Agricultural health monitoring system and method fusing multispectrum and AI model

Abstract

The invention relates to the technical field of agricultural information, in particular to an agricultural health monitoring system and method fusing multispectral and AI models. The system comprises a multispectral acquisition unit, a double-feedback linkage control circuit and an AI processing unit, wherein the double-feedback linkage control circuit senses the ambient light intensity and the AI load in real time, generates a cooperative control instruction, and dynamically adjusts the acquisition frequency and the spectral band; and the AI processing unit executes health state reasoning based on the optimized multispectral data, and feeds back a performance index through a closed-loop mechanism to iteratively update a sensing threshold. The invention provides an agricultural health monitoring system and method fusing multispectrum and an AI model, and solves or at least alleviates the problems of low monitoring precision, high response delay, high omission ratio and the like caused by ambient illumination fluctuation and limited edge computing resources.

Description

Technical Field

[0001] This invention relates to the field of agricultural information technology, and in particular to an agricultural health monitoring system and method that integrates multispectral and AI models. Background Technology

[0002] With the deepening practice of precision agriculture and the continuous evolution of intelligent sensing technology, crop health monitoring systems integrating multispectral imaging and artificial intelligence (AI) models have become a key infrastructure for modern farmland management. These systems capture the reflectance spectral characteristics of crops in the visible to near-infrared bands and combine them with deep learning algorithms to non-contactly assess physiological indicators such as chlorophyll content, water stress, and pest and disease infestation, significantly improving the scientific rigor and timeliness of agricultural decision-making. The current mainstream architecture generally consists of three parts: multiple sensors, edge computing units, and a dedicated AI health assessment model. Its initial design purpose was to achieve a closed-loop logic of "perception-processing-feedback," and it has already demonstrated considerable application value under specific lighting and load conditions.

[0003] However, as field applications become increasingly complex, the demands on the real-time performance and robustness of monitoring systems are constantly rising, revealing some deep-seated structural contradictions in existing technologies at the fundamental level. Specifically, while multispectral sensors can provide rich spectral dimension information, their data quality is highly dependent on ambient lighting conditions; edge AI inference can reduce reliance on the cloud, but it is limited by the limited computing resources of embedded platforms. In this situation, traditional systems typically adopt a static parameter configuration strategy, i.e., a fixed acquisition frequency and synchronous acquisition of data across the entire spectrum. Although this can barely maintain performance under simulation conditions, its shortcomings become apparent in real-world farmland scenarios, especially when the combined requirements of low illumination (such as in cloudy or rainy weather, or in seedling greenhouses) and high sampling frequencies across the entire spectrum (such as 10Hz) suddenly become apparent. Low illumination leads to a sharp decrease in the signal-to-noise ratio, resulting in a sharp decrease in key spectral signals. Chlorophyll estimation error rates reach 38.7% when light intensity is below 500 lux. The large amount of data generated by the high sampling frequency across the entire band exceeds the processing capacity of edge devices, with CPU utilization often exceeding 85% and delays exceeding 2 seconds, seriously affecting the timeliness of monitoring.

[0004] Furthermore, existing improvement solutions fail to address the comprehensive trade-off between environmental adaptability and resource efficiency. Some patents attempt to dynamically adjust parameters at the software layer, such as CN114021632A adjusting transmission volume based on network bandwidth, or CN113873456B adjusting acquisition frequency based on light intensity. However, these methods still treat environmental perception and system load as isolated variables, resulting in fragmented and lagging parameter control logic. Software judgment involves three steps: acquisition, algorithm analysis, and command issuance, introducing an average delay of over 120ms across all scenarios. The multispectral data stream and AI inference process are out of sync, and under high load, additional computational overhead creates a vicious cycle. To make matters worse, the one-dimensional adjustment logic of these methods cannot adapt to low light + high load conditions. Reducing the frequency to reduce computational pressure makes it easy to miss low-concentration signals in the early stages of disease, meaning the data timeliness requirement cannot be met; increasing the band to improve the signal-to-noise ratio still results in system overload and crash under high-frequency acquisition, meaning the requirement for switching operating conditions cannot be met. The current solution has a health detection miss rate of 42.3% under conditions of light intensity ≤500 lux and CPU utilization ≥80%. Hardware solutions that rely on manually preset modes (such as "sunny mode" and "cloudy mode") lack adaptive capabilities, and mode switching inevitably leads to monitoring interruption and loss of critical growth period data.

[0005] Fundamentally, the aforementioned challenges arise because current systems lack a linkage mechanism capable of simultaneously sensing both environmental and system load variables in real time at the hardware level, and further driving the optimization of the acquired frequency spectrum. Purely software-based control methods inevitably lead to high latency and competition for computational resources, while independent, modular parameter adjustment methods cannot achieve Pareto optimality under multiple system constraints. Therefore, there is an urgent need to research a low-latency, highly coupled hardware-level reaction control architecture that enables multispectral acquisition strategies to adjust the dynamic coupling state under the reaction of illumination conditions and AI inference load, and to combine this with a closed-loop threshold optimization mechanism to cope with such long-term environmental changes. This would break through the performance bottleneck of current multispectral agricultural health monitoring systems, achieving a trinity of high precision, low latency, and high energy efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide an agricultural health monitoring system and method that integrates multispectral and AI models, which can solve or at least alleviate the problems of decreased monitoring accuracy, system response delay and high data miss rate caused by fluctuations in ambient light conditions and limited edge computing resources in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an agricultural health monitoring system integrating multispectral and AI models, comprising a multispectral acquisition unit, a dual-feedback linkage control circuit, and an AI processing unit; the multispectral acquisition unit is used to acquire reflectance spectral data of crops in the visible to near-infrared band; the dual-feedback linkage control circuit is used to sense ambient light intensity and the computational load of the AI ​​processing unit in real time, and generate collaborative control instructions based on preset logic rules; the AI ​​processing unit is used to receive multispectral data after parameter optimization, perform agricultural health status inference, and provide feedback performance indicators to dynamically adjust the sensing threshold; wherein, the dual-feedback linkage control circuit is composed entirely of discrete logic devices such as combinational logic gates and sequential logic circuits (such as flip-flops, registers, etc.), and does not rely on microcontroller or operating system scheduling, and the end-to-end delay from the change of ambient light intensity or CPU utilization to the completion of adjustment of multispectral acquisition parameters does not exceed 8ms.

[0008] To further realize the present invention, the following technical solutions may be preferred:

[0009] Preferably, the multispectral acquisition unit includes a multispectral sensor, a clock control circuit, and a band selection circuit. The multispectral sensor is a silicon-based CMOS image sensor with multiple built-in optical filter channels. The independent controllability of each channel is achieved by the band selection circuit, and the adjustable frame rate supported by the multispectral sensor is achieved by the clock control circuit. The clock control circuit is composed of a programmable frequency divider, with its input connected to the main crystal oscillator and its output connected to the frame synchronization pin of the multispectral sensor. The acquisition frequency is controlled by adjusting the frequency division coefficient. The band selection circuit consists of multiple MOSFET switches, each MOSFET switch being connected in series on the power line of the corresponding spectral channel. It is controlled by the band control signal output by the dual feedback linkage control circuit to achieve independent conduction or deactivation of the channel.

[0010] Preferably, the dual-feedback linkage control circuit includes a scene perception module, a load monitoring module, and a collaborative control module; the scene perception module includes a digital ambient light sensor and its peripheral circuits, as well as a threshold comparison circuit; the digital ambient light sensor periodically samples the ambient light intensity; the threshold comparison circuit compares the real-time light intensity with a preset low-light threshold and outputs a scene signal; the load monitoring module includes a register reading circuit and a load judgment logic unit; the register reading circuit periodically reads the CPU utilization rate of the AI ​​processing unit; the load judgment logic unit compares the CPU utilization rate with a preset high-load threshold and outputs a load signal.

[0011] Preferably, the collaborative control module includes multiple two-input AND gates, a data selector, and an instruction encoder; the inputs of the two-input AND gates are logical combinations of scene signals and load signals, respectively; the outputs of each AND gate are connected to the selection terminal of the data selector; the data input terminal of the data selector is preset with instruction encoding; the instruction encoder decodes the selected instruction into a frequency control signal and a band control signal.

[0012] Preferably, the collaborative control module outputs corresponding frequency control signals and band control signals based on the combined state of the scene signal and the load signal; when in a low-light and high-load condition, it outputs a control signal to reduce the acquisition frequency and turn off some spectral channels; when in a low-light and low-load condition, it outputs a control signal to maintain the current acquisition frequency but turn off some spectral channels; when in a normal-light and high-load condition, it outputs a control signal to reduce the acquisition frequency and maintain full-band acquisition; when in a normal-light and low-load condition, it outputs a control signal to maintain the current acquisition frequency and full-band acquisition.

[0013] Preferably, the AI ​​processing unit includes an edge computing chip, an agricultural health monitoring AI model, and a threshold feedback optimization module; the edge computing chip communicates with the multispectral acquisition unit via a high-speed serial bus; the agricultural health monitoring AI model is a convolutional neural network, the input layer receives a normalized multispectral image tensor, and the network structure sequentially includes multiple convolutional blocks, a global average pooling layer, and a fully connected output layer, outputting crop health status indicators.

[0014] Preferably, each convolutional block of the agricultural health monitoring AI model consists of a convolutional layer, a batch normalization layer, and an activation function in sequence; the fully connected output layer is trained using a weighted loss function to optimize the prediction accuracy of chlorophyll content, water stress index, and disease probability.

[0015] Preferably, the threshold feedback optimization module periodically calculates historical inference performance indicators; if the average inference accuracy in low-light scenarios is lower than a predetermined threshold, the low-light threshold is adjusted; if the average system response latency under high load exceeds a predetermined threshold, the high-load threshold is adjusted; the threshold update is written to the threshold registers in the scene perception module and the load monitoring module through the communication bus, and the update process is completed within a predetermined time and does not affect the normal operation of the system.

[0016] Preferably, multiple systems are deployed in cascade via a communication bus; one is a master node and the rest are slave nodes; the master node performs threshold feedback optimization and sends the updated threshold parameters to all slave nodes via a broadcast frame; upon receiving the broadcast, each slave node immediately updates its local threshold register to ensure the consistency of the monitoring strategy.

[0017] An agricultural health monitoring method applicable to the above system includes the following steps:

[0018] S1. Real-time acquisition of ambient light intensity and CPU utilization of the AI ​​processing unit;

[0019] S2. Based on the comparison results between ambient light intensity and preset low light threshold, and CPU utilization rate and preset high load threshold, generate corresponding frequency control signals and band control signals;

[0020] S3. Adjust the acquisition frequency according to the frequency control signal, and dynamically enable or disable the spectral channel according to the band control signal;

[0021] S4. Receive the optimized multispectral data, perform forward inference, and output the crop health status;

[0022] S5. Periodically analyze historical performance indicators. If preset conditions are met, update the low light threshold and high load threshold to achieve closed-loop adaptive optimization.

[0023] The beneficial effects of this invention are:

[0024] This invention, by constructing a dual-feedback linkage control circuit, achieves real-time perception, logical linkage, and coordinated parameter adjustment of ambient light intensity and AI inference load, solving the performance bottleneck caused by high software control latency and fragmented parameter adjustment in existing technologies. Under combined conditions of low light intensity and high CPU utilization, the crop health status false negative rate and average system latency of this invention are both low. Furthermore, the closed-loop threshold optimization mechanism enables the system to have long-term adaptive capabilities, coping with continuous environmental and load evolution without manual intervention, thus improving robustness and energy efficiency. Attached Figure Description

[0025] Figure 1 A schematic diagram of the overall structure of the system of the present invention;

[0026] Figure 2 A circuit block diagram of the multispectral acquisition unit of the present invention;

[0027] Figure 3 A schematic diagram of the logic architecture of the dual-feedback linkage control circuit of the present invention;

[0028] Figure 4 A schematic diagram showing the connection relationship between the logic gates and data selectors of the collaborative control module of the present invention;

[0029] Figure 5 A table showing the correspondence between control instruction codes and hardware output signals under four scenarios and load combinations according to the present invention;

[0030] Figure 6 A block diagram of the AI ​​processing unit of the present invention;

[0031] Figure 7A schematic diagram of the network structure of the agricultural health monitoring AI model of the present invention;

[0032] Figure 8 A flowchart illustrating the threshold feedback optimization module of this invention;

[0033] Figure 9 A schematic diagram of the working state of the system of the present invention under the combined conditions of low light and high load;

[0034] Figure 10 A schematic diagram of the system topology for multi-device cascaded deployment according to the present invention. Detailed Implementation

[0035] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1

[0038] This embodiment discloses an agricultural health monitoring system that integrates multispectral and AI models. (Refer to...) Figure 1 The system block diagram shown is composed of three main parts: a multispectral acquisition unit, a dual feedback linkage control circuit, and an AI processing unit. Each part is tightly coupled through a hardware-level low-latency path to form an adaptive control system.

[0039] First, the multispectral acquisition unit is used to acquire the reflectance spectral information of crops in the visible to near-infrared range. (Refer to...) Figure 2 The unit includes a multispectral sensor, a clock control circuit, and a band selection circuit.

[0040] The multispectral sensor employs a silicon-based CMOS image sensor chip, with a photosensitive area covering a wavelength range of 400nm to 1000nm. It incorporates a six-channel optical filter array, corresponding to six independent spectral channels: blue light (center wavelength 450nm, bandwidth ±20nm), green light (550±20nm), red light (650±20nm), red edge (710±10nm), near-infrared I (800±30nm), and near-infrared II (900±30nm). Each channel has an independent photoelectric conversion and signal readout path. The pixel array is 1280×960, supporting a maximum frame rate of 10Hz. The sensor output is quantized by an onboard 12-bit ADC and then sent to subsequent processing links via a parallel data bus.

[0041] The clock control circuit consists of a programmable frequency divider. Its input signal is a connected temperature-compensated crystal oscillator, and the output of the divider is connected to the frame synchronization (FSYNC) pin of the multispectral sensor. The divider is a 16-bit register, and the division factor N can be set between 100 and 10000, corresponding to an output frame synchronization signal frequency f = 10⁷ / NHz. This means the sampling frequency can be continuously adjusted between 1Hz and 10Hz. For example, when N = 1000, f = 10Hz; when N = 2000, f = 5Hz; and when N = 3333, f ≈ 3Hz. The divider receives a frequency control signal F_ctrl from the dual-feedback linkage control circuit. This signal is a pulse signal; to reach a target division factor, it indicates the factor that the current division factor should be multiplied by. After receiving a valid control signal, the divider completes the change within the next crystal oscillator cycle.

[0042] The band selection circuit consists of six N-channel MOSFET switches. The drain of each MOSFET is connected to the power rail (VDD_channel) of the corresponding spectral channel, the source is grounded, and the gate receives the corresponding signal bit of the band control signal B_ctrl from the dual feedback linkage control circuit. B_ctrl is a six-bit parallel level signal, with each bit independently controlling the power supply state of a channel: when a bit of B_ctrl is high (3.3V), the corresponding MOSFET is turned on, and the channel is powered on; when it is low (0V), the MOSFET is turned off, and the channel is powered off. This architecture allows arbitrary spectral channels to be dynamically turned on or off during system operation, thereby changing the system's dimensionality and signal-to-noise ratio.

[0043] Reference Figure 3, The dual-feedback linkage control circuit is used to sense the external light environment and internal computing load in real time, and generate cooperative control instructions accordingly, and generate linkage control signals. The dual-feedback linkage control circuit is composed of discrete logic devices, does not rely on any microcontroller or operating system scheduling, and makes the end-to-end response delay less than 8 ms. There are three sub-units in the dual-feedback linkage control circuit: a scene sensing module, a load monitoring module, and a cooperative control module.

[0044] The scene sensing module consists of a digital ambient light sensor BH1750 and its peripheral circuit. Its sampling period is 1 second. After each sampling, 12-bit valid data is written into the internal data register. The output of this register is connected to a threshold comparison circuit, which consists of an 8-bit non-volatile register (storing the preset low-light threshold L_th) and a 12-bit digital comparator (such as 74HC85). The digital comparator numerically compares the received real-time light intensity L with L_th in real time: if L ≤ L_th, then the output S_light = 0 (low level, indicating a low-light scene); if L > L_th, then S_light = 1 (high level, indicating a normal-light scene). In the initial state, L_th is configured as 500 lux and can be dynamically adjusted through subsequent closed-loop optimization mechanisms.

[0045] The load monitoring module is used to monitor the real-time computing load of the AI processing unit. This module includes a register reading circuit and a load judgment logic unit. The register reading circuit reads the CPU idle time (idle) and total time (total) every 500 ms, and calculates the CPU occupancy rate U = (total - idle) / total × 100%. This calculation is implemented by a dedicated hardware state machine, and the calculation result U is sent to the load judgment logic unit in 8-bit binary form. The latter also consists of an 8-bit threshold register (storing the high-load threshold U_th) and an 8-bit digital comparator. When U ≥ U_th, the output S_load = 0 (low level, indicating high load); when U < U_th, S_load = 1 (high level, indicating low load). The initial U_th is set to 80% and can also be updated through the closed-loop mechanism.

[0046] Refer to Figure 4The collaborative control module receives two binary signals, S_light and S_load, and generates corresponding control commands based on their combination. This module consists of four two-input AND gates (e.g., 74HC08), one 4-to-1 data selector (e.g., 74HC153), and two command encoders. The inputs to the four AND gates are: first AND gate (¬S_light, ¬S_load), second AND gate (¬S_light, S_load), third AND gate (S_light, ¬S_load), and fourth AND gate (S_light, S_load). Each AND gate outputs a high level only when its corresponding scenario-load combination is true, serving as the selection signal S1S0 for the data selector. (Refer to...) Figure 5 The four data input terminals D0–D3 of the data selector are preset with instruction codes “00”, “01”, “10”, and “11”, respectively, corresponding to four operating conditions:

[0047] D0 (S1S0=00): Normal light + low load (S_light=1, S_load=1)

[0048] D1 (S1S0=01): Normal light + high load (S_light=1, S_load=0)

[0049] D2 (S1S0=10): Low light + low load (S_light=0, S_load=1)

[0050] D3 (S1S0=11): Low light + high load (S_light=0, S_load=0)

[0051] The data selector outputs a two-bit instruction code, which is then fed into the instruction encoder. The instruction encoder is implemented using combinational logic circuitry, mapping the two-bit instruction code into two independent outputs: a frequency control signal F_ctrl (used to set the frequency division factor) and a band control signal B_ctrl (a six-bit parallel signal).

[0052] In the system, different commands correspond to different operations. Command "00" keeps F_ctrl at its current division ratio while B_ctrl is set to "111111", meaning all channels are enabled. Command "01" causes F_ctrl to output a pulse sequence, setting the division ratio to 2000 (5Hz), and B_ctrl is also set to "111111". Command "10" keeps F_ctrl at its current division ratio and sets B_ctrl to "001111", meaning only the red light, red edge, near-infrared 1, and near-infrared 2 channels are enabled. Command "11" causes F_ctrl to output a pulse sequence, changing the division ratio to 3333 (3Hz), and B_ctrl is set to "001111".

[0053] Through this logic, the system can automatically switch to the optimal acquisition strategy under different environmental and load conditions. For example, in complex situations with low light and high load (corresponding to instruction "11"), the system will reduce the acquisition frequency to 3Hz, shut down the blue and green light channels, and retain only the red and near-infrared bands that are sensitive to chlorophyll and water. This ensures the acquisition of key physiological information while reducing the burden of data transmission and processing.

[0054] The AI ​​processing unit is responsible for receiving optimized multispectral data, performing crop health status inference, and participating in closed-loop threshold optimization. (Refer to...) Figure 6 This unit includes an edge computing chip, an agricultural health monitoring AI model, and a threshold feedback optimization module.

[0055] The edge computing chip communicates with the multispectral acquisition unit via the SPI bus. The SPI is configured in master mode with a clock frequency of 25MHz, CPOL=0, CPHA=0, and a data frame length of 32 bits. Multispectral image data is transmitted in DMA mode, with a single frame of 1280×960×12bit data transmission taking approximately 980μs, meeting the latency requirement of less than 1ms.

[0056] Reference Figure 7 The agricultural health monitoring AI model is a customized convolutional neural network, and its structure is as follows: Figure 7 As shown. The input layer receives a normalized multispectral image tensor with a size of 128×128×C, where C is the number of active channels (2, 4, or 6, depending on the B_ctrl configuration). A preprocessing module is set up in the DMA buffer of the AI ​​processing unit, and its execution flow is as follows: The active channel index is determined based on the B_ctrl value; when C=4, channels 3 (red light), 4 (red edge), 5 (near-infrared I), and 6 (near-infrared II) are extracted from the original six-channel data, and the remaining channels are discarded; when C=2, channels 3 (red light) and 6 (near-infrared II) are extracted, and the pixel values ​​of channels 4 (red edge) and 5 (near-infrared I) are weighted and averaged 1:1 to merge into a new channel; bilinear interpolation is performed on the retained channels to scale the image size to 128×128. Normalization uses channel-level Z-score standardization, and the mean and standard deviation are derived from training set statistics.

[0057] The main body of the network consists of three convolutional blocks. The first convolutional block contains 32 3×3 convolutional kernels with a stride of 1 and padding of 1, followed by batch normalization (BatchNorm) and ReLU activation. The second convolutional block contains 64 3×3 convolutional kernels with the same configuration. The third convolutional block contains 128 3×3 convolutional kernels. A global average pooling layer is then applied to compress the spatial dimension to 1×1×128. Finally, the fully connected layer outputs a three-dimensional vector: chlorophyll content (SPAD value, range 0–80), water stress index (WSI, range 0–1), and disease probability (0–1).

[0058] The model was trained using a dataset containing 120,000 annotated field images, covering major crops such as corn, wheat, and rice. The annotations included measured SPAD values, soil moisture sensor readings, and expert diagnoses. The loss function was a weighted combination.

[0059]

[0060] Where MSE is the mean squared error; CE is the cross-entropy loss; The actual SPAD value is the measured chlorophyll content index (SPAD value) obtained from the field-labeled images. The SPAD value predicted by the model is the model's estimated output of chlorophyll content; This is the actual WSI value, representing the soil moisture content; The WSI value predicted by the model is the model's estimated output of soil moisture. This is a true disease label, i.e., the diagnosis result of plant protection experts; The disease classification probabilities predicted by the model are the disease category probability distribution output by the model. For training, we used the Adam optimizer with an initial learning rate of 0.01% (1e-4), a batch size of 32 per training iteration, and a total of 150 training rounds. On the validation set, SPAD achieved an R² of 0.89, WSI achieved an R² of 0.85, and the disease classification accuracy was 92.3%.

[0061] Reference Figure 8 and Figure 9The threshold feedback optimization module automatically starts at 24:00 every day to perform closed-loop threshold iteration. This module reads the system logs from the past 24 hours and extracts three types of metrics. The first is the average inference accuracy A_light during the period S_light=0, which is based on the accuracy of the validation set; the second is the system response latency D_load during the period S_load=0, which is the time taken from data acquisition to MQTT upload completion; and the third is the ratio of effective inference frames to total acquisition frames under each working condition. If A_light is less than 90%, L_th is updated to L_th×0.9; if D_load exceeds 1000ms, U_th is updated to U_th×0.95. The update command is written to the threshold register (address 0x40) of the scene awareness module and the threshold register (address 0x41) of the load monitoring module via the I²C bus. The write operation is completed within 50ms. During this period, the system continues to run according to the original parameters to ensure uninterrupted service.

[0062] As a preferred implementation, the entire dual-feedback linkage control circuit adopts an integrated design, and the logic gates, comparators, and registers are all made of industrial-grade 74HC series components. The power supply is provided by an LDO regulator at 3.3V, with a ripple of less than 10mV.

[0063] The band selection strategy is designed based on the spectral reflectance characteristics of crops. Under low light conditions (illuminance less than 500 lux), the signal-to-noise ratio (SNR) of the sensor will significantly decrease in the blue (450 nm) and green (550 nm) bands due to the low solar radiation flux. According to laboratory calibration data (illuminance 300 lux, sensor gain = 1), the SNR of the blue channel is 8.3 ± 0.5 dB, and the SNR of the red channel is 21.7 ± 0.8 dB. Field trials also show that the error in estimating chlorophyll using all six channels is 38.7%, while using only the red / red-edge / near-infrared channels can reduce the error to 12.4% (sample size n = 12, p < 0.01). At the same time, the data volume is reduced from approximately 10 MB / frame for 6 channels × 1280 × 960 × 12 bits to approximately 7.1 MB / frame for 4 channels × the same specification, a reduction of 33%, which greatly alleviates the pressure on the SPI bus and memory bandwidth.

[0064] The frequency adjustment strategy is well-matched to the computational complexity of the AI ​​model. When running this CNN model on an embedded platform, a single frame inference averages 78% of CPU resources. If data is collected at a frequency of 10Hz, the data arrival rate is 10 frames / second, but the processing capacity is only about 5.5 frames / second (183ms / frame), which will lead to queue backlog and the latency will quickly rise to more than 2 seconds. When the frequency is reduced to 5Hz (corresponding to instruction "01"), the data rate and processing capacity can be basically balanced, and the latency can be stabilized at 480-520ms. Under extreme combined conditions (low light plus high load), further reducing the frequency to 3Hz (corresponding to instruction "11") can control the latency to within 420ms while still ensuring continuous monitoring of key indicators (SPAD, WSI).

[0065] The system supports multi-device deployment. (Refer to...) Figure 10 Multiple monitoring nodes are connected via an RS-485 bus (using a MAX3485 transceiver). The master node (ID=0) is responsible for performing threshold feedback optimization. After updating daily, it broadcasts the new L_th and U_th values ​​to all slave nodes (ID=1-N) via a broadcast frame. Upon receiving the broadcast, the slave nodes immediately update their local threshold registers. This mechanism ensures consistent strategy across the entire field, preventing monitoring deviations caused by differences in local environments.

[0066] Example 2

[0067] This embodiment discloses an agricultural health monitoring method applicable to the system in Embodiment 1, the method comprising the following steps:

[0068] Step 1: Real-time acquisition of ambient light intensity and CPU utilization of the AI ​​processing unit

[0069] First, the ambient light intensity is acquired in real time through the scene perception module. The digital ambient light sensor starts sampling the light intensity every 1 second, reading the current illuminance value. After D / A conversion, an analog voltage signal is generated, which is then filtered by an RC low-pass filter with a cutoff frequency of 0.5Hz to remove high-frequency noise before being fed into the non-inverting input of a comparator. The inverting input of the comparator is connected to a reference voltage set by a threshold register. When the ambient illuminance is less than or equal to the low-light threshold, the comparator outputs a low-level signal as the scene signal S_light=0; when the ambient illuminance is greater than the low-light threshold, it outputs a high-level signal as S_light=1.

[0070] Meanwhile, the CPU utilization of the AI ​​processing unit is acquired in real time through the load monitoring module. The register read circuit reads the CPU idle time and total time every 500ms and calculates the CPU utilization U=(total-idle) / total×100%. This calculation result is sent to the load judgment logic unit. When the CPU utilization is greater than or equal to the high load threshold, a low-level signal is output as S_load=0; when the CPU utilization is less than the high load threshold, a high-level signal is output as S_load=1.

[0071] Step 2: Based on the comparison results of ambient light intensity and preset low light threshold, and CPU utilization and preset high load threshold, generate frequency control signal and band control signal.

[0072] After acquiring ambient light intensity and CPU utilization data, this method compares and judges based on preset thresholds to generate corresponding control signals. The collaborative control module receives two binary signals, scene signal S_light and load signal S_load, and generates corresponding control commands based on their combined state.

[0073] The collaborative control module consists of combinational logic gates and data selectors, and internally contains four linkage rules. When S_light=0 and S_load=0, the output F_ctrl indicates that the division factor is set to 3333 (corresponding to 3Hz), and B_ctrl outputs "001111" (turning off the blue and green light channels and turning on the other four channels); when S_light=0 and S_load=1, F_ctrl maintains the current division factor, and B_ctrl outputs "001111"; when S_light=1 and S_load=0, F_ctrl indicates that the division factor is set to 2000 (corresponding to 5Hz), and B_ctrl outputs "111111"; when S_light=1 and S_load=1, both F_ctrl and B_ctrl maintain their original states, and the system collects data from all six bands at a frequency of 10Hz.

[0074] Step 3: The multispectral acquisition unit adjusts the acquisition frequency according to the frequency control signal and dynamically enables or disables spectral channels according to the band control signal.

[0075] After generating the control signal, this method performs parameter adjustments through a multispectral acquisition unit. Specifically, the clock control circuit receives the frequency control signal F_ctrl and sets the frequency division coefficient N according to the instruction, so that the output frame synchronization signal period T_frame = N / 10. 7Seconds. When F_ctrl corresponds to 3Hz, N=3333, T_frame≈333.3ms; when it corresponds to 5Hz, N=2000, T_frame=200ms; when it corresponds to 10Hz, N=1000, T_frame=100ms. The frame synchronization signal is directly connected to the FSYNC pin of the multispectral sensor, triggering the exposure and readout operations for each frame of the image.

[0076] Meanwhile, the band selection circuit receives the band control signal B_ctrl. When a certain bit is high (3.3V), the corresponding relay is turned on, and the channel of that band is activated; when it is low (0V), the relay is turned off, and the channel is closed.

[0077] Step 4: The AI ​​processing unit receives the optimized multispectral data, performs forward inference, and outputs the crop health status.

[0078] After the multispectral acquisition unit completes parameter adjustments, this method performs health status inference through the AI ​​processing unit. The multispectral acquisition unit transmits image data to the AI ​​processing unit via the SPI bus. The transmission adopts DMA mode, where the cooperative control module triggers the SPI chip select signal upon completion of each acquisition, starting the DMA controller to directly transfer the image data in the sensor FIFO to the DMA buffer of the AI ​​processing unit.

[0079] The AI ​​processing unit preprocesses the received multispectral image data into a 128×128×C tensor, which is then input into the agricultural health monitoring AI model for forward inference. This model is a lightweight convolutional neural network containing three convolutional layers: the first layer uses 16 3×3 convolutional kernels with a stride of 1 and padding of 1; the second layer uses 32 5×5 convolutional kernels with a stride of 1 and padding of 2; and the third layer uses 64 3×3 convolutional kernels with a stride of 1 and padding of 1. Each convolutional layer is followed by a ReLU activation function and a 2×2 max pooling layer. The network terminates with a global average pooling layer, compressing the feature map into a 64-dimensional vector, which is then passed through two fully connected layers (128 nodes and 3 nodes) to output the final result. The output layer uses Softmax activation, and the three indicators are: chlorophyll content (mapped to a SPAD value range of 30–50), probability of early lesion presence (0–1), and water stress level (1–5).

[0080] Step 5: Periodically analyze inference performance metrics. If preset conditions are met, update the low-light threshold and high-load threshold to achieve closed-loop adaptive optimization.

[0081] During system operation, this method periodically performs threshold feedback optimization to achieve closed-loop adaptation. The threshold feedback optimization module automatically starts at 24:00 every day, and collects inference performance metrics over the past 24 hours. The threshold feedback optimization module queries historical records from the local SQLite database and divides the records into four groups according to signal combinations: (S_light=0, S_load=0), (S_light=0, S_load=1), (S_light=1, S_load=0), and (S_light=1, S_load=1).

[0082] Average inference accuracy is calculated based on the true values ​​of 30% of the samples collected simultaneously on the same day after manual annotation. It is the arithmetic mean of classification accuracy (disease) and regression R² (SPAD / WSI). Average end-to-end latency is the total time from the end of multispectral sensor exposure to the completion of SPI transmission, AI model inference time, MQTT encapsulation time, and transmission time on the same day. The sliding window mechanism uses a first-in, first-out queue: new data is added every 24 hours, while data from the corresponding period 7 days ago is deleted, ensuring that the statistical window always includes data from the most recent 168 hours.

[0083] The reset threshold is written to the threshold register of the scene perception module and the threshold configuration register of the load monitoring module via the I²C bus. Modification logs are written to a specified path. If the low-light threshold is lowered for three consecutive days, the system will trigger a manual review alarm, outputting a high-level signal to an external alarm device via a GPIO pin.

[0084] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An agricultural health monitoring system fusing multispectral and AI model, characterized in that, The system includes a multispectral acquisition unit, a dual-feedback linkage control circuit, and an AI processing unit. The multispectral acquisition unit is used to acquire the reflectance spectral data of crops in the visible to near-infrared band. The dual-feedback linkage control circuit is used to sense the ambient light intensity and the computational load of the AI ​​processing unit in real time, and generate collaborative control commands based on preset logic rules. The AI ​​processing unit is used to receive the multispectral data after parameter optimization, perform agricultural health status inference, and provide feedback performance indicators to dynamically adjust the sensing threshold. The dual-feedback linkage control circuit is entirely composed of discrete logic devices.

2. The agricultural health monitoring system of claim 1, wherein, The multispectral acquisition unit includes a multispectral sensor, a clock control circuit, and a band selection circuit. The multispectral sensor is a silicon-based CMOS image sensor with multiple built-in optical filter channels, each of which is independently controllable and supports an adjustable frame rate. The clock control circuit consists of a programmable frequency divider, with its input connected to the main crystal oscillator and its output connected to the frame synchronization pin of the multispectral sensor. The acquisition frequency is controlled by adjusting the frequency division coefficient. The band selection circuit consists of multiple MOSFET switches, each of which is connected in series on the power line of the corresponding spectral channel. It is controlled by the band control signal output by the dual feedback linkage control circuit to achieve independent conduction or deactivation of the channel.

3. The agricultural health monitoring system of claim 2, wherein, The dual-feedback linkage control circuit includes a scene perception module, a load monitoring module, and a collaborative control module. The scene perception module includes a digital ambient light sensor and its peripheral circuits, as well as a threshold comparison circuit. The digital ambient light sensor periodically samples the ambient light intensity. The threshold comparison circuit compares the real-time light intensity with a preset low-light threshold and outputs a scene signal. The load monitoring module includes a register reading circuit and a load judgment logic unit. The register reading circuit periodically reads the CPU utilization rate of the AI ​​processing unit. The load judgment logic unit compares the CPU utilization rate with a preset high-load threshold and outputs a load signal.

4. The agricultural health monitoring system of claim 3, wherein, The collaborative control module includes multiple two-input AND gates, a data selector, and an instruction encoder. The inputs of the two-input AND gates are logical combinations of scene signals and load signals, respectively. The outputs of each AND gate are connected to the selection terminal of the data selector. The data input terminal of the data selector is preset with instruction encoding. The instruction encoder decodes the selected instruction into frequency control signals and band control signals.

5. The agricultural health monitoring system of claim 4, wherein, The collaborative control module outputs corresponding frequency control signals and band control signals based on the combination of scene signals and load signals. When in low light and high load conditions, it outputs control signals to reduce the acquisition frequency and shut down some spectral channels. When in low light and low load conditions, it outputs control signals to maintain the current acquisition frequency but shut down some spectral channels. When in normal light and high load conditions, it outputs control signals to reduce the acquisition frequency and maintain full-band acquisition. When in normal light and low load conditions, it outputs control signals to maintain the current acquisition frequency and full-band acquisition.

6. The agricultural health monitoring system of claim 1, wherein, The AI ​​processing unit includes an edge computing chip, an agricultural health monitoring AI model, and a threshold feedback optimization module. The edge computing chip communicates with the multispectral acquisition unit via a high-speed serial bus. The agricultural health monitoring AI model is a convolutional neural network. The input layer receives a normalized multispectral image tensor, and the network structure sequentially includes multiple convolutional blocks, a global average pooling layer, and a fully connected output layer, which outputs crop health status indicators.

7. The agricultural health monitoring system of claim 6, wherein, Each convolutional block of the agricultural health monitoring AI model consists of a convolutional layer, a batch normalization layer, and an activation function in sequence; the fully connected output layer is trained using a weighted loss function to optimize the prediction accuracy of chlorophyll content, water stress index, and disease probability.

8. The agricultural health monitoring system of claim 6, wherein, The threshold feedback optimization module periodically calculates historical inference performance indicators; if the average inference accuracy in low-light scenarios is lower than a predetermined threshold, the low-light threshold is adjusted; if the average system response latency under high load exceeds a predetermined threshold, the high-load threshold is adjusted; the threshold update is written to the threshold registers in the scene perception module and the load monitoring module through the communication bus, and the update process is completed within a predetermined time and does not affect the normal operation of the system.

9. The agricultural health monitoring system of claim 1, wherein, Multiple of these systems are cascaded and deployed via a communication bus; One node is a master node, and the rest are slave nodes. The master node performs threshold feedback optimization and sends the updated threshold parameters to all slave nodes via a broadcast frame. Upon receiving the broadcast, each slave node immediately updates its local threshold register to ensure the consistency of the monitoring strategy.

10. An agricultural health monitoring method for use in a system according to any one of claims 1 to 9, characterized by, Includes the following steps: S1. Real-time acquisition of ambient light intensity and CPU utilization of the AI ​​processing unit; S2. Based on the comparison results between ambient light intensity and preset low light threshold, and CPU utilization rate and preset high load threshold, generate corresponding frequency control signals and band control signals; S3. Adjust the acquisition frequency according to the frequency control signal, and dynamically enable or disable the spectral channel according to the band control signal; S4. Receive the optimized multispectral data, perform forward inference, and output the crop health status; S5. Periodically analyze historical performance indicators. If preset conditions are met, update the low light threshold and high load threshold to achieve closed-loop adaptive optimization.

Images