first chlorophyll content detection model corresponding to the clear background, a second chlorophyll content detection model corresponding to the turbid background, and a third chlorophyll content detection model corresponding to the green algae covered background. Specifically as follows:

[0072] The first chlorophyll content detection model includes a standardization scaling and a non-informative variable elimination method spectral data preprocessing method, and a partial least squares regression chlorophyll content detection model;

[0073] The second chlorophyll content detection model includes a multi-source scattering correction and a genetic algorithm spectral data preprocessing method, and a support vector regression chlorophyll content detection model;

[0074] The third chlorophyll content detection model includes a detrend correction and a genetic algorithm spectral data preprocessing method, and a convolutional neural network chlorophyll content detection model.

[0075] Step 5: Raspberry Pi development board 8 processes the spectral image collected by the miniature spectrometer 6 in step 3 using the selected chlorophyll content detection model to calculate the chlorophyll content of the crop;

[0076] Step 6: Raspberry Pi development board 8 sends the result information to the touch screen 1 for result visualization.

[0077] The above embodiments are preferred embodiments of the present application, but the present application is not limited to the above embodiments, and any obvious improvements, replacements or modifications made by those skilled in the art without departing from the essential content of the present application shall fall within the protection scope of the present application.

Claims

1. A method for crop growth environment sensing and chlorophyll content detection based on map fusion, utilizing a map fusion-based crop growth environment sensing and chlorophyll content detection device, characterized in that, The device includes a touch screen (1) fixed inside a waterproof housing (2), which is used to display the detection results and operation interface. The waterproof housing (2), the aluminum waterproof housing (3), and the handheld aluminum waterproof housing (4) together constitute the outer shell of the detection device. It also includes a miniature camera (5), a miniature spectrometer (6), a driver board assembly (7), a Raspberry Pi development board (8), a connection assembly (9), a power supply (10), and a supplementary lighting system mounted on the handheld aluminum waterproof housing (4). The Raspberry Pi development board (8) is equipped with a mobile lightweight crop background classification model MLCBC-net and a crop chlorophyll content detection model for different backgrounds, which is used to classify the background of the collected images and select the corresponding detection model to calculate the chlorophyll content based on the classification results. The method for crop growth environment perception and chlorophyll content detection based on map fusion includes the following process: Step 1: In the touch screen (1) operation interface, set the integration time, light intensity, and average number of times; Step 2: Hold the handheld aluminum waterproof case (4), use the miniature camera (5) and miniature spectrometer (6) to collect crop color images and spectral image data respectively, and transfer the data to the Raspberry Pi development board (8); Step 3: The Raspberry Pi development board (8) classifies the collected crop color images through MLCBC-net to obtain the crop background type. At the same time, it uses the crop chlorophyll adaptive spatial reduction grouping query attention module in MLCBC-net, namely the CCA-SR-GQA module, to process the spectral image. Step 4: The Raspberry Pi development board (8) selects the corresponding chlorophyll content detection model stored internally to detect chlorophyll content based on the obtained crop background type; Step 5: The Raspberry Pi development board (8) uses the selected chlorophyll content detection model to process the spectral images collected by the miniature spectrometer (6) and calculate the crop chlorophyll content; Step 6: The Raspberry Pi development board (8) sends the result information to the touch screen (1) for result visualization; In step 3, the structure of MLCBC-net consists of 4 parts, namely 4 stages, each stage including different modules. The crop color image is processed through the four stages of MLCBC-net in sequence. The specific deployment of MLCBC-net is as follows: Phase 1: Input is 448 2 A Conv2D module of size ×3, with an input of 224. 2 A 32-bit FusedIB module with input 112. 2 ×48 FusedIB modules; Phase 2: Input is 56 2 The ExtraDW module with a size of ×80 and an input of 28. 2 ×160 Inverted Bottleneck module, input 28 2 The CCA-SR-GQA module has a size of ×160 and an input of 28. 2 ×160 Inverted Bottleneck module, input 28 2 The CCA-SR-GQA module has a size of ×160 and an input of 28. 2 ×160 Inverted Bottleneck module, input 28 2 The CCA-SR-GQA module has a size of ×160 and an input of 28. 2 ×160 Inverted Bottleneck module, input 28 2 The CCA-SR-GQA module has a size of ×160 and an input of 28. 2 ConvNext module with a size of 160; Phase 3: Input is 28 2 The ExtraDW module has a size of ×160 and an input of 14. 2 The ExtraDW module has a capacity of 256 and an input of 14. 2 The CCA-SR-GQA module has a capacity of 256 and an input of 14. 2 A 256-bit Inverted Bottleneck module with 14 inputs. 2 The CCA-SR-GQA module has a capacity of 256 and an input of 14. 2 A 256-bit Inverted Bottleneck module with 14 inputs. 2 The CCA-SR-GQA module has a capacity of 256 and an input of 14. 2 A 256-bit Inverted Bottleneck module with 14 inputs. 2 The CCA-SR-GQA module has a capacity of 256 and an input of 14. 2 ×256 Inverted Bottleneck module; Phase 4: Input is 14 2 A Conv2D module with a size of 256 and an input of 14. 2 AvgPool module with a size of 256 and an input of 1. 2 Conv2D module of ×960, input is 1 2 Conv2D module with a resolution of ×1280; Based on MLCBC-net, crop backgrounds are categorized into clear backgrounds, murky backgrounds, and green algae-covered backgrounds. In step 3, the formula for the CCA-SR-GQA module is: Where Concat(·) is a multi-attention head. i Feature concatenation, i = 1, ..., n; B is the background type encoding vector; W O , The weight matrix is ​​a learnable weight matrix; Where Q, K, and V are the query, key, and value of the image feature; S is the spectral feature; d k Adjust parameters for dimensions; Attention(·) represents the attention mechanism; Here is the learnable weight matrix for each part; (·) T This is the matrix transpose operation; SR(·) is the space-reducing attention, defined as follows: Where Norm(·) is layer normalization; Reshape(·,·) is dimension adjustment operation; R, R S These are the reduction ratios for image features and spectral features, respectively. is the learnable weight matrix; ⊙ represents element-wise multiplication.

2. The method for crop growth environment perception and chlorophyll content detection based on map fusion according to claim 1, characterized in that, In step 4, the chlorophyll content detection model includes a first chlorophyll content detection model corresponding to a clear background, a second chlorophyll content detection model corresponding to a turbid background, and a third chlorophyll content detection model corresponding to a background covered by green algae; specifically as follows: The first chlorophyll content detection model includes a standardized scaling and non-information variable elimination method for spectral data preprocessing, as well as a partial least squares regression chlorophyll content detection model. The second chlorophyll content detection model includes a multi-source scattering correction and genetic algorithm spectral data preprocessing method, as well as a support vector regression chlorophyll content detection model; The third chlorophyll content detection model includes detrending correction and genetic algorithm spectral data preprocessing methods, as well as a convolutional neural network chlorophyll content detection model.

Images