Early identification and damage grading method of rhizome traditional Chinese medicinal materials in heavy clay soil conditions

Abstract

The application discloses a kind of rhizome Chinese herbal medicine early identification and damage reduction grading method under heavy soil condition, it is related to S1, to mark as suspicious plant at seedling stage to whole field plant phenotype preliminary screening, only to suspicious plant and its peripheral area are carried out hyperspectral and near-infrared nondestructive testing and extract spectral characteristics, combined with disease severity regression model and spatial interpolation method generate continuous plot health risk heat map, output root segmentation mask and estimated damage risk chart, according to the boundary depth under mask, estimated damage risk chart and current effective soil moisture content, adjust the digging depth of harvester, soil pulverizing roller speed and conveying chain speed;Adjust the collection of medicinal materials by harvester, grade medicinal materials after harvesting, calculate the conditional probability of mechanical damage, disease damage and transportation damage.The application can realize early accurate identification of rhizome Chinese herbal medicine, targeted partition improvement, and grading damage reduction method of intelligent avoidance mechanical damage in harvesting period.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture, specifically to a method for early identification and loss reduction grading of rhizomes and medicinal herbs under heavy clay soil conditions. Background Technology

[0002] Rhizome-based medicinal herbs (such as Salvia miltiorrhiza, Scutellaria baicalensis, Atractylodes lancea, and Atractylodes macrocephala) are important raw materials for my country's traditional Chinese medicine industry. Their medicinal parts are the underground rhizomes, and their quality and yield are highly dependent on soil environment and harvesting methods. However, most major producing areas of rhizome-based medicinal herbs in my country are located in hilly and mountainous areas or areas with heavy clay soils. These soils are characterized by heavy texture, poor aeration, and a tendency to compact, posing unique challenges to the cultivation and harvesting of medicinal herbs. Existing technologies for heavy clay soils mainly suffer from the following problems:

[0003] Heavy clay soils have high water retention and poor drainage, making them highly susceptible to soil-borne diseases such as root rot and nematode infestations. Due to the strong absorption and scattering of spectral signals by soil clay particles, traditional visible light remote sensing is unable to penetrate the soil surface to obtain root information. On the other hand, manual inspections on the ground are inefficient and highly subjective, often only discovering diseases when obvious above-ground symptoms such as wilting and yellowing appear. By this time, the root system has already been severely damaged, missing the optimal window for prevention and control, leading to large-scale yield reduction or even crop failure.

[0004] For soil improvement in heavy clay soils, methods such as deep tillage, application of organic fertilizers, and pH adjustment are commonly employed, with a uniform operation pattern across the entire field. However, heavy clay soils exhibit significant spatial heterogeneity; low-lying areas are prone to waterlogging, oxygen deficiency, and disease outbreaks, while higher areas remain relatively dry and healthy. A uniform deep tillage depth and soil amendment dosage across the entire field not only wastes agricultural inputs but also excessively disturbs soil structure in low-risk areas, accelerating soil compaction; while in high-risk areas, insufficient soil improvement fails to effectively suppress diseases, creating a vicious cycle.

[0005] Heavy clay soils are extremely hard when dry and highly adhesive when wet, making digging difficult and breaking up clods challenging during harvesting. Existing harvesting machinery is mostly designed for ordinary loam soils; when operating in heavy clay soils, it either fails to dig deep enough to miss areas or break roots, or causes root and stem damage and breakage due to clogging of the vibrating screen and compression of soil clods. Damaged roots and stems are susceptible to pathogen infection, dehydration, and shriveling, directly downgrading them to substandard or waste products, severely impacting the commercial value and efficacy stability of medicinal herbs.

[0006] Therefore, there is an urgent need for a graded loss reduction method specifically designed for heavy clay soil conditions, which can enable early and accurate identification of rhizomes and medicinal herbs, targeted regional improvement, and intelligent avoidance of mechanical damage during harvest, in order to solve the above problems. Summary of the Invention

[0007] The present invention aims to solve the above-mentioned technical problems and provide a method for early identification and loss reduction grading of rhizomes and Chinese medicinal materials under heavy clay soil conditions.

[0008] To achieve the above-mentioned objectives, this invention provides a method for early identification and loss reduction grading of rhizomes and other medicinal materials under heavy clay soil conditions, comprising the following steps:

[0009] S1. During the seedling stage, all plants in the field were initially screened phenotypically and marked as suspected plants. Only suspected plants and their surrounding areas were subjected to non-destructive hyperspectral and near-infrared spectral detection to extract spectral features. Combined with a disease severity regression model and spatial interpolation, continuous plot health risk heat maps were generated. ,in, These are the field plane coordinates;

[0010] In step S1, during the initial phenotypic screening of all plants in the field during the seedling stage, medicinal materials that meet preset abnormality conditions are marked as suspicious plants. The preset abnormality conditions include: relative plant height. Leaf SPAD value and the rate of wilting that did not recover 2 hours after watering At least one of them.

[0011] A continuous heatmap of land health risk is generated by combining a disease severity regression model with spatial interpolation. Specifically, this includes: using a preset grid as the unit, spatially interpolating discrete risk points using ordinary Kriging interpolation, normalizing the predicted severity values ​​of the disease to an interval, and generating a continuous plot health risk heatmap. ,according to The risk value is used to divide the land parcel into high-risk, medium-risk, and low-risk zones, and pre-defined soil improvement measures are taken for each zone. When a high-risk area is identified, deep plowing to 40cm is implemented, along with the application of quicklime, Bacillus subtilis, and well-rotted organic fertilizer; when When the area is identified as a medium-risk area, deep plowing to 35cm and application of organic fertilizer and microbial agents are implemented; when At that time, the area was determined to be a low-risk area, and deep plowing to 30cm and returning straw to the field were implemented.

[0012] S2. 15-20 days before the planned harvest date, collect field orthophotos and take the moisture content reading at a depth of 20cm as the current effective soil moisture content. K-means clustering was performed on the collected field orthophotos to extract the color and texture centroids of each cluster, which were then encoded into visual cue vectors. K-means clustering was performed on the collected field orthophotos to extract the color and texture centroids of each cluster, which were then encoded into visual cue vectors. The text embedding vector Pt, which contains semantics about heavy clay soil occlusion, is used for cross-modal feature alignment and fusion through a cross-attention mechanism to output a root segmentation mask. and estimated damage risk map ,in, , representing the root pixel, To indicate the background, the ground surface is used as a reference. Reference point for calculating the lower boundary depth of the mask. , The range of values ​​is ;

[0013] In step S2, when Lock the harvester's start permission and trigger a drainage prompt; when If the condition is deemed a suitable harvest window, proceed to step S3; when This triggers an irrigation reminder.

[0014] S3, based on the depth of the lower boundary of the mask. Damage risk prediction map and current effective soil moisture content Adjust the digging depth of the harvester Rotation speed of soil crushing roller and conveyor chain speed ;

[0015] Dig deeper ;

[0016] Soil crushing roller speed ;

[0017] Conveyor chain speed ;

[0018] The depth of the lowest point of the rhizome is calculated from the image. This is the moisture content compensation coefficient. Based on the base speed, For risk speed gain, For moisture rotation speed gain, This represents the actual soil moisture content. For optimal moisture content, Based on the chain speed, and All are chain speed adjustment coefficients;

[0019] when hour, The first compensation value is preset; when hour, This is the preset second compensation value. When the damage risk map is estimated... When the value is greater than the preset high-risk threshold of 0.8, the corresponding plant is marked, and the operator is prompted to use a non-metallic tool to gently pry along the outside of the plant to assist in harvesting.

[0020] S4, adjust the harvester according to step S3 to collect medicinal materials, grade the medicinal materials after harvesting, and combine the grading results with the excavation parameters and seedling health tags. Constructing feature vectors Input the damage attribution network to calculate the conditional probabilities of mechanical damage, disease damage, and transportation damage. When the probability of mechanical damage exceeds a preset threshold and the actual damage location overlaps with the high-risk area of ​​the estimated damage risk map to a certain extent, update the moisture content compensation coefficient for the next batch of excavation according to the following formula:

[0021] ;in, For learning rate, To allow for damage threshold, This represents the actual extent of damage. (Updated) The value range is [3cm, 12cm]. When the probability of three consecutive batches of medicinal materials being determined to have mechanical damage is less than 0.3, the update of the moisture content compensation coefficient will be suspended.

[0022] In step S4, feature vectors are constructed. Specifically, this includes: adding seedling health labels Estimated damage risk value Actual excavation depth Rotation speed of soil crushing roller Conveyor chain speed Actual soil moisture content and the actual location of the injury The combination forms a feature vector X.

[0023] In step S4, when the probability of mechanical damage exceeds a preset threshold and the seedling health label... If the strain is suspected, the corresponding region will be generated in the next season. The base risk weight is increased by a preset bonus value of 0.2.

[0024] First, a non-destructive phenotypic scan was performed on all plants in the field during the seedling stage, with pre-set abnormal conditions including: relative plant height. Leaf SPAD value and the rate of wilting that did not recover 2 hours after watering At least one of the following criteria must be met for a plant to be identified as a suspected plant: impaired water conduction (wilt) often precedes decreased photosynthetic capacity (reduced SPAD) and inhibited biomass accumulation (delayed plant height), forming a complementary early warning indicator system. Secondly, hyperspectral and near-infrared detection is performed only on suspected plants and their surrounding areas to avoid the signal-to-noise ratio degradation problem caused by full-field scanning in heavy clay soils with high background noise. Subsequently, using a pre-defined grid, ordinary Kriging interpolation is employed to transform discrete risk points into continuous plot health risk heatmaps. This method utilizes a semi-variogram to characterize the spatial autocorrelation characteristics of disease patchy aggregation in heavy clay soils. Finally, based on the heat map, three levels of risk zones are divided and differentiated improvement measures are implemented: in high-risk zones, deep plowing to 40cm breaks up the plow pan and is combined with the application of quicklime, Bacillus subtilis, and well-rotted organic fertilizer to achieve a triple synergy of disease suppression, antagonism, and soil improvement; in medium-risk zones, deep plowing to 35cm is combined with the application of organic fertilizer and microbial agents to maintain microecological homeostasis; and in low-risk zones, deep plowing to 30cm is combined with straw return to the field to protect the original soil structure.

[0025] Existing technologies often rely on a single vegetation index or uniform application of pesticides across the entire field, which are easily affected by soil brightness and lack physiological specificity in heavy clay soils. This invention uses "wilt recovery rate after watering" as a primary screening index specific to heavy clay soils, directly characterizing the water conduction function of roots in viscous pores. This is a physiological signal that traditional purely optical indicators such as NDVI cannot capture, avoiding excessive intervention in medium-risk areas and insufficient treatment in high-risk areas, and is particularly suitable for the strong spatial heterogeneity of heavy clay soils.

[0026] This invention enables the identification window for soil-borne diseases such as root rot and nematode disease in heavy clay soils to be 7–10 days earlier than that of visible light remote sensing alone, effectively solving the problem of late detection.

[0027] Step S2: During the critical window period of 15-20 days before the planned harvest date, take a sample of the current effective soil moisture content at a depth of 20cm. As a basis for decision-making, this depth corresponds to the distribution layer of the main root of most rhizome-type Chinese medicinal herbs. When At that time, the heavy clay soil is in a plastic flow state, and mechanical operations are prone to sticking and clogging, locking the start permission and triggering a drainage prompt; To ensure a suitable harvest window, proceed to the next stage of the process; At this time, a surge in soil hardness can easily lead to mechanical damage, triggering irrigation prompts. At the perception level, K-means clustering is used to extract color and texture centroids from field orthophotos to encode visual cue vectors. Simultaneously, a text embedding vector Pt containing semantics of heavy clay soil occlusion, such as "clay cover," "root concealment," and "soil crust," is constructed. Cross-modal feature alignment and fusion are achieved through a cross-attention mechanism to output a root segmentation mask. and estimated damage risk map Calculate the depth of the lower boundary of the mask using the ground surface as a reference. This provides precise geometric anchor points for adaptive adjustment of excavation depth.

[0028] Existing root sensing methods are mostly purely visual or purely radar-based, and their performance deteriorates sharply under the interference of surface crusts, cracks, and crop residues in heavy clay soils. This invention injects "heavy clay soil occlusion semantics" as a text modality into the sensing model, enabling the system to possess experience-based reasoning capabilities similar to those of agronomic experts. Even with partial visual occlusion, it can infer the possible location and risk level of the roots based on soil texture and contextual priors—a cognitive dimension that purely data-driven models lack.

[0029] Therefore, this step prevents problems such as machinery getting stuck and sieve clogging caused by forced harvesting in excessively wet clay, as well as in excessively dry and hard soil, ensuring operational continuity and safety. Simultaneously, this invention achieves higher root segmentation accuracy than purely visual models under heavy clay soil conditions, making damage risk prediction more reliable. Furthermore, this invention transforms fuzzy experience into a range of moisture content values, improving the certainty and schedulability of the harvest plan.

[0030] Furthermore, step S3 of the present invention converts the sensing results of step S2 into the execution parameters of the harvester in real time, and the digging depth is calculated according to... Adjustment, among which, The moisture content compensation coefficient is used in... and Two different values ​​are used to match the nonlinear resistance characteristics of heavy clay soils to the digging shovel at different moisture contents. The speed of the soil-crushing roller is set according to... Adjustments are made based on risk level and soil moisture deviation from optimal values, increasing rotational speed to enhance soil breaking and reduce large clod compression; the conveyor chain speed is adjusted accordingly. Adjustments are made based on risk level and soil moisture: higher risk and lower moisture content result in slower chain speeds, extending separation time to reduce collision kinetic energy. At that time, the system automatically sprays fluorescent markers and prompts operators to use non-metallic tools to gently pry along the outer side of the plant to assist in harvesting. This threshold is 0.7 higher than the disease threshold of S1, better reflecting the effect of machines ensuring efficiency and manual labor ensuring quality.

[0031] Existing harvesters mostly use fixed parameters or rely on simple feedback from a single sensor, which cannot cope with the nonlinear mechanical properties of heavy clay soils. This invention couples soil moisture content, damage risk, and root depth into a continuous mechanical control surface, ensuring that the harvester always operates within a safe envelope under varying conditions in heavy clay soils.

[0032] Step S4 of this invention establishes postharvest grading, damage attribution, and parameter adjustment, and incorporates seedling health labels. Actual excavation depth Rotation speed of soil crushing roller Conveyor chain speed Actual soil moisture content and the actual location of the injury The combined feature vector X is input into the damage attribution network to calculate the conditional probabilities of mechanical damage, disease damage, and transportation damage. When the probability of being identified as mechanical damage exceeds a preset threshold and the actual damage location overlaps with the high-risk area of ​​the estimated damage risk map to a certain extent, then... Update the moisture content compensation coefficient, where the learning rate is... Allowable damage threshold The updated The value range is [3cm, 12cm]; updates are paused when the probability of mechanical damage is below 0.3 for three consecutive batches to prevent overfitting. Furthermore, when the probability of mechanical damage exceeds the threshold and the seedling health label... If it is a suspected strain, it will be generated in the next season. The basic risk weight for the region increases by 0.2, enabling cross-seasonal migration.

[0033] Existing technologies mostly employ open-loop control or only provide feedback within the current season, lacking the ability to identify damage types and learn across seasons. This invention incorporates seedling health tags into attribution features, enabling the identification of the interaction effect that "weak and diseased plants are more susceptible to mechanical damage," thus avoiding misjudging vulnerability caused by disease as improper mechanical parameters. The design with a 0.2 weighting for cross-seasonal damage possesses a "plot memory" similar to that of a seasoned planting expert, achieving a leap from single-season optimization to long-term evolution. The risk weighting transforms lessons learned from single-season damage into key preventative measures for the following season.

[0034] Corresponding to the method in this invention, this invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for early identification and loss reduction grading of rhizomes in heavy clay soil conditions.

[0035] Corresponding to the method in this invention, this invention also provides a storage medium storing a computer-readable storage medium, which, when executed by a processor, implements the steps of the above-described method for early identification and loss reduction grading of rhizomes in heavy clay soil conditions.

[0036] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0037] This invention solves the problem of difficult identification of medicinal herbs in heavy clay soil by integrating data from various stages, such as seedling phenotypic screening, targeted soil improvement, pre-harvest moisture gating, multi-parameter harvesting, and damage attribution, with each stage's data serving as input and verification.

[0038] This invention relies on real-time data from sensors and models, and incorporates symbolic representations of agronomical knowledge of heavy clay soils and cross-seasonal experience, enabling it to make robust and interpretable decisions even under conditions of sparse data or noise interference.

[0039] Through synergistic loss reduction throughout the entire cycle, the rate of first-grade medicinal herbs from rhizomes in heavy clay soil areas can be effectively increased, and the harvest damage rate can be reduced. At the same time, the structure of heavy clay soil is protected, the degradation process is slowed down, and solid technical support is provided for the sustainable utilization of traditional Chinese medicinal herbs in their authentic producing areas. Attached Figure Description

[0040] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0041] Figure 1 This is a flowchart illustrating a method for early identification and loss reduction grading of rhizomes in heavy clay soil conditions. Detailed Implementation

[0042] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0044] Those skilled in the art should understand that, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention.

[0045] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0046] Example 1

[0047] Please refer to Figure 1 , Figure 1This invention provides a method for early identification and loss reduction grading of rhizome-based Chinese medicinal materials under heavy clay soil conditions. The method includes the following steps: taking Salvia miltiorrhiza as an example.

[0048] S1, 55-65 days after transplanting, on a sunny, windless morning from 9:00-11:00, a ground inspection robot equipped with a multispectral camera and LiDAR scanned the entire field. Simultaneously, the plant height (H1), leaf SPAD value, and midday wilting recovery status of each plant were acquired. The wilting recovery rate was determined as follows: 2 hours after irrigation, visible light images were taken of plants initially suspected of wilting. The leaf projection area was calculated through image segmentation and compared with the baseline area before irrigation. The wilting failure rate 2 hours after irrigation was calculated as (baseline area - area after 2 hours) / baseline area × 100%. A threshold was set to calculate the relative plant height. ,in, The height of each medicinal plant. The average height of healthy plants in the field is used. The image segmentation specifically adopts an adaptive threshold segmentation method based on the HSV color space: First, the visible light image is converted from RGB space to HSV space, and the S (saturation) channel is extracted. The optimal segmentation threshold of the S channel is automatically calculated using the Otsu algorithm to generate a binary mask. Then, morphological opening operation (structuring element is a 3×3 circular kernel) is performed on the mask to remove soil noise. Then, non-leaf areas with an area of ​​less than 500 pixels are removed by connected component analysis, and finally the effective projected area of ​​a single leaf is obtained.

[0049] Will or SPAD value Or the rate of wilting that did not recover within 2 hours after watering The plants were marked as suspicious and their geographic coordinates were recorded. In this initial screening, a total of 326 suspected plants were marked, accounting for 4.1% of the total number of plants in the field.

[0050] Only 326 suspected plants and their surrounding area were sampled. Within the specified area, in-situ canopy spectral data were acquired using a portable hyperspectral imager, with each measurement repeated three times and the average value taken. After Savitzky-Golay smoothing and first-derivative preprocessing of the raw spectra, PCA was used to extract the scores of the first five principal components as feature vectors. These vectors were then input into a pre-trained PLSR model based on local heavy clay soil samples of *Salvia miltiorrhiza*, and the output was the original disease severity index for each detection point. The original disease severity index The values ​​are dimensionless quantizations ranging from 0 to 100, and their calibration benchmarks are established using laboratory extreme samples: the mean value of the canopy spectrum of healthy Salvia miltiorrhiza plants output by the same PLSR model is used as the standard. The mean spectral output of the limiting sample, artificially inoculated with wet rot pathogens and cultured at a constant temperature and humidity of 28℃ until the rhizomes were completely soft rotten, was used as the baseline. The subsequent normalization formula and These correspond to the calibration endpoints of 0 and 100 mentioned above, respectively. Indicates the first The original measured values ​​of each monitoring indicator at the current detection point; in this embodiment, when hour, This refers to the original disease severity index output by the aforementioned PLSR model. ;when hour, The percentage of tuber porosity measured in situ using a portable soil porosimeter; when hour, w represents the percentage of soil saturated moisture content simultaneously measured by the TDR sensor.

[0051] , They represent the first The physical lower limit and physical upper limit values ​​of each monitoring indicator in the calibration dataset are fixed benchmark parameters calculated by normalization, and are not the real-time statistical extreme values ​​of the current plot. Indicates the first The dimensionless values ​​of each monitoring indicator after range normalization.

[0052] Discrete risk value The calculation includes the following steps: First, the range normalization process is performed on each monitoring indicator, that is... First, eliminate dimensional differences; then, based on the current damage mechanisms of medicinal materials, soil, and disease combinations, determine the fusion weights of each normalized index. , and according to The comprehensive discrete risk value is calculated. In this embodiment, the normalized values ​​of disease severity, tuber porosity, and soil saturated moisture content are denoted as follows: , , The corresponding disease severity weight tuber porosity weight Soil saturation . , , .

[0053] Specifically, the effective sensitive range for disease severity weighting The critical threshold and saturation threshold of mechanical failure due to wet rot are jointly defined; the weighted interval of the root diameter variation coefficient is also considered. The probability-deterministic transition zone corresponding to fine root buckling instability; upper limit of soil compaction weight. This is limited by the physical ceiling of the additional bending moment transfer efficiency of the flexible root material. Within the scope defined by the above mechanism, through multiple regression analysis of 32 sets of field orthogonal experiments, the optimal weight combination in this embodiment is: , , Therefore .

[0054] This is the input data for spatial interpolation. Because hyperspectral detection is only performed on the "suspected strain" and its surroundings, the results are discrete points. This invention needs to convert these discrete points... The values, as known nodes, are input into the Ordinary Kriging interpolation algorithm to deduce and generate continuous values ​​for the entire field. Heat map.

[0055] Spatial interpolation was performed on 326 discrete risk points using a 5m × 5m grid, employing ordinary kriging interpolation. A spherical model was used for the semi-variogram, with a nugget value of 0.02, a sill value of 0.18, a range of 35m, and a coefficient of determination for fitting. Generate a heatmap of health risks for continuous land parcels. 5m resolution. According to... Divide the region into three levels and perform variable operations:

[0056] when When the area is identified as a high-risk area, a smart deep tillage machine is used to cultivate the soil to a depth of 40cm, and quicklime (75kg / mu), Bacillus subtilis granules (5kg / mu), and well-rotted cow manure organic fertilizer (800kg / mu) are applied simultaneously.

[0057] when At that time, it was determined to be a medium-risk area. The soil was deep-plowed to 35cm, and 500kg / mu of well-rotted cow manure organic fertilizer and 3kg / mu of compound microbial agent were applied.

[0058] when At that time, the area was determined to be a low-risk area, and deep plowing to 30cm was carried out, with 300kg / mu of straw returned to the field. All operations were executed through the Beidou navigation automatic driving system, and the deviation between the operation trajectory and the boundary of the heat map was ≤10cm.

[0059] S2. 15-20 days before the planned harvest date, use a TDR soil moisture sensor to bury probes in a 20m×20m grid throughout the field, read the moisture content at a depth of 20cm, and take the average value of the entire field as the current effective soil moisture content. The measurements taken this time... It is in a suitable harvest window ( ,) allows entry into the harvest process. If Then lock the harvester's startup permissions and push a drainage reminder via the mobile app; if Then, a reminder to replenish irrigation moisture will be sent.

[0060] A DJI M300 RTK drone equipped with an RGB camera was used to acquire orthophotos of the field at a flight altitude of 30m and a ground sampling distance of 0.8cm / pixel. K-means clustering (k=8) was performed on the images, and the centroids of the color mean (LAB space) and texture features (GLCM energy, contrast, entropy) of each cluster were extracted and encoded into a 128-dimensional visual cue vector. .

[0061] Construct a text embedding vector Pt with the content "Heavy clay soil, slight surface crusting, Salvia miltiorrhiza roots may be hidden under soil clods, be careful of occlusion". and Inputting a pre-trained cross-modal alignment network (fine-tuned based on CLIP architecture) and fusing it through a cross-attention mechanism, the output is a root segmentation mask. and estimated damage risk map Using the Earth's surface as the reference point (y=0), calculate the depth of the lower boundary of the mask. The forecast results show that the average for the whole field... The fine-tuning process of the cross-modal alignment network is as follows: A training set of 2000 quadruplets is constructed, each consisting of an orthophoto slice, text description, root pixel-level mask, and damage risk annotation map; the damage risk annotation map is generated by spatial registration of damage data obtained from field measurements. During fine-tuning, the first 6 layers of the CLIP image encoder and text encoder are frozen, while only the last 6 layers and the newly added multi-task decoder are unfrozen; a joint loss function is used. End-to-end training is performed, where Used to monitor the segment overlap of the root mask. The regression accuracy was used to supervise the risk map, with a training epoch of 50 epochs and an initial learning rate of 1× It also employs a cosine annealing strategy for attenuation.

[0062] In this embodiment, the damage risk map is estimated. The high-value areas with median values ​​greater than 0.8 are concentrated in the high-risk area defined in step S1 (i.e. ) and medium-risk areas ( The transition zone between these two areas, known as the "edge of the original S1 high-risk zone," is analyzed from an agronomic and soil mechanics perspective. The cause of this phenomenon is that the root systems of plants in the center of the high-risk zone are often severely rotten or dead, resulting in low soil resistance and no intact rhizomes to damage during excavation. In contrast, the root systems of plants in the transition zone are in a mildly infected or sub-healthy state; although the rhizomes exist, their toughness is significantly reduced and their texture becomes brittle. Simultaneously, the heavy clay soil exhibits a density gradient in this transition zone, and the lateral squeezing force generated when the harvester's shovel cuts into the soil, combined with the brittle rhizomes, easily leads to mechanical breakage or bark damage. Therefore, cross-modal alignment networks are crucial for integrating visual cue vectors. After embedding the text vector Pt, it can accurately identify the region where stress concentration is coupled with physiological vulnerability and output a high damage risk prediction value.

[0063] Based on the generation in step S1 Heatmap, extract all The grid was used as the core patch of the high-risk area; then, based on the boundary of this patch, a 5m buffer zone was extended outward as a candidate area for the edge transition zone; finally, the grid was used as the core area of ​​the high-risk area. Spatial overlay analysis was performed between the raster and the buffer zone, and the intersection was taken as the actual high-risk edge area. According to GIS statistics, the total area of ​​the plot in this embodiment is 20 mu, and the core area of ​​the high-risk zone is 2.46 mu (accounting for 12.3%), within its 5m buffer zone. The total area of ​​the grid is 1.34 mu, accounting for 6.7% of the total field area.

[0064] S3, the harvester is equipped with an industrial control computer, receiving... , and real-time moisture content The optimum moisture content is known. Base speed Basic chain speed Risk speed gain Moisture speed gain Chain speed adjustment coefficient , .because , In this embodiment, the preset second compensation value is used. .

[0065] Substitute into the formula to calculate:

[0066] Dig deeper ;

[0067] Soil crushing roller ;

[0068] Conveyor chain speed ;

[0069] The above parameters are transmitted in real time to the harvester's hydraulic and motor controller via the CAN bus, with a refresh rate of 10Hz.

[0070] when At that moment, the industrial control computer triggered the fluorescent spraying device (orange water-based fluorescent paint) to spray a 3cm diameter dot 10cm above the plant. Simultaneously, the voice announcement in the control cabin stated, "High risk 2 meters ahead, please assist." The operator then used a bamboo stick to gently pry the plant at a 30° angle along the outside of the marked plant, helping to detach the roots from the heavy soil clods. A total of 89 plants were marked during this harvest, with a total manual assistance time of 42 minutes.

[0071] After harvesting, the medicinal herbs are manually graded into three grades: Grade 1 (intact and undamaged), Grade 2 (minor abrasions), and Grade 3 (broken / damaged). 200 samples are randomly selected, and the actual locations of damage are recorded. Actual degree of damage (Measured by the longest crack length, in cm). Seedling health label. ( This is normal. (Suspected strain), estimated damage risk value Actual excavation depth Rotation speed of soil crushing roller Conveyor chain speed Actual soil moisture content and the actual location of the injury The combination results in a 7-dimensional feature vector X.

[0072] S4, input the pre-trained damage attribution network (three-layer MLP, Softmax output) to calculate the mechanical damage probability. Of the 200 samples collected this time, And the location of the injury is related to A total of 17 plants had an 82% overlap in their regions. These 17 plants were then analyzed using the formula... Update the compensation coefficient. For example, a certain plant ,but The updated values ​​were cropped to the [3, 12] cm range. After continuous monitoring of the next three batches, the probability of mechanical damage was consistently below 0.3, and updates were paused.

[0073] Labels for the above 17 seedlings Five plants, in their corresponding GPS coordinate areas, will produce next season's [crops / plants]. The basic risk weight is increased by 0.2. This information is stored in the land parcel's digital archive for use in the next quarter's S1 step.

[0074] Among them, learning rate This value is determined by both the convergence speed and stability of the compensation coefficient update. Too small a value (<0.1) will cause parameter adjustment to lag, making it unable to respond promptly to changes in soil moisture; An excessively large value (>0.5) will cause the compensation coefficient to oscillate wildly around its optimal value. This was demonstrated through 15 different... Comparative experiments to determine the values At that time, the compensation coefficient can converge stably within 3-5 batches, and the system mean square error is minimized.

[0075] Mechanical damage probability threshold This threshold is a balance between the "damage underreporting rate" and the "misreporting rate". When the threshold is below 0.5, a large number of defects are misclassified as mechanical damage, leading to incorrect updates of the compensation coefficients; when the threshold is above 0.7, some real mechanical damage samples are missed, reducing the model's learning efficiency. 0.6 is the optimal cutoff value corresponding to the maximum Youden exponent on the ROC curve.

[0076] A risk area overlap of >80% is a threshold used to ensure that damage actually occurs within the predicted high-risk area. An overlap of less than 80% means the damage may be caused by other unmodeled factors (such as rock impacts) and should not be attributed to the current risk prediction model. The 80% threshold effectively filters out these noisy samples, ensuring the accuracy of parameter updates.

[0077] Next season generated The baseline risk weight is increased by 0.2, based on the agronomic principle that "weak and diseased plants are more susceptible to mechanical damage." Seedling stage label. Plants with poor root system vigor and resistance to disease have weaker root systems and, even if the disease does not recur in the following season, the mechanical strength of their rhizomes is lower than that of healthy plants. By increasing the base risk weight by 0.2, early warning and intervention can be carried out in this area.

[0078] According to the "Standards for the Specifications and Grades of Chinese Medicinal Herbs," if the length of damage or cracks on the surface of rhizome-type medicinal herbs (such as Salvia miltiorrhiza and Atractylodes macrocephala) exceeds 1.0 cm, it will directly lead to the downgrading of the medicinal material from "first-class" to "second-class" or "general grade." Heavy clay soils are rich in Pythium and Fusarium fungi. Field experiments show that when the length of damage to the rhizome epidermis is less than 1.0 cm, the plant's own callus tissue can seal the wound within 24 hours, blocking pathogen infection; however, when the length of damage exceeds 1.0 cm, the wound healing speed lags behind the pathogen reproduction rate, easily inducing post-harvest rot. Therefore... .

[0079] Postharvest damage to rhizomes and other medicinal herbs originates from complex sources. Judging solely by the morphology of the damage without considering the probability of disease and transportation risks can easily lead to misdiagnosis of pathological autolysis fractures or damage from transport collisions as damage caused by excavation machinery. Calculations... and When the probability of these two is higher than the probability of mechanical damage At this time, the system will automatically block parameter update commands to prevent incorrect fluctuations in harvester parameters caused by misjudgments due to non-mechanical factors.

[0080] Feature vector Includes seedling health labels .like (Suspicious strains) The network tends to output a higher probability of disease damage; if (For healthy plants) and where the damage is located at the stress point of the rootstock, the network tends to output a higher probability of mechanical damage. Calculating the joint probability distribution of these three factors can fully utilize the data association throughout the entire lifecycle and improve the accuracy of attribution.

[0081] Example 2

[0082] Based on the above embodiments, the specific calculation method for calculating the conditional probabilities of mechanical damage, disease damage, and transportation damage is as follows:

[0083] A multi-class neural network model is used to achieve decoupled calculation of conditional probabilities. Specifically, this includes seedling health labels. Estimated damage risk value Actual excavation depth Rotation speed of soil crushing roller Conveyor chain speed Actual soil moisture content and the actual location of the injury The 7-dimensional feature vector X is input into a pre-trained three-layer multilayer perceptron (MLP). The hidden layers of the MLP extract the aforementioned heterogeneous features (including discrete variables representing historical states) through a nonlinear activation function. Continuous variables characterizing real-time operating conditions , The network's output layer contains three neurons that output unnormalized scores for mechanical damage, disease damage, and transport damage, respectively. , , .

[0084] Subsequently, the output scores are normalized using the Softmax function, transforming the three scores into a conditional probability distribution that sums to 1.

[0085] During the post-harvest grading process, the system obtained damage data from a *Salvia miltiorrhiza* sample. The seedling health label of this sample was then established. (i.e., healthy strain), estimated damage risk value Actual excavation depth cm, speed of soil crushing roller Conveyor chain speed Actual soil moisture content and the actual location of the injury The seven parameters are combined to form a seven-dimensional feature vector. The data is then fed into a pre-trained three-layer multilayer perceptron (MLP).

[0086] Feature vector After nonlinear feature extraction and weight matrix mapping in the MLP hidden layer, three unnormalized real scores (Logits) are generated in the output layer, corresponding to mechanical damage ( ), disease damage ( ) and transport damage ( In this embodiment, it is assumed that the Logits calculated by the network forward propagation for this sample are as follows:

[0087] ;

[0088] Calculate the natural index value for each score; ; ;

[0089] Calculate the sum of the index values: 24.5325 + 2.2255 + 2.7183 = 29.4763.

[0090] Calculate the conditional probability for each damage type. ; ; = .

[0091] Upon verification, the sum of the probabilities of the three factors is 1, which conforms to the mathematical constraints of the joint probability distribution.

[0092] Based on the calculation results, the conditional probability of mechanical damage for this sample .because (The preset mechanical damage trigger threshold), and the actual damage location of the sample is located at the "force point in the middle section of the rootstock", which corresponds to the estimated damage risk map in step S2. The high-risk area spatial overlap rate met the standard, and the system ultimately attributed the damage to the sample to "mechanical damage," triggering the next batch of excavation to adjust the moisture content compensation coefficient. Update.

[0093] Example 3

[0094] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method for early identification and loss reduction grading of rhizomes in heavy clay soil conditions.

[0095] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0096] The memory can be used to store the computer program and / or modules. The processor, by running or executing the data stored in the memory, realizes various functions of the method for early identification and loss reduction grading of rhizomatous Chinese medicinal materials under heavy clay soil conditions. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback function, image playback function, etc.). In addition, the memory may include high-speed random access memory and non-volatile memory, such as hard disk, memory, plug-in hard disk, smart memory card, secure digital card, flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0097] Example 4

[0098] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for early identification and loss reduction grading of rhizomes in heavy clay soil conditions.

[0099] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory, erasable programmable read-only memory or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0100] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0101] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for early identification and loss reduction grading of rhizomes and medicinal materials under heavy clay soil conditions, characterized in that, Includes the following steps: S1. During the seedling stage, all plants in the field were initially screened phenotypically and marked as suspected plants. Only suspected plants and their surrounding areas were subjected to non-destructive hyperspectral and near-infrared spectral detection to extract spectral features. Combined with a disease severity regression model and spatial interpolation, continuous plot health risk heat maps were generated. ,in, These are the field plane coordinates; S2. 15-20 days before the planned harvest date, collect field orthophotos and take the moisture content reading at a depth of 20cm as the current effective soil moisture content. K-means clustering was performed on the collected field orthophotos to extract the color and texture centroids of each cluster, which were then encoded into visual cue vectors. K-means clustering was performed on the collected field orthophotos to extract the color and texture centroids of each cluster, which were then encoded into visual cue vectors. The text embedding vector Pt, which contains semantics about heavy clay soil occlusion, is used for cross-modal feature alignment and fusion through a cross-attention mechanism to output a root segmentation mask. and estimated damage risk map ,in, , representing the root pixel, To indicate the background, the ground surface is used as a reference. Reference point for calculating the lower boundary depth of the mask. , The range of values ​​is ; S3, based on the depth of the lower boundary of the mask. Damage risk prediction map and current effective soil moisture content Adjust the digging depth of the harvester Rotation speed of soil crushing roller and conveyor chain speed ; Dig deeper ; Soil crushing roller speed ; Conveyor chain speed ; The depth of the lowest point of the rhizome is calculated from the image. This is the moisture content compensation coefficient. Based on the base speed, For risk speed gain, For moisture rotation speed gain, This represents the actual soil moisture content. For optimal moisture content, Based on the chain speed, and All are chain speed adjustment coefficients; S4, adjust the harvester according to step S3 to collect medicinal materials, grade the medicinal materials after harvesting, and combine the grading results with the excavation parameters and seedling health tags. Constructing feature vectors Input the damage attribution network to calculate the conditional probabilities of mechanical damage, disease damage, and transportation damage. When the probability of mechanical damage exceeds a preset threshold and the actual damage location overlaps with the high-risk area of ​​the estimated damage risk map to a certain extent, update the moisture content compensation coefficient for the next batch of excavation according to the following formula: ;in, For learning rate, To allow for damage threshold, This represents the actual extent of damage.

2. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, Step S1: During the initial phenotypic screening of all plants in the field at the seedling stage, medicinal materials that meet the preset abnormality conditions are marked as suspicious plants. The preset abnormality conditions include: relative plant height. Leaf SPAD value and the rate of wilting that did not recover 2 hours after watering At least one of them.

3. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, A continuous heatmap of land health risk is generated by combining a disease severity regression model with spatial interpolation. Specifically, this includes: using a preset grid as the unit, spatially interpolating discrete risk points using ordinary Kriging interpolation, normalizing the predicted severity values ​​of the disease to an interval, and generating a continuous plot health risk heatmap. ,according to The risk value is used to divide the plot into high-risk, medium-risk and low-risk areas, and pre-set soil improvement is carried out in the high-risk, medium-risk and low-risk areas.

4. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 3, characterized in that, when When a high-risk area is identified, deep plowing to 40cm is implemented, along with the application of quicklime, Bacillus subtilis, and well-rotted organic fertilizer; when When the area is identified as a medium-risk area, deep plowing to 35cm and application of organic fertilizer and microbial agents are implemented; when At that time, the area was determined to be a low-risk area, and deep plowing to 30cm and returning straw to the field were implemented.

5. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, In step S2, when Lock the harvester's start permission and trigger a drainage prompt; when If the condition is deemed a suitable harvest window, proceed to step S3; when This triggers an irrigation reminder.

6. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, when hour, The first compensation value is preset; when hour, This is the preset second compensation value.

7. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, In step S4, feature vectors are constructed. Specifically, this includes: adding seedling health labels Estimated damage risk value Actual excavation depth Rotation speed of soil crushing roller Conveyor chain speed Actual soil moisture content and the actual location of the injury The combination forms a feature vector X.

8. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, In step S4, when the probability of mechanical damage exceeds a preset threshold and the seedling health label... If the strain is suspected, the corresponding region will be generated in the next season. The base risk weight is increased by a preset bonus value of 0.

2.

9. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, Updated The value range is [3cm, 12cm]. When the probability of three consecutive batches of medicinal materials being determined to have mechanical damage is less than 0.3, the update of the moisture content compensation coefficient will be suspended.

10. The method for early identification and loss reduction grading of rhizomatous medicinal materials under heavy clay soil conditions according to claim 1, characterized in that, When the damage risk map is estimated When the value is greater than the preset high-risk threshold of 0.8, the corresponding plant is marked, and the operator is prompted to use a non-metallic tool to gently pry along the outside of the plant to assist in harvesting.

Images