Tomato precision water and fertilizer decision method and system based on feature extraction

Abstract

The application discloses a kind of tomato precision water and fertilizer decision-making method and system based on feature extraction, comprising: by drip irrigation belt, cold water pulse with water temperature lower than current soil temperature is injected into tomato planting soil;Temperature time series output by vertically distributed soil temperature sensor array is collected and temperature recovery curve feature is extracted;It is input into pre-trained root depth positioning model to output the migration position depth of real active root system;Real-time soil moisture content is obtained, and the migration position depth is input into the water and fertilizer infiltration calculation model based on strategy optimization with real-time soil moisture content, to output target infiltration flow rate and target irrigation duration and drive intelligent agricultural power machinery to execute irrigation.The application solves the problem of water and fertilizer deep seepage waste and plant latent hunger caused by root system spatial migration in traditional blind irrigation.

Description

Technical Field

[0001] This invention relates to the field of agricultural and computer technology integration, and more specifically, to a method and system for precise water and fertilizer decision-making for tomatoes based on feature extraction. Background Technology

[0002] The continuous development of modern agriculture has made precision fertigation technology a core means to improve crop yield and quality. Tomatoes, as a highly valuable solanaceous vegetable, are extremely sensitive to water and fertilizer conditions during their growth process. Furthermore, the root system's ability to absorb water and the direction of nutrient distribution change drastically at different growth stages and even at different times of the day. Currently, many agricultural IoT systems are attempting to use various sensors and algorithm models to achieve automated irrigation and fertilization decisions. For example, Chinese invention patent document CN111552253A discloses an intelligent cloud irrigation integrated platform architecture and its control method. This method uses an intelligent cloud integrated platform to collect and process basic agricultural and crop data, thereby outputting decision information and controlling irrigation equipment. Another example is Chinese invention patent document CN112450056A, which discloses an intelligent irrigation system integrating water, fertilizer, and pesticides based on machine learning algorithms. This system uses machine learning models to process environmental and crop health data, thereby dynamically generating an integrated irrigation plan. However, these existing technologies still face many deep-seated bottlenecks when practically applied to drip irrigation scenarios for crops such as tomatoes.

[0003] In greenhouse or field drip irrigation scenarios, tomato capillary roots exhibit strong hydrotropism and spatial mobility. Shallow irrigation in the early stages often leads to the entire root system rising to the surface. Existing systems typically bury soil moisture sensors at a fixed depth. When the system issues standard water and fertilizer irrigation commands based on sensor data at this depth, nutrient-rich water and fertilizer easily penetrate the already risen shallow root network and seep into the deeper, rootless soil. This results in the sensors detecting adequate moisture content, but the plants are actually in a state of latent hunger, leading to significant waste of water and fertilizer. Furthermore, during the hot, sunny days of summer, tomato plants often activate a photosynthetic midday depression mechanism, forcibly closing stomata and ceasing water absorption. Blindly irrigating at this time can instantly cause root hypoxia and root rot. During the normal high transpiration period of the day, the strong leaf pull draws most of the nutrients absorbed by the roots to the upper leaves. Applying excessive fertilizer at this time can cause excessive leaf growth, leaving the fruit, which is truly in its expansion stage, without sufficient nutrition. Therefore, accurately locating the spatial position of active root systems and precisely grasping the timing of irrigation and fertilization are technical challenges that need to be addressed in existing technologies. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for precise water and fertilizer decision-making for tomatoes based on feature extraction, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A precision water and fertilizer decision-making method for tomatoes based on feature extraction includes: A preset volume of cold water pulses is injected into the tomato planting soil via drip irrigation tape. The temperature of the cold water pulses is lower than the current soil temperature of the tomato planting soil. The temperature time series output by the soil temperature sensor array vertically distributed in the tomato planting soil was collected; Feature extraction is performed on the temperature time series to obtain the temperature recovery curve features corresponding to each soil depth; The temperature rise curve features corresponding to each soil depth are input into the pre-trained root depth localization model, and the migration depth of the current real active root system is output. Obtain the real-time soil moisture content of the tomato planting soil; The migration location depth and the real-time soil moisture content are input into a pre-trained water and fertilizer infiltration estimation model based on strategy optimization, and the target infiltration velocity and target irrigation duration that make the nutrient fertilizer solution remain at the migration location depth are output. The corresponding control commands are generated to drive the intelligent agricultural power machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and the target irrigation duration.

[0006] Specifically, the steps of extracting features from the temperature time series to obtain the temperature recovery curve features corresponding to each soil depth include: Within a preset time window after the injection of the cold water pulse, the temperature minimum point in the temperature time series is extracted as the bottoming feature point. Extract sequence data from the bottoming feature point to the initial temperature range before the cold water pulse was injected; The first derivative of the sequence data is performed to obtain the maximum temperature rise slope at the corresponding soil depth. The maximum temperature rise slope is combined with the bottoming feature point to form the temperature rise curve feature.

[0007] Specifically, the root system depth localization model is a multilayer perceptron network; the training process of the root system depth localization model includes: Acquire a historical sample dataset, which includes historical temperature rise curve features and actual active root depth labels calibrated by an in-situ root observation device. The historical temperature recovery curve features are input into the initial multilayer perceptron network, and the predicted root depth is output. Calculate the mean squared error loss value between the predicted root depth and the actual active root depth label; The weight parameters of the initial multilayer perceptron network are updated by backpropagation using the gradient descent algorithm until the mean squared error loss value converges, thus obtaining the pre-trained root system depth localization model.

[0008] Specifically, the pre-trained water and fertilizer infiltration estimation model based on policy optimization is a deep deterministic policy gradient reinforcement learning model. The state space input of the deep deterministic policy gradient reinforcement learning model includes the migration location depth and the real-time soil moisture content; The action space output of the deep deterministic policy gradient reinforcement learning model is the target infiltration velocity and the target irrigation duration; The reward function of the deep deterministic policy gradient reinforcement learning model is set as the weighted sum of the overlap between the actual residence depth of the nutrient solution and the migration location depth and the amount of water and fertilizer consumed.

[0009] Specifically, before the step of injecting a predetermined volume of cold water pulses into the tomato planting soil via drip irrigation, the process also includes: The time series of microcavitation acoustic emission was collected by an acoustic sensor attached to the base of the tomato plant stem. Collect ambient light intensity and ambient temperature data of the environment in which the tomato plants are located; The microcavitation acoustic emission time series, the ambient light intensity data, and the ambient temperature data are concatenated into a multidimensional state feature vector. The multidimensional state feature vector is input into a pre-trained photosynthetic dormancy recognition model, which outputs the physiological state category of the tomato plant. When the physiological state category is photosynthetic midday rest state, intercept the cold water pulse injection command and the water and fertilizer irrigation command; When the physiological state category is active transpiration, the step of injecting a preset volume of cold water pulse into the tomato planting soil through the drip irrigation tape is triggered.

[0010] Specifically, the step of concatenating the microcavitation acoustic emission time series, the ambient light intensity data, and the ambient temperature data into a multidimensional state feature vector includes: The microcavitation acoustic emission time series was converted to the frequency domain using Fast Fourier Transform; Extract the frequency and average amplitude of acoustic emission pulses within a preset frequency band; The acoustic emission pulse frequency, the average amplitude, the ambient light intensity data, and the ambient temperature data are combined to form the multidimensional state feature vector.

[0011] Specifically, the photosynthetic dormancy recognition model is a support vector machine classifier; the training process of the photosynthetic dormancy recognition model includes: Acquire historical multidimensional state feature vector samples, as well as the corresponding stomatal closure state labels measured synchronously by a stomatal conductance meter; The historical multidimensional state feature vector sample is used as input, and the stoma closure state label is used as output. The hyperplane parameters of the support vector machine classifier are solved using the sequence minimum optimization algorithm to complete the training of the photosynthetic dormancy recognition model.

[0012] Specifically, the step of generating corresponding control commands to drive intelligent agricultural power machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and the target irrigation duration includes: Obtain the diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk, which are collected by a micro-displacement sensor. The diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk are input into a pre-trained timing optimization judgment model to output the suction switching time node. When the system reaches the suction switching time node, the corresponding control command is generated to drive the intelligent agricultural power machinery to inject the nutrient fertilizer solution into the drip irrigation tape according to the target infiltration flow rate and the target irrigation duration, thereby completing the water and fertilizer irrigation operation.

[0013] Specifically, the timing optimization judgment model is a long short-term memory network model; the step of inputting the diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk into the pre-trained timing optimization judgment model and outputting the suction switching time node includes: The diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk are fed into the hidden layer unit of the long short-term memory network to extract temporal dynamic features. The temporal dynamic features are mapped into an output probability distribution sequence through a fully connected layer; Obtain the timestamp in the output probability distribution sequence that indicates the tomato main stem has stopped contracting and the target fruit stalk has begun to swell continuously; The timestamp is output as the suction switching time node.

[0014] This invention also discloses a precision water and fertilizer decision-making system for tomatoes based on feature extraction, comprising: The pulse injection module is used to inject a preset volume of cold water pulse into the tomato planting soil through a drip irrigation tape. The temperature of the cold water pulse is lower than the current soil temperature of the tomato planting soil. The temperature acquisition module is used to acquire the temperature time series output by the soil temperature sensor array that is vertically distributed in the tomato planting soil. The feature extraction module is used to extract features from the temperature time series to obtain the temperature recovery curve features corresponding to each soil depth. The depth localization module is used to input the temperature rise curve features corresponding to each soil depth into the pre-trained root depth localization model and output the migration location depth of the current real active root system. A moisture content acquisition module is used to acquire the real-time soil moisture content of the tomato planting soil; The flow velocity estimation module is used to input the migration location depth and the real-time soil moisture content into a pre-trained water and fertilizer infiltration estimation model based on strategy optimization, and output the target infiltration flow velocity and target irrigation duration that make the nutrient fertilizer solution remain at the migration location depth. The irrigation execution module is used to generate corresponding control commands to drive intelligent agricultural power machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and the target irrigation duration.

[0015] The advantages of this invention compared to existing technologies lie in its effective solution to the problems of deep water and fertilizer leakage and low utilization rates caused by root migration in traditional drip irrigation scenarios. This invention creatively utilizes a cold water pulse thermal shock response mechanism. By injecting a preset cold water pulse lower than the current soil temperature into the soil, and combining this with vertically distributed temperature sensors to extract temperature rise curve characteristics at various depths, the system leverages the modulation patterns of heat transport caused by water absorption and physiological metabolic activities in the active root zone. The system can accurately locate the current migration position of the truly active roots by analyzing the temperature rise slope. Based on this depth positioning and real-time soil moisture content, the system then uses a strategy-optimized water and fertilizer infiltration calculation model to accurately deduce the target infiltration velocity and target irrigation duration. This technique allows the nutrient solution to be precisely suspended and retained in the true root absorption layer of the plant, completely preventing fertilizer loss to ineffective soil layers and effectively alleviating the plant's latent hunger.

[0016] This invention also addresses the potential for root rot and oxygen deficiency in plants caused by blind irrigation decisions under high temperature and strong light conditions. Before formal detection and irrigation, this invention introduces the acquisition of time-series data on the acoustic emission from the microcavitation at the base of the plant stem, fusing it with ambient light and temperature data to determine the plant's physiological state. When plants forcibly close their stomata and enter photosynthetic dormancy due to high temperature and strong light, the sudden drop in sap flow pull within the plant causes a sharp reduction in high-frequency acoustic pulses to a silent state. This invention utilizes this hidden plant hydraulic characteristic as a pre-verification method, accurately intercepting cold water pulses and irrigation commands during dormancy, constructing a robust safety interlock, and fundamentally preventing irreversible root damage caused by incorrect water supply timing.

[0017] This invention also solves the problem of developmental imbalance caused by nutrient competition between plant vegetative and reproductive organs. Instead of blindly applying high-concentration fertilizers during the daytime transpiration period when water demand is high, this invention extracts the micro-diameter deformation sequences of the tomato main stem and the target fruit stalk for temporal dynamic analysis. By identifying specific probabilistic time points where the main stem contraction stops and the fruit stalk begins continuous expansion, the system precisely pinpoints the period of hydraulic switching within the plant, from leaf transpiration to suction pressure in the fruit phloem. At this opportune moment, mechanical injection of nutrient solution is achieved, realizing targeted nutrient delivery to the target fruit over time, effectively avoiding excessive vegetative growth caused by ineffective fertilizer absorption by the leaves. Attached Figure Description

[0018] Figure 1 This is the main flowchart of the precision water and fertilizer decision-making process of this invention; Figure 2 This is a diagram of the closed-loop structure of reinforcement learning for water and fertilizer infiltration calculation in this invention. Figure 3 This is a schematic diagram of the tomato growing environment and underground sensor array arrangement of the present invention; Figure 4 This is a schematic diagram of the installation of the acoustic wave sensor at the base of the tomato plant stem according to the present invention; Figure 5 This is a schematic diagram of the environmental monitoring device of the present invention; Figure 6 This is a diagram showing the contact state between the root system in situ observation device of this invention and the soil. Detailed Implementation

[0019] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0020] This invention provides a method and system for precise water and fertilizer decision-making for tomatoes based on feature extraction. It is applicable to drip irrigation scenarios for greenhouse tomatoes and field tomatoes, and is especially suitable for cultivation environments where the depth of the root active layer dynamically changes with irrigation history, soil structure and transpiration status.

[0021] like Figure 1 As shown, the core idea of ​​this invention is not to directly give the water and fertilizer formula based solely on the readings of a fixed depth sensor, but rather to first actively construct an observable thermal disturbance, and then use the recovery characteristics of this thermal disturbance in different soil layers to infer the spatial location of the actual active root system. Subsequently, the infiltration rate of the nutrient fertilizer solution and the irrigation duration are inferred around this location, so that the nutrient fertilizer solution stays as much as possible within the soil layer that can be efficiently absorbed by the current root system.

[0022] In one embodiment, the tomato planting area is equipped with drip irrigation tape, a pulse injection device, a fertilizer mixing device, a soil temperature sensor array, a soil moisture sensor, an acoustic sensor attached to the base of the tomato plant stem, micro-displacement sensors located on the main stem and the fruit stalk of the target fruit, and an edge controller for executing control logic. The edge controller can be an industrial computer, an agricultural gateway, or an embedded control host, internally running a root depth positioning model, a water and fertilizer infiltration calculation model, a photosynthetic dormancy identification model, and a timing optimization judgment model. The edge controller is also communicatively connected to solenoid valves, variable frequency pumps, fertilizer pumps, and power machinery control units to execute irrigation and fertilization actions according to the calculated target parameters.

[0023] The "true active root system" described in this invention does not refer to the geometric distribution range of the entire plant root system, but rather to the active root layer that contributes the most to water and nutrient absorption and significantly influences local heat and water migration behavior within the current time period. This definition needs to be explicitly stated in the implementation because tomato roots are not static; shallow irrigation can induce roots to rise to the surface, and alternating soil moisture levels will guide the roots back to deeper layers. If traditional fixed-depth sensors are still used as the sole basis, it is easy to encounter situations where the sensor indicates adequate absorption, but the plant's actual absorption is insufficient.

[0024] In one embodiment, before formal irrigation, a predetermined volume of cold water pulse is injected into the tomato planting soil via drip irrigation. A cold water pulse refers to the injection of a certain volume of water at a temperature lower than the current soil temperature over a short period, used to create a short-term, identifiable temperature drop and rise around the roots. This process is equivalent to artificially applying a standardized stimulus, allowing the differences in the response of soil layers at different depths to thermal disturbances to be collected and quantified. The cold water pulse is not primarily for replenishment but for detection; therefore, its volume is preferably smaller than that of a single conventional irrigation. To balance the identifiability of the disturbance with gentleness on root physiology, the injection volume of the cold water pulse can be set to 50 mL to 300 mL per plant, preferably 80 mL to 180 mL per plant; the cold water temperature can be 3°C to 12°C lower than the current soil temperature, preferably 4°C to 8°C lower. When the temperature difference is too small, the temperature recovery curve is not easily distinguishable and is easily drowned out by environmental noise; when the temperature difference is too large, although the signal is more obvious, it may cause excessive stimulation to the local rhizosphere environment. Therefore, a moderate temperature difference is more conducive to stable feature extraction.

[0025] To accurately obtain the thermal response process at different depths, such as Figure 3As shown, a soil temperature sensor array is vertically buried near the tomato root zone. The sensor array can be positioned 3cm to 15cm away from the drip irrigation tape to avoid direct contact with the irrigation flow while still reflecting actual temperature changes in the root zone. The sensors can be spaced at depths of 1cm to 5cm, preferably with one measuring point at depths of 5cm, 10cm, 15cm, 20cm, 25cm, and 30cm. In soil layers where tomato root activity is more concentrated, the number of measuring points can be further increased. Each sensor preferably uses a digital temperature probe or a high-stability thermistor, and the sampling frequency can be set to 0.5Hz to 5Hz, preferably 1Hz to 2Hz, to ensure complete capture of the temperature bottoming and recovery process after the cold water pulse injection.

[0026] After the cold water pulse is injected, the edge controller continuously acquires the temperature time series output by the depth sensors. To ensure consistency in feature extraction, it is preferable to first record the baseline temperature from 30 to 180 seconds before injection as the initial temperature reference value for each depth. Subsequently, starting from the injection moment, the temperature series is acquired within a preset time window of 2 to 20 minutes, preferably 5 to 12 minutes. The purpose of setting the preset time window is to cover the entire process from the temperature bottoming out to the temperature recovering to near the pre-injection temperature. If the window is too short, the temperature may not have fully recovered, resulting in incomplete features; if the window is too long, it will introduce more environmental temperature drift and external disturbances.

[0027] When extracting features from the temperature time series, the minimum temperature point is first located within a preset time window after the cold water pulse is injected, and this minimum point is used as the bottoming feature point. This bottoming feature point contains at least two pieces of information: first, the lowest temperature value, reflecting the strength of the cold water disturbance at that depth; and second, the time position when the lowest temperature is reached, reflecting the dynamic process of the cold water arrival and residence. Subsequently, the sequence data from the bottoming feature point to the temperature recovery to the initial temperature range before the cold water pulse is injected is extracted. Here, the recovery to the initial temperature range preferably refers to the temperature rising to a fluctuation range of 0.2℃ to 0.8℃ above and below the initial temperature, more preferably 0.3℃ to 0.5℃ above and below. Using a range rather than a single point as the recovery termination condition can avoid the difficulty in stably determining the recovery endpoint due to sensor quantization errors and small noises.

[0028] After obtaining the sequence data, its first derivative is calculated to obtain the maximum temperature recovery slope at the corresponding soil depth. The maximum temperature recovery slope essentially reflects the speed at which the soil layer at that depth recovers thermal equilibrium after being disturbed by cold water. However, the thermal recovery kinetics around active roots typically differ from those of rootless or weakly rooted soil layers due to stronger water absorption, root respiration, and local heat exchange. Therefore, this invention combines the maximum temperature recovery slope with the bottoming-out feature point to form a temperature recovery curve feature. This combination can be a direct splicing into a two-dimensional feature, or it can further include auxiliary quantities such as the bottoming-out time, minimum temperature difference, and recovery duration to form a multi-dimensional feature vector. In a preferred embodiment, each depth forms a local feature vector containing at least four features: the bottoming-out temperature value, the bottoming-out timestamp, the maximum temperature recovery slope, and the recovery duration. By splicing the local feature vectors of all depths in depth order, a comprehensive input feature reflecting the current thermal response state of the root zone is formed.

[0029] To reduce the impact of outlier sampling points on slope calculation, in one embodiment, the temperature time series is preprocessed before performing the first derivative calculation. Preprocessing may include median filtering, moving average filtering, and outlier removal. The moving window length can be set to 3 to 9 sampling points, preferably 5. Isolated outliers that significantly deviate from the neighborhood mean can be removed using a threshold method and interpolated with adjacent values; the threshold is preferably 1°C to 3°C. After this preprocessing, the calculation of the temperature rise slope is more stable, which is beneficial for improving the generalization ability of the subsequent root depth localization model.

[0030] The temperature recovery curve features corresponding to each soil depth are input into a pre-trained root depth localization model, which outputs the migration depth of the current actual active root system. The root depth localization model is preferably a multilayer perceptron network. The multilayer perceptron network is chosen because this invention processes structured numerical vectors after feature extraction at this stage. Its main task is to learn the nonlinear mapping relationship between thermal recovery features at different depths and the actual active root depth. The multilayer perceptron network can achieve a good balance between model complexity, training efficiency, and edge deployment cost.

[0031] In one embodiment, the multilayer perceptron network includes an input layer, 2 to 5 hidden layers, and an output layer, preferably 3 hidden layers. The number of neurons in each hidden layer can be set to 128, 64, and 32, respectively, or scaled according to the number of sensor channels. The activation function for the hidden layers can be ReLU, Leaky ReLU, or ELU, with ReLU being preferred to improve training stability. The output layer is a single-node linear output used to provide the predicted depth value of the currently active root system, in cm. To avoid different dimensional features affecting training, the input features are preferably normalized or standardized, and the output depth label can also be mapped to a preset range before being denormalized. For common tomato root depths, the output range can be set to 3cm to 35cm, preferably 5cm to 30cm.

[0032] The training process of the root system depth localization model is as follows. First, a historical sample dataset is acquired, which includes historical temperature rise curve features and labels of the actual active root depth calibrated using in-situ root observation equipment. For example... Figure 6 As shown, in-situ root observation equipment can employ transparent root windows, micro-root canal observation systems, or visual root boxes, combined with manual annotation or image analysis to determine the dominant depth of the current water-absorbing active layer. The true active root depth label mentioned here is preferably the peak depth of root density, the peak depth of rhizosphere water absorption intensity, or a weighted average of both. In-situ calibration is necessary because only by mapping thermal response characteristics to the true active root layer can the model learn a mapping relationship with actual physiological significance, rather than merely fitting soil heat diffusion phenomena.

[0033] During training, the historical temperature rise curve features are input into the initial multilayer perceptron network, which outputs the predicted root depth. The mean squared error loss between the predicted root depth and the actual active root depth label is calculated, and the network weight parameters are updated using the gradient descent algorithm through backpropagation until the loss value converges, resulting in a pre-trained root depth localization model. The training optimizer can use Adam or stochastic gradient descent algorithms, with a learning rate of 0.0001 to 0.01, preferably 0.0005 to 0.002; a batch size of 16 to 256, preferably 32 to 64; and the number of training epochs determined based on the validation set error, typically 50 to 500 epochs. To avoid overfitting, Dropout can be added after the hidden layers, with a deactivation ratio of 0.1 to 0.5, preferably 0.2 to 0.3. Alternatively, an early stopping strategy can be used to terminate training when the validation set error does not decrease for several consecutive epochs. After training, the converged network parameters are fixed and deployed to the edge controller for online inference.

[0034] After obtaining the current migration depth of the active root system, the real-time soil moisture content of the tomato planting soil is further acquired. Soil moisture content can be expressed as volumetric moisture content or a converted relative moisture content; volumetric moisture content measured at multiple points near the root zone is preferred. Sensors can be placed near the temperature array to ensure consistency. The effective monitoring range for volumetric moisture content is 10% to 45%, with a focus on the actual control range of 15% to 35%. Real-time soil moisture content is used as input for the subsequent extrapolation model because, at the same target root depth, the propulsion speed, lateral diffusion, and retention depth of water and fertilizer solutions will differ under different initial humidity conditions. If the initial moisture content is ignored and irrigation parameters are extrapolated solely from root depth, it is easy to under-irrigate in dry soil or over-irrigate in wet soil.

[0035] This invention further inputs the migration location depth and real-time soil moisture content into a pre-trained water and fertilizer infiltration estimation model based on strategy optimization, outputting a target infiltration velocity and target irrigation duration that ensures the nutrient solution remains at the migration location depth. Figure 2 As shown, the term "residence" here does not mean absolute stillness, but rather that during a period of effective absorption after irrigation, the main wetting front and high nutrient concentration area of ​​the nutrient solution remain near the target root layer, allowing active roots to preferentially contact and absorb nutrients.

[0036] In one embodiment, the pre-trained water and fertilizer infiltration estimation model based on strategy optimization is a deep deterministic policy gradient reinforcement learning model. The reason for using this type of model is that the target infiltration velocity and target irrigation duration are essentially continuous control variables, while the result formed by the combined effects of soil moisture movement, capillary action, gravity infiltration, and root absorption exhibits significant nonlinearity and time-varying characteristics. Simply relying on static empirical formulas is insufficient to adapt to different soil types and root states. The deep deterministic policy gradient reinforcement learning model can learn optimized strategies from states to continuous actions based on simulation environments or historical interaction data, making it more suitable for the online control scenario of this invention.

[0037] In one embodiment, such as Figure 2As shown, the deep deterministic policy gradient reinforcement learning model includes a policy network and a value network. The policy network corresponds to the Actor, and the value network corresponds to the Critic. It also includes a target policy network, a target value network, and an experience replay cache. The state space input includes at least the migration location depth and real-time soil moisture content. In a more refined implementation, extended variables such as soil type parameters, drip irrigation tape flow rate, recent irrigation history, and environmental evaporation intensity can also be input to improve adaptability to complex plots. The action space output consists of the target infiltration velocity and the target irrigation duration. The target infiltration velocity can be achieved by adjusting the variable frequency pump speed, valve opening, or drip irrigation branch pressure. Its value range can be converted to a single dripper equivalent flow rate of 0.1 L / h to 4 L / h, preferably 0.3 L / h to 2.5 L / h. The target irrigation duration can be set from 10 s to 1800 s, preferably 60 s to 900 s. The reason for limiting this range is to achieve precise control while avoiding unstable control due to extremely low flow rates or excessive infiltration of water and fertilizer over extremely long periods of time.

[0038] The reward function of a deep deterministic strategy gradient reinforcement learning model can be set as a weighted combination of a reward term for the overlap between the actual retention depth of the nutrient solution and the migration depth, and a penalty term for water and fertilizer consumption. Positive rewards are given for overlap, while negative penalties are given for excessively deep seepage, shallow retention, and excessive water and fertilizer consumption. To make the model more aligned with field control objectives, in one embodiment, higher rewards are given for deviations of the actual retention depth within 1 to 3 cm from the target depth, while significant penalties are imposed for deviations exceeding 5 cm. Resource constraints are set for water and fertilizer consumption per unit irrigation cycle to prevent the model from solely relying on increasing irrigation volume to pursue depth accuracy. The weights can be set according to production objectives; for example, water control safety is emphasized during the seedling and seedling establishment stages, while retention accuracy and fruit nutrient supply are emphasized during the fruit enlargement stage. During actual training, the reward weights are preferably searched and determined within the range of 0.1 to 0.9 to achieve an acceptable balance between hit rate and resource consumption.

[0039] To train the water and fertilizer infiltration estimation model, a soil moisture transport simulation environment or an interactive environment based on historical irrigation test data can be constructed. In the simulation environment, different initial soil moisture contents, different target root depths, and different actions can be input to obtain the post-irrigation wetting front position, nutrient peak position, and resource consumption, which are then used to provide feedback rewards. During training, both the policy network and the value network can use fully connected neural networks, each network including 2 to 4 hidden layers, with the number of neurons in a single layer ranging from 64 to 256, preferably 128. The learning rate can be set from 0.0001 to 0.005, with a lower learning rate preferred for the policy network and a slightly higher learning rate for the value network; the soft update coefficient of the target network can be set from 0.001 to 0.05, and the experience replay cache capacity can be set from 10,000 to 1,000,000 records. Through repeated interactive training, the model can output a better combination of flow rate and duration for a given root depth and soil moisture.

[0040] Before the aforementioned root system location and water and fertilizer parameter calculation, this invention preferably includes a pre-judgment step based on physiological state to prevent detection or irrigation during periods when the plant is not suitable for water absorption, especially to prevent accidental irrigation when tomatoes enter their photosynthetic dormancy period during the high temperature and strong light of summer. Specifically, before injecting a preset volume of cold water pulses into the tomato planting soil through drip irrigation, a sonic sensor attached to the base of the tomato plant stem collects a microcavitation acoustic emission time series, and simultaneously collects ambient light intensity and ambient temperature data of the tomato plant's environment, using these as a joint characterization of the plant's current hydraulic state and external driving conditions. Figure 4 and Figure 5 As shown, the acoustic sensor is located at the base of the tomato plant stem, and the ambient light intensity and ambient temperature data are collected by the corresponding environmental monitoring device.

[0041] The microcavitation acoustic emission described here refers to transient acoustic pulses generated by the formation, disintegration, or continuous changes in the sap column in the hydroponic tissues of plants under high-tension water transport conditions. These acoustic emission signals are more pronounced when plant water transport is active and transpiration pull is significant, but show significant attenuation or even become silent during the midday photosynthetic period when stomata close and sap flow pull decreases. Therefore, this invention uses it as an important preliminary signal for determining whether a plant is in a safe irrigation condition. Compared to relying solely on ambient temperature or time period, this signal better reflects the plant's immediate physiological feedback, thus improving decision-making safety.

[0042] In one embodiment, before concatenating the microcavitation acoustic emission time series, ambient light intensity data, and ambient temperature data into a multidimensional state feature vector, frequency domain analysis is performed on the acoustic emission time series. Specifically, a Fast Fourier Transform (FFT) can be used to convert the microcavitation acoustic emission time series to the frequency domain, and the acoustic emission pulse frequency and average amplitude can be extracted within a preset frequency band. The preset frequency band can be set according to the sensor type and field noise characteristics, for example, 20kHz to 150kHz, or a narrower high signal-to-noise ratio band can be selected based on hardware capabilities. The acoustic emission pulse frequency reflects the frequency of plant hydraulic events per unit time, and the average amplitude reflects the event intensity. Together with ambient light intensity and ambient temperature, they form a multidimensional state feature vector. The ambient light intensity can be measured in the range of 1000 lx to 120000 lx, and the ambient temperature can be measured in the range of 5°C to 50°C. The edge controller preferably normalizes each feature before inputting it into the model.

[0043] The photosynthetic dormancy identification model is preferably a support vector machine (SVM) classifier. The reason for using SVM is that this model has good robustness in agricultural field classification tasks with relatively limited sample size and low feature dimensionality, and is particularly suitable for binary or multi-class classification after fusing acoustic features with environmental data. The physiological state categories include at least photosynthetic midday rest and active transpiration states. In further embodiments, this can be expanded to include multiple output categories such as morning recovery state, normal transpiration state, stress transpiration state, and dormancy state.

[0044] The training process of the photosynthetic dormancy recognition model is as follows. First, historical multidimensional state feature vector samples and corresponding stomatal closure status labels measured synchronously by a stomatal conductance meter are acquired. Stomatal closure status labels can be manually labeled based on conductance thresholds, or classified into three categories: open, partially closed, and closed, based on continuous observation results. Then, using the historical multidimensional state feature vector samples as input and the stomatal closure status labels as output, the hyperplane parameters of the support vector machine classifier are solved using a sequence minimum optimization algorithm to complete the training of the photosynthetic dormancy recognition model. The kernel function can be a radial basis function kernel, a polynomial kernel, or a linear kernel, with a radial basis function kernel being preferred to better fit the nonlinear boundary between plant physiological states and multi-source features. The penalty parameter can be set from 0.1 to 100, and the kernel parameter can be set from 0.0001 to 10, with the optimal combination determined through cross-validation. After training, in actual operation, when the physiological state category is photosynthetic midday dormancy, the system intercepts cold water pulse injection commands and water and fertilizer irrigation commands; when the physiological state category is active transpiration, the cold water pulse injection step and subsequent detection process are triggered. This creates a proactive safety interlock, preventing ineffective or harmful irrigation during periods when the root system's water absorption capacity is significantly reduced.

[0045] After completing the root depth positioning and target infiltration parameter calculation, this invention further optimizes the timing of fertilization. The aim is to address the nutrient competition between leaves and fruits during the peak daytime transpiration period in tomatoes. Simply keeping the fertilizer solution in the active root layer does not necessarily mean that nutrients will preferentially flow to the fruit, because when transpiration pull is strong, water and solutes within the plant are more easily captured by the upper leaves. Therefore, this invention sets up a timing optimization judgment model to identify the time points when the plant's internal suction gradually shifts from being leaf-dominated to meeting the needs of fruit development. Nutrient fertilizer is injected at these points to increase the probability of the fruit obtaining nutrients.

[0046] In one embodiment, generating corresponding control commands to drive intelligent agricultural machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and target irrigation duration includes first acquiring the diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk through a micro-displacement sensor. The minute changes in the diameter of the main stem and the stalk can reflect the water redistribution process within the plant. Generally, during periods of strong transpiration during the day, the main stem tends to shrink, while when the fruit begins to receive more sustained vascular support, the stalk diameter shows a more stable expansion trend. This invention utilizes this microscale dynamic difference to identify the so-called suction switching time point. The suction switching described here is not a physical valve switching, but rather refers to the physiological process in which the driving force dominating the direction of water and nutrient distribution within the plant gradually shifts from leaf transpiration pull to the demand for fruit expansion.

[0047] The optimal timing judgment model is preferably a Long Short-Term Memory (LSTM) network model. The reason for choosing the LSTM network is that the deformation of the main stem and fruit stalk is not a single-point signal that instantaneously determines the irrigation timing, but rather a sequence characteristic with obvious time dependence, requiring judgment based on trend changes over several preceding and following moments. The LSTM network has a memory gating structure, which can better capture the relationship between short-term fluctuations and medium- to long-term trends in the deformation sequence.

[0048] In one embodiment, the Long Short-Term Memory (LSTM) network model includes an input layer, 1 to 3 LSTM hidden layers, and a fully connected output layer, preferably a 2-layer LSTM structure. The number of hidden units in each layer can be set to 32 to 128, preferably 64. The input sequence window length can be set to 10 min to 240 min, preferably 30 min to 120 min; the sampling period can be set to 10 s to 300 s, preferably 30 s to 60 s. The diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk are synchronized in time and input into the hidden layer units of the LSTM network to extract temporal dynamic features. Then, the temporal dynamic features are mapped to an output probability distribution sequence through a fully connected layer. The probability distribution sequence is used to indicate the probability of each time point being in the target timing state. Subsequently, the time point with the highest probability in the output probability distribution sequence indicating that the tomato main stem has stopped contracting and the target fruit stalk has begun to continuously expand is obtained, and this time point is output as the suction switching time node.

[0049] To ensure training quality, the timing optimization judgment model can be trained using historical main stem deformation sequences, fruit stalk deformation sequences, and corresponding manually calibrated time nodes. Manual calibration can be performed by combining fruit turgor pressure changes, observation of continuous deformation curves, or other synchronously collected physiological indicators. The loss function can be cross-entropy loss, the optimizer can be Adam, and the learning rate can be set to 0.0001 to 0.01, preferably around 0.001. To improve model robustness, the input sequence can be detrended, normalized, and time-synchronized corrected. During online system operation, when the current time reaches the suction switching time node, a corresponding control command is generated to drive the intelligent agricultural power machinery to inject nutrient fertilizer solution into the drip irrigation tape according to the target infiltration flow rate and the target irrigation duration, completing the water and fertilizer irrigation operation.

[0050] At the execution level, in one embodiment, the control commands generated by the edge controller include at least pump start / stop commands, variable frequency target values, solenoid valve opening / closing commands, fertilizer pump injection ratio commands, and execution duration parameters. Control of the target infiltration velocity can be achieved by adjusting the variable frequency pump speed and branch valve opening; control of the target irrigation duration can be precisely executed by the timer module. The fertilizer mixing device can mix the mother liquor with irrigation water according to a set ratio. The fertilizer concentration can be set according to the tomato growth stage; for example, the conductivity control range can be set from 1.0 mS / cm to 3.5 mS / cm. To avoid sudden high-concentration input stimulating the rhizosphere, in one embodiment, a low-concentration induction section of 5 to 60 seconds can be set at the beginning of irrigation, and a clear water follow-up section of 5 to 60 seconds can be set at the end of irrigation to make the fertilizer solution front and back edges smoother.

[0051] To ensure long-term stable operation of the system, in one embodiment, a sensor self-check is performed before each detection or irrigation. For soil temperature sensors, the system checks whether baseline drift exceeds a preset threshold, which can be set between 0.5°C and 2°C; for acoustic sensors, it checks whether background noise is above acceptable levels; and for micro-displacement sensors, it checks for zero-point drift and signal interruption. When any critical sensor malfunctions, the system can pause automatic decision-making and switch to a conservative control mode. In conservative control mode, only low-risk irrigation can be performed, without precision fertilization, or the system can simply wait for the next round of valid data collection. This avoids amplifying erroneous decisions due to the inaccuracy of a single sensor.

[0052] The method of this invention can operate in its entirety or be tailored as needed. In a simplified embodiment, only cold water pulse detection, root depth positioning, and water and fertilizer infiltration calculation functions can be used to achieve precise retention irrigation targeting active root layers. In a more complete embodiment, photosynthetic dormancy identification and suction switching timing judgment are further superimposed to simultaneously address irrigation safety and nutrient distribution directionality issues. The former mainly prevents accidental irrigation during periods unsuitable for water absorption, while the latter mainly increases the probability of nutrient transport to the fruit during fertilization. Together with precise root depth positioning, these two methods constitute a decision-making system that is synergistically optimized from spatial, temporal, and physiological safety dimensions.

[0053] This invention also discloses a precision water and fertilizer decision-making system for tomatoes based on feature extraction. The system can be composed of a pulse injection module, a temperature acquisition module, a feature extraction module, a depth positioning module, a moisture content acquisition module, a flow rate calculation module, and an irrigation execution module. In a further embodiment, it also includes a physiological state recognition module and a timing optimization module.

[0054] The pulse injection module is used to inject a preset volume of cold water pulses into the tomato planting soil via drip irrigation tape. This module can consist of a cold water storage unit, a metering pump, a temperature control unit, and an injection valve assembly, ensuring that the cold water temperature and pulse volume meet preset conditions. The temperature control unit can achieve injection temperatures lower than the current soil temperature through a heat exchanger, a small refrigeration unit, or a cold water mixing method.

[0055] The temperature acquisition module is used to collect the temperature time series output by a soil temperature sensor array vertically distributed in the tomato planting soil. This module preferably has a time synchronization function to ensure that the data from each depth measurement point has a unified time reference. The significance of synchronous sampling lies in the fact that the temporal relationship of thermal responses at multiple depths is itself an important clue reflecting cold water propagation and root activity.

[0056] The feature extraction module is used to extract features from the temperature time series to obtain the temperature rise curve features corresponding to each soil depth. This module can integrate a smoothing filter unit, an extreme value detection unit, a rise interval truncation unit, and a slope calculation unit. Modular implementation ensures fast online processing, enabling the system to complete analysis in a short time after a single detection.

[0057] The deep localization module inputs the temperature rise curve features corresponding to each soil depth into a pre-trained root depth localization model, outputting the migration depth of the currently active root system. This module can be deployed locally for inference on edge devices, or deployed in the cloud and return the results to the field controller. In agricultural production sites, edge deployment is preferred to reduce latency caused by network fluctuations.

[0058] The moisture content acquisition module is used to obtain the real-time soil moisture content of the tomato planting soil. This module can use a single-point sensor to read the moisture content, or it can fuse the moisture content from multiple depths and locations. For plots with significant soil heterogeneity, a weighted average or confidence level screening method is preferred to obtain more stable state values.

[0059] The flow velocity estimation module inputs the migration location depth and real-time soil moisture content into a pre-trained water and fertilizer infiltration estimation model based on strategy optimization, and outputs the target infiltration flow velocity and target irrigation duration corresponding to the nutrient fertilizer solution remaining at the migration location depth. Preferably, this module also receives equipment execution constraint parameters, such as the minimum stable speed of the pump, the minimum controllable opening of the valve, and the response delay of the fertilizer pump, so as to output control results that satisfy both agronomic objectives and mechanical capabilities.

[0060] The irrigation execution module generates corresponding control commands to drive intelligent agricultural machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and target irrigation duration. This module may further include a safety interlock unit, a flow closed-loop correction unit, and an execution feedback unit. The flow closed-loop correction unit can fine-tune the frequency converter control value based on flow meter feedback, making the actual flow rate approximate the target flow rate. The execution feedback unit records the actual execution parameters after irrigation, providing a data basis for subsequent model updates.

[0061] In a further embodiment, the physiological state recognition module is used to determine whether the tomato plant is in a photosynthetic midday dormancy state based on the microcavitation acoustic emission time series, ambient light intensity data, and ambient temperature data. When the recognition result indicates a photosynthetic midday dormancy state, the system directly prohibits the pulse injection module and irrigation execution module from operating; when the recognition result indicates an active transpiration state, the system allows subsequent detection and irrigation processes to proceed. The timing optimization module is used to identify suction switching time nodes based on the diameter deformation sequence of the main stem and fruit stalk, thereby determining when to actually issue fertilization and irrigation commands.

[0062] When implementing this invention, it is preferable to operate according to the following timing sequence. For example... Figure 1 As shown, the physiological state recognition module first performs preliminary screening to confirm that the plant is in an active transpiration state, and then the pulse injection module injects cold water pulses. Subsequently, the temperature acquisition module collects temperature time series at various depths, the feature extraction module extracts the temperature rise curve features, and the depth positioning module outputs the current migration depth of the actual active root system. Next, the moisture content acquisition module reads the real-time soil moisture content, and the flow rate calculation module outputs the target infiltration flow rate and target irrigation duration. Based on this, the timing optimization module continuously monitors the deformation of the main stem and fruit stalk. Once the current time reaches the suction switching time node, the irrigation execution module drives the power machinery to implement precise water and fertilizer irrigation according to the aforementioned parameters.

[0063] It should be noted that the above parameter ranges and model structures are preferred embodiments of the present invention, intended to make the technical solution feasible and stable, and do not constitute a limitation on the scope of protection. Without altering the core concept of the present invention—locating active roots through cold water pulse thermal shock response, deducing infiltration control parameters based on root location and water content, and implementing precise water and fertilizer irrigation in conjunction with plant physiological state and internal suction switching timing—equivalent substitutions and conventional adjustments to sensor placement, model layer number, parameter range, training sample source, and actuator form should all fall within the scope of protection of the present invention.

Claims

1. A method for precise water and fertilizer decision-making in tomatoes based on feature extraction, characterized in that, include: A preset volume of cold water pulses is injected into the tomato planting soil via drip irrigation tape. The temperature of the cold water pulses is lower than the current soil temperature of the tomato planting soil. The temperature time series output by the soil temperature sensor array vertically distributed in the tomato planting soil was collected; Feature extraction is performed on the temperature time series to obtain the temperature recovery curve features corresponding to each soil depth; The temperature rise curve features corresponding to each soil depth are input into the pre-trained root depth localization model, and the migration depth of the current real active root system is output. Obtain the real-time soil moisture content of the tomato planting soil; The migration location depth and the real-time soil moisture content are input into a pre-trained water and fertilizer infiltration estimation model based on strategy optimization, and the target infiltration velocity and target irrigation duration that make the nutrient fertilizer solution remain at the migration location depth are output. The corresponding control commands are generated to drive the intelligent agricultural power machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and the target irrigation duration.

2. The method for precise water and fertilizer decision-making in tomatoes based on feature extraction according to claim 1, characterized in that, The steps for extracting features from the temperature time series to obtain the temperature recovery curve features corresponding to each soil depth include: Within a preset time window after the injection of the cold water pulse, the temperature minimum point in the temperature time series is extracted as the bottoming feature point. Extract sequence data from the bottoming feature point to the initial temperature range before the cold water pulse was injected; The first derivative of the sequence data is performed to obtain the maximum temperature rise slope at the corresponding soil depth; The maximum temperature rise slope is combined with the bottoming feature point to form the temperature rise curve feature.

3. The method for precise water and fertilizer decision-making for tomatoes based on feature extraction according to claim 1, characterized in that, The root system depth localization model is a multilayer perceptron network; the training process of the root system depth localization model includes: Acquire a historical sample dataset, which includes historical temperature rise curve features and actual active root depth labels calibrated by an in-situ root observation device. The historical temperature recovery curve features are input into the initial multilayer perceptron network, and the predicted root depth is output. Calculate the mean squared error loss value between the predicted root depth and the actual active root depth label; The weight parameters of the initial multilayer perceptron network are updated by backpropagation using the gradient descent algorithm until the mean squared error loss value converges, thus obtaining the pre-trained root system depth localization model.

4. The method for precise water and fertilizer decision-making for tomatoes based on feature extraction according to claim 1, characterized in that, The pre-trained water and fertilizer infiltration estimation model based on policy optimization is a deep deterministic policy gradient reinforcement learning model. The state space input of the deep deterministic policy gradient reinforcement learning model includes the migration location depth and the real-time soil moisture content; The action space output of the deep deterministic policy gradient reinforcement learning model is the target infiltration velocity and the target irrigation duration; The reward function of the deep deterministic policy gradient reinforcement learning model is set as the weighted sum of the overlap between the actual residence depth of the nutrient solution and the migration location depth and the amount of water and fertilizer consumed.

5. The method for precise water and fertilizer decision-making for tomatoes based on feature extraction according to claim 1, characterized in that, Prior to the step of injecting a predetermined volume of cold water pulses into the tomato planting soil via drip irrigation, the process also includes: The time series of microcavitation acoustic emission was collected by an acoustic sensor attached to the base of the tomato plant stem. Collect ambient light intensity and ambient temperature data of the environment in which the tomato plants are located; The microcavitation acoustic emission time series, the ambient light intensity data, and the ambient temperature data are concatenated into a multidimensional state feature vector. The multidimensional state feature vector is input into a pre-trained photosynthetic dormancy recognition model, which outputs the physiological state category of the tomato plant. When the physiological state category is photosynthetic midday rest state, intercept the cold water pulse injection command and the water and fertilizer irrigation command; When the physiological state category is active transpiration, the step of injecting a preset volume of cold water pulse into the tomato planting soil through the drip irrigation tape is triggered.

6. The method for precise water and fertilizer decision-making for tomatoes based on feature extraction according to claim 5, characterized in that, The step of concatenating the microcavitation acoustic emission time series, the ambient light intensity data, and the ambient temperature data into a multidimensional state feature vector includes: The microcavitation acoustic emission time series was converted to the frequency domain using Fast Fourier Transform; Extract the frequency and average amplitude of acoustic emission pulses within a preset frequency band; The acoustic emission pulse frequency, the average amplitude, the ambient light intensity data, and the ambient temperature data are combined to form the multidimensional state feature vector.

7. The method for precise water and fertilizer decision-making for tomatoes based on feature extraction according to claim 5, characterized in that, The photosynthetic dormancy recognition model is a support vector machine classifier; the training process of the photosynthetic dormancy recognition model includes: Acquire historical multidimensional state feature vector samples, as well as the corresponding stomatal closure state labels measured synchronously by a stomatal conductance meter; The historical multidimensional state feature vector sample is used as input, and the stoma closure state label is used as output. The hyperplane parameters of the support vector machine classifier are solved using the sequence minimum optimization algorithm to complete the training of the photosynthetic dormancy recognition model.

8. The method for precise water and fertilizer decision-making for tomatoes based on feature extraction according to claim 1, characterized in that, The step of generating corresponding control commands to drive intelligent agricultural power machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and the target irrigation duration includes: Obtain the diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk, which are collected by a micro-displacement sensor. The diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk are input into a pre-trained timing optimization judgment model to output the suction switching time node. When the system reaches the suction switching time node, the corresponding control command is generated to drive the intelligent agricultural power machinery to inject the nutrient fertilizer solution into the drip irrigation tape according to the target infiltration flow rate and the target irrigation duration, thereby completing the water and fertilizer irrigation operation.

9. A method for precise water and fertilizer decision-making in tomatoes based on feature extraction according to claim 8, characterized in that, The timing optimization judgment model is a long short-term memory network model; the step of inputting the diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk into the pre-trained timing optimization judgment model and outputting the suction switching time node includes: The diameter deformation sequence of the tomato main stem and the diameter deformation sequence of the target fruit stalk are fed into the hidden layer unit of the long short-term memory network to extract temporal dynamic features. The temporal dynamic features are mapped into an output probability distribution sequence through a fully connected layer; Obtain the timestamp in the output probability distribution sequence that indicates the tomato main stem has stopped contracting and the target fruit stalk has begun to swell continuously; The timestamp is output as the suction switching time node.

10. A feature extraction-based precision water and fertilizer decision-making system for tomatoes for implementing the method of claim 1, characterized in that, include: The pulse injection module is used to inject a preset volume of cold water pulse into the tomato planting soil through a drip irrigation tape. The temperature of the cold water pulse is lower than the current soil temperature of the tomato planting soil. The temperature acquisition module is used to acquire the temperature time series output by the soil temperature sensor array that is vertically distributed in the tomato planting soil. The feature extraction module is used to extract features from the temperature time series to obtain the temperature recovery curve features corresponding to each soil depth. The depth localization module is used to input the temperature rise curve features corresponding to each soil depth into the pre-trained root depth localization model and output the migration location depth of the current real active root system. A moisture content acquisition module is used to acquire the real-time soil moisture content of the tomato planting soil; The flow velocity estimation module is used to input the migration location depth and the real-time soil moisture content into a pre-trained water and fertilizer infiltration estimation model based on strategy optimization, and output the target infiltration flow velocity and target irrigation duration that make the nutrient fertilizer solution remain at the migration location depth. The irrigation execution module is used to generate corresponding control commands to drive intelligent agricultural power machinery to perform water and fertilizer irrigation operations according to the target infiltration flow rate and the target irrigation duration.

Images