Bioremediation and intelligent salt control decision generation method for crop root zone soil

Abstract

The invention relates to a bioremediation and intelligent salt control decision generation method for crop root zone soil, and belongs to the technical field of intelligent agriculture. The method comprises the steps that root zone bioremediation knowledge and saline-alkali soil multi-source data are acquired, a multi-modal data association system is built, structured processing is completed, and an initial model of saline-alkali control and water and fertilizer management decision is generated; constructing a multi-dimensional dynamic analysis framework based on real-time environmental perception data and crop growth period characteristics, and generating a collaborative scheme containing bioremediation components, salt control measures and nutrient configuration; the regulation and control benchmark is corrected in combination with feedback data, the scheme collaboration is verified, and scheme execution and data management are achieved through a root zone restoration and salt control collaborative management platform; and a model dynamic iteration mechanism is constructed based on data space-time association, and a continuous optimization closed loop of model-scheme-platform linkage is formed. The transformation of saline-alkali soil treatment from whole field management to root zone precise regulation and control is realized.

Description

Technical Field

[0001] This invention belongs to the field of smart agriculture technology, specifically relating to a method for bioremediation and intelligent salt control decision generation for crop root zone soil. Background Technology

[0002] Saline-alkali land is widely distributed and is an important reserve of arable land. For a long time, saline-alkali land management has mainly relied on engineering measures and chemical amendments, such as large-scale water pressure treatment and underground drainage. However, these methods have significant limitations. Large-scale water pressure treatment consumes large amounts of freshwater resources, making it difficult to implement in arid and water-scarce areas; underground drainage projects involve large investments and high maintenance costs, and are prone to drainage system failure due to soil subsidence. While chemical amendments are more effective, long-term use may lead to soil compaction and ecological damage, and they fail to fundamentally improve soil structure.

[0003] With the development of agricultural technology, integrated water and fertilizer management technology has been gradually applied to the improvement of saline-alkali land. However, the existing technology system still has significant shortcomings. First, traditional water and fertilizer management mainly relies on manual experience, lacking precise perception and control of the crop root zone microenvironment. It often adopts a method of uniform fertilization and irrigation across the entire field, failing to target the root activity layer for improvement. Second, existing systems mostly remain at the data monitoring level, failing to deeply integrate and intelligently analyze multi-source data such as soil salinity indicators, crop growth stages, and climate conditions, resulting in insufficient scientific decision-making. Third, there is a lack of coordination mechanisms between saline-alkali improvement, nutrient supply, and water management, often resulting in each acting independently and even causing mutual cancellation between measures. For example, if irrigation for salt control and fertilization for alkali adjustment are not precisely coordinated in terms of timing and dosage, it may exacerbate salt accumulation in the root zone.

[0004] Furthermore, existing technology platforms often suffer from data loop gaps. Model decision-making is disconnected from field execution, and execution results data fail to be effectively fed back to the decision-making system, leading to lagging model optimization and difficulty in adapting to dynamically changing field environments. Particularly noteworthy is the significant difference in crop responses to salt stress and water and fertilizer requirements at different growth stages, while existing systems lack the ability to adaptively adjust to these dynamic changes.

[0005] Therefore, a new technology for saline-alkali land management is needed that can achieve precise root zone perception, intelligent fusion of multi-source data, coordinated regulation of water, fertilizer and salt, and continuous optimization capabilities, in order to break through the current technical bottlenecks and achieve efficient utilization of saline-alkali land. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a method for bioremediation and intelligent salt control decision generation for crop rhizosphere soil. The objective of this invention can be achieved through the following technical solutions: S1: Acquire knowledge of root zone bioremediation and multi-source data on saline-alkali land regulation, and build a multimodal data association system; perform structured processing on the associated data and sort out the internal logical relationships of the data; analyze and train crop knowledge data, saline-alkali land soil data and crop growth data to generate an initial model for application scenarios of saline-alkali control and water and fertilizer management decisions. S2: Based on real-time environmental sensing data of the root zone and dynamic characteristics of crop growth period, a multi-dimensional dynamic analysis framework is constructed; the soil condition of the root zone and crop needs are analyzed to generate a synergistic scheme containing bioremediation components, salt control measures and nutrient allocation; the linkage relationship in the synergistic scheme is sorted out at the same time, and a dynamic mapping rule between scheme parameters and root zone soil characteristics is established. S3: Combine multi-source sensing feedback data of the root region to revise the preset root region environment regulation benchmark; adopt the root region repair and salt control synergy verification mechanism to verify the root region application adaptability of the synergy scheme and identify potential deviations of the synergy scheme; optimize the implementation parameters of the synergy scheme according to the verification results and collect root region status data simultaneously. S4: Build a collaborative management platform architecture for root region repair and salt control, store scheme execution data and root region status data and perform structured management; analyze the implicit regulation rules of data based on the spatiotemporal correlation of data, and build a dynamic iteration mechanism for the model; establish scheme parameter self-tuning logic through the platform to form a continuous iteration link of model-scheme-platform linkage.

[0007] As a preferred embodiment of the present invention, the specific method for establishing the multimodal data association system is as follows: Crop knowledge data consisting of literature, books and historical management records is collected and integrated, and real-time field data is acquired simultaneously. Using a multimodal large model based on the Transformer architecture, the text information in the crop knowledge data is aligned and semantically associated with the numerical and image information in the real-time field data, thus constructing an interactive mapping network between knowledge data and real-time data and forming a unified data association system.

[0008] Specifically, the method for structuring the associated data is as follows: Natural language processing technology is used to extract key parameters and interrelationships of crop growth from unstructured text knowledge and transform them into standardized data units with unified semantic definitions. At the same time, real-time image information is transformed into quantitative indicators that characterize crop growth status and soil environmental features through computer vision technology. Through unified spatiotemporal labels and semantic encoding, heterogeneous data are mapped to the same vector space to form a structured dataset with complete semantic information and spatiotemporal correlation.

[0009] Specifically, the method for generating the initial model for the application scenario of salinity control and water and fertilizer management decisions is as follows: Based on the structured dataset, a three-level architecture of encoder-fusionist-predictor is adopted for transfer learning and domain adaptation. With crop growth period as the core organizational dimension, crop physiological needs, soil salinity dynamics, and climatic environmental factors are coupled and modeled in multiple dimensions. Simultaneously, unmodeled key factors are transformed into quantifiable correction coefficients, threshold constraints, or adaptation rules to reduce the interference of unmodeled factors on decision accuracy. Through an attention mechanism that embeds agronomic features to anchor quantities, the model learns the water and fertilizer absorption patterns in the crop root zone and soil ion transport characteristics under salt stress. Simultaneously, the prediction accuracy of root zone acidification regulation, sodium ion replacement efficiency, and salt leaching effect targets is optimized, forming an initial application model capable of outputting a comprehensive water and fertilizer salt control decision scheme that includes the amount of acidic organic raw materials used, the ratio of salt-tolerant microbial agents, the precise application amount of ammonium nitrogen, and the timing of irrigation and fertilization.

[0010] Specifically, the method for constructing the multi-dimensional dynamic analysis framework is as follows: The analysis framework is based on time, environment, and crop physiology dimensions. In the time dimension, a dynamic analysis time series based on the crop growth cycle is established. In the environment dimension, a multi-parameter analysis space including soil salinity and alkalinity indicators, water status, and nutrient content is constructed. In the crop physiology dimension, a feature map covering crop salt tolerance, water requirement patterns, and nutrient requirements is formed. By designing data interfaces and coupling rules between multiple dimensions, an analysis framework capable of simultaneously analyzing the relationship between environmental dynamics and crop physiological responses is constructed.

[0011] Specifically, the method for generating the synergistic solution containing bioremediation components, salt control measures, and nutrient configuration is as follows: Based on the degree of soil salinization in the root zone, the application ratio of acidic organic raw materials to phosphorus ion activators is calculated to target the acidification of the root zone micro-domain and activate the fixed phosphorus element. Simultaneously, based on the soil sodium adsorption ratio and crop salt tolerance characteristics, salt-resistant microbial agents are formulated, which are made by mixing multiple salt-resistant strains. In terms of nutrient configuration, the application amount of ammonium nitrogen is controlled, and the acidification effect generated by the nitrification process in the soil is used to replace sodium ions on soil colloids. The bioremediation components, salt-controlling substances, and nutrients are integrated in a specific ratio and application sequence to form a comprehensive and synergistic regulation scheme for root zone acidification and alkali reduction, salt replacement and leaching, and relief of salt-alkali stress.

[0012] Specifically, the method for establishing the dynamic mapping rule between scheme parameters and root zone soil properties is as follows: Based on the various components and dosage parameters in the aforementioned synergistic scheme, a quantitative response relationship is established with the salt concentration, pH value, sodium adsorption ratio, and soil texture characteristics of the root zone soil. By analyzing historical regulation data and soil response data, a quantitative relationship between the input ratio of acidic organic raw materials and changes in soil pH value is determined. Correspondence rules are established between the dosage of phosphorus ion activator and the increase in available phosphorus content and calcium and magnesium ion activity in the soil. An empirical function is formed for the inoculation amount of salt-alkali resistant microbial agents and the reduction in sodium adsorption ratio (SAR). Furthermore, a dynamic relationship is constructed between the application amount of ammonium nitrogen and the degree of root zone acidification and sodium ion replacement efficiency, forming a mapping rule system for adjusting the application ratio of each component in the scheme.

[0013] Specifically, the method for modifying the preset root region environment control benchmark is as follows: Based on real-time monitoring of soil salinity and moisture in the root zone and crop growth stage data, the dynamic deviation between preset thresholds and actual values ​​is compared. When soil salinity remains above the threshold, the upper limit of salt tolerance is dynamically lowered based on the crop's current growth stage salt tolerance and water requirement characteristics. At the same time, combined with meteorological evaporation data and root development depth, the water requirement for soil salinity leaching in the root zone and irrigation quota are recalculated to form a dynamic control benchmark value adapted to the current environment and crop growth.

[0014] Specifically, the root region repair and salt control synergy verification mechanism includes: By using a pre-set material compatibility rule library, the chemical compatibility of acidic organic raw materials and salt-resistant microbial agents in the synergistic scheme is automatically compared and an early warning is issued. A synergy evaluation mechanism is constructed, and based on the matching degree analysis of water and fertilizer transport models and root distribution characteristics, the consistency of irrigation salt control measures and nutrient allocation in the spatiotemporal action of the root zone is evaluated. Key indicator thresholds are preset to verify the expected pH adjustment range, sodium ion replacement rate and salt leaching efficiency of the scheme in multiple dimensions.

[0015] Specifically, the method for building the root region repair and salt control collaborative management platform architecture is as follows: The cloud platform establishes a unified data middleware platform to perform spatiotemporal alignment and fusion processing of multimodal sensing data, scheme execution data, and root zone state data; the edge side deploys model inference services to sink the trained decision-making model to the near end of the field for low-latency real-time analysis and decision generation; the terminal layer integrates intelligent fertilizer application equipment and irrigation control system, and interacts with the edge side and cloud side for command and data feedback through standardized communication protocols; by constructing a data-driven three-level collaborative mechanism of cloud-edge-terminal, a root zone repair and salt control collaborative management platform architecture integrating perception, decision-making, execution, and feedback is formed.

[0016] Specifically, the dynamic iteration mechanism for building the model is implemented using the following method: Data on changes in soil salinity in the root zone, microbial community activity, crop yield, and stress mitigation after the implementation of the plan are obtained and combined with historical remediation case data to form an iterative data source. The actual implementation effect data is compared with the decision expectation data initially output by the model. Based on the verification results of the synergy between root zone remediation and salt control, the calculation weights of the bioremediation component ratio and the intensity of salt control measures in the model are adjusted. At the same time, the parameter optimization effect and root zone adaptation rules under the new scenario are added to the root zone remediation knowledge base, and the decision logic of the model for different saline-alkali soil types and crop growth stages is updated to form a closed-loop iteration of data feedback-parameter optimization-knowledge update-model upgrade.

[0017] Specifically, the method for establishing a continuous iterative link between the model, solution, and platform is as follows: A direct mapping channel is established between model decision output and scheme execution. The water, fertilizer and salt control decision scheme generated by the initial model of the application scenario is automatically converted into executable instructions through the cloud platform. The platform collects real-time data on changes in root zone soil parameters and crop growth response to construct an execution effect evaluation dataset. Based on the dataset, an automatic optimization process for model parameters is initiated, and the optimized model is regenerated to form a continuous iterative link of model decision-making, scheme execution and platform linkage.

[0018] The beneficial effects of this invention are as follows: (1) By constructing a multimodal data association system and intelligent decision-making model, the transformation from whole-field management to precise regulation of crop root zone has been realized, which has improved water and fertilizer utilization efficiency and saved water resources and amendment dosage.

[0019] (2) By establishing a biological synergistic regulation mechanism of water, fertilizer and salt, acidic organic raw materials, salt-resistant microbial agents and precise water and fertilizer programs are organically combined to achieve soil acidification and alkali reduction, sodium ion replacement and salt leaching in the root zone, thereby improving the rhizosphere microenvironment.

[0020] (3) By constructing a closed-loop iterative link of model-scheme-platform, real-time feedback of decision-making effect and model optimization are realized, which can adapt to the changes in different crop growth stages and soil conditions, and improve the adaptability and stability of the technology. Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating the bioremediation and intelligent salt control decision generation method for crop root zone soil according to the present invention. Figure 2 This is an architecture diagram of the bioremediation and intelligent salt control decision generation method for crop root zone soil in this invention. Detailed Implementation

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0024] Please see Figure 1-2 A method for generating bioremediation and intelligent salt control decisions for crop rhizosphere soil: S1: Acquire knowledge of root zone bioremediation and multi-source data on saline-alkali land regulation, and build a multimodal data association system; perform structured processing on the associated data and sort out the internal logical relationships of the data; analyze and train crop knowledge data, saline-alkali land soil data and crop growth data to generate an initial model for application scenarios of saline-alkali control and water and fertilizer management decisions. S2: Based on real-time environmental sensing data of the root zone and dynamic characteristics of crop growth period, a multi-dimensional dynamic analysis framework is constructed; the soil condition of the root zone and crop needs are analyzed to generate a synergistic scheme containing bioremediation components, salt control measures and nutrient allocation; the linkage relationship in the synergistic scheme is sorted out at the same time, and a dynamic mapping rule between scheme parameters and root zone soil characteristics is established. S3: Combine multi-source sensing feedback data of the root region to revise the preset root region environment regulation benchmark; adopt the root region repair and salt control synergy verification mechanism to verify the root region application adaptability of the synergy scheme and identify potential deviations of the synergy scheme; optimize the implementation parameters of the synergy scheme according to the verification results and collect root region status data simultaneously. S4: Build a collaborative management platform architecture for root region repair and salt control, store scheme execution data and root region status data and perform structured management; analyze the implicit regulation rules of data based on the spatiotemporal correlation of data, and build a dynamic iteration mechanism for the model; establish scheme parameter self-tuning logic through the platform to form a continuous iteration link of model-scheme-platform linkage.

[0025] As a preferred embodiment of the present invention, the specific method for establishing the multimodal data association system is as follows: Crop knowledge data, consisting of literature, books, and historical management records, is collected and integrated, while real-time field data is acquired simultaneously. This real-time field data includes soil salinity and pH data, crop canopy image data, and meteorological time-series data. Based on a multimodal large model with a Transformer architecture, the textual information in the crop knowledge data is cross-modal aligned and semantically associated with the numerical and image information in the real-time field data, thus constructing an interactive mapping network between knowledge data and real-time data and forming a unified data association system.

[0026] In this embodiment, spring wheat cultivation in saline-alkali land is used as the application scenario. The specific implementation process is as follows: First, multi-source data was collected and integrated. Regarding crop knowledge data, the system compiled monographs such as "Chinese Wheat Cultivation" and "Water and Fertilizer Requirements of Spring Wheat in Gansu Irrigation Areas," extracting key parameters such as fertilizer requirements at different growth stages of spring wheat (e.g., emergence, jointing, and heading stages), suitable soil moisture thresholds (e.g., the lower limit of soil moisture content at the seedling stage is 65% of field capacity), and salt tolerance thresholds (e.g., soil EC value at the seedling stage does not exceed 3.5 dS / m). Simultaneously, historical records of water and fertilizer management, soil testing reports, and yield data from the past five years in this region were integrated. For real-time field data, an integrated sensor network for soil moisture and salinity was deployed to collect data on volumetric water content and electrical conductivity in the 0-20cm and 20-40cm soil layers in real time; data on air temperature, humidity, wind speed, and evaporation were obtained through field weather stations; multispectral drones were used to conduct aerial photography of the field once a week to obtain crop growth information such as normalized difference vegetation index (NDVI) and leaf area index (LAI); supplemented by close-up images of crop roots taken manually on a regular basis.

[0027] Next, cross-modal alignment and semantic association were performed. A multimodal large model based on the Transformer architecture was used to deeply fuse the heterogeneous data. Specifically, natural language processing techniques were employed to perform entity recognition and relation extraction on the textual knowledge, transforming textual descriptions such as "nitrogen requirement at the jointing stage is 8-10 kg / mu" into structured triples (spring wheat, jointing stage, nitrogen requirement, 8-10 kg / mu). Simultaneously, computer vision algorithms were used to analyze UAV imagery, identifying crop growth stages and quantifying growth characteristics; sensor readings were serialized according to spatiotemporal dimensions. Based on this, a cross-modal attention mechanism was used to establish semantic associations between different modalities of data. For example, the textual description of "salt stress symptoms" was aligned with leaf yellowing image features captured by the UAV; the literature-recorded "suitable soil moisture threshold" was associated with real-time sensor-monitored water content data; and a correspondence was established between historical "fertilization effects" and current crop growth images.

[0028] Ultimately, a unified data association system is constructed. Through the above processing, all data is integrated into a unified semantic space, forming a map containing multi-dimensional information such as crop physiology, soil environment, meteorological conditions, and management measures.

[0029] Specifically, the method for structuring the associated data is as follows: Natural language processing technology is used to extract key parameters and interrelationships of crop growth from unstructured text knowledge and transform them into standardized data units with unified semantic definitions. At the same time, real-time image information is transformed into quantitative indicators that characterize crop growth status and soil environmental features through computer vision technology. Through unified spatiotemporal labels and semantic encoding, heterogeneous data are mapped to the same vector space to form a structured dataset with complete semantic information and spatiotemporal correlation.

[0030] Specifically, the method for generating the initial model for the application scenario of salinity control and water and fertilizer management decisions is as follows: Based on the structured dataset, a three-level architecture of encoder-fusionist-predictor is adopted for transfer learning and domain adaptation. Using crop growth stage as the core organizational dimension, crop physiological needs, soil salinity dynamics, and climatic environmental factors are coupled and modeled in multiple dimensions. Simultaneously, unmodeled key factors are transformed into quantifiable correction coefficients, threshold constraints, or adaptation rules. Through multimodal data complementarity and attention weight redirection, the interference of unmodeled factors on decision accuracy is reduced. By embedding agronomic features to anchor the attention mechanism, the water and fertilizer absorption patterns in the crop root zone and soil ion transport characteristics under salt stress are learned. Root zone optimization is performed concurrently. The prediction accuracy of acidification regulation, sodium ion replacement efficiency, and salt leaching effect targets is improved, forming an initial application model that can output a comprehensive water and fertilizer salt control decision scheme that includes the amount of acidic organic raw materials, the ratio of salt-resistant microbial agents, the precise application amount of ammonium nitrogen, and the timing of irrigation and fertilization. This includes a parameterization transformation step for unmodeled factors: transforming unmodeled key factors such as soil texture, initial state of microbial community, climate prediction uncertainty, and crop root dynamics into quantifiable correction coefficients, threshold constraints, or adaptation rules. Through multimodal data complementarity and attention weight redirection guidance, the interference of unmodeled factors on decision accuracy is reduced.

[0031] This embodiment describes the model construction process using the planting of 'Cuiguan' honey pears in a moderately saline-alkali area (soil salinity 3.2 g / kg, pH 8.5) as an application scenario: 1. Methods for constructing structured datasets The training data is constructed by integrating publicly available or standardly simulated historical knowledge data with real-time perception data.

[0032] (1) The composition and acquisition methods of historical basic data: Crop physiological parameters: These are derived from publicly published crop cultivation manuals, variety approval reports, and agricultural technology extension materials. For example, the daily nitrogen requirement of the 'Cuiguan' honey pear during its first fruit enlargement stage is approximately 1.49 g / plant / day. Parameters such as water requirements and salt tolerance thresholds at each growth stage can also be extracted from the aforementioned publicly available data.

[0033] Initial parameters of soil salinity and climatic data: Soil type, initial salinity (e.g., 3.2 g / kg), pH range (e.g., 8.5), and typical climatic characteristics (temperature, precipitation, evaporation) of the target area can be obtained from local soil records, meteorological yearbooks, and publicly available agricultural environmental monitoring reports.

[0034] Mapping relationship between management measures and effects: By collecting experimental cases on saline-alkali land improvement from publicly available literature, expert experience rule base (e.g., "On saline-alkali soil with pH 8.5, applying X kg of humic acid per acre can reduce the pH of the root zone by about 0.3-0.5 units"), and some verifiable field trial data (for calibration), a paired sample logic of "salt-alkali stress status (input) -> regulation measures and dosage (output) -> expected improvement effect (verification)" is constructed.

[0035] (2) Simulation generation of real-time sensing data: In order to enrich the training scenario, based on the above historical data and agronomic mechanism, sensor data (soil EC, pH, water content), UAV canopy image features (NDVI, LAI) and synchronous meteorological data are simulated and generated under different salinity gradients, weather conditions and growth stages to form training samples covering various possible scenarios.

[0036] (3) Specific construction of the dataset: Align the above data according to "growth period-soil depth-timestamp" to form a sample. Example of a single sample: The input is the text description "first swelling period, moderate salt stress", the soil EC time series sequence of the past 7 days, and the current canopy image features; the output labels are humic acid application rate of 12 kg / mu, inoculant ratio type B, ammonium nitrogen application rate of 8 kg / mu, and irrigation required in the next three days. Finally, a multimodal structured dataset containing about 8,000 such samples is constructed, and the training, validation and test sets are divided in a 7:2:1 ratio.

[0037] 2. Specific generation of the initial model: architecture, training, and core mechanisms 2.1 Detailed Definition of Model Architecture (MSDT-Model) The model adopts an encoder-fusionist-predictor architecture, as follows: Input encoder: Text encoder: Uses the BERT-base model (12 layers, 768 hidden layers, 12 attention heads, 30522 vocabulary), pre-trained weights use BERT-base-Chinese, and the transfer learning strategy is "freeze the first 6 layers, fine-tune the last 6 layers"; Numerical encoder: Soil and meteorological time-series data are processed through a one-dimensional convolutional layer (3 kernels, 128 channels, stride 1) and a bidirectional LSTM layer (256 hidden units, 2 layers, dropout=0.3). The image processing unit (ASU) outputs a 256-dimensional feature vector V_num. The image encoder uses a ResNet-34 model pre-trained on ImageNet (34 layers, 3 / 7 kernel size, 64-512 channels). The pre-trained weights are from the ImageNet dataset. The transfer learning strategy is to freeze the first 20 layers and fine-tune the last 14 layers. The ASU extracts canopy image features and pools them into a 512-dimensional vector V_img. The multimodal feature fusion unit concatenates the three features into a 1536-dimensional vector and inputs it into a 3-layer Transformer encoder for deep fusion. The encoder has a hidden layer dimension of 1024, 8 attention heads, a feedforward network dimension of 2048, and dropout=0.1. The multi-task predictor takes the [CLS] label vector from the fusion output and generates decisions through four parallel sub-networks: FC_Sigmoid (dosage regression layer): predicts the dosage of acidic organic raw materials (0-100 kg / mu); FC_Softmax (ratio classification layer): predicts the selection probability of microbial agents among 5 predefined ratios; FC_Sigmoid (dosage regression layer): predicts the dosage of ammonium nitrogen (0-30 kg / mu); FC_SeqSigmoid (temporal decision layer): predicts the daily irrigation decision (0 / 1) for the next 7 days.

[0038] In other words: Input and output tensor definitions. Model input: Each training sample consists of three tensors: X_text: [batch_size, 128, 768] (text description word vector sequence); X_num: [batch_size, 168, 8] (soil / meteorological time series data, 168 for 7 days × 24 hours, 8 for 8 parameters such as EC / pH / water content); X_img: [batch_size, 3, 1024, 1024] (crop canopy RGB image); Model output: The model simultaneously... Four decision tensors are generated: Y_acid: [batch_size,1] (acidic raw material dosage, unit kg / mu, range 0-100); Y_microbe: [batch_size,5] (probability distribution of 5 microbial agent ratios); Y_nitrogen: [batch_size,1] (ammonium nitrogen dosage, unit kg / mu, range 0-30); Y_irrig: [batch_size,7] (daily irrigation decision for the next 7 days, 0=no irrigation, 1=irrigation).

[0039] The correlation between model inputs (soil EC, pH, images, text, and historical data) and outputs (material application rate, irrigation decisions) is not a simple statistical correlation, but rather a clear causal relationship based on soil chemistry and crop physiology, as detailed below: The causal mechanism between ammonium nitrogen application rate and sodium ion replacement: Ammonium nitrogen undergoes an oxidation reaction under the action of soil nitrifying bacteria (NH4⁺ + 2O2 → NO3⁻ + 2H⁺ + H2O). The generated H⁺ can lower the pH in the root zone, disrupt the adsorption balance of sodium ions (Na⁺) by soil colloids, and promote the replacement of Na⁺ by H⁺, which is then leached out with irrigation water. The reaction equilibrium constant is K = 1.8 × 10⁻⁶. 5 (25℃), the model calculates the total amount of H⁺ required to reach the target pH using this reaction formula, and then inversely calculates the amount of ammonium nitrogen to be applied; Synergistic effect of humic acid and soil acidification: The carboxyl group (-COOH) and phenolic hydroxyl group (-OH) of humic acid can dissociate H⁺, directly reducing soil pH. At the same time, its macromolecular structure can adsorb Ca²⁺ and Mg²⁺ on the surface of soil colloids, indirectly promoting the replacement of Na⁺. The model combines the degree of dissociation of humic acid (65% at pH 8.5) with the soil buffer capacity to calculate the amount of humic acid required to achieve the target acidification effect. The correlation between microbial inoculants and soil structure improvement: The organic acids secreted by salt-tolerant bacterial strains (such as Bacillus subtilis) can aid in acidification, while polysaccharides can promote soil aggregate formation (increasing the proportion of aggregates with a diameter ≥0.25mm by 20%), increase soil porosity, and improve salt leaching efficiency. The model quantifies the relationship between microbial activity and aggregate structure improvement (for every 10% increase in activity...). 6 (CFU / g, porosity increased by 3%), and the bacterial agent ratio was optimized.

[0040] 2.2 The core mechanism for learning the patterns of salt stress: parameterized attention guidance + embedding agronomic knowledge In order to enable the model to explicitly capture the water and fertilizer absorption patterns and soil ion transport characteristics of crop root zone under salt stress, a dual module of "parametric attention guidance + agronomic feature anchoring" is embedded in the first layer Transformer of the feature fusion machine. The specific implementation steps are as follows: (1) Stress intensity quantification and attention triggering: extract the mean value of soil EC and pH deviation value from the numerical feature V_num, and map them to the scalar salt stress intensity s through a two-layer perceptron (structure: 2→8→1, ReLU activation) (value range 0-1, s≥0.6 is judged as severe stress). When s≥0.3, the dedicated attention branch is automatically activated to enhance the capture of water and fertilizer absorption and ion transport related features; (2) Agronomic feature anchoring: Construct a “water and fertilizer-ion transport feature dictionary” containing 200+ core agronomic features (such as “root zone ammonium nitrogen concentration”, “sodium ion leaching rate”, “crop transpiration coefficient”, “soil ion diffusion coefficient”, etc.), and convert the dictionary features into 128-dimensional embedding vectors as “anchor vectors” for the attention mechanism; (3) Attention weight calculation: In the Transformer attention head, the cosine similarity calculation between the anchor vector and the input features is introduced, and the weight update formula is: , where Anchor is the agronomic feature anchoring vector, s is the stress intensity, and d_k is the attention head dimension (128). This formula ensures that the model prioritizes features related to water and fertilizer absorption and ion transport under high stress scenarios; (4) Ion transport law reinforcement learning: During the training process, "ion transport constraint loss" is added, and a theoretical model is constructed based on the convection-diffusion equation (CDE) in soil physics. The expression of the CDE equation is: ∂C / ∂t=D∂²C / ∂x²-v∂C / ∂x, where C is the ion concentration, t is the time, D is the diffusion coefficient, and v is the average pore water flow velocity. The theoretical ion concentration distribution C_cde under different spatiotemporal conditions is calculated by this equation. The deviation between the ion concentration distribution C_pred and C_cde predicted by the model is used as the constraint loss, i.e., L_diff=|C_pred-C_cde|. This loss is incorporated into the total loss function with a weight of 0.3. The model learns the transport characteristics that conform to physical laws; (5) Robust design of unmodeled factors: For unmodeled factors such as soil texture profile and initial state of microbial community, the interference is offset by the "multimodal data complementarity + uncertainty quantification" mechanism: Multimodal data complementarity: textual knowledge (literature records that "the ion diffusion coefficient of sandy loam is 30% higher than that of clay loam"), numerical data (real-time EC / pH from sensors), and image data (crop root distribution phenotype) are aligned across modalities. When a certain factor is not modeled, the model can correct the decision by the correlation patterns of other modal data. Uncertainty quantification: A Bayesian neural network is used to extend the model architecture and generate confidence intervals (e.g., acidic raw material dosage = 12 kg / mu ± 0.8 kg / mu) when outputting decisions. When unmodeled factors cause the prediction variance to exceed the threshold (±15%), a "conservative decision-making mechanism" is automatically triggered (e.g., increasing the acidic raw material dosage by 10% and shortening the irrigation interval) to ensure the safety of field application; (6) Mechanism effectiveness verification: Through attention visualization heatmap, it is verified that the model's attention weight for core features such as "sodium ion replacement", "water leaching", and "ion diffusion" in high salt stress samples is increased by more than 40% compared with the baseline model; at the same time, on test sets with different soil textures such as sandy loam and clay loam, the predicted R² of ion transport law is ≥ 0.85, proving that the model can stably learn the target law.

[0041] 2.3 Training Strategy and Loss Function Loss function: The total loss is "weighted multi-task loss + constraint penalty loss + ion transport constraint loss": L_total=1.0×L_acid+0.8×L_microbe+1.2×L_nitrogen+0.5×L_irrig+λ×L_constraint+0.3×L_diff Where: L_acid: Mean squared error loss for predicting acidic organic feedstock usage, using MSE to optimize continuous value regression accuracy; L_microbe: Cross-entropy loss for predicting microbial agent ratios, suitable for single-label multi-classification tasks; L_nitrogen: Huber loss for predicting ammonium nitrogen usage, balancing accuracy and resistance to outlier data interference; L_irrig: Binary cross-entropy loss for irrigation decision prediction, suitable for sequential binary decision-making; L_constraint: Agronomic constraint penalty loss, constructed based on the following 3 types of material balance / compatibility equations, generating a penalty value when the constraints are not met: ① Nitrogen balance constraint equation Equations: ① N_input = N_crop + N_leach + N_denit (N_input is the amount of ammonium nitrogen applied, N_crop is the amount absorbed by the crop, N_leach is the amount of leaching loss, and N_denit is the amount of denitrification loss); ② Material compatibility constraint equation: pH_min ≤ pH_soil - 0.1 × M_acid + 0.03 × M_microbe ≤ pH_max (M_acid is the amount of acidic raw material used, M_microbe is the amount of microbial agent used, and pH_min / pH_max is the critical pH for microbial agent survival); ③ Salt leaching water requirement constraint equation: I ≥ (EC_soil - EC_target) × 10 × γ (I is the irrigation amount, EC_target is the target EC value, and γ is the soil infiltration coefficient); L_diff: Ion transport constraint loss, ensuring that the model learns transport characteristics that conform to physical laws; λ is the penalty weight (set to 2.0 after training optimization), ensuring that the constraint conditions are satisfied first.

[0042] The weight parameters (α, β, γ, δ) are set based on the following criteria: Acidic raw materials (weight 1.0): As the core improver for root zone acidification and alkali reduction, the accuracy of dosage directly affects the pH control target; Microbial agents (weight 0.8): Belong to discrete category decision-making, and have a synergistic relationship with raw material dosage, with a weight slightly lower than the continuous dosage prediction task; Ammonium nitrogen (weight 1.2): Nitrification is the key driving force for root zone acidification and sodium ion replacement, and is related to yield and environmental safety, requiring the highest accuracy; Irrigation decision (weight 0.5): Time-series decision-making has greater flexibility, and has been considered in conjunction with material dosage to learn basic laws.

[0043] Transfer learning and training: Initialization: Load general pre-trained weights for text and image encoders; Two-stage fine-tuning: Stage 1 (feature adaptation): Freeze the underlying parameters of BERT and ResNet, train the remaining parts for 10 epochs, with a learning rate of 3e-5; Stage 2 (global tuning): Unfreeze all parameters, fine-tune training for 20 epochs, reduce the learning rate to 1e-5, and apply gradient clipping (threshold 1.0). 3. Model Validation Quantitative performance evaluation: The model performance was quantitatively evaluated on an independent test set (approximately 800 samples, representing 10% of the dataset). The mean absolute error (MAE) for predicting the amount of acidic feedstock used is 2.1 ± 0.3 kg / mu.

[0044] The accuracy rate for classifying microbial inoculant ratios reached 91.5%. This accuracy rate was calculated by the model's prediction results for inoculant ratio categories on the test set, specifically as the percentage of correctly predicted samples out of the total number of test samples.

[0045] The mean absolute error (MAE) for predicting ammonium nitrogen application rate is 0.8 ± 0.1 kg / mu.

[0046] The F1 score for the irrigation decision sequence prediction is 0.89.

[0047] The predicted ion transport patterns have an R² ≥ 0.85.

[0048] Mechanism effectiveness verification: Attention visualization shows that when high-salt data is input, the model's attention weight for regulatory keywords in text features increases by more than 40% on average compared to the baseline scenario, which intuitively confirms the successful operation of the "salt perception attention guidance mechanism".

[0049] Specifically, the method for constructing the multi-dimensional dynamic analysis framework is as follows: The analysis framework is based on time, environment, and crop physiology dimensions. In the time dimension, a dynamic analysis time series based on the crop growth cycle is established. In the environment dimension, a multi-parameter analysis space including soil salinity and alkalinity indicators, water status, and nutrient content is constructed. In the crop physiology dimension, a feature map covering crop salt tolerance, water requirement patterns, and nutrient requirements is formed. By designing data interfaces and coupling rules between multiple dimensions, an analysis framework capable of simultaneously analyzing the relationship between environmental dynamics and crop physiological responses is constructed.

[0050] Specifically, the method for generating the synergistic solution containing bioremediation components, salt control measures, and nutrient configuration is as follows: Based on the degree of soil salinization in the root zone, the application ratio of acidic organic raw materials to phosphorus ion activators is calculated to target the acidification of the root zone micro-domain and activate the fixed phosphorus element. Simultaneously, based on the soil sodium adsorption ratio (SAR) and crop salt tolerance characteristics, salt-resistant microbial agents are formulated, which are made by mixing multiple salt-resistant strains. In terms of nutrient configuration, the application rate of ammonium nitrogen is controlled, and the acidification effect generated by the nitrification process in the soil is used to replace sodium ions on the soil colloid. The bioremediation components, salt-controlling substances, and nutrients are integrated in a specific ratio and application sequence to form a comprehensive and synergistic regulation scheme for root zone acidification and alkali reduction, salt replacement and leaching, and relief of salt-alkali stress.

[0051] In this embodiment, a moderately saline-alkali crop planting area in a certain region is used as the application scenario. The process of generating a synergistic solution containing bioremediation components, salt control measures and nutrient allocation is implemented for the root zone soil characteristics of crops (such as potatoes, wheat, fruit trees and other crops) during the key growth period (the period of vigorous fertilizer and water demand). The initial salinization data of the 0-30cm root zone soil in this region are: salt content Ag / kg, pH B, sodium adsorption ratio (SAR) C, the salt tolerance threshold of the crop during the key growth period is SAR≤C1, the root zone available phosphorus requirement is ≥Dmg / kg, and the initial available phosphorus content is D2mg / kg (D2<D).

[0052] First, based on the analysis of the soil salinization level in the root zone, the application ratio of acidic organic raw material (humic acid, HA) and phosphorus ion activator (citric acid, CA) was determined. Combining the soil pH buffering capacity and phosphorus fixation coefficient in the root zone, the application mass ratio of HA to CA was calculated to be E:F, with specific dosages of HAG kg / mu and CAH kg / mu. This ratio can directionally acidify the 0-20cm micro-zone soil in the root zone, reducing the root zone pH from the initial pH B to the target pH B1 (B1 < B) within one day. Simultaneously, through the chelating effect of CA, phosphorus fixed by calcium and magnesium ions in the soil is activated, increasing the available phosphorus content in the root zone from the initial D2 mg / kg to D-D3 mg / kg (D3 is the available phosphorus fluctuation value, and D-D3 ≥ D), meeting the crop's requirement of ≥D mg / kg of available phosphorus during critical growth stages.

[0053] Secondly, based on the initial soil salinity (SARC) and the crop's critical growth stage salinity (SAR) ≤ C1, a salinity-tolerant microbial agent was formulated. The model screened out Bacillus subtilis and Bacillus mucilaginosa at a viable count ratio of 1:5 to produce an agent with an effective viable count ≥ L billion CFU / g, applied at a rate of M kg / acre. After application, the agent helps reduce soil alkalinity in the rhizosphere through the organic acids produced by the strain's metabolism (helping the pH drop from B to B1), while simultaneously secreting polysaccharides to improve the rhizosphere soil aggregate structure. The synergistic effect of these two factors reduces the rhizosphere SAR from the initial C to the target C1 within N days, simultaneously improving the crop roots' absorption efficiency of HA, CA, and subsequent nutrients.

[0054] In the nutrient allocation stage, the application rate of ammonium nitrogen is the key focus. The model combines the nitrogen requirements of crops during key growth stages with the soil nitrification rate, and constrains the application rate through the nitrogen balance equation: N_input = N_crop + N_leach + N_denit, where N_input is the application rate of ammonium nitrogen (model output), N_crop is the amount absorbed by the crop in the current season (calculated based on the characteristics of the crop growth stage, such as N_crop = 8-10 kg / mu at the jointing stage), N_leach is the leaching loss (positively correlated with irrigation amount, N_leach = 0.05 × I, where I is the irrigation quota), and N_denit is the denitrification loss (related to soil moisture content; when soil moisture content > 70% of field capacity, N_denit = 0.1 × N_input). The final ammonium sulfate application rate was determined. After its application, nitrifying bacteria in the soil convert ammonium nitrogen into nitrate nitrogen, producing trace amounts of nitrate. This helps maintain the pH of the root zone at B1 and replaces sodium ions adsorbed on the soil colloid surface, resulting in a higher sodium ion leaching rate in the root zone compared to conventional fertilization. At the same time, it meets the nitrogen requirements of crops during their critical growth stages.

[0055] Meanwhile, material compatibility constraints are embedded in the scheme generation process: pH_min≤pH_soil-0.1×M_acid+0.03×M_microbe≤pH_max, where M_acid is the amount of acidic raw material, M_microbe is the amount of microbial agent, and pH_min / pH_max is the critical pH for microbial agent survival (e.g., the suitable pH for Bacillus subtilis is 6.5-7.5), ensuring that the acidification effect of acidic raw materials does not destroy the activity of microbial agents; the irrigation amount is determined by the salt leaching water demand constraint equation: I≥(EC_soil-EC_target)×10×γ, where EC_target is the target EC value, and γ is the soil infiltration coefficient (γ=1.2 for sandy loam and γ=0.8 for clay loam), ensuring that the irrigation amount meets the salt leaching requirements.

[0056] Finally, the above-mentioned bioremediation components (salt-resistant microbial agent, dosage M kg / mu), salt-controlling substances (HAG kg / mu, CAH kg / mu), and nutrients (ammonium sulfate R kg / mu) were integrated according to a specific time sequence: On the first day, the HA and CA mixture was evenly spread, followed by drip irrigation of U m³ / mu to allow the substances to penetrate into the 0-20cm root zone, laying the foundation for the pH to drop to B1; on the fourth day, the salt-resistant microbial agent was spread, combined with shallow tillage (depth W-W1cm) to promote the contact between the agent and the rhizosphere soil, accelerating the conversion of SAR from C to C1; on the sixth day, ammonium sulfate solution was applied through the drip irrigation system (R kg / mu of ammonium sulfate was dissolved in U1 m³ / mu of water) to avoid excessively high local salt concentrations. Fifteen days after the implementation of this synergistic program, the soil pH in the root zone stabilized at B1, the SAR remained at C1, and the available phosphorus content remained at D-D3 mg / kg. The dry weight of the crop roots increased by Z% compared with conventional management, and the chlorophyll content (SPAD value) of the leaves increased from Z1 to Z2, effectively alleviating the salt and alkali stress in the root zone and matching the growth needs of the crop during its key growth period.

[0057] Specifically, the method for establishing the dynamic mapping rule between scheme parameters and root zone soil properties is as follows: Based on the various components and dosage parameters in the synergistic scheme, a quantitative response relationship is established with the salt concentration, pH value, sodium adsorption ratio, and soil texture characteristics of the root zone soil. By analyzing historical control data and soil response data, a quantitative relationship between the input ratio of acidic organic raw materials and changes in soil pH value is determined. Correspondence rules are established between the dosage of phosphorus ion activator and the increase in available phosphorus content and calcium and magnesium ion activity in the soil. An empirical function is formed for the inoculation amount of salt-alkali resistant microbial agents and the decrease in sodium adsorption ratio. A dynamic relationship is constructed between the application amount of ammonium nitrogen and the degree of root zone acidification and sodium ion replacement efficiency. A mapping rule system is formed that can dynamically adjust the application ratio of each component in the scheme according to the real-time monitored soil characteristic parameters. The phosphorus ion activator includes organic acid activators, microbial activators, and inorganic chelating activators.

[0058] Specifically, the method for modifying the preset root region environment control benchmark is as follows: Based on real-time monitoring of soil salinity and moisture in the root zone and crop growth stage data, the dynamic deviation between preset thresholds and actual values ​​is compared. When soil salinity remains above the threshold, the upper limit of salt tolerance is dynamically lowered based on the crop's current growth stage salt tolerance and water requirement characteristics. At the same time, combined with meteorological evaporation data and root development depth, the water requirement for soil salinity leaching in the root zone and irrigation quota are recalculated to form a dynamic control benchmark value adapted to the current environment and crop growth.

[0059] Specifically, the root region repair and salt control synergy verification mechanism includes: By using a pre-set material compatibility rule library, the chemical compatibility of acidic organic raw materials and salt-tolerant microbial agents in the synergistic scheme is automatically compared and an early warning is issued. A synergy evaluation mechanism is constructed, and based on the matching degree analysis of water and fertilizer transport models and root distribution characteristics, the consistency of irrigation salt control measures and nutrient allocation in the spatiotemporal action of the root zone is evaluated. Key indicator thresholds are preset to verify the expected pH adjustment range, sodium ion replacement rate and salt leaching efficiency of the scheme in multiple dimensions. The key indicator thresholds include root zone soil chemical regulation thresholds: soil salinity index, soil pH value, sodium adsorption ratio (SAR) available phosphorus content; crop physiological adaptation thresholds: crop salt tolerance threshold, root zone water threshold, and nutrient synergy threshold.

[0060] Specifically, the method for building the root region repair and salt control collaborative management platform architecture is as follows: The cloud platform establishes a unified data middleware platform to perform spatiotemporal alignment and fusion processing of multimodal sensing data, scheme execution data, and root zone state data; the edge side deploys model inference services to sink the trained decision-making model to the near end of the field for low-latency real-time analysis and decision generation; the terminal layer integrates intelligent fertilizer application equipment and irrigation control system, and interacts with the edge side and cloud side for command and data feedback through standardized communication protocols; by constructing a data-driven three-level collaborative mechanism of cloud-edge-terminal, a root zone repair and salt control collaborative management platform architecture integrating perception, decision-making, execution, and feedback is formed.

[0061] In this embodiment, a winter wheat planting area in a moderately saline-alkali land in North China is used as the application scenario. The construction and operation of a root zone restoration and salt control collaborative management platform are implemented to meet the needs of root zone restoration and salt control throughout the entire wheat growth period (sowing period - greening period - jointing period - grain filling period). The specific functions and collaborative actions of each level of the platform are explained in detail.

[0062] As the "data hub" of the architecture, the cloud platform's core task is to process three types of core data based on a unified data platform. This perfectly matches the document's requirement to "perform spatiotemporal alignment and fusion processing of multimodal sensing data, scheme execution data, and root zone status data." In actual operation, the cloud first receives multi-source data from the field via the network: multimodal sensing data comes from soil sensors (uploading 0-30cm soil salinity, pH, and moisture data every 5 minutes), drones (collecting crop canopy NDVI images twice a week), and weather stations (uploading temperature, humidity, and precipitation data hourly); scheme execution data includes the daily fertilizer application ratio of the terminal layer's intelligent fertilizer application equipment and the irrigation flow rate and duration of the irrigation control system; and root zone status data covers laboratory test results and field observation results such as root distribution depth and effective phosphorus content in the root zone. Next, the data platform aligns heterogeneous data spatiotemporally using "timestamp (accurate to the second) + plot coordinates (accurate to the square meter)". For example, it binds soil pH data of a plot at a specific time with irrigation execution data from the same period. Furthermore, it uses algorithms to fuse crop growth stress levels converted from image data with numerical soil data, generating a structured dataset. Finally, this data is stored in a distributed database, supporting on-demand access from both edge and terminal layers. For instance, if an edge layer needs the salinity trend of a plot over the past 7 days, the data platform can return the required data within one second, providing support for subsequent decision-making.

[0063] The edge server assumes the role of "near-field decision-making," strictly adhering to the design outlined in the document: "deploying model inference services to bring the trained decision-making model down to the near-field level." Edge servers with GPU computing power are deployed at field management stations to locally deploy the "initial model for the application scenario of salt-alkali control and water and fertilizer management decision-making" defined in the document, focusing on low-latency analysis and ad-hoc decision-making. During daily operation, the edge server receives high-frequency data from the terminal layer via a local area network, directly calling the local model for analysis without transmitting it to the cloud. For example, when the soil salinity in a plot suddenly exceeds the crop's salt tolerance threshold, the edge server can complete data verification and anomaly detection within 100ms and generate a temporary "start temporary irrigation" command, directly sending it to the terminal layer, avoiding control delays caused by cloud transmission latency. Furthermore, during the daily early morning hours when there is no field operation, the edge server synchronizes model parameters with the cloud, updating the local model with decision coefficients optimized by the cloud based on full-region data (such as the weighting of acidic organic raw material usage), ensuring consistency in decision accuracy with the cloud and meeting the document's requirement for "low-latency real-time analysis and decision generation."

[0064] The terminal layer, acting as the "execution feedback end," establishes hardware and communication links according to the document's requirement to "integrate intelligent fertilizer blending equipment and irrigation control systems, interacting through standardized communication protocols." Intelligent fertilizer blending machines (with flow and ratio sensors) and drip irrigation controllers (with solenoid valves) are deployed in the planting area at a density of one set per 50 mu (approximately 3.3 hectares). All devices connect to the edge and cloud via the MQTT standardized protocol. When receiving temporary instructions from the edge or periodic instructions from the cloud (such as "humic acid: phosphorus ion activator = 5:1, fertilizer application rate 200 kg / mu"), the intelligent fertilizer blending machine automatically mixes the raw materials according to the ratio and delivers them to the field through pipelines; the drip irrigation controller precisely controls the solenoid valve switching to execute irrigation duration and flow rate. Simultaneously, the terminal layer collects and transmits execution data in real time—the fertilizer blending machine transmits "actual fertilizer application rate and deviation value" (usually ≤2%), and the drip irrigation controller transmits "actual irrigation duration and instruction deviation" (usually ≤1 minute), ensuring that the cloud and edge sides have real-time control over instruction execution, achieving the functional closed loop of "instruction interaction and data feedback" as described in the document.

[0065] Specifically, the dynamic iteration mechanism for building the model is implemented using the following method: Data on changes in root zone soil salinity, microbial community activity, crop yield, and stress mitigation after the implementation of the scheme, combined with historical remediation case data, form an iterative data source. The actual implementation effect data is compared with the decision expectation data initially output by the model. Based on the verification results of the synergy between root zone remediation and salt control, the calculation weights of the bioremediation component ratio and the intensity of salt control measures in the model are adjusted. Simultaneously, the parameter optimization effect under the new scenario and the root zone adaptation law are added to the root zone remediation knowledge base to update the model's decision logic for different saline-alkali soil types and crop growth stages, forming a closed-loop iteration of data feedback, parameter optimization, knowledge update, and model upgrade.

[0066] In this embodiment, the implicit regulation law based on the spatiotemporal correlation of data and the construction of a dynamic iteration mechanism for the model are discussed. I. Core Methods for Analyzing Implicit Regulatory Patterns Through Spatiotemporal Data Correlation Based on self-collected multi-source data (real-time soil sensor data, program execution records, field measurement data, historical remediation case data, etc.), analysis is achieved through a self-developed analysis mode of "spatiotemporal anchoring + correlation mining". Specific steps include: Time-dimensional correlation analysis: Data is serialized according to crop growth stages (sowing stage - jointing stage - heading stage - grain filling stage), and the "implementation time of control measures" is precisely aligned with the "response time of root zone status" (such as changes in soil salinity and pH 24 hours, 48 ​​hours, and 72 hours after irrigation) to construct a time-series data sequence; the delayed correlation between control measures and effects at different growth stages is captured, such as identifying the pattern that the sodium ion replacement rate reaches its peak 36 hours after ammonium nitrogen application at the jointing stage.

[0067] Spatial Dimension Correlation Analysis: Spatial units are defined by root zone soil layers (0-20cm, 20-40cm) and plot zoning (divided by soil texture or topography), and the "measure application space" and "soil response space" are matched (such as the correspondence between drip irrigation coverage areas and root zone salinity reduction areas). Through spatial interpolation and zonal comparative analysis, the differences in the control effect of different spatial units are explored. For example, it was found that the amount of salt control measures in sandy loam areas needs to be 15% higher than that in clay loam areas to achieve the same effect.

[0068] Multi-dimensional cross-correlation analysis: Soil salinity data, water and fertilizer application data, climate data, and crop growth data are automatically integrated and aligned according to "spatiotemporal tags". Pearson correlation analysis, decision tree mining and other algorithms are used to screen strong correlation factors in multi-source data. For example, cross-analysis found that when the average daily temperature is higher than 28℃ and the soil moisture content is lower than 15%, the acidification efficiency of acidic organic raw materials will decrease by 20%, and the application rate needs to be adjusted.

[0069] II. Core Categories of Implicit Regulation Patterns Implicit regulatory patterns are the inherent logics that do not directly manifest themselves but govern regulatory effects, and they mainly include four categories: Spatiotemporal coupling law: the synergistic effect of the temporal combination and spatial distribution of control measures, such as "applying phosphorus ion activator 36 hours after irrigation during the heading stage, the effective phosphorus in the 0-20cm soil layer increases by 22% compared to random timing" and "slope plots need to be zoned along contour lines to control salt, otherwise the salt leaching will be uneven, and the effect will be biased by 18%".

[0070] Environmental adaptation rules: the adaptation relationship between environmental factors such as climate and soil texture and control parameters, such as "more than 3 consecutive days of rainfall, the activity of salt-alkali resistant microbial agents is enhanced, and the inoculation amount can be reduced by 10%" and "when the soil organic matter is less than 20g / kg, the acidification efficiency of acidic organic raw materials is reduced, and the application amount needs to be increased by 20%".

[0071] Threshold effect law: the critical value of key parameters and the abrupt change in effect after exceeding the threshold, such as "when the root zone pH is below 7.0, continuing to apply acidic raw materials will lead to a 30% decrease in crop phosphorus absorption efficiency" and "when the soil salinity concentration is above 4.0 g / kg, the effect of simple water and fertilizer regulation is saturated, and bioremediation components need to be added to further reduce salinity".

[0072] Synergistic and antagonistic patterns: the interaction between different regulatory measures, such as "when acidic organic raw materials and ammonium nitrogen are applied in combination, the sodium ion replacement rate increases by 16% compared with the application alone (synergistic effect)" and "high concentration of phosphorus ion activator and Bacillus subtilis have antagonistic effects, which will lead to a 12% decrease in the activity of the inoculant (antagonistic effect)".

[0073] III. Construction process of the model dynamic iteration mechanism (no platform dependency) Iterative data source integration: Collect self-collected and stored scheme execution data (water and fertilizer application, application timing, etc.), root zone status data (salt content, pH, microbial activity, etc.), and crop growth data (yield, stress relief indicators, etc.). Combine this with historical remediation case data and archive them uniformly according to "spatiotemporal tags" to form an iterative data source, ensuring that the data covers different soil types, growth stages, and climatic conditions.

[0074] Implicit pattern mining and verification: Implicit regulatory patterns are extracted using the above spatiotemporal correlation analysis method. The reliability of the patterns is verified using independent test data (accounting for 10% of the total data) (e.g., verifying that the adaptability of the "spatiotemporal coupling pattern" is ≥85% in different plots). The verified patterns are then transformed into quantitative rules (e.g., mathematical functions, decision tables).

[0075] Model parameter optimization: Adjust model weights based on implicit laws, such as increasing the calculation weight of climate factors on the ratio of microbial agents according to the "environmental adaptation law"; and setting parameter critical value constraints in the model according to the "threshold effect law" to avoid decisions exceeding the effective range.

[0076] Knowledge base update and model upgrade: Newly discovered implicit patterns are added to the root zone repair knowledge base, and the model's decision-making logic for different salinity types and growth stages is updated; through local-terminal data linkage, the optimized model is deployed to the field terminal, and continuous iteration is carried out based on real-time data to form a closed loop of "data collection-pattern mining-parameter optimization-model upgrade".

[0077] Specifically, the method for establishing a continuous iterative link between the model, solution, and platform is as follows: A direct mapping channel is established between model decision output and scheme execution. The water, fertilizer and salt control decision scheme generated by the initial model of the application scenario is automatically converted into executable instructions through the cloud platform. The platform collects real-time data on changes in root zone soil parameters and crop growth response to construct an execution effect evaluation dataset. Based on the dataset, an automatic optimization process for model parameters is initiated, and the optimized model is regenerated to form a continuous iterative link of model decision-making, scheme execution and platform linkage.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for bioremediation and intelligent salt control decision generation for crop rhizosphere soil, characterized in that, include: S1: Acquire knowledge of root zone bioremediation and multi-source data on saline-alkali land regulation, and build a multimodal data association system; Structure the related data and clarify the inherent logical relationships between them; Analyze and train crop knowledge data, saline-alkali soil data, and crop growth data to generate an initial model for application scenarios of saline-alkali control and water and fertilizer management decisions. S2: Based on real-time environmental sensing data of the root zone and dynamic characteristics of crop growth period, a multi-dimensional dynamic analysis framework is constructed; the soil condition of the root zone and crop needs are analyzed to generate a synergistic scheme containing bioremediation components, salt control measures and nutrient allocation; the linkage relationship in the synergistic scheme is sorted out at the same time, and a dynamic mapping rule between scheme parameters and root zone soil characteristics is established. S3: Combine multi-source sensing feedback data of the root region to revise the preset root region environment regulation benchmark; adopt the root region repair and salt control synergy verification mechanism to verify the root region application adaptability of the synergy scheme and identify potential deviations of the synergy scheme; optimize the implementation parameters of the synergy scheme according to the verification results and collect root region status data simultaneously. S4: Build a collaborative management platform architecture for root region repair and salt control, store scheme execution data and root region status data and perform structured management; analyze the implicit regulation rules of data based on the spatiotemporal correlation of data, and build a dynamic iteration mechanism for the model; establish scheme parameter self-tuning logic through the platform to form a continuous iteration link of model-scheme-platform linkage.

2. The method according to claim 1, characterized in that, The specific method for building the multimodal data association system is as follows: Crop knowledge data composed of literature, books and historical management records is acquired, and real-time field data is acquired simultaneously. Multimodal modeling technology is used to perform cross-modal alignment and semantic association between the text information in the crop knowledge data and the numerical and image information in the real-time field data, thereby constructing an interactive mapping network between knowledge data and real-time data and forming a unified data association system.

3. The method according to claim 1, characterized in that, The specific method for structuring the associated data is as follows: Natural language processing technology is used to extract key parameters and interrelationships of crop growth from unstructured text knowledge and transform them into standardized data units with unified semantic definitions. At the same time, real-time image information is transformed into quantitative indicators that characterize crop growth status and soil environmental features through computer vision technology. Through unified spatiotemporal labels and semantic encoding, heterogeneous data are mapped to the same vector space to form a structured dataset with complete semantic information and spatiotemporal correlation.

4. The method according to claim 1, characterized in that, The initial model for the application scenario of generating salinity control and water and fertilizer management decisions is specifically modeled as follows: Based on the structured dataset, a three-level architecture of encoder-fusionist-predictor is adopted for transfer learning and domain adaptation; crop growth period is taken as the core organizational dimension, and crop physiological needs, soil salinity and alkali dynamics and climate environmental factors are coupled and modeled in multiple dimensions. Simultaneously, unmodeled key factors are transformed into quantifiable correction coefficients, threshold constraints, or adaptation rules; by embedding agronomic features to anchor the attention mechanism of quantities, the water and fertilizer absorption patterns of crop rhizosphere and soil ion transport characteristics under salt stress are learned. Simultaneously optimize the prediction accuracy of root zone acidification regulation, sodium ion replacement efficiency, and salt leaching effect targets, and form an initial application model that can output a comprehensive water and fertilizer salt control decision scheme including the amount of acidic organic raw materials, the ratio of salt-resistant microbial agents, the precise application amount of ammonium nitrogen, and the timing of irrigation and fertilization.

5. The method according to claim 1, characterized in that, The specific method for constructing the multi-dimensional dynamic analysis framework is as follows: The analysis framework is based on time, environment, and crop physiology dimensions. In the time dimension, a dynamic analysis time series based on the crop growth cycle is established. In the environment dimension, a multi-parameter analysis space including soil salinity and alkalinity indicators, water status, and nutrient content is constructed. In the crop physiology dimension, a feature map covering crop salt tolerance, water requirement patterns, and nutrient requirements is formed. By designing data interfaces and coupling rules between multiple dimensions, an analysis framework capable of simultaneously analyzing the relationship between environmental dynamics and crop physiological responses is constructed.

6. The method according to claim 1, characterized in that, The specific method for generating the synergistic solution containing bioremediation components, salt control measures, and nutrient configuration is as follows: Based on the degree of soil salinization in the root zone, the application ratio of acidic organic raw materials to phosphorus ion activators is calculated to target the acidification of the root zone micro-domain and activate the fixed phosphorus element. Simultaneously, based on the soil sodium adsorption ratio and crop salt tolerance characteristics, salt-resistant microbial agents are formulated, which are made by mixing multiple salt-resistant strains. In terms of nutrient configuration, the application amount of ammonium nitrogen is controlled, and the acidification effect generated by the nitrification process in the soil is used to replace sodium ions on soil colloids. The bioremediation components, salt-controlling substances, and nutrients are integrated in a specific ratio and application sequence to form a comprehensive and synergistic regulation scheme for root zone acidification and alkali reduction, salt replacement and leaching, and relief of salt-alkali stress.

7. The method according to claim 1, characterized in that, The specific method for establishing the dynamic mapping rule between scheme parameters and root zone soil properties is as follows: Based on the various components and dosage parameters in the aforementioned synergistic scheme, a quantitative response relationship is established with the salt concentration, pH value, sodium adsorption ratio, and soil texture characteristics of the root zone soil. By analyzing historical regulation data and soil response data, a quantitative relationship between the input ratio of acidic organic raw materials and changes in soil pH value is determined. Correspondence rules are established between the dosage of phosphorus ion activator and the increase in available phosphorus content and calcium and magnesium ion activity in the soil. An empirical function is formed for the inoculation amount of salt-alkali resistant microbial agents and the reduction in sodium adsorption ratio. Furthermore, a dynamic relationship is constructed between the application amount of ammonium nitrogen and the degree of root zone acidification and sodium ion replacement efficiency, forming a mapping rule system for adjusting the application ratio of each component in the scheme.

8. The method according to claim 1, characterized in that, The specific method for correcting the preset root region environment control benchmark is as follows: Based on real-time monitoring of soil salinity and moisture in the root zone and crop growth stage data, the dynamic deviation between preset thresholds and actual values ​​is compared. When soil salinity remains above the threshold, the upper limit of salt tolerance is dynamically lowered based on the crop's current growth stage salt tolerance and water requirement characteristics. At the same time, combined with meteorological evaporation data and root development depth, the water requirement for soil salinity leaching in the root zone and irrigation quota are recalculated to form a dynamic control benchmark value adapted to the current environment and crop growth.

9. The method according to claim 1, characterized in that, The synergistic verification mechanism for root zone repair and salt control includes: By using a pre-set material compatibility rule library, the chemical compatibility of acidic organic raw materials and salt-resistant microbial agents in the synergistic scheme is automatically compared and an early warning is issued. A synergy evaluation mechanism is constructed, and based on the matching degree analysis of water and fertilizer transport models and root distribution characteristics, the consistency of irrigation salt control measures and nutrient allocation in the spatiotemporal action of the root zone is evaluated. Key indicator thresholds are preset to verify the expected pH adjustment range, sodium ion replacement rate and salt leaching efficiency of the scheme in multiple dimensions.

10. The method according to claim 1, characterized in that, The specific method for building the root region repair and salt control collaborative management platform architecture is as follows: The cloud platform establishes a unified data middleware platform to perform spatiotemporal alignment and fusion processing of multimodal perception data, scheme execution data, and root zone state data; the edge side deploys model inference services to sink the trained decision-making model to the near end of the field for low-latency real-time analysis and decision generation; the terminal layer integrates intelligent fertilizer application equipment and irrigation control system, and uses standardized communication protocols to interact with the edge side and the cloud for command and data feedback. By constructing a data-driven three-level collaborative mechanism of cloud-edge-terminal, a collaborative management platform architecture for root zone repair and salt control that integrates perception, decision-making, execution, and feedback is formed.

11. The method according to claim 1, characterized in that, The specific method of the dynamic iteration mechanism for building the model is as follows: Data on changes in root zone soil salinity, microbial community activity, crop yield, and stress mitigation after the implementation of the scheme, combined with historical remediation case data, form an iterative data source. The actual implementation effect data is compared with the decision expectation data initially output by the model. Based on the verification results of the synergy between root zone remediation and salt control, the calculation weights of the bioremediation component ratio and the intensity of salt control measures in the model are adjusted. Simultaneously, the parameter optimization effect under the new scenario and the root zone adaptation law are added to the root zone remediation knowledge base to update the model's decision logic for different saline-alkali soil types and crop growth stages, forming a closed-loop iteration of data feedback, parameter optimization, knowledge update, and model upgrade.

12. The method according to claim 1, characterized in that, The specific method for establishing a continuous iterative link between the model, solution, and platform is as follows: A direct mapping channel is established between model decision output and scheme execution. The water, fertilizer and salt control decision scheme generated by the initial model of the application scenario is automatically converted into executable instructions through the cloud platform. The platform collects real-time data on changes in root zone soil parameters and crop growth response to construct an execution effect evaluation dataset. Based on the dataset, an automatic optimization process for model parameters is initiated, and the optimized model is regenerated to form a continuous iterative link of model decision-making, scheme execution and platform linkage.

Images