Farmland full-process management method and system based on deep learning

Abstract

The invention relates to the technical field of deep learning, and discloses a farmland full-process management method and system based on deep learning, and the method comprises the steps: collecting image data, environment data and crop physiological indexes of a target farmland; generating noise reduction data of the environmental data based on a moving average algorithm; generating a farmland image based on the conditional generative adversarial network and the image data; extracting time features of the noise reduction data by using a crop evaluation model, and generating a dependency relationship of the noise reduction data based on the time features and a long-short term memory network; extracting multi-level spatial features of the farmland image according to the crop evaluation model, and further generating fusion features of the target farmland; and performing layer-by-layer processing on the fusion features by using a full connection layer and an output layer of the crop evaluation model to obtain a crop state value of the target farmland, generating a management decision of the target farmland according to the crop state value, and managing the target farmland. According to the invention, the efficiency of farmland whole-process management can be improved based on deep learning.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a full-process farmland management method and system based on deep learning. Background Art

[0002] As the global population grows and food demand rises, improving the efficiency of farmland management throughout the entire process has become the key to the development of modern agriculture. Efficient management can optimize crop growth, accurately control resources such as water, fertilizer, and pesticides, increase yield and quality, stabilize food supply, resist natural disasters and pests, and reduce the risk of production reduction.

[0003] For example, the Chinese patent document with publication number CN201910136570.6 discloses a wide-area monitoring system and method based on image style transfer; the Chinese patent document with publication number CN113239877A discloses a farmland monitoring method based on computer vision and related equipment.

[0004] However, many existing farmland management systems lack raw data and have relatively single data sources. They often rely only on regular manual field observations and a small number of simple sensor monitoring, making it difficult to obtain comprehensive and continuous farmland information. On the other hand, in terms of data processing and analysis, traditional methods mainly rely on manual experience and simple statistical analysis, lacking efficient data analysis models and algorithms. The agricultural production environment is complex and changeable, involving many interrelated factors, and manual analysis is difficult to quickly and accurately mine the potential laws and trends behind the data. Summary of the invention

[0005] The present invention provides a method and system for full-process farmland management based on deep learning, the main purpose of which is to solve the problem of low efficiency in full-process farmland management.

[0006] To achieve the above objectives, the present invention provides a method for managing farmland in the entire process based on deep learning, comprising: Collect image data, environmental data and crop physiological indicators of target farmland; Performing noise reduction processing on the environmental data based on a sliding average algorithm to obtain noise-reduced data of the environmental data; Generate a farmland image of the target farmland based on a conditional generative adversarial network and the image data; Extracting the time features of the denoised data using a pre-built crop evaluation model, and generating a dependency relationship of the denoised data based on the time features and a long short-term memory network; Extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model; Determining the fusion characteristics of the target farmland based on the dependency relationship, the crop physiological index, the multi-level spatial characteristics and the temporal characteristics; The fusion features are processed layer by layer using the fully connected layer and the output layer of the crop evaluation model to obtain the crop status value of the target farmland, a management decision of the target farmland is generated according to the crop status value, and the target farmland is managed according to the management decision.

[0007] Optionally, the acquisition of image data, environmental data and crop physiological indicators of the target farmland includes: Using satellite sensors to collect image data of target farmland, wherein the image data is multispectral image data and hyperspectral image data; Collecting environmental data of the target farmland according to ground sensors, wherein the environmental data includes soil moisture, soil temperature, light intensity and carbon dioxide concentration; Acquire crop physiological indicators of the target farmland, wherein the crop physiological indicators are plant height, leaf area and fruit size.

[0008] Optionally, the performing denoising processing on the environmental data based on a sliding average algorithm to obtain denoised data of the environmental data includes: Performing outlier detection on the environmental data based on the 3σ principle, and eliminating the abnormal data in the environmental data according to the detection result of the outlier detection to obtain standard data of the environmental data; The standard data is smoothed based on a sliding average algorithm to obtain noise reduction data of the environmental data.

[0009] Optionally, generating the farmland image of the target farmland based on the conditional generative adversarial network and the image data includes: The conditional generative adversarial network is trained to obtain a trained conditional generative adversarial network, wherein the algorithm formula of the generator in the conditional generative adversarial network is: in, is the image data, is the condition information corresponding to the image data, Is the given condition information The generated image, is the generator in the conditional generative adversarial network, is the discriminator in the conditional generative adversarial network, It means that the goal of the generator is to minimize its loss function, It means that the goal of the discriminator is to maximize its loss function. is the expected value, Indicates the image data And its corresponding condition information The joint distribution of is the logarithm of the discriminator’s judgment value on the image data and its condition information, is random noise, is random noise The probability distribution of The generator generates images based on given conditional information and random noise. is the logarithm of the generated image that is misclassified as a fake image by the discriminator, is the discriminator’s evaluation of the generated image; Based on the image data and pre-acquired condition information, a farmland image of the target farmland is generated using the trained conditional generative adversarial network.

[0010] Optionally, the extracting the temporal features of the noise reduction data using a pre-built crop evaluation model includes: The temporal features of the denoised data are extracted according to the temporal convolution layer in the pre-constructed crop evaluation model, wherein the convolution algorithm of the temporal convolution layer is: in, is the temporal feature of the denoised data, is the noise reduction data, is a one-dimensional convolution kernel, is the bias term, is the timing step length, is the current time step.

[0011] Optionally, the generating the dependency relationship of the denoised data based on the time feature and the long short-term memory network includes: The time feature is input into a long short-term memory network, and the time feature is processed layer by layer according to an input gate, a forget gate, an output gate, a cell state update layer, and a hidden state update layer of the long short-term memory network to obtain the dependency relationship of the denoised data.

[0012] Optionally, extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model includes: S21, using the spatial convolution layer in the crop evaluation model to perform a convolution operation on the farmland image to obtain a convolution result of the farmland image, wherein the convolution algorithm of the spatial convolution layer is: in, is in spatial position The multi-level spatial features of the farmland image extracted from is the size of the convolution kernel in the width direction, is the size of the convolution kernel in the height direction, is the size of the convolution kernel in the depth direction, represents the number of rows of the multi-level spatial features, is the number of columns of the multi-level spatial features, is the number of channels of the multi-level spatial features, is the three-dimensional convolution kernel, is the bias term, is the farmland image at location The pixel value on Represents the coordinate position in three-dimensional space; S22, performing pooling processing on the convolution result to obtain a dimensionality reduction feature of the farmland image; S23, repeating steps S21-S22 until a plurality of spatial feature maps at different levels are generated according to the dimensionality reduction features; S24. Determine the multi-level spatial features of the farmland image according to the spatial feature map.

[0013] Optionally, the determining the fusion characteristics of the target farmland based on the dependency, the crop physiological index, the multi-level spatial characteristics and the time characteristics includes: Processing the time characteristics by using the dependency relationship to obtain the time dependency characteristics of the target farmland; Based on the time-dependent features, the crop physiological indicators and the multi-level spatial features, an attention mechanism is used to perform feature fusion to obtain fusion features of the target farmland.

[0014] Optionally, the performing model training on the pre-built crop evaluation model according to the fusion feature to obtain a trained crop evaluation model includes: Generating a data set of the pre-built crop evaluation model according to the fusion features, wherein the data set is divided into a training set, a validation set, and a test set; The pre-built crop evaluation model is trained using a preset loss function and the data set, wherein the preset loss function is: in, is the loss value, is the number of samples in the dataset, is the sample identifier, It means The true labels of samples, The model is The predicted value of samples; The pre-constructed crop evaluation model is model verified according to the training result of the model training, and the crop evaluation model that passes the model verification is determined to be the trained crop evaluation model.

[0015] In order to solve the above problems, the present invention also provides a farmland full-process management system based on deep learning, the system comprising: Data acquisition module, used to collect image data, environmental data and crop physiological indicators of target farmland; A data denoising module, used for performing denoising processing on the environmental data based on a sliding average algorithm to obtain denoised data of the environmental data; An image generation module, used for generating a farmland image of the target farmland based on a conditional generative adversarial network and the image data; A data dependency relationship generation module, used to extract the time characteristics of the denoised data using a pre-built crop evaluation model, and generate the dependency relationship of the denoised data based on the time characteristics and a long short-term memory network; A multi-level feature extraction module, used for extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model; A feature fusion module, used for determining the fusion feature of the target farmland based on the dependency, the crop physiological index, the multi-level spatial feature and the time feature; The farmland management module is used to process the fusion features layer by layer using the fully connected layer and the output layer of the crop evaluation model to obtain the crop status value of the target farmland, generate the management decision of the target farmland according to the crop status value, and manage the target farmland according to the management decision.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention uses a crop evaluation model to simultaneously process multimodal data from different sensors, including environmental data that has been denoised and multispectral and hyperspectral farmland image data. For environmental data, its temporal convolution layer can extract temporal features based on a specific algorithm, and can collaborate with a long short-term memory network to generate dependencies and mine the rules of data in time series. For farmland image data, the spatial convolution layer uses a three-dimensional convolution kernel to perform convolution operations in the time and space dimensions. Through multiple convolution and pooling processes, multi-level spatial features such as low-level texture and color features to high-level crop distribution and farmland regional features can be effectively extracted, achieving comprehensive feature capture of complex farmland data. 2. The crop evaluation model of the present invention can not only obtain the static characteristics of the farmland at a certain moment through comprehensive analysis in the two dimensions of time and space, but also understand the dynamic change process of the farmland status in combination with time series information; 3. The present invention uses conditional generative adversarial networks to improve the quality of image data, generate more detailed and usable farmland images, and enhance the diversity and expressiveness of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a process flow of a method for full-process farmland management based on deep learning provided by an embodiment of the present invention; Figure 2 A functional module diagram of a farmland full-process management system based on deep learning provided by an embodiment of the present invention; The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0018] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0019] The embodiment of the present application provides a method for managing the entire process of farmland based on deep learning. The execution subject of the method for managing the entire process of farmland based on deep learning includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided in the embodiment of the present application. In other words, the method for managing the entire process of farmland based on deep learning can be executed by software or hardware installed on a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 FIG. 1 is a flow chart of a method for managing a farmland in a whole process based on deep learning according to an embodiment of the present invention. In this embodiment, the method for managing a farmland in a whole process based on deep learning includes: S1. Collect image data, environmental data and crop physiological indicators of target farmland.

[0021] In the embodiment of the present invention, the acquisition of image data, environmental data and crop physiological indicators of the target farmland includes: Using satellite sensors to collect image data of target farmland, wherein the image data is multispectral image data and hyperspectral image data; Collecting environmental data of the target farmland according to ground sensors, wherein the environmental data includes soil moisture, soil temperature, light intensity and carbon dioxide concentration; Acquire crop physiological indicators of the target farmland, wherein the crop physiological indicators are plant height, leaf area and fruit size.

[0022] In detail, the image data is collected using satellite sensors and includes multispectral and hyperspectral image data. Multispectral images can obtain reflected light information in multiple specific wavelength ranges, while hyperspectral images can provide more continuous and detailed spectral information. Both reflect the surface characteristics of farmland crops and soil from different spectral dimensions, providing basic data for subsequent analysis of crop growth conditions, pest and disease conditions, and soil fertility distribution.

[0023] For example, the multispectral image data of a farmland was collected. In the 400-700nm band, the reflectances at different wavelengths were: 0.2 at 450nm, 0.3 at 550nm, 0.4 at 650nm, etc.; the hyperspectral image data in the 400-1000nm band had a reflectance value every 10nm, such as 0.18 at 410nm, 0.19 at 420nm, etc. These data can reflect the growth status of crops.

[0024] In detail, environmental data are collected by ground sensors, covering soil moisture, soil temperature, light intensity and carbon dioxide concentration. Among them, soil moisture affects crop water absorption and metabolism; soil temperature affects microbial activity and crop root growth; light intensity is a key factor in photosynthesis; carbon dioxide concentration is related to the photosynthesis carbon assimilation process. These data comprehensively reflect the ecological environment status of farmland and are crucial to the assessment of the crop growth environment.

[0025] For example, the environmental data of the farmland collected by ground sensors showed that the soil moisture was 30%, the soil temperature was 25℃, the light intensity was 1000lux, and the carbon dioxide concentration was 400ppm. Soil moisture affects crop water absorption and metabolism. Different crops have different requirements for soil moisture at different growth stages. For example, the suitable soil moisture for wheat during the jointing period is 60% to 70%. This data can be used to determine whether the current humidity is suitable. Soil temperature affects microbial activity and crop root growth. For example, corn roots grow more actively at 20-25℃. Light intensity is a key factor in photosynthesis. Generally, vegetable crops can perform photosynthesis better under a light intensity of 800-1200lux. Carbon dioxide concentration is related to the carbon assimilation process of photosynthesis. The appropriate concentration can ensure the normal photosynthesis of crops.

[0026] In detail, crop physiological indicators include plant height, leaf area and fruit size. Among them, plant height reflects the growth stage and growth potential of crops; leaf area affects the photosynthesis area and efficiency; fruit size reflects the yield-related characteristics of crops. These indicators intuitively reflect the growth and development process and health level of crops, and are the direct basis for evaluating the growth status of crops.

[0027] For example: the plant height of the crop in this farmland is 50cm, the leaf area is 200cm², and the fruit size (assuming the fruit diameter) is 5cm. The plant height reflects the growth stage and growth potential of the crop. For example, the plant height of rice during the tillering stage is generally 30~50cm; the leaf area affects the photosynthesis area and efficiency. The larger the leaf area, the stronger the photosynthesis; the fruit size reflects the characteristics related to crop yield, and the fruit size can evaluate the growth status of the crop to a certain extent.

[0028] S2. Perform noise reduction processing on the environmental data based on a sliding average algorithm to obtain noise-reduced data of the environmental data.

[0029] In the embodiment of the present invention, the noise reduction processing is performed on the environmental data based on the sliding average algorithm to obtain the noise reduction data of the environmental data, including: Performing outlier detection on the environmental data based on the 3σ principle, and eliminating the abnormal data in the environmental data according to the detection result of the outlier detection to obtain standard data of the environmental data; The standard data is smoothed based on a sliding average algorithm to obtain noise reduction data of the environmental data.

[0030] In detail, firstly, the outliers of environmental data are detected based on the 3σ principle, where the 3σ principle is based on the normal distribution characteristics of data. About 99.7% of the data should be within the range of the mean ± 3 times the standard deviation. Data outside this range are regarded as outliers and eliminated to obtain standard data.

[0031] Furthermore, the sliding average algorithm is used to smooth the standard data. By calculating the average value of the data within a certain window to replace the original data points, the data fluctuation noise is reduced, making the environmental data more stable and reliable, which is conducive to the subsequent analysis and mining of data patterns.

[0032] S3. Generate a farmland image of the target farmland based on a conditional generative adversarial network and the image data.

[0033] In an embodiment of the present invention, the generating of the farmland image of the target farmland based on the conditional generative adversarial network and the image data includes: The conditional generative adversarial network is trained to obtain a trained conditional generative adversarial network, wherein the algorithm formula of the generator in the conditional generative adversarial network is: in, is the image data, is the condition information corresponding to the image data, Is the given condition information The generated image, is the generator in the conditional generative adversarial network, is the discriminator in the conditional generative adversarial network, It means that the goal of the generator is to minimize its loss function, It means that the goal of the discriminator is to maximize its loss function. is the expected value, Indicates the image data And its corresponding condition information The joint distribution of is the logarithm of the discriminator’s judgment value on the image data and its condition information, is random noise, is random noise The probability distribution of The generator generates images based on given conditional information and random noise. is the logarithm of the generated image that is misclassified as a fake image by the discriminator, is the discriminator’s evaluation of the generated image; Based on the image data and pre-acquired condition information, a farmland image of the target farmland is generated using the trained conditional generative adversarial network.

[0034] Specifically, when training the conditional generative adversarial network, a large amount of image data of different farmlands and their corresponding condition information (such as crop type, growth stage, weather conditions, etc.) are input. For example, for a piece of farmland image data of wheat in the grain filling period and sunny weather, the trained conditional generative adversarial network can generate clearer and more detailed farmland images based on these conditional information and random noise, highlighting the wheat filling situation, wheat ear fullness and other features, which helps to more accurately evaluate the growth status of crops.

[0035] Specifically, in the training phase of the conditional generative adversarial network, it is necessary to collect a large amount of farmland image data covering different scenes, different crop growth stages and different environmental conditions, and carefully annotate each image with its corresponding condition information. This condition information includes but is not limited to detailed parameters such as crop types (such as wheat, corn, rice, etc.), growth stages (such as sowing period, jointing period, filling period, etc.), weather conditions (sunny, cloudy, rainy, etc.) and soil types.

[0036] In detail, take a large-scale farmland image database as an example, which contains thousands of farmland images from different regions and seasons. For each image, professional agricultural technicians or with the help of advanced image recognition algorithms accurately identify the crop type, and combine the growth stage recorded in the field survey, weather data provided by the local meteorological station, soil test reports and other information to attach complete condition information labels to the image.

[0037] Furthermore, when it is actually applied to the generation of target farmland images, the image data of the target farmland is first obtained, and the condition information corresponding to the farmland image is collected through various channels such as field surveys, sensor data, and historical records. For example, for a farmland located in North China, which is currently in the grain filling period and has sunny weather recently, its image data and these detailed condition information are input into the trained conditional generative adversarial network.

[0038] In detail, the generator will generate a new farmland image based on these conditional information and random noise. During the generation process, the generator will refer to the characteristic patterns of farmland images under different conditions learned during the training process. For the image generation of wheat filling period, it will focus on the fullness of wheat ears, the filling of wheat grains, and the overall color and morphological characteristics of wheat plants. Compared with the original image, the generated image is richer in details. For example, it can more clearly show the texture of wheat grains and subtle changes in stems, thereby providing agricultural experts or farmers with more accurate and intuitive farmland crop growth information, which effectively assists in the subsequent farmland management decision-making, such as the implementation of precision irrigation, reasonable fertilization, and pest and disease control measures.

[0039] S4. Using a pre-built crop evaluation model to extract the time features of the denoised data, and generating a dependency relationship of the denoised data based on the time features and a long short-term memory network.

[0040] In an embodiment of the present invention, before extracting the time characteristics of the noise reduction data using the pre-built crop evaluation model, the method further includes: Building a crop evaluation model to be trained for the target farmland; Generating a data set for the crop evaluation model to be trained, wherein the data set is divided into a training set, a validation set, and a test set; The crop evaluation model to be trained is trained using a preset loss function and the data set, wherein the preset loss function is: in, is the loss value, is the number of samples in the dataset, is the sample identifier, It means The true labels of samples, The model is The predicted value of samples; The crop evaluation model to be trained is model verified according to the training result of the model training, and the crop evaluation model to be trained that passes the model verification is determined to be a pre-constructed crop evaluation model.

[0041] In detail, the network parameters are continuously adjusted during training to minimize the loss value, and the pre-constructed crop evaluation model is verified and determined based on the training results. This model generates target farmland management decisions based on input data, such as irrigation water volume, fertilization timing and amount, pest and disease control measures, etc., to achieve precise and intelligent full-process management of farmland.

[0042] In detail, during the training process, the network parameters are constantly adjusted to ensure that the loss value continues to decrease, so that the model can predict the growth status of crops with high accuracy. For example, for a field planted with wheat, the model can accurately determine whether the growth condition of wheat is good during the filling period, whether there is water shortage, fertilizer shortage or pest and disease threats based on environmental data, image features and crop physiological indicators at different growth stages.

[0043] Furthermore, the data set is divided into a training set, a validation set and a test set, and the division ratio is [specific ratio, such as 70:15:15]. For example, 100 groups of the above collected data are selected and divided into a training set (70 groups), a validation set (15 groups) and a test set (15 groups).

[0044] For example, for a set of data in the training set, including the processed environmental data time features, image multi-level spatial features, etc., such as the time features are [0.3, 0.4, 0.5] (representing the environmental data features at different time points), the spatial features are [0.2, 0.3, 0.4] (image spatial features at a certain level), and the crop physiological indicators are plant height 55cm, leaf area 220cm², and fruit size 5.5cm, these data are combined into a sample.

[0045] In detail, the loss function is determined according to the task type of the model to be trained, wherein the trained crop evaluation model needs to be used for regression tasks, for example: using the crop evaluation model to generate the crop growth status of the target farmland.

[0046] In detail, assuming that during the training process, for a sample, the model predicts a growth status of 0.87, while the actual yield (true label) is 0.91, the loss value is calculated and the network parameters are adjusted according to the loss value. After multiple iterative training, the loss value is continuously reduced, so that the model can more accurately predict the crop growth status.

[0047] Furthermore, during the training process, the validation set is used to monitor the model performance. For example, when the mean square error on the validation set gradually decreases from the initial 100 to 20 and no longer improves, when the performance indicators on the validation set (such as accuracy and mean square error) no longer improve, the early stopping method is used to stop training to prevent overfitting.

[0048] In detail, the trained model is evaluated using the test set to calculate the mean square error (MSE), mean absolute error (MAE), and coefficient of determination; confusion matrices and ROC curves can also be generated simultaneously to more comprehensively evaluate model performance.

[0049] In an embodiment of the present invention, the extracting the temporal features of the noise reduction data using a pre-built crop evaluation model includes: The temporal features of the denoised data are extracted according to the temporal convolution layer in the pre-constructed crop evaluation model, wherein the convolution algorithm of the temporal convolution layer is: in, is the temporal feature of the denoised data, is the noise reduction data, is a one-dimensional convolution kernel, is the bias term, is the timing step length, is the current time step.

[0050] In detail, for the time dimension of the spatiotemporal environmental data and the spatiotemporal image data processed by spatial convolution, a 1D convolution layer is used to extract temporal features, and 1D convolution kernels of different lengths (such as 3 and 5) are set to extract temporal features through the convolution algorithm of the temporal convolution layer.

[0051] In the embodiment of the present invention, the dependency relationship of generating the denoised data based on the time feature and the long short-term memory network includes: The time feature is input into a long short-term memory network, and the time feature is processed layer by layer according to an input gate, a forget gate, an output gate, a cell state update layer, and a hidden state update layer of the long short-term memory network to obtain the dependency relationship of the denoised data.

[0052] In detail, a long short-term memory network (LSTM) is used to process the temporal features to capture the long-term dependencies in the time series, and a gated recurrent unit (GRU) can also be used instead of the long short-term memory network (LSTM).

[0053] In detail, the time features are input into the long short-term memory network (LSTM), and the LSTM is processed through the input gate, forget gate, output gate, cell state update layer and hidden state update layer. Among them, the input gate determines the input information retention ratio based on the product of the current input and the hidden state weight matrix of the previous moment plus the bias term after the activation function; the forget gate controls the degree of retention of the cell state information of the previous moment; and the output gate determines the output of the current cell state information.

[0054] In detail, based on the processed temporal features and LSTM, denoised data dependencies are generated, and then the temporal dependency features of the temporal features are processed using the dependencies, which are then combined with multi-level spatial features through an attention mechanism (self-attention or multi-head attention mechanism). The attention mechanism can automatically focus on the key parts of the data, enhance the weights of important features, and fuse them into fusion features that comprehensively reflect the spatiotemporal correlation between crops and the environment in the target farmland.

[0055] In general, the crop evaluation model is used to efficiently process multimodal data. The crop evaluation model plays a key role in the present invention, and it can simultaneously accept multiple data from different sensors. For environmental data, its time convolution layer accurately extracts time features based on a unique algorithm formula. In this process, by setting one-dimensional convolution kernels of different lengths (such as 3, 5), the changing patterns of environmental data in time series can be captured in detail. Moreover, it works closely with the long short-term memory network, which processes the time features layer by layer in depth by virtue of its input gate, forgetting gate, output gate, cell state update layer, and hidden state update layer, thereby successfully generating the dependency relationship of environmental data. This synergy enables the present invention to deeply explore the potential laws of environmental data in the time dimension, providing a solid data foundation for a comprehensive understanding of the dynamic changes in the farmland ecological environment.

[0056] S5. Extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model.

[0057] In an embodiment of the present invention, the step of extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model includes: S21, using the spatial convolution layer in the crop evaluation model to perform a convolution operation on the farmland image to obtain a convolution result of the farmland image; S22, performing pooling processing on the convolution result to obtain a dimensionality reduction feature of the farmland image; S23, repeating steps S21-S22 until a plurality of spatial feature maps at different levels are generated according to the dimensionality reduction features; S24. Determine the multi-level spatial features of the farmland image according to the spatial feature map.

[0058] In detail, the convolution result is pooled, using maximum pooling or average pooling, and the pooling kernel size is set to (2x2x2) to reduce the size of the feature map, enhance the robustness of the feature, and reduce the amount of calculation; In detail, the above convolution and pooling operations are repeated to form multiple spatial feature maps at different levels, from low-level texture and color features to high-level crop distribution and farmland area features.

[0059] In detail, the convolution algorithm of the spatial convolution layer is: in, is in spatial position The multi-level spatial features of the farmland image extracted from is the size of the convolution kernel in the width direction, is the size of the convolution kernel in the height direction, is the size of the convolution kernel in the depth direction, represents the number of rows of the multi-level spatial features, is the number of columns of the multi-level spatial features, is the number of channels of the multi-level spatial features, is the three-dimensional convolution kernel, is the bias term, is the farmland image at location The pixel value on Represents the coordinate position in three-dimensional space.

[0060] In detail, for the spatiotemporal image data, it is regarded as three-dimensional data (time, width, height), and is processed using a 3D convolution layer. Multiple 3D convolution kernels of different sizes (such as 3x3x3, 5x5x5) are used to perform convolution operations on the image data in the spatiotemporal dimension to extract spatial features.

[0061] In detail, taking the above-mentioned farmland image collected as an example, the spatial convolution layer in the crop evaluation model performs convolution operation on the image. Using a 3x3x3 three-dimensional convolution kernel, at the spatial position (10, 10, 1), the multi-level spatial features are calculated by the convolution algorithm. Assuming that the weights of the convolution kernel in the width, height, and depth directions are [0.1, 0.2, 0.3], [0.2, 0.3, 0.1], [0.3, 0.1, 0.2], respectively, and the bias term is 0.1, the pixel value at this position is [0.5, 0.6, 0.7], and the multi-level spatial feature value of this position is calculated. Through multiple convolution and pooling operations (such as using a 2x2x2 pooling kernel for pooling), multiple spatial feature maps of different levels are generated, from low-level texture and color features to high-level crop distribution and farmland area features, providing rich image information for subsequent analysis.

[0062] In general, for multispectral and hyperspectral farmland image data, the spatial convolution layer of the crop evaluation model uses a three-dimensional convolution kernel to perform convolution operations in the spatiotemporal dimension. Taking a certain farmland image collected as an example, at the spatial position (10, 10, 1), through the carefully designed convolution kernel weights and bias terms, combined with the pixel value at that position, the multi-level spatial features can be accurately calculated. After multiple such operations, a full range of multi-level spatial features, from low-level texture and color features to high-level crop distribution, farmland area features, etc., can be efficiently extracted. This process greatly enriches the analysis dimension of farmland image data, realizes the comprehensive and detailed feature capture of complex farmland data, and provides rich and accurate information support for subsequent precise management decisions.

[0063] S6. Determine the fusion characteristics of the target farmland based on the dependency relationship, the crop physiological index, the multi-level spatial characteristics and the time characteristics.

[0064] In the embodiment of the present invention, the step of determining the fusion feature of the target farmland based on the dependency relationship, the crop physiological index, the multi-level spatial feature and the time feature includes: Processing the time characteristics by using the dependency relationship to obtain the time dependency characteristics of the target farmland; Based on the time-dependent features, the crop physiological indicators and the multi-level spatial features, an attention mechanism is used to perform feature fusion to obtain fusion features of the target farmland.

[0065] In detail, the attention mechanism is a self-attention mechanism.

[0066] In detail, the spatial features and temporal features are fused and the attention mechanism is used. First, the similarity scores of spatial features and temporal features are calculated, and then the dot product operation is performed. , where Q is the query matrix (can be spatial features or temporal features), K is the key matrix, is the feature dimension; then the attention weight is obtained through the softmax function; finally, the attention weight is multiplied and added with the feature to obtain the fusion feature ,in is a value matrix (can be spatial features or temporal features).

[0067] In general, the crop evaluation model has powerful functions through in-depth comprehensive analysis in both time and space. It can not only accurately obtain the static characteristics of farmland at a specific moment, such as clear imaging and feature extraction of the growth form of farmland crops and soil surface conditions at a certain point in time, but also closely combine time series information to fully present the dynamic changes in farmland status. Through long-term environmental data monitoring and image comparison, it can clearly show the subtle changes in crop growth, the dynamic evolution of soil fertility, and the occurrence and development trends of pests and diseases.

[0068] S7. Use the fully connected layer and output layer of the crop evaluation model to process the fused features layer by layer to obtain the crop status value of the target farmland, generate a management decision for the target farmland according to the crop status value, and manage the target farmland according to the management decision.

[0069] In the embodiment of the present invention, the fusion features are processed layer by layer using the fully connected layer and the output layer of the crop evaluation model, starting from the input features (fusion features), and gradually extracting more abstract and high-level features through the processing of each layer of the neural network until the output is finally generated. The specific process is as follows: the fusion features are input into the fully connected layer, and the high-level features are extracted through the weighted and activated functions, and then the high-level features are mapped to the specific state values ​​of the crop through the output layer.

[0070] In detail, the fully connected layer is one of the most common layers in the neural network, responsible for processing the input features and transmitting information in the neural network. Its main function is to connect the output features of the previous layer with all neurons to generate a higher-level feature representation.

[0071] In detail, the output layer is the last layer of the neural network, which is responsible for generating the final result based on the information processed by the fully connected layer. In the crop evaluation model, the output layer converts the high-dimensional feature representation of the fully connected layer into a specific prediction result.

[0072] In an embodiment of the present invention, a crop evaluation model is deployed on a cloud server or an edge computing device (such as a smart sensor or an unmanned aerial vehicle airborne device) to collect spatiotemporal data in real time, including satellite images and ground sensor data, and input them into a trained model to output the crop status value of the target farmland.

[0073] For example, taking a wheat field as an example, the multispectral image data collected by the satellite sensor has a reflectivity of 0.3 at 550nm, and the hyperspectral image data has a reflectivity of 0.19 at 420nm. The environmental data collected by the ground sensor shows that the soil moisture is 28%, the soil temperature is 23℃, the light intensity is 900lux, and the carbon dioxide concentration is 380ppm. The wheat plant height in the field is 45cm, the leaf area is 180cm², and the fruit (wheat ear) size is 4cm in diameter.

[0074] Furthermore, for the above environmental data, after 3σ principle detection and sliding average algorithm processing, it is assumed that the time characteristics of the denoised data are [0.25, 0.32, 0.38] (representing the characteristics of the environmental data after processing at different time points); after processing the farmland image, the spatial characteristics of a certain level are [0.18, 0.25, 0.32]. The fusion characteristics obtained by further fusion and other operations are vectors containing comprehensive information of the environment and image, such as [0.22, 0.30, 0.35] (here is only an example). The model outputs the crop growth status value of 0.77 for the wheat farmland, corresponding to the crop growth status that the growth can be intervened and there are slight signs of water shortage.

[0075] In detail, based on the model output, the management decision is that the recommended irrigation water volume is 40m³ / mu, the timing of fertilization is 10 days later (when the wheat is in the early stage of filling, according to the growth stage and soil fertility analysis), the fertilizer dosage is 8kg / mu of nitrogen fertilizer, 4kg / mu of phosphorus fertilizer, and 6kg / mu of potassium fertilizer, and the pest and disease control measures are that no special treatment is required for the time being, but continuous monitoring is required (because the overall health of the crop is good).

[0076] In general, with the help of collected crop physiological indicators (plant height, leaf area, fruit size), processed image data and environmental data characteristics, we can comprehensively analyze the health of crops. Based on the evaluation results, we can accurately make management decisions such as irrigation, fertilization, pest and disease control to ensure good growth of crops.

[0077] like Figure 2 As shown, it is a functional module diagram of a farmland full-process management system based on deep learning provided by one embodiment of the present invention.

[0078] The farmland full-process management system 100 based on deep learning of the present invention can be installed in an electronic device. According to the functions implemented, the farmland full-process management system 100 based on deep learning can include a data acquisition module 101, a data noise reduction module 102, an image generation module 103, a data dependency generation module 104, a multi-level feature extraction module 105, a feature fusion module 106 and a farmland management module 107. The module of the present invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and can complete fixed functions, which are stored in the memory of the electronic device.

[0079] In this embodiment, the functions of each module / unit are as follows: The data acquisition module 101 is used to collect image data, environmental data and crop physiological indicators of the target farmland; The data denoising module 102 is used to perform denoising on the environmental data based on a sliding average algorithm to obtain denoised data of the environmental data; The image generation module 103 is used to generate a farmland image of the target farmland based on a conditional generative adversarial network and the image data; The data dependency relationship generating module 104 is used to extract the time characteristics of the denoised data using a pre-built crop evaluation model, and generate the dependency relationship of the denoised data based on the time characteristics and the long short-term memory network; The multi-level feature extraction module 105 is used to extract the multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model; The feature fusion module 106 is used to determine the fusion feature of the target farmland based on the dependency relationship, the crop physiological index, the multi-level spatial feature and the time feature; The farmland management module 107 is used to use the fully connected layer and the output layer of the crop evaluation model to process the fused features layer by layer to obtain the crop status value of the target farmland, generate the management decision of the target farmland according to the crop status value, and manage the target farmland according to the management decision.

[0080] In the several embodiments provided by the present invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are only illustrative, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0081] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0082] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0083] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0084] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A full-process farmland management method based on deep learning, characterized in that: The method comprises: Collect image data, environmental data and crop physiological indicators of target farmland; Performing noise reduction processing on the environmental data based on a sliding average algorithm to obtain noise-reduced data of the environmental data; Generate a farmland image of the target farmland based on a conditional generative adversarial network and the image data; Extracting the time features of the denoised data using a pre-built crop evaluation model, and generating a dependency relationship of the denoised data based on the time features and a long short-term memory network; Extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model; Determining the fusion characteristics of the target farmland based on the dependency relationship, the crop physiological index, the multi-level spatial characteristics and the temporal characteristics; The fusion features are processed layer by layer using the fully connected layer and the output layer of the crop evaluation model to obtain the crop status value of the target farmland, a management decision of the target farmland is generated according to the crop status value, and the target farmland is managed according to the management decision.

2. The method for managing farmland in the whole process based on deep learning according to claim 1, characterized in that: The image data, environmental data and crop physiological indicators of the target farmland are collected, including: Using satellite sensors to collect image data of target farmland, wherein the image data is multispectral image data and hyperspectral image data; Collecting environmental data of the target farmland according to ground sensors, wherein the environmental data includes soil moisture, soil temperature, light intensity and carbon dioxide concentration; Acquire crop physiological indicators of the target farmland, wherein the crop physiological indicators are plant height, leaf area and fruit size.

3. The farmland full-process management method based on deep learning according to claim 1, characterized in that: The performing denoising processing on the environmental data based on a sliding average algorithm to obtain denoised data of the environmental data includes: based on According to the principle, outlier detection is performed on the environmental data, and the abnormal data in the environmental data is eliminated according to the detection result of the outlier detection to obtain standard data of the environmental data; The standard data is smoothed based on a sliding average algorithm to obtain noise reduction data of the environmental data.

4. The method for managing farmland in its entirety based on deep learning according to claim 1, characterized in that: The method of generating a farmland image of the target farmland based on a conditional generative adversarial network and the image data includes: The conditional generative adversarial network is trained to obtain a trained conditional generative adversarial network, wherein the algorithm formula of the generator in the conditional generative adversarial network is: in, is the image data, is the condition information corresponding to the image data, Is the given condition information The generated image, is the generator in the conditional generative adversarial network, is the discriminator in the conditional generative adversarial network, It means that the goal of the generator is to minimize its loss function, It means that the goal of the discriminator is to maximize its loss function. is the expected value, Indicates the image data And its corresponding condition information The joint distribution of is the logarithm of the discriminator’s judgment value on the image data and its condition information, is random noise, is random noise The probability distribution of The generator generates images based on given conditional information and random noise. is the logarithm of the generated image that is misclassified as a fake image by the discriminator, is the discriminator’s evaluation of the generated image; Based on the image data and pre-acquired condition information, a farmland image of the target farmland is generated using the trained conditional generative adversarial network.

5. The method for managing farmland in the whole process based on deep learning according to claim 1, characterized in that: The extracting the temporal characteristics of the noise reduction data by using the pre-built crop evaluation model includes: The temporal features of the denoised data are extracted according to the temporal convolution layer in the pre-constructed crop evaluation model, wherein the convolution algorithm of the temporal convolution layer is: in, is the temporal feature of the denoised data, is the noise reduction data, is a one-dimensional convolution kernel, is the bias term, is the timing step length, is the current time step.

6. The method for managing farmland in the whole process based on deep learning according to claim 1, characterized in that: The generating the dependency relationship of the noise reduction data based on the time feature and the long short-term memory network includes: The time feature is input into a long short-term memory network, and the time feature is processed layer by layer according to an input gate, a forget gate, an output gate, a cell state update layer, and a hidden state update layer of the long short-term memory network to obtain the dependency relationship of the denoised data.

7. The method for managing farmland in the whole process based on deep learning according to claim 1, characterized in that: The step of extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model includes: S21, using the spatial convolution layer in the crop evaluation model to perform a convolution operation on the farmland image to obtain a convolution result of the farmland image, wherein the convolution algorithm of the spatial convolution layer is: in, is in spatial position The multi-level spatial features of the farmland image extracted from is the size of the convolution kernel in the width direction, is the size of the convolution kernel in the height direction, is the size of the convolution kernel in the depth direction, represents the number of rows of the multi-level spatial features, is the number of columns of the multi-level spatial features, is the number of channels of the multi-level spatial features, is the three-dimensional convolution kernel, is the bias term, is the farmland image at location The pixel value on Represents the coordinate position in three-dimensional space; S22, performing pooling processing on the convolution result to obtain a dimensionality reduction feature of the farmland image; S23, repeating steps S21-S22 until a plurality of spatial feature maps at different levels are generated according to the dimensionality reduction features; S24. Determine the multi-level spatial features of the farmland image according to the spatial feature map.

8. The method for managing farmland in the whole process based on deep learning according to claim 1, characterized in that: The step of determining the fusion characteristics of the target farmland based on the dependency relationship, the crop physiological index, the multi-level spatial characteristics and the time characteristics includes: Processing the time characteristics by using the dependency relationship to obtain the time dependency characteristics of the target farmland; Based on the time-dependent features, the crop physiological indicators and the multi-level spatial features, an attention mechanism is used to perform feature fusion to obtain fusion features of the target farmland.

9. The method for managing farmland in its entirety based on deep learning according to any one of claims 1 to 8, characterized in that: Before extracting the time characteristics of the noise reduction data using the pre-built crop evaluation model, the method further includes: Building a crop evaluation model to be trained for the target farmland; Generating a data set for the crop evaluation model to be trained, wherein the data set is divided into a training set, a validation set, and a test set; The crop evaluation model to be trained is trained using a preset loss function and the data set, wherein the preset loss function is: in, is the loss value, is the number of samples in the dataset, is the sample identifier, It means The true labels of samples, The model is The predicted value of samples; The crop evaluation model to be trained is model verified according to the training result of the model training, and the crop evaluation model to be trained that passes the model verification is determined to be a pre-constructed crop evaluation model.

10. A farmland full-process management system based on deep learning, characterized in that: The system comprises: Data acquisition module, used to collect image data, environmental data and crop physiological indicators of target farmland; A data denoising module, used for performing denoising processing on the environmental data based on a sliding average algorithm to obtain denoised data of the environmental data; An image generation module, used for generating a farmland image of the target farmland based on a conditional generative adversarial network and the image data; A data dependency relationship generation module, used to extract the time characteristics of the denoised data using a pre-built crop evaluation model, and generate the dependency relationship of the denoised data based on the time characteristics and a long short-term memory network; A multi-level feature extraction module, used for extracting multi-level spatial features of the farmland image according to the spatial convolution layer in the crop evaluation model; A feature fusion module, used for determining the fusion feature of the target farmland based on the dependency, the crop physiological index, the multi-level spatial feature and the time feature; The farmland management module is used to process the fusion features layer by layer using the fully connected layer and the output layer of the crop evaluation model to obtain the crop status value of the target farmland, generate the management decision of the target farmland according to the crop status value, and manage the target farmland according to the management decision.

Images