Intelligent agricultural fusion system based on deep learning

Abstract

The application discloses a kind of intelligence agricultural fusion systems based on deep learning, the present application relates to intelligent agricultural management technical field, including: data acquisition module monitors the canopy spectral reflectance of rice plant and related environmental factors, calculates canopy chlorophyll index;Photosynthetic response module extracts photosynthetic response sensitive coefficient by linear regression;State monitoring module extracts the data of stem flow rate and soil temperature, carries out cross correlation analysis to obtain transport lag phase;Repulsion state analysis module determines whether state vector satisfies critical repulsion constraint;Agricultural distribution module calls deep learning model according to the judgment result to output actual nitrogen amount or detect pest type;The system realizes the intelligent monitoring and regulation of rice growth state, improves the yield and quality of rice, also promotes the rational use of resources, reduces agricultural production cost.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural management technology, specifically to a smart agricultural fusion system based on deep learning. Background Technology

[0002] Smart agriculture is an important direction for the development of modern agriculture. Its goal is to improve the yield and quality of crops by using advanced technologies such as the Internet of Things, big data, and artificial intelligence to achieve precise monitoring and control of the crop growth environment. As an important food crop, rice is affected by a variety of environmental factors during its growth process, including photosynthetically active radiation, stem flow rate, and soil temperature. These factors have a significant impact on rice's photosynthesis, nutrient absorption efficiency, and growth and development process.

[0003] Currently, existing technologies often monitor rice growth status using single indicators or simple models. For example, chlorophyll index is calculated using canopy spectral reflectance to assess photosynthetic efficiency; or soil environment is monitored based on soil sensor data. However, these methods have limitations and cannot comprehensively capture the complex dynamic relationships between multiple factors during rice growth. Specifically, monitoring a single indicator is difficult to accurately reflect the changes in rice status at different growth stages, especially the rapid response to environmental regulation needs during critical growth windows. In addition, most existing methods are based on static analysis and lack comprehensive modeling of time series data, making it difficult to predict future growth trends of rice or determine the specific impact of environmental conditions on plant growth. The shortcomings of existing technologies lie in their inability to achieve precise management of key growth stages of rice, especially during the transitional periods between different growth stages. Nutrient requirements during this period are significantly time-sensitive, and management measures that are too early or too late may lead to nutrient waste or hinder crop growth. In addition, existing methods lack the ability to deeply mine multidimensional time series data, making it difficult to predict and determine in real time whether the current state of the plant falls within the critical rejection window, and even more so, unable to provide accurate nitrogen application rates or pest and disease control recommendations based on specific circumstances.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a smart agriculture fusion system based on deep learning to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based smart agriculture integration system specifically includes: The data acquisition module is used to set several rice plants at different growth stages as monitoring samples, continuously collect the time series of canopy spectral reflectance, photosynthetically active radiation, stem flow rate and soil temperature at each sampling time for each monitoring sample, and calculate the canopy chlorophyll index based on the spectral reflectance. The photosynthetic response module is used to construct a sliding time window for any monitoring sample, with the current day as the endpoint. It performs linear regression on the average change of the canopy chlorophyll index and the average change of photosynthetically active radiation within the sliding time window, and extracts the slope coefficient of the regression as the photosynthetic response sensitivity coefficient for the current day. The state monitoring module is used to extract the stem flow rate and soil temperature at each sampling time within the day, perform cross-correlation analysis on the two, solve for the time shift that makes the cross-correlation function reach its maximum value, use the time shift as the transport lag phase of the day, and combine the transport lag phase of the day with the photosynthetic response sensitivity coefficient to form a state vector. The exclusion state analysis module is used to record the growth stage of each monitored sample when the state vector satisfies the critical exclusion constraint, and form a critical exclusion window. It calculates the state vector of the rice plant under monitoring within the current sliding time window, determines whether the state vector satisfies the critical exclusion constraint, and whether the growth stage of the rice plant under monitoring falls into the critical exclusion window. The agricultural allocation module is used to execute corresponding control schemes based on the judgment results. If the judgment is correct, the first deep learning model is called to output the actual amount of nitrogen applied; otherwise, the second deep learning model is called to detect the types of pests and diseases of the rice plants to be monitored.

[0007] Furthermore, for any monitoring sample, the specific calculation method of the canopy chlorophyll index is as follows: for any sampling time, obtain the near-infrared band reflectance and red edge band reflectance of the rice canopy at that sampling time, and use the difference between the ratio of the near-infrared band reflectance and the red edge band reflectance and the constant 1 as the canopy chlorophyll index at that sampling time. The same laboratory environment was set up for each monitoring sample, and data was collected from the monitoring samples in the laboratory environment, which was set up based on the rice growing environment.

[0008] Furthermore, the specific method for constructing the sliding time window is as follows: taking the current day as the end point and tracing back a fixed number of days to form a sliding time window, for any monitoring sample, the average change in the canopy chlorophyll index and the average change in photosynthetically active radiation are extracted daily within the sliding time window. For any given monitoring sample, the average change in canopy chlorophyll index and the average change in photosynthetically active radiation within a sliding time window are determined using the following method: Except for the first day of the sliding time window, for any day within the sliding time window, the canopy chlorophyll index at each sampling time on that day is calculated, and the average value of the canopy chlorophyll index at each sampling time is used as the canopy chlorophyll index for that day. The canopy chlorophyll index of the previous day within the sliding time window is subtracted from the canopy chlorophyll index of that day to obtain the change in the canopy chlorophyll index for that day. Using the same method, the change in photosynthetically active radiation for that day was calculated. Except for the first day of the sliding time window, based on the changes in photosynthetically active radiation and canopy chlorophyll index for each day, the average change in canopy chlorophyll index and the average change in photosynthetically active radiation for the monitoring sample within the sliding time window were calculated.

[0009] Furthermore, the specific method for determining the photosynthetic response sensitivity coefficient on a given day is as follows: the average change in the two canopy chlorophyll indices and the average change in photosynthetically active radiation on the same day are combined into a data point pair. Least square linear regression is performed on all data point pairs within the sliding time window, and the slope obtained from the regression is used as the photosynthetic response sensitivity coefficient.

[0010] Furthermore, for any monitoring sample, the specific method for calculating its transport lag phase is as follows: taking multiple sampling times within the day as a reference, extract the measured values ​​of stem flow rate and soil temperature at each sampling time, calculate the deviation sequences of the stem flow rate sequence and soil temperature sequence relative to their mean values, perform a sliding inner product operation of the soil temperature deviation sequence relative to the stem flow rate deviation sequence within a preset lag time range, obtain the time shift corresponding to the maximum inner product value, and use this time shift as the transport lag phase of the monitoring sample for that day; By mapping the transport lag phase of each day to the photosynthetic response sensitivity coefficient, the state vector of the monitored sample at the corresponding time is obtained. This process is repeated for all monitored samples, and their state vectors are continuously monitored.

[0011] Furthermore, for any monitoring sample, if any state vector of the monitoring sample simultaneously satisfies the following conditions, then the monitoring sample is considered to satisfy the critical repulsion constraint in the current growth stage. Condition 1: In the current state vector of the monitored sample, the photosynthetic response sensitivity coefficient is not greater than the preset threshold, and the rate of change of the photosynthetic response sensitivity coefficient is negative; Condition 2: In the current state vector of the monitored sample, the transmission lag phase is not less than the preset lag threshold; When the critical exclusion constraint is met, the current growth stage of the monitored sample is recorded, and the growth stage is specifically characterized by the number of planting days. Traverse all monitoring samples to determine the planting days corresponding to all those that meet the critical exclusion constraint. For all planting days, perform clustering without specifying clusters based on the absolute difference between the planting days. Based on the clustering results, form several clusters of planting days. Extract the number of planting days in each cluster, remove clusters whose number of planting days is not greater than a preset size threshold, and record the remaining clusters as the critical exclusion window.

[0012] Furthermore, it is determined whether the state vector of the rice plant under monitoring within the current sliding time window satisfies the critical repulsion constraint, and whether the growth stage of the rice plant under monitoring falls within the critical repulsion window. The specific logic is as follows: If the current state vector of the rice plant to be monitored does not satisfy the critical repulsion constraint, then the amount of nitrogen to be applied to the rice plant to be monitored is determined based on the classical nitrogen application amount for different growth stages. If the current state vector of the rice plant to be monitored satisfies the critical repulsion constraint, and the growth stage of the rice plant to be monitored falls within the critical repulsion window, then the first deep learning model is called to output the actual amount of nitrogen applied; otherwise, the second deep learning model is called to detect the types of diseases and pests affecting the rice plant to be monitored.

[0013] Furthermore, the first deep learning model is constructed based on the LSTM model. The input of the first deep learning model is the growth stage of the rice plant to be monitored, the nitrogen content of the plant, the nitrogen content of the soil, and the current state vector. The output of the model is the current actual amount of nitrogen applied to the rice plant to be monitored. The second deep learning model is built based on a convolutional neural network, which consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The activation function in the convolutional layers is... For the fully connected layer, the function sets the number of neurons in the fully connected layer to 32, the initial neural network learning rate to 0.001, and the number of training rounds to 100. The input to the second deep learning model is the canopy and leaf images of the rice plant to be monitored, and the output is the type of pests and diseases of the rice plant to be monitored.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention establishes a comprehensive, multi-dimensional monitoring platform by setting rice plants at different growth stages as monitoring samples and continuously collecting data such as canopy spectral reflectance, photosynthetically active radiation, stem flow rate, and soil temperature. This platform not only provides rich physiological characteristic information but also provides valuable training data for subsequent deep learning models, enabling the models to more accurately understand the performance of rice under different growth conditions.

[0015] This invention also employs a sliding time window method and linear regression analysis, making the extraction of photosynthetic response sensitivity coefficients more flexible and dynamic. By analyzing the relationship between daily canopy chlorophyll index and photosynthetically active radiation, the sensitivity of rice to environmental changes can be accurately identified. This feature enables agricultural managers to develop targeted fertilization and irrigation plans based on real-time data, maximizing the photosynthetic efficiency and growth rate of rice.

[0016] Meanwhile, by combining critical exclusion constraints with the calculation of state vectors, the system establishes an adaptive control mechanism for different growth stages. When the state vector of the monitored sample satisfies the critical exclusion constraint, the system can quickly identify the growth stage of rice and determine the state of the rice plant to be monitored. By calling different deep learning models, the system can not only accurately output the actual amount of nitrogen applied, but also detect the types of pests and diseases in rice plants in a timely manner, effectively realizing precision fertilization and pest and disease control. This not only improves the yield and quality of rice, but also promotes the rational use of resources and reduces agricultural production costs. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Example: Please see Figure 1 The present invention provides a technical solution: A deep learning-based smart agriculture integration system specifically includes: The data acquisition module is used to set several rice plants at different growth stages as monitoring samples, continuously collect the canopy spectral reflectance and the time series of photosynthetically active radiation, stem flow rate and soil temperature at each sampling time for each monitoring sample, and calculate the canopy chlorophyll index based on the spectral reflectance.

[0021] The different growth stages specifically include the germination stage, seedling stage, jointing stage, heading stage, grain-filling stage, and maturity stage; The canopy spectral reflectance is obtained using a spectrometer or portable spectral sensor. The reflectance spectrum of vegetation can be measured within a specific wavelength range. A spectrometer suitable for agricultural applications is selected, typically operating in the 400-2500 nm band to cover the visible and near-infrared regions. This effectively captures spectral information related to chlorophyll content. The ASD FieldSpec4 portable spectrometer is commonly used, as it can perform high-precision measurements in the 350-2500 nm range, making it suitable for agricultural research. Photosynthetically active radiation (PAR) can be obtained using a quantum optical sensor or a PAR sensor. A dedicated PAR sensor can be selected to measure light intensity in the wavelength range of 400-700 nanometers. The sensor is placed above the canopy of the rice plant and light intensity data is continuously recorded. The sampling frequency is set to generate time-series data. Specifically, the LI-COR LI-250A PAR sensor can be selected for measuring PAR, which is suitable for agricultural and ecological research. Stem flow rate sensors are used to acquire the stem flow rate of each monitoring sample. A flow sensor based on the thermal pulse method is selected to measure the water flow rate in the plant stem. The sensor is installed on the stem of the rice plant to record changes in water flow. Periodic sampling can be set to record time-series data. Specifically, the Dynamax TDP30 thermal pulse stem flow sensor can be selected, which can accurately measure the water flow rate in the plant stem. Soil temperature is obtained using a soil temperature sensor. By selecting a soil temperature probe, soil temperature can be measured at a specified depth, such as 5cm or 10cm. The temperature probe is buried in the soil and connected to a data logger. The sampling frequency is set to obtain time-series data. The Decagon Devices EM50 can connect to multiple soil temperature sensors, making it suitable for soil environmental monitoring.

[0022] For any given monitoring sample, the specific calculation method for the canopy chlorophyll index is as follows: For any sampling time, obtain the near-infrared band reflectance and red-edge band reflectance of the rice canopy at that sampling time. The difference between the ratio of the near-infrared band reflectance to the red-edge band reflectance and the constant 1 is taken as the canopy chlorophyll index at that sampling time; the specific formula used is: In the formula, Let be the canopy chlorophyll index of the i-th monitoring sample at the j-th sampling time. Let be the near-infrared reflectance of the i-th monitoring sample at the j-th sampling time. Let i be the red-edge band reflectance of the i-th monitoring sample at the j-th sampling time, where i is the index of the monitoring sample and j is the index of the sampling time. It should be noted that rice plants typically have high reflectivity in the near-infrared band because plant cell structures have a strong scattering effect on near-infrared light. Therefore, near-infrared reflectivity can effectively reflect the plant's growth status and water content. Its reflectivity is mainly determined by the intercellular structure of the spongy tissue in the leaves and is not sensitive to changes in chlorophyll content. The red-edge band, located between red and near-infrared light, is generally considered a key band for measuring chlorophyll content in plants. Chlorophyll's absorption characteristics result in relatively low reflectance in the red light band, while absorption in the red-edge band increases with chlorophyll content. Reflectance is strongly regulated by chlorophyll absorption; higher chlorophyll content leads to stronger absorption and lower reflectance in this band. By calculating the ratio of near-infrared reflectance to red-edge reflectance, the influence of factors such as light conditions and vegetation background can be eliminated, thus more accurately reflecting the physiological state of the plant. Subtracting 1 from the formula primarily aims to make the calculation results theoretically reflect more intuitive changes. The larger the value, the greener the rice canopy and the denser the canopy structure.

[0023] The same laboratory environment is set up for each monitoring sample, and data is collected from the monitoring samples in the laboratory environment. The laboratory environment is set up based on the rice growth environment. The laboratory environment generally includes temperature, humidity and light conditions.

[0024] The photosynthetic response module is used to construct a sliding time window for any monitoring sample, with the current day as the endpoint. It performs linear regression on the average change of the canopy chlorophyll index and the average change of photosynthetically active radiation within the sliding time window, and extracts the slope coefficient of the regression as the photosynthetic response sensitivity coefficient for the current day.

[0025] The specific method for constructing the sliding time window is as follows: taking the current day as the end point and tracing back a fixed number of days to form a sliding time window, for any monitoring sample, the average change in the canopy chlorophyll index and the average change in photosynthetically active radiation are extracted daily within the sliding time window; the sliding time window generally includes 3-5 days; For any given monitoring sample, the average change in canopy chlorophyll index and the average change in photosynthetically active radiation within a sliding time window are determined using the following method: Except for the first day of the sliding time window, for any day within the sliding time window, the canopy chlorophyll index at each sampling time on that day is calculated, and the average value of the canopy chlorophyll index at each sampling time is used as the canopy chlorophyll index for that day. The canopy chlorophyll index of the previous day within the sliding time window is subtracted from the canopy chlorophyll index of that day to obtain the change in the canopy chlorophyll index for that day. Using the same method, the change in photosynthetically active radiation for that day was calculated. Except for the first day of the sliding time window, based on the changes in photosynthetically active radiation and canopy chlorophyll index for each day, the average change in canopy chlorophyll index and the average change in photosynthetically active radiation for the monitoring sample within the sliding time window were calculated. It should be noted that the sliding time window allows for real-time monitoring of plant responses to changes in photosynthetically active radiation, and can promptly reflect the dynamic process of how environmental changes affect plant physiological characteristics. By calculating the average change, it can effectively reduce data fluctuations caused by single sampling errors or occasional environmental changes, thereby enhancing data reliability.

[0026] The method used to determine the photosynthetic response sensitivity coefficient for a given day is as follows: A data point pair is formed by combining the average change in the chlorophyll index of two canopies on the same day with the average change in photosynthetically active radiation. Least squares linear regression is performed on all data point pairs within the sliding time window, and the slope obtained from the regression is used as the photosynthetic response sensitivity coefficient. The specific formula used to calculate the photosynthetic response sensitivity coefficient is as follows: In the formula, Let be the photosynthetic response sensitivity coefficient of the i-th monitored sample. Let be the change in photosynthetically active radiation of the i-th monitoring sample on day t within the sliding time window. Let be the change in canopy chlorophyll index of the i-th monitoring sample on day t within the sliding time window. This is the index of the number of days within the sliding time window. This represents the total number of days within the sliding time window. Let be the average change in photosynthetically active radiation for the i-th monitoring sample within the sliding time window. denoted as the average change in canopy chlorophyll index for the i-th monitoring sample within the sliding time window; It should be noted that the sliding time window allows for the analysis of dynamic changes over a short period of time; the plant's response to photosynthetically active radiation is not instantaneous, but has a certain cumulative effect over time; therefore, selecting a dynamic window can better capture changes in the plant's state. The linear regression slope, also known as the photosynthetic response sensitivity coefficient, quantifies the linear relationship between changes in photosynthetically active radiation and changes in canopy chlorophyll index. By analyzing the magnitude of the slope, we can understand the plant's sensitivity to changes in light intensity. The larger the value, the more significant the response of the canopy chlorophyll index to changes in photosynthetically active radiation; The formula is essentially the calculation of the slope of the least squares linear regression, which is a classic method for studying the linear relationship between two variables. The regression slope can be used to quantitatively describe the intensity of the influence of changes in photosynthetically active radiation on changes in the canopy chlorophyll index. During the normal vegetative growth stage, the value is positive and relatively large, indicating that the stronger the light, the faster the chlorophyll synthesis. When rice enters the critical rejection window, the marginal efficiency of chlorophyll synthesis in leaves drops sharply because photosynthetic products are preferentially transported to the young panicles. It will quickly approach zero or even turn negative.

[0027] The state monitoring module is used to extract the stem flow rate and soil temperature at each sampling time within the day, perform cross-correlation analysis on the two, solve for the time shift that makes the cross-correlation function reach its maximum value, use the time shift as the transport lag phase of the day, and combine the transport lag phase of the day with the photosynthetic response sensitivity coefficient to form a state vector.

[0028] The specific method for calculating the transport lag phase of any monitoring sample is as follows: Using multiple sampling times within the day as a reference, the measured values ​​of stem flow rate and soil temperature at each sampling time are extracted. The deviation sequences of the stem flow rate sequence and soil temperature sequence relative to their respective means are calculated. The soil temperature deviation sequence is then subjected to a sliding inner product operation with respect to the stem flow rate deviation sequence within a preset lag time range, and the time shift corresponding to the maximum inner product value is obtained. This time shift is used as the transport lag phase of the monitoring sample for that day. The specific formula for determining the time shift is as follows: In the formula, The inner product value corresponding to the time shift d. This represents the average stem flow rate for the day. This represents the average soil temperature for the day. Let m be the stem flow rate at the m-th sampling time. For the m-th The soil temperature at each sampling time, where m is the index of the sampling time within that day. The total number of sampling times within the day; d is the time shift variable; It should be noted that by calculating the deviation series of stem flow rate and soil temperature relative to their mean, the influence of the mean on the data can be eliminated, and the focus can be placed on the changes or fluctuations in the data. and The purpose of this setting is to eliminate the dimensional differences and baseline shifts in the absolute amplitudes of the two sequences. The absolute value of stem flow rate is greatly affected by factors such as plant size and weather conditions, and there is a lack of comparability between different dates. The deviation sequence, on the other hand, represents the fluctuation pattern of each sequence around the mean, retaining the waveform shape and phase information, while eliminating the interference caused by amplitude differences. The formula further performs a product summation operation on the soil temperature deviation sequence relative to the stem flow rate deviation sequence within a preset time shift range, performing a shift-by-shift operation on the corresponding misaligned elements. When the two waveforms match best at a certain shift, the positive and negative fluctuations are matched, the product term is mainly positive and has a large absolute value, and the inner product value reaches its peak. Make the inner product value The time shift to the maximum value characterizes the time delay of the stem flow rate waveform relative to the soil temperature waveform on the time axis. During the normal growth period, the root water conduction resistance is small, and the stem flow increases rapidly with the increase of temperature. The lag time is usually small and generally stable at 0.3 to 0.6 hours. When rice enters the critical rejection window period, the activity of root aquaporins is actively downregulated, the water conduction resistance increases sharply, and the response of stem flow to temperature changes is significantly slowed. The lag time jumps to between 0.8 and 2.0 hours. Cross-correlation analysis is a standard method for estimating the time delay between two signals. Treating soil temperature as the input excitation signal of the system and stem flow rate as the system's response output signal to that excitation, the peak point of the cross-correlation function between the two is an unbiased estimate of the system's pure time delay.

[0029] By mapping the transport lag phase of each day to the photosynthetic response sensitivity coefficient, the state vector of the monitored sample at the corresponding time is obtained. This process is repeated for all monitored samples, and their state vectors are continuously monitored.

[0030] The exclusion state analysis module is used to record the growth stage of each monitored sample when the state vector satisfies the critical exclusion constraint, and to form a critical exclusion window. It calculates the state vector of the rice plant under monitoring within the current sliding time window, determines whether the state vector satisfies the critical exclusion constraint, and whether the growth stage of the rice plant under monitoring falls into the critical exclusion window.

[0031] For any monitoring sample, if any state vector of the monitoring sample simultaneously satisfies the following conditions, then the monitoring sample is considered to satisfy the critical repulsion constraint in the current growth stage. Condition 1: In the current state vector of the monitored sample, the photosynthetic response sensitivity coefficient is not greater than the preset threshold, and the rate of change of the photosynthetic response sensitivity coefficient is negative; Condition 2: In the current state vector of the monitored sample, the transmission lag phase is not less than the preset lag threshold; It should be noted that in condition 1, the photosynthetic response sensitivity coefficient is not greater than the preset threshold, indicating that the leaf response to chlorophyll synthesis driven by enhanced light is saturated or even desensitized, indicating that the rice may be entering a critical rejection state at the current growth stage. At the same time, a trend criterion of negative change rate of photosynthetic response sensitivity coefficient is added. When the coefficient is already at a low value and is still decreasing, it indicates that the photosynthetic system is actively regressing rather than being passively inhibited, excluding the possibility of false triggering caused by instantaneous environmental disturbances. The efficiency of leaf chlorophyll synthesis drops sharply, thus judging that the monitored sample is currently in a critical rejection state. In condition 2, the transport lag phase is not less than the preset lag threshold. The water conduction capacity of rice roots is mainly regulated by the activity of aquaporins. During the vigorous vegetative growth period, the root system is full of vitality, and the stem flow rate follows soil temperature changes well. The delay time between the peak values ​​of the two is short and stable. When the transport lag phase is not less than the preset lag threshold, it indicates that the response process of water transport from the soil to the canopy through the roots is significantly delayed. That is, the plant actively downregulates the root metabolic activity to inhibit excessive nutrient growth, and the root water conduction resistance increases. Therefore, it may indicate that the monitored sample is currently in a critical rejection state. The two conditions were used to verify the state of rice from two completely independent physiological axes: the photosynthetic energy conversion channel and the water transport channel. If both photosynthetic response was continuously desensitized and transport response was continuously delayed, the possibility of a single channel being disturbed was basically ruled out. Therefore, it was determined that the monitored sample was currently in a critical rejection state.

[0032] The preset threshold is obtained by monitoring the photosynthetic response sensitivity coefficient of rice at different growth stages in the field. Based on historical data, the average value and standard deviation of the photosynthetic response sensitivity coefficient at each growth stage are calculated. The average value is usually reduced by a certain multiple of the standard deviation to cover the normal fluctuation range. The 25th percentile or other percentiles of the normal fluctuation range are used as the threshold. By monitoring rice at different growth stages, analyzing the delayed relationship between root conduction and stem flow rate, and extracting the lag time distribution of each stage, including the mean and standard deviation, an appropriate lag threshold can be set. Generally, it is set as the average lag time during normal growth plus 1.5 times the standard deviation to ensure that the growth status under normal conditions is fully covered.

[0033] When the critical exclusion constraint is met, the current growth stage of the monitored sample is recorded, and the growth stage is specifically characterized by the number of planting days. Traverse all monitoring samples to determine the planting days corresponding to all those that meet the critical exclusion constraint. For all planting days, perform clustering without specifying clusters based on the absolute difference between the planting days. Based on the clustering results, form several clusters of planting days. Extract the number of planting days in each cluster, remove clusters with a number of planting days not exceeding a preset size threshold, and record the remaining clusters as the critical exclusion window; The specific steps for clustering without defined clusters include: collecting the planting days corresponding to all monitoring samples that satisfy the critical exclusion constraint, forming a dataset; for each planting day, calculating the absolute difference between it and other planting days to construct a distance matrix; using the DBSCAN method to perform clustering without defined clusters, where Euclidean distance is typically used to measure the similarity between samples; calculating the distance matrix between all planting days; defining the radius of the neighborhood, which determines the range within which points are considered dense, and defining the minimum number of points in a cluster, which determines the condition for a core point; for each planting day, determining the number of points in its neighborhood by calculating its distance from other days. If the number of points in the neighborhood is greater than or equal to the defined minimum number of points, the point is marked as a core point. Starting from the core point, all points in its neighborhood are searched and grouped into the same cluster; this process is repeated until all points that can be clustered have been grouped; processing all points, ultimately forming several clusters. The DBSCAN output divides the planting days into different clusters; each cluster contains similar planting days. Clusters with a planting number of days not exceeding a preset size threshold are removed, that is, only those clusters that meet the preset size threshold are retained, and the retained clusters are recorded as critical rejection windows; the planting number of these clusters is relatively concentrated, representing the time period during which critical rejection may occur under a specific growth stage. The preset size threshold is derived by analyzing the number of samples required for successful clustering in historical monitoring data, and is generally set between 5 and 15.

[0034] To determine whether the state vector satisfies the critical exclusion constraint, and whether the growth stage of the rice plant to be monitored falls within the critical exclusion window, the dual-channel verification of the critical exclusion constraint and growth stage is used to confirm whether the rice plant to be monitored is in a critical exclusion state. In actual monitoring, rice plants may be affected by instantaneous environmental factors, such as sudden climate changes, fertilization, irrigation, etc., which may cause temporary fluctuations in physiological indicators. Introducing the dual verification of the growth stage can effectively eliminate false alarms caused by disturbances in a single channel. If the rice plant to be monitored satisfies the critical exclusion constraint, but its growth stage does not fall within the critical exclusion window, it indicates that there may be other factors causing the rice plant to enter a critical exclusion state, and corresponding measures should be taken.

[0035] The agricultural allocation module is used to execute corresponding control schemes based on the judgment results. If the judgment is correct, the first deep learning model is called to output the actual amount of nitrogen applied; otherwise, the second deep learning model is called to detect the types of pests and diseases of the rice plants to be monitored.

[0036] The logic for determining whether the state vector of the rice plant under monitoring within the current sliding time window satisfies the critical repulsion constraint, and whether the growth stage of the rice plant falls within the critical repulsion window, is as follows: If the current state vector of the rice plant to be monitored does not satisfy the critical repulsion constraint, then the amount of nitrogen to be applied to the rice plant to be monitored is determined based on the classical nitrogen application amount for different growth stages. If the current state vector of the rice plant to be monitored satisfies the critical repulsion constraint, and the growth stage of the rice plant to be monitored falls into the critical repulsion window, then the first deep learning model is called to output the actual amount of nitrogen applied; otherwise, the second deep learning model is called to detect the types of diseases and pests of the rice plant to be monitored. The first deep learning model is built based on the LSTM model. The input of the first deep learning model is the growth stage of the rice plant to be monitored, the nitrogen content of the plant, the nitrogen content of the soil, and the current state vector. The output of the model is the current actual amount of nitrogen applied to the rice plant to be monitored. The growth stage of the rice plant is specifically described by the number of days of planting. For the LSTM model, the activation function and optimization algorithm are selected. The Tanh function is chosen as the activation function, and Adam is chosen as the optimization algorithm for the LSTM model. The formula for the Tanh function is: In the formula, Represents the Tanh function, with the independent variable... This represents the weighted sum of the neuron's inputs, that is, the result of the weighted sum of the inputs received by the neuron from the previous layer; Simultaneously, the hyperparameters of the LSTM model are set, including: number of network layers, number of iterations, learning rate, batch size, number of training iterations, number of batches, and number of hidden layer neurons. The network is set to a 4-layer structure, the number of iterations is set to 200, the learning rate is set to 0.001, the batch size is set to 32, the number of training iterations is set to 100, the number of batches is set to 256, and the number of hidden layer neurons is set to 32. The specific logic for training the LSTM model is as follows: Historical data related to rice growth is collected, including the growth stage of the rice plant to be monitored, plant nitrogen content, soil nitrogen content, and historical state vectors. Experimental verification is conducted, using maximum yield as the verification target, and combined with expert experience to determine the actual nitrogen application amount for the corresponding state. Several historical data points are collected as training samples. The features of the selected training samples are standardized to ensure that data of different dimensions are within the same range. Common methods include Z-score standardization or Min-Max normalization. The normalized data... The dataset is divided into training, validation, and test sets, typically in a 70% training, 15% validation, and 15% test ratio. Multiple rounds of training are performed on the training set, with parameters updated using mini-batch data in each round. Model performance is tested on the validation set, and model parameters, such as the learning rate and the number of LSTM units, are adjusted. Mean squared error (MSE) is used as the loss function to measure the difference between predicted and actual values. The test set is used to evaluate the model's performance, including calculating the MSE as an evaluation metric and generating a comparison chart of the true and predicted values ​​to visualize the model's predictive ability. This completes the training of the LSTM model.

[0037] The second deep learning model is built based on a convolutional neural network, which consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The activation function in the convolutional layers is... For the fully connected layer, the function sets the number of neurons in the fully connected layer to 32, the initial neural network learning rate to 0.001, and the number of training rounds to 100. The input to the second deep learning model is the canopy and leaf images of the rice plant to be monitored, and the output is the type of pests and diseases of the rice plant to be monitored. The training of the second deep learning model specifically includes: collecting canopy and leaf image data of the rice plants to be monitored. These images should include samples of different types of pests and diseases as well as samples of healthy plants; each image is manually labeled to identify the corresponding type of pest or disease. In the data preparation stage, image cleaning is performed to remove blurry or low-quality images. Image enhancement techniques, such as rotation, flipping, scaling, and cropping, are applied to increase the diversity of the data and enhance the generalization ability of the model. The collected image data is divided into training set, validation set, and test set, typically using 70% for training, 15% for validation, and 15% for testing. When building a convolutional neural network model, an input layer is set to receive image data of a fixed size, such as a 224x224 pixel RGB image. Multiple convolutional layers are constructed, and each layer uses the ReLU activation function to extract the features of the image. Each convolutional layer is usually followed by a pooling layer, which uses max pooling to reduce the size of the feature map and reduce computational complexity.

[0038] In the latter part of the network, add one or more fully connected layers to map the extracted features to the output; depending on the requirements, set the number of neurons in the fully connected layer to 32, build the output layer, the number of output neurons should be the same as the number of pest types, and use the Softmax activation function for classification; During the model training phase, an appropriate loss function, such as cross-entropy, and an optimizer, such as Adam, are selected, and the initial learning rate is set to 0.001. The model is trained using the training set, while the model performance is monitored using the validation set to prevent overfitting. The number of training epochs is set to 100 epochs, and the training process is optimized by continuously adjusting the batch size and learning rate. After training is completed, the model performance is evaluated on the test set, and the model's accuracy and other metrics are calculated to achieve the training of the second deep learning model.

[0039] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0040] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0041] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0042] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A smart agriculture integration system based on deep learning, characterized in that, Specifically, it includes: The data acquisition module is used to set several rice plants at different growth stages as monitoring samples, continuously collect the time series of canopy spectral reflectance, photosynthetically active radiation, stem flow rate and soil temperature at each sampling time for each monitoring sample, and calculate the canopy chlorophyll index based on the spectral reflectance. The photosynthetic response module is used to construct a sliding time window for any monitoring sample, with the current day as the endpoint. It performs linear regression on the average change of the canopy chlorophyll index and the average change of photosynthetically active radiation within the sliding time window, and extracts the slope coefficient of the regression as the photosynthetic response sensitivity coefficient for the current day. The state monitoring module is used to extract the stem flow rate and soil temperature at each sampling time within the day, perform cross-correlation analysis on the two, solve for the time shift that makes the cross-correlation function reach its maximum value, use the time shift as the transport lag phase of the day, and combine the transport lag phase of the day with the photosynthetic response sensitivity coefficient to form a state vector. The exclusion state analysis module is used to record the growth stage of each monitored sample when the state vector satisfies the critical exclusion constraint, and form a critical exclusion window. It calculates the state vector of the rice plant under monitoring within the current sliding time window, determines whether the state vector satisfies the critical exclusion constraint, and whether the growth stage of the rice plant under monitoring falls into the critical exclusion window. The agricultural allocation module is used to execute corresponding control schemes based on the judgment results. If the judgment is correct, the first deep learning model is called to output the actual amount of nitrogen applied; otherwise, the second deep learning model is called to detect the types of pests and diseases of the rice plants to be monitored.

2. The intelligent agriculture fusion system based on deep learning according to claim 1, characterized in that, For any monitoring sample, the specific calculation method of the canopy chlorophyll index is as follows: for any sampling time, obtain the near-infrared band reflectance and red edge band reflectance of the rice canopy at that sampling time, and use the difference between the ratio of the near-infrared band reflectance and the red edge band reflectance and the constant 1 as the canopy chlorophyll index at that sampling time. The same laboratory environment was set up for each monitoring sample, and data was collected from the monitoring samples in the laboratory environment, which was set up based on the rice growing environment.

3. The intelligent agriculture fusion system based on deep learning according to claim 2, characterized in that, The specific method for constructing the sliding time window is as follows: taking the current day as the end point and tracing back a fixed number of days to form a sliding time window, for any monitoring sample, the average change in the canopy chlorophyll index and the average change in photosynthetically active radiation are extracted daily within the sliding time window. For any given monitoring sample, the average change in canopy chlorophyll index and the average change in photosynthetically active radiation within a sliding time window are determined using the following method: Except for the first day of the sliding time window, for any day within the sliding time window, the canopy chlorophyll index at each sampling time on that day is calculated, and the average value of the canopy chlorophyll index at each sampling time is used as the canopy chlorophyll index for that day. The canopy chlorophyll index of the previous day within the sliding time window is subtracted from the canopy chlorophyll index of that day to obtain the change in the canopy chlorophyll index for that day. Using the same method, the change in photosynthetically active radiation for that day was calculated. Except for the first day of the sliding time window, based on the changes in photosynthetically active radiation and canopy chlorophyll index for each day, the average change in canopy chlorophyll index and the average change in photosynthetically active radiation for the monitoring sample within the sliding time window were calculated.

4. The intelligent agriculture fusion system based on deep learning according to claim 3, characterized in that, The specific method for determining the photosynthetic response sensitivity coefficient on a given day is as follows: the average change in the two canopy chlorophyll indices and the average change in photosynthetically active radiation on the same day are combined into a data point pair. Least square linear regression is performed on all data point pairs within the sliding time window, and the slope obtained from the regression is used as the photosynthetic response sensitivity coefficient.

5. The intelligent agriculture fusion system based on deep learning according to claim 4, characterized in that, For any monitoring sample, the specific method for calculating its transport lag phase is as follows: taking multiple sampling times within the day as a reference, extract the measured values ​​of stem flow rate and soil temperature at each sampling time, calculate the deviation sequences of the stem flow rate sequence and soil temperature sequence relative to their mean values, perform a sliding inner product operation on the soil temperature deviation sequence relative to the stem flow rate deviation sequence within a preset lag time range, obtain the time shift corresponding to the maximum inner product value, and use this time shift as the transport lag phase of the monitoring sample for that day; By mapping the transport lag phase of each day to the photosynthetic response sensitivity coefficient, the state vector of the monitored sample at the corresponding time is obtained. This process is repeated for all monitored samples, and their state vectors are continuously monitored.

6. The intelligent agriculture fusion system based on deep learning according to claim 1, characterized in that, For any monitoring sample, if any state vector of the monitoring sample simultaneously satisfies the following conditions, then the monitoring sample is considered to satisfy the critical repulsion constraint in the current growth stage. Condition 1: In the current state vector of the monitored sample, the photosynthetic response sensitivity coefficient is not greater than the preset threshold, and the rate of change of the photosynthetic response sensitivity coefficient is negative; Condition 2: In the current state vector of the monitored sample, the transmission lag phase is not less than the preset lag threshold; When the critical exclusion constraint is met, the current growth stage of the monitored sample is recorded, and the growth stage is specifically characterized by the number of planting days. Traverse all monitoring samples to determine the planting days corresponding to all those that meet the critical exclusion constraint. For all planting days, perform clustering without specifying clusters based on the absolute difference between the planting days. Based on the clustering results, form several clusters of planting days. Extract the number of planting days in each cluster, remove clusters whose number of planting days is not greater than a preset size threshold, and record the remaining clusters as the critical exclusion window.

7. A deep learning-based smart agriculture fusion system according to claim 6, characterized in that, The logic for determining whether the state vector of the rice plant under monitoring within the current sliding time window satisfies the critical repulsion constraint, and whether the growth stage of the rice plant falls within the critical repulsion window, is as follows: If the current state vector of the rice plant to be monitored does not satisfy the critical repulsion constraint, then the amount of nitrogen to be applied to the rice plant to be monitored is determined based on the classical nitrogen application amount for different growth stages. If the current state vector of the rice plant to be monitored satisfies the critical repulsion constraint, and the growth stage of the rice plant to be monitored falls within the critical repulsion window, then the first deep learning model is called to output the actual amount of nitrogen applied; otherwise, the second deep learning model is called to detect the types of diseases and pests affecting the rice plant to be monitored.

8. A deep learning-based smart agriculture fusion system according to claim 7, characterized in that, The first deep learning model is built based on the LSTM model. The input of the first deep learning model is the growth stage of the rice plant to be monitored, the nitrogen content of the plant, the nitrogen content of the soil, and the current state vector. The output of the model is the current actual amount of nitrogen applied to the rice plant to be monitored. The second deep learning model is built based on a convolutional neural network, which consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The activation function in the convolutional layers is... For the fully connected layer, the function sets the number of neurons in the fully connected layer to 32, the initial neural network learning rate to 0.001, and the number of training rounds to 100. The input to the second deep learning model is the canopy and leaf images of the rice plant to be monitored, and the output is the type of pests and diseases of the rice plant to be monitored.

Images