Citrus orchard soil fertility query method and system based on artificial intelligence

Abstract

The present application relates to the technical field of soil fertility query and analysis, and discloses a soil fertility query method and system for a citrus orchard based on artificial intelligence. The method comprises: collecting surface and deep soil samples in the citrus orchard for physical and chemical testing, and using spatiotemporal dual interpolation to fill in missing values to obtain a data set; extracting physical and chemical and biological activity characteristics, and using a fusion strategy to form a feature vector; inputting a cascade forest model for five-level classification; generating a distribution map for the classification of sampling points; performing intelligent diagnosis, calculating the amount of fertilizer to be applied, and generating a differentiated fertilization recommendation map. The present application realizes accurate soil fertility assessment, spatial distribution visualization, and differentiated fertilization recommendations based on multi-source data, thereby improving fertilizer utilization, reducing environmental risks, and increasing citrus yield and quality.

Description

Technical Field

[0001] The present application relates to the technical field of soil fertility query and analysis, and in particular to a citrus orchard soil fertility query method and system based on artificial intelligence. Background Art

[0002] At present, soil fertility management in citrus orchards mainly relies on traditional soil testing and empirical fertilization methods, including conventional physical and chemical index testing, empirical fertilizer amount determination, and unified fertilization operations. Traditional soil testing methods usually use grid sampling and laboratory analysis to obtain soil physical and chemical parameters at limited sampling points; fertilizer amount determination is mostly based on regional averages or empirical formulas, lacking refined differential management; fertilization operations are mainly based on unified fertilization throughout the orchard, making it difficult to precisely regulate the spatial heterogeneity of soil fertility. In addition, existing soil fertility assessment methods are mostly based on physical and chemical indicators, such as soil organic matter, total nitrogen, available phosphorus, available potassium, and pH value, and lack comprehensive consideration of soil biological activity, resulting in incomplete fertility assessment.

[0003] However, traditional methods have numerous shortcomings: First, the limited number of sampling points makes it difficult to accurately characterize the spatial variation of soil fertility. Second, soil testing cycles are long and costly, hindering real-time monitoring and dynamic management. Third, fixed fertilization plans lack specificity and cannot meet the differentiated needs of different regions, soil types, and growing periods. Finally, relying solely on physical and chemical indicators to assess fertility ignores the important role of soil microbial activity in nutrient transformation and supply, resulting in a lack of scientific basis for fertilization decisions and low fertilizer utilization, which increases production costs and creates an environmental burden. Precise soil fertility management is particularly important in the cultivation of high-value, high-quality, and permanent crops such as citrus. Summary of the Invention

[0004] This application provides an artificial intelligence-based soil fertility query method and system for citrus orchards, which is used to achieve accurate soil fertility assessment, spatial distribution visualization and differentiated fertilization recommendations based on multi-source data, thereby improving fertilizer utilization, reducing environmental risks, and increasing citrus yield and quality.

[0005] In a first aspect, the present application provides a method for querying soil fertility in a citrus orchard based on artificial intelligence, the method comprising: collecting surface and deep soil samples in a citrus orchard, performing physical and chemical index testing and on-site rapid testing on the soil samples, filling missing values using a spatiotemporal dual interpolation algorithm, and obtaining a standardized multimodal soil dataset; extracting physical and chemical characteristics and biological activity characteristics from the standardized multimodal soil dataset, and performing multimodal feature fusion using a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism to obtain a fused feature vector;

[0006] The fused feature vector is input into a cascade forest model containing a 4-layer cascade structure, and a five-level classification is performed based on the comprehensive soil fertility index to obtain a soil fertility classification model; the soil fertility levels of the citrus orchard sampling points are classified according to the soil fertility classification model to generate a soil fertility distribution map; based on the soil fertility distribution map, an intelligent soil fertility diagnosis is performed, the amount of fertilizer is calculated according to the nutrient balance principle, and a differentiated fertilization recommendation map is generated.

[0007] In a second aspect, the present application provides an artificial intelligence-based citrus orchard soil fertility query system, the artificial intelligence-based citrus orchard soil fertility query system comprising:

[0008] A detection module is used to collect surface and deep soil samples in citrus orchards, perform physical and chemical index testing and on-site rapid testing on the soil samples, and use a spatiotemporal dual interpolation algorithm to fill missing values to obtain a standardized multimodal soil dataset;

[0009] an extraction module for extracting physical and chemical characteristics and biological activity characteristics from the standardized multimodal soil dataset, and fusing multimodal features using a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism to obtain a fused feature vector;

[0010] An input module is used to input the fused feature vector into a cascade forest model with a 4-layer cascade structure, perform five-level classification based on the soil fertility comprehensive index, and obtain a soil fertility classification model;

[0011] A classification module, configured to classify the soil fertility levels of the citrus orchard sampling points according to the soil fertility classification model and generate a soil fertility distribution map;

[0012] The diagnosis module is used to perform intelligent soil fertility diagnosis based on the soil fertility distribution map, calculate the amount of fertilizer to be applied according to the nutrient balance principle, and generate a differentiated fertilization recommendation map.

[0013] In a third aspect, an artificial intelligence-based citrus orchard soil fertility query device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the artificial intelligence-based citrus orchard soil fertility query device executes the above-mentioned artificial intelligence-based citrus orchard soil fertility query method.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned artificial intelligence-based citrus orchard soil fertility query method.

[0015] In the technical solution provided by the present application, surface and deep soil samples are collected in citrus orchards for physical and chemical index detection and on-site rapid detection, and a spatiotemporal dual interpolation algorithm is used to fill in missing values to obtain a standardized multimodal soil data set. The physical and chemical characteristics and biological activity characteristics of the soil data are extracted, and a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism are used for multimodal feature fusion. The fused feature vector is input into a cascade forest model containing a 4-layer cascade structure for soil fertility classification, a soil fertility distribution map is generated and intelligent diagnosis is performed, and finally the amount of fertilizer is calculated to generate a differentiated fertilization recommendation map. This effectively solves the technical problems of limited sampling points, long detection cycle, fixed fertilization scheme, and single fertility assessment in traditional citrus orchard soil fertility management, and realizes accurate soil fertility assessment and differentiated fertilization recommendation based on multi-source data. Specifically, the beneficial effects of this solution are reflected in: First, the hierarchical hexagonal sampling strategy and multimodal data collection method are adopted to expand the sampling coverage, improve the representativeness and timeliness of soil data, and solve the problem of insufficient spatial representation caused by limited sampling points in traditional methods; secondly, through the multimodal feature fusion strategy , integrating soil physical and chemical characteristics with biological activity characteristics to comprehensively characterize soil fertility status, overcoming the one-sidedness of relying solely on physical and chemical indicators to assess fertility; third, applying the cascade forest algorithm, an artificial intelligence model, to replace traditional soil classification methods. The multi-level cascade structure of the cascade forest algorithm can extract deep features layer by layer, while the combined design of the two-type forests enhances the model's generalization ability and robustness to outliers. These algorithmic characteristics directly improve the accuracy and adaptability of soil fertility classification, especially in the application scenario of citrus orchards with limited sample size. Fourth, based on spatial interpolation technology, a continuous soil fertility distribution map is generated, which intuitively displays the spatial variation characteristics of fertility and provides a decision-making basis for precise management; fifth, the two artificial intelligence technologies of rule reasoning and case reasoning are comprehensively applied for intelligent soil fertility diagnosis, which not only utilizes expert knowledge but also takes into account historical experience, resulting in more comprehensive and accurate diagnostic results; finally, through multi-factor correction and soil-fertilizer-crop system coupling model to optimize fertilization plans, a precise match of fertilizer type, amount, time and method is achieved, which greatly improves fertilizer utilization efficiency and reduces environmental pollution risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 This is a schematic diagram of an embodiment of a method for querying soil fertility in a citrus orchard based on artificial intelligence in an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of an embodiment of an artificial intelligence-based citrus orchard soil fertility query system in an embodiment of the present application;

[0019] Figure 3 It is a schematic block diagram of the structure of a citrus orchard soil fertility query device based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a method and system for querying soil fertility in a citrus orchard based on artificial intelligence. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or apparatus.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the artificial intelligence-based citrus orchard soil fertility query method includes:

[0022] Step S101: Collect surface and deep soil samples in a citrus orchard, perform physical and chemical index testing and on-site rapid testing on the soil samples, and use a spatiotemporal dual interpolation algorithm to fill in missing values to obtain a standardized multimodal soil dataset;

[0023] Step S102: extracting physical and chemical characteristics and biological activity features from the standardized multimodal soil dataset, and fusing the multimodal features using a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism to obtain a fused feature vector;

[0024] Step S103: input the fused feature vector into a cascade forest model with a four-layer cascade structure, perform five-level classification based on the soil fertility comprehensive index, and obtain a soil fertility classification model;

[0025] Step S104: classifying the soil fertility levels of the citrus orchard sampling points according to the soil fertility classification model to generate a soil fertility distribution map;

[0026] Step S105: Perform intelligent soil fertility diagnosis based on the soil fertility distribution map, calculate the amount of fertilizer to be applied according to the nutrient balance principle, and generate a differentiated fertilization recommendation map.

[0027] It is understandable that the execution subject of this application can be the artificial intelligence-based citrus orchard soil fertility query system, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0028] Specifically, a layered hexagonal sampling strategy was used to divide an irregular sampling network in the citrus orchard. First, the sampling density was determined according to the area of the citrus orchard. For a 10-hectare citrus orchard, the hexagonal sampling network was designed as a hexagonal unit with a side length of 20 meters, forming a honeycomb structure covering the entire park, ensuring that the sampling points were evenly distributed and avoiding the directional deviation of regular grid sampling. At each sampling node, soil samples of 0-20 cm in the surface layer and 20-40 cm in the deep layer were collected, and the coordinate information of the sampling points was recorded using a high-precision GNSS receiver with a coordinate accuracy of better than ±0.5 meters. The collected samples were divided into three equal parts, and the first part was sent to the laboratory for testing of 16 conventional physical and chemical indicators such as pH value, organic matter content, total nitrogen, available phosphorus, available potassium, trace element content and cation exchange capacity to obtain a data set containing complete physical and chemical parameters. The second portion was rapidly tested on-site using a four-channel near-infrared spectrometer (NIR) analyzer. The spectrometer operates in the 400-2500nm wavelength range with a 2nm scanning interval. Data from 200 bands was collected for each sample and directly converted into physicochemical parameters using a pre-established PLSR spectral model, enabling rapid on-site assessment. The third portion was used to analyze soil particle composition and structural properties, including physical properties such as soil texture, bulk density, porosity, and aggregate stability. A multi-parameter sensor array with self-calibration capabilities, including soil moisture, temperature, conductivity, and pH sensors, was deployed at the sampling sites. Data were collected at 4-hour intervals for 30 days, and transmitted via a LoRaWAN low-power wide-area network. Quality assessment of the collected multi-source data was performed, using a modified Mahalanobis distance algorithm to identify outliers, with a threshold set at a 99.5% confidence interval. Identified outliers and missing data were filled using a dual interpolation algorithm combining tensor decomposition with spatiotemporal autoregression to ensure data integrity. Finally, the cleaned multi-source data were normalized using the adaptive weight Z-score standardization method. The weights were determined according to the contribution of each parameter to soil fertility assessment. The standardized indicators were reconstructed according to four dimensions: nutrient status, acid-base balance, physical structure, and biological activity, forming a structured standardized multimodal soil dataset.

[0029] Physicochemical and biological activity features were extracted from a standardized multimodal soil dataset. Physicochemical feature extraction was based on the calculation of the soil fertility index. The weighted summation of the organic matter index, total nitrogen index, available phosphorus index, available potassium index, and pH suitability index was performed by assigning different weights to each index. Each index was calculated by the ratio of the corresponding parameter to the optimal range for citrus growth. Biological activity features were extracted based on soil dehydrogenase activity and microbial carbon / nitrogen ratio parameters. A nine-dimensional feature vector representing the dynamic characteristics of the soil microbial community was constructed through nonlinear transformation. Feature fusion employed a hierarchical strategy. First, dimensionality reduction was performed on intra-modal features. Kernel principal component analysis (KPCA) was used to extract key features. Gaussian radial basis functions (RBFs) were used as the kernel function, and principal components with a cumulative variance contribution of 95% were selected. Weight coefficients were calculated for the reduced features. The information gain method was used to assess the contribution of each feature to classification. The weight vector was then normalized. An inter-modal feature fusion matrix was then constructed. Adaptive tensor decomposition was used to map features from different modalities into a shared latent space. The optimal mapping parameters were determined through iterative optimization to minimize reconstruction error and maximize inter-modal mutual information. Finally, the multimodal interactive feature representation is concatenated with the original dimensionality reduction features, and the nonlinear relationship between features is captured through a high-order feature cross network to form a twelve-dimensional fusion feature vector including nutrient status, acid-base balance, nutrient supply capacity and microbial activity.

[0030] A cascade forest model with a four-layer cascade structure was constructed. Each cascade layer consisted of two random forests and two completely randomized tree forests, with each forest containing 100 decision trees. While random forests select the optimal feature from a subset of features when splitting a node, completely randomized tree forests select features and split points completely randomly, improving model diversity. Training data was partitioned using five-fold cross-validation. The original fused feature vector was input to the first cascade layer. Each forest independently processed the data and generated a class probability distribution, forming the first-layer feature representation. The first-layer feature representation was concatenated with the original features as input to the second layer, and so on, until the fourth layer outputted the final classification result. Each node split was calculated based on the Gini impurity measure, selecting the split point that maximized the purity of the child node. After training, the model categorized soil fertility into five levels based on the comprehensive soil fertility index: very low (<0.2), low (0.2-0.4), medium (0.4-0.6), high (0.6-0.8), and very high (>0.8), forming a soil fertility classification model.

[0031] The characteristic vectors of all sampling points in the citrus orchard were input into the soil fertility classification model to obtain the fertility grade classification results of each sampling point. A spatial distribution layer was established based on the geographic coordinate information of the sampling points, and spatial autocorrelation analysis was performed on the sampling points. The Moran's I index was calculated to evaluate spatial clustering, and a semivariogram model was constructed to characterize the spatial structural characteristics of soil fertility. An improved Kriging interpolation algorithm was used to estimate the unsampled area. The weight coefficient was calculated based on the spatial relationship between each point and the soil fertility attribute value to perform the optimal linear unbiased estimation. Kriging interpolation combines the results of the spherical model, exponential model, and Gaussian model. The weight of each model is determined by cross-validation error to generate a continuous fertility distribution field. The interpolation results are rasterized, and the grid unit is set to 5 meters × 5 meters to ensure that the expression of details while maintaining computational efficiency. Each grid unit is assigned a corresponding soil fertility grade attribute. Finally, the data were rendered using a geographic information system, using a five-color grading scheme ranging from red (very low), orange (low), yellow (medium), light green (high) to dark green (very high) to intuitively display the spatial variation characteristics of soil fertility and generate a standardized soil fertility distribution map with a legend, scale, and coordinate grid.

[0032] Intelligent diagnosis based on soil fertility distribution maps begins with the establishment of a diagnostic rule base, encoding expert knowledge into structured rules encompassing nutrient deficiency symptoms, excess symptoms, and balance relationships. These rules are expressed in an "if-then" format, such as "If pH < 5.5 and available potassium < 80 mg / kg, then the diagnosis is acidic potassium deficiency." A case library is also constructed to collect typical cases and diagnostic results. Rule-based reasoning is then applied to regional data within the soil fertility distribution map. A forward chaining mechanism is employed to derive conclusions from known facts through rule chains, identifying regional fertility issues. The inference results are matched with typical cases in the case library, and similarities between cases are calculated. The most similar cases are then extracted as auxiliary references. Based on the diagnostic results and target yield, fertilization requirements are calculated based on the principle of nutrient balance. Specific fertilizer rates for each region are calculated, taking into account factors such as nutrient requirements per unit yield, soil nutrient supply, and nutrient utilization efficiency. Fertilizer rates are adjusted for soil type, fertilizer form, and climatic conditions. A coupled soil-fertilizer-crop system model is used to simulate nutrient dynamics and determine the optimal time and method for fertilization. Finally, a differentiated fertilization recommendation map is generated, which includes a zoning fertilization plan and intuitively displays the type, amount, time and method of fertilizer application in each area.

[0033] For example, in a real-world citrus orchard application, 47 sampling points were identified using a hexagonal sampling network. Soil samples were collected and analyzed, resulting in a multimodal dataset containing 16 physical and chemical parameters and 9 biological activity indicators. Feature extraction and fusion yielded a 12-dimensional fused feature vector. These features were then fed into a cascade forest model for training, achieving 99.60% classification accuracy on the test set. The resulting soil fertility map revealed areas of medium to low fertility in the southeastern part of the orchard and high fertility in the northwest. The intelligent diagnostic system identified acidity and potassium deficiency in the southeastern region and developed a targeted, differentiated fertilization plan. For the southeastern region, potassium sulfate compound fertilizer was recommended, applied in three stages: a base application 10 days before flowering, a topdressing application in holes during the fruit expansion phase, and a recovery application in furrows after harvest. For the northwest region, nitrogen fertilizer use was controlled, phosphorus fertilizer was appropriately increased, and a single, hole-applied, slow-release fertilizer was used to prevent excessive growth and quality loss caused by nutrient overload. After implementing this plan, the soil fertility balance in the park was improved, and the fruit quality and yield were significantly improved compared with traditional fertilization methods.

[0034] In an embodiment of the present application, surface and deep soil samples are collected in a citrus orchard for physical and chemical index detection and on-site rapid detection, and a spatiotemporal dual interpolation algorithm is used to fill in missing values to obtain a standardized multimodal soil data set. The physical and chemical characteristics and biological activity characteristics of the soil data are extracted, and a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism are used for multimodal feature fusion. The fused feature vector is input into a cascade forest model containing a 4-layer cascade structure for soil fertility classification, a soil fertility distribution map is generated, and intelligent diagnosis is performed. Finally, the amount of fertilizer is calculated to generate a differentiated fertilization recommendation map, which effectively solves the technical problems of limited sampling points, long detection cycle, fixed fertilization scheme, and single fertility assessment in traditional citrus orchard soil fertility management, and realizes accurate soil fertility assessment and differentiated fertilization recommendation based on multi-source data. Specifically, the beneficial effects of this solution are reflected in: First, the hierarchical hexagonal sampling strategy and multimodal data collection method are adopted to expand the sampling coverage, improve the representativeness and timeliness of soil data, and solve the problem of insufficient spatial representation caused by limited sampling points in traditional methods; secondly, through the multimodal feature fusion strategy, The integration of soil physical and chemical characteristics with biological activity characteristics comprehensively characterizes soil fertility, overcoming the one-sidedness of relying solely on physical and chemical indicators to assess fertility. Third, the Cascade Forest Algorithm (CFAL), an artificial intelligence model, replaces traditional soil classification methods. The CFAL's multi-level cascade structure extracts deep features layer by layer, while the dual-type forest combination design enhances the model's generalization and robustness to outliers. These algorithmic features directly improve the accuracy and adaptability of soil fertility classification, showing significant advantages in citrus orchards with limited sample sizes. Fourth, a continuous soil fertility distribution map is generated based on spatial interpolation technology, visually displaying the spatial variation of fertility and providing a decision-making basis for precise management. Fifth, the combined application of rule-based reasoning and case-based reasoning, two artificial intelligence technologies, is used for intelligent soil fertility diagnosis. This utilizes both expert knowledge and historical experience, resulting in more comprehensive and accurate diagnostic results. Finally, through multi-factor correction and a soil-fertilizer-crop system coupling model to optimize fertilization plans, a precise match of fertilizer type, amount, timing, and method is achieved, significantly improving fertilizer utilization efficiency and reducing environmental pollution risks.

[0035] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0036] An irregular sampling network was divided in the citrus orchard according to the stratified hexagonal sampling strategy. Soil samples from the surface layer (0-20 cm) and the depth (20-40 cm) were collected at each sampling node. The precise geographic coordinates of the sampling points were recorded to obtain geo-tagged soil samples.

[0037] The geographically marked soil samples were separated into three equal parts. The first part was sent to the laboratory for testing of conventional physical and chemical indicators such as pH value, organic matter content, total nitrogen, available phosphorus, available potassium, trace element content and cation exchange capacity to obtain a full physical and chemical indicator data set;

[0038] The second aliquot of the geotagged soil sample was rapidly scanned using a four-channel near-infrared spectrometer to collect spectral reflectance data in the wavelength range of 400-2500 nm. The data were converted into rapid soil physical and chemical parameters using a pre-built spectral model to obtain a spectral conversion parameter set.

[0039] The third aliquot of the geotagged soil sample was analyzed for soil particle composition and structural properties to construct a soil physical properties database for subsequent fertilization program design and correction coefficient determination, thus obtaining a soil physical properties dataset;

[0040] A multi-parameter sensor array with self-calibration function is placed at the sampling node. The sensor data collection interval is set to 4 hours and the collection period is 30 days. The data is transmitted to the edge computing unit through a low-power wide area network to obtain a time-series continuous monitoring data stream.

[0041] Data quality assessment was performed on the physical and chemical index dataset, spectral conversion parameter set, soil physical property dataset, and time-series continuous monitoring data stream. Multidimensional outliers were identified using a modified Mahalanobis distance algorithm, and missing and outliers were filled using a dual interpolation algorithm combining tensor decomposition with spatiotemporal autoregression to obtain a cleaned multi-source dataset.

[0042] The cleaned multi-source dataset was normalized using the adaptive weighted Z-score standardization method. The weight coefficient was determined by the objective function of minimizing the variance within each parameter group and maximizing the variance between groups. The standardized indicators were reconstructed according to the four dimensions of nutrient status, acid-base balance, physical structure and biological activity to obtain a structured standardized multimodal soil dataset.

[0043] Specifically, the sampling density is calculated based on the area of the citrus orchard. For a 10-hectare citrus orchard, a regular hexagonal grid with a side length of 20 meters is set to form a continuously covered honeycomb structure. Hexagonal sampling has the characteristic of equal spacing in six directions compared to traditional square grid sampling, avoiding the defect that the square grid has a high sampling density in the orthogonal direction and a low sampling density in the diagonal direction. During the sampling process, stratification is carried out according to soil type and terrain conditions. The same sampling density is used for homogeneous areas, while the sampling density is increased in areas with large soil type variation or complex terrain. After the sampling nodes are determined, a special soil sampler is used to collect soil samples from the surface layer of 0-20 cm and the deep layer of 20-40 cm. About 500 grams of soil are collected at each sampling point. At the same time, high-precision GNSS equipment is used to record the latitude and longitude coordinates of the sampling points with an accuracy better than ±0.5 meters to form geo-tagged soil samples. After homogenization, the geo-tagged soil samples are separated into three equal parts, each weighing about 150 grams. The first aliquot was sent to the laboratory for physical and chemical index testing, including pH value determination (using the potentiometric method, 1:2.5 soil-water ratio), organic matter content determination (using potassium dichromate oxidation method), total nitrogen (using the Kjeldahl method), available phosphorus (using the Olsen method, sodium bicarbonate extraction-molybdenum antimony countercolorimetry), fast-acting potassium (using ammonium acetate extraction-flame photometry), trace element content (using DTPA extraction-atomic absorption spectrophotometry) and cation exchange capacity (using ammonium acetate exchange method), forming a full physical and chemical index data set, which contains 16 physical and chemical parameter values for each sampling point.

[0044] The second aliquot was rapidly scanned using a four-channel near-infrared spectrometer with a continuous wavelength range of 400-2500 nm and a scanning interval of 2 nm, yielding 1050 reflectance data points per sample. The spectral data were converted to physicochemical parameters using a pre-built partial least squares regression (PLSR) spectral model. The PLSR model analyzes the correlation between spectral data and physicochemical parameters to establish prediction equations, enabling rapid parameter prediction without the need for chemical analysis. The resulting spectral conversion parameter set contains the same physicochemical parameters as the first aliquot, but is acquired more quickly.

[0045] The third aliquot was used to analyze soil particle composition and structural properties. Soil particle distribution was measured using a laser particle size analyzer, soil bulk density was determined using the knife ring method, soil porosity was determined using the knife ring saturation method, and aggregate stability was determined using the wet sieving method. This data set, encompassing physical parameters such as soil texture, bulk density, porosity, and aggregate stability, was constructed to inform subsequent fertilization program design. A multi-parameter sensor array with self-calibration capabilities was deployed at each sampling node. This array included a soil moisture sensor (using frequency domain reflectometry, with an accuracy of ±3%), a temperature sensor (with a measurement range of -40°C to +85°C and an accuracy of ±0.5°C), a conductivity sensor (with a measurement range of 0-20 mS / cm and an accuracy of ±2%), and a pH sensor (with a measurement range of 3-10 and an accuracy of ±0.1%). Sensor data was collected at a 4-hour interval to capture daily variations in soil parameters. The data collection period lasted 30 days, covering the critical growth stages of citrus. The data is transmitted to the edge computing unit via the LoRaWAN low-power wide-area network, forming a time-series continuous monitoring data stream that includes multi-parameter monitoring data for each node every four hours within 30 days.

[0046] Data quality assessment was performed on the full physical and chemical index dataset, spectral conversion parameter set, soil physical property dataset, and time-series continuous monitoring data stream. Multidimensional outliers were first identified using a modified Mahalanobis distance algorithm. This algorithm detects outliers by calculating the standardized distance between a data point and the center of a multidimensional data center. A modified version of the Mahalanobis distance algorithm takes into account the covariance structure of the data, calculating the distance between a sample point and the sample mean and dividing it by the covariance matrix. This effectively identifies data points that exhibit anomalous behavior across multiple variables. Identified outliers and missing values were imputed using a dual interpolation algorithm combining tensor decomposition with spatiotemporal autoregression. This algorithm first organizes the multidimensional data into a tensor structure, decomposes the tensor into a core tensor and a factor matrix using Tucker decomposition, and then reconstructs missing values using the decomposition results. Furthermore, a spatiotemporal autoregressive model was used to predict missing values using the time series characteristics and spatial correlation of the data points. Finally, the results of these two methods were combined to generate imputed values. After data quality assessment and processing, a cleaned multi-source dataset was obtained.

[0047] The cleaned multi-source dataset was normalized using the adaptive weighted Z-score standardization method. Z-score standardization converts each indicator into a standard distribution with a mean of 0 and a standard deviation of 1. The calculation method is to subtract the mean from the original value and divide it by the standard deviation. The adaptive weight is determined by the objective function of minimizing the variance within each parameter group and maximizing the variance between groups. This method comprehensively considers the degree of variation and classification ability of the indicator, giving higher weights to indicators with strong discrimination ability. The standardized indicators are reconstructed according to four dimensions: nutrient status (including organic matter, total nitrogen, available phosphorus, available potassium, etc.), acid-base balance (including pH, cation exchange capacity, etc.), physical structure (including texture, bulk density, porosity, etc.), and biological activity (including dehydrogenase activity, microbial carbon-nitrogen ratio, etc.), forming a structured, standardized multimodal soil dataset with clear physical meaning.

[0048] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0049] The soil fertility index was calculated based on the physicochemical parameters in the standardized multimodal soil dataset. The organic matter index, total nitrogen index, available phosphorus index, available potassium index, and pH suitability index were weighted and summed according to different weights. Each index was calculated by the ratio of the corresponding parameter to the suitable range for citrus growth, and the physicochemical characteristic matrix was obtained.

[0050] Nonlinear data transformation was performed on soil dehydrogenase activity and microbial carbon-nitrogen ratio in a standardized multimodal soil dataset to construct a nine-dimensional feature vector representing the dynamic characteristics of soil microbial communities and obtain a biological activity feature matrix.

[0051] The physicochemical feature matrix and the biological activity feature matrix were subjected to intra-modal feature dimensionality reduction, and the kernel principal component analysis method was used to extract the main features of each modality. The number of principal components was selected by maximizing the variance contribution rate to obtain the intra-modal feature set after dimensionality reduction.

[0052] The feature weight coefficient is calculated for the feature set within the modal after dimensionality reduction. The information gain method is used to measure the contribution of each feature to the soil fertility classification. The information gain rate of each feature is divided by the sum of the information gain rates of all features to obtain the feature weight vector.

[0053] An inter-modal feature fusion matrix is constructed based on the feature weight vector. Adaptive tensor decomposition technology is used to map different modal features to a shared latent space. The optimal mapping parameters are determined by joint optimization of minimizing reconstruction error and maximizing inter-modal mutual information to obtain a multi-modal interactive feature representation.

[0054] The multimodal interactive feature representation is cascaded and spliced with the original dimensionality reduction features. The nonlinear interaction relationship between features is captured through a high-order feature cross network. A residual connection structure is designed to avoid gradient problems during training. A twelve-dimensional fusion feature vector is constructed that contains comprehensive information on four aspects: soil nutrient status, acid-base balance, nutrient supply capacity, and microbial activity.

[0055] Specifically, the soil fertility index is calculated from physical and chemical parameters. The soil fertility index is a comprehensive indicator reflecting soil fertility status and is obtained by weighted summation of the organic matter index, total nitrogen index, available phosphorus index, available potassium index, and pH suitability index. During the calculation process, each individual index is obtained by comparing the measured parameter value with the suitable range for citrus growth. Specifically, the organic matter index is the ratio of the measured organic matter content to the organic matter content suitable for citrus growth (25-30g / kg); the total nitrogen index is the ratio of the measured total nitrogen content to the suitable content (1.5-2.0g / kg); the available phosphorus index is the ratio of the measured available phosphorus content to the suitable content (15-20mg / kg); the available potassium index is the ratio of the measured available potassium content to the suitable content (150-200mg / kg); and the pH suitability index is calculated based on the optimal pH range for citrus growth (5.5-6.5) and the degree of agreement with the measured pH. The final soil fertility index was obtained by weighted summation of each index, with weights assigned as 0.3 for organic matter, 0.25 for total nitrogen, 0.2 for available phosphorus, 0.15 for available potassium, and 0.1 for pH suitability. Weights were determined based on the importance of each nutrient to citrus growth. After calculating the soil fertility index for each sampling site, it was aggregated into a physicochemical characteristic matrix with the number of rows equal to the number of sampling sites and the number of columns equal to the number of physicochemical indicators. Nonlinear data transformation was performed on soil dehydrogenase activity and microbial carbon-nitrogen ratio from a standardized multimodal soil dataset to construct a feature vector representing the dynamic characteristics of the soil microbial community. Soil dehydrogenase activity is an important indicator of overall soil microbial activity, while the microbial carbon-nitrogen ratio reflects the structural characteristics and nutrient conversion capacity of the soil microbial community. The nonlinear data transformation used a modified hyperbolic tangent function to map the original data to the interval [-1, 1]. The transformed data retained the relative magnitude relationships of the original data while eliminating dimensional differences and the influence of extreme values. The converted dehydrogenase activity and microbial carbon-nitrogen ratio data were combined with environmental factors such as soil texture, humidity, and temperature to generate interactive features through polynomial expansion, ultimately forming a nine-dimensional feature vector to comprehensively characterize the dynamic characteristics of soil microbial communities and constitute a biological activity feature matrix.

[0056] The physicochemical and bioactivity feature matrices were subjected to intra-modal feature dimensionality reduction, and kernel principal component analysis (KPCA) was used to extract the main features of each modality. KPCA is a nonlinear extension of traditional principal component analysis. It uses a kernel function to map data into a high-dimensional feature space, within which linear principal component analysis is performed. Specifically, a kernel matrix is first constructed, using a Gaussian radial basis function as the kernel function to calculate the similarity between sample points. Eigenvalue decomposition is then performed on the kernel matrix to obtain eigenvalues and eigenvectors. The eigenvectors are sorted by their corresponding eigenvalues, and the first few eigenvectors with the largest contribution are selected as principal components. The number of principal components is selected based on the cumulative variance contribution rate, with a threshold of 95%. Thus, the first few principal components with a cumulative variance contribution rate of 95% are selected. Through KPCA, the physicochemical and bioactivity feature matrices are each reduced to a more compact representation, forming a reduced set of intra-modal features.

[0057] The feature weight coefficients were calculated for the set of intra-modal features after dimensionality reduction, and the information gain method was used to measure the contribution of each feature to the soil fertility classification. The information gain method is a commonly used feature selection metric in decision tree algorithms that can assess the ability of features to distinguish between categories. During the calculation process, the entropy of the dataset is first calculated to represent the uncertainty of the dataset. Then, the conditional entropy after partitioning using a certain feature is calculated to represent the remaining uncertainty after partitioning using that feature. The difference between the two is the information gain, which represents the reduction in uncertainty brought about by partitioning using that feature. Finally, the information gain is obtained by dividing the information gain by the entropy of the feature itself, thus avoiding the disadvantage of information gain being biased towards features with a large number of values. After calculating the information gain for each feature, the information gain of each feature is normalized by dividing it by the sum of the information gain rates of all features to obtain a feature weight vector. The weight value reflects the importance of the feature.

[0058] An inter-modal feature fusion matrix is constructed based on the feature weight vector, and adaptive tensor decomposition technology is used to map different modal features to a shared latent space. A tensor is a high-dimensional extension of the matrix concept and can naturally represent the structure of multimodal data. Adaptive tensor decomposition first constructs the physicochemical and bioactive features into a third-order tensor, with the three dimensions corresponding to sampling points, feature type, and modality. This third-order tensor is then decomposed into the product of a core tensor and three factor matrices through Tucker decomposition. A modal weight adjustment mechanism is introduced during the decomposition process to dynamically adjust the decomposition parameters based on the feature weight vector, giving higher weights to important features. The optimization goal is to minimize the reconstruction error and maximize the inter-modal mutual information. The alternating least squares method is used to iteratively solve the problem, obtaining a parameter setting that effectively fuses multimodal information. Finally, based on the decomposition results, the different modal features are mapped to a shared latent space to obtain a multimodal interaction feature representation. The multimodal interaction feature representation is cascaded with the original dimensionality reduction features, and a high-order feature cross network is used to capture the nonlinear interactions between features. The high-order feature cross-network is a deep learning structure specifically designed to capture complex interactions between features. The network consists of multiple layers of feature cross-networks, each of which implements second-order or higher-order interactions between features. Residual connections are designed into the network structure to directly add the output of the previous layer to the output of the current layer, effectively avoiding the vanishing gradient problem in deep network training. After processing through the high-order feature cross-network, the output is a twelve-dimensional fused feature vector representing the comprehensive status of soil fertility. This vector contains comprehensive information on four aspects: soil nutrient status, acid-base balance, nutrient supply capacity, and microbial activity. Each aspect corresponds to three feature components, forming a structured feature expression.

[0059] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0060] A cascade forest model framework with four cascade structures was constructed. Two random forest classifiers and two completely random tree forest classifiers were set in each cascade structure. The number of decision trees in the random forest classifier was 100, and the number of decision trees in the completely random tree forest classifier was 100, to obtain the initial cascade forest structure.

[0061] Perform a five-fold cross-validation on the fused feature vector, divide the dataset into a training set and a validation set, divide the training samples and the test samples in an 8:2 ratio, and use stratified sampling to maintain the distribution consistency of each type of sample to obtain the model training dataset;

[0062] The model training data set is input into the first layer of the initial cascade forest structure. Each forest processes the input features independently and generates a class probability distribution by voting to obtain the first layer feature representation.

[0063] The first-layer feature representation is concatenated with the original fused feature vector as the input of the second-layer cascade. Through random feature selection and decision tree split threshold optimization, the Gini impurity of each node is calculated, and the optimal split point is selected to construct a decision tree to obtain the second-layer feature representation.

[0064] The second-layer feature representation is concatenated with the original fused feature vector and sequentially input into the third and fourth cascade layers. Hierarchical soil fertility features are extracted through the multi-layer cascade structure. The last layer outputs the probability distribution of five categories. The final classification result is determined by majority voting, resulting in the trained cascade forest model.

[0065] Based on the trained cascade forest model and soil fertility comprehensive index, the soil fertility level is divided into five levels: very low, low, medium, high, and very high. The soil fertility level division rules are constructed, and the cascade forest model and the soil fertility level division rules are combined to form a soil fertility classification model.

[0066] Specifically, a model framework containing a four-layer cascade structure was designed, with two random forest classifiers and two completely random tree forest classifiers set in each cascade structure. The random forest classifier is an ensemble learning method composed of multiple decision trees. Its characteristic is that when constructing a decision tree, the optimal features are selected from the feature subset for node splitting, while the completely random tree forest selects features and splitting points completely randomly when splitting the node, without considering the importance of the features. In this method, the number of decision trees in each forest is set to 100, ensuring the complexity and generalization ability of the model. The combination of these two different types of forests enhances the diversity and robustness of the model, and the complementarity between the two helps to capture the complex patterns in soil fertility data. The four-layer cascade structure forms a deep learning framework, and each cascade layer sequentially processes and extracts higher-level feature representations, which is similar to the hierarchical feature extraction mechanism of deep neural networks, but does not require backpropagation training. A five-fold cross-validation partitioning of the fused feature vector is performed to assess the stability of model performance. Specifically, the dataset is randomly divided into five parts, four of which are used for training and one for validation. Different subsets are used alternately as validation sets, and training and validation are performed five times. Before each training session, the four training data sets are further divided into training and test samples in an 8:2 ratio, with 80% used for model training and 20% for model testing. Stratified sampling is used in the partitioning process to ensure that the distribution of samples of different fertility levels in each subset is consistent with the original dataset, avoiding model bias caused by uneven sample distribution. For example, if the proportion of fertility levels in the original dataset is very low (10%), low (20%), medium (40%), high (20%), and very high (10%), each subset after partitioning will maintain the same proportion, thus forming a representative model training dataset.

[0067] When the model training dataset is input into the first cascade of the initial cascade forest structure, each forest processes the input features independently. The specific processing process is to input the fused feature vector into each forest, and the 100 decision trees in the forest will classify and predict the samples separately to generate a category probability distribution. The random forest classifier adopts a feature random selection strategy when constructing the decision tree. Each node selects the optimal feature from the feature subset for splitting; while the completely random tree forest selects features and split points completely randomly to increase the diversity of the model. The output of each tree is the probability that the sample belongs to each category. By averaging the outputs of all trees, the category probability distribution at the forest level is obtained. The category probability distributions of the four forests are merged to form a first-level feature representation with high-dimensional representation capabilities. This feature representation contains new information extracted from the original features after analyzing them from different angles.

[0068] The first-layer feature representation is concatenated with the original fused feature vector and used as the input for the second-layer cascade. The concatenation operation concatenates two feature vectors along their dimensions to form a longer feature vector. The forest in the second-layer cascade constructs a decision tree based on this concatenated feature vector and trains it through random feature selection and decision tree splitting threshold optimization. Random feature selection involves randomly selecting a subset of features from the square root of the total number of features at each node in the decision tree. Decision tree splitting threshold optimization involves calculating the Gini impurity of each potential split point and selecting the split point that maximizes the purity of the child node. Gini impurity measures the degree of class confusion within a node; smaller values indicate greater class purity within the node. It is calculated as the difference between the sum of the squared probabilities of each class and 1. Node splits are performed by minimizing the Gini impurity, gradually building a decision tree. Ultimately, the four forests collectively generate the second-layer feature representation.

[0069] The second-layer feature representation is concatenated with the original fused feature vector and fed into the third and fourth cascade layers, respectively. The cascade process is similar to the second layer, but the dimension of the feature vectors processed increases with each layer. Each cascade layer further explores patterns in the data based on the features extracted in the previous layer, extracting hierarchical soil fertility features through a multi-layer cascade structure. The final layer outputs a probability distribution for five categories (very low, low, medium, high, and very high), and the final classification is determined by majority voting. Majority voting involves counting the classification results of all decision trees and selecting the most frequently occurring category as the final prediction. The entire training process does not require backpropagation; each cascade layer directly generates feature representations through a forward pass, significantly reducing computational complexity while maintaining high model accuracy.

[0070] Based on the trained cascade forest model and the comprehensive soil fertility index, soil fertility levels were classified into five categories. The comprehensive soil fertility index is a previously calculated comprehensive indicator reflecting soil fertility, with a value range of 0-1. The specific classification rules are: an index <0.2 indicates very low fertility, 0.2-0.4 indicates low fertility, 0.4-0.6 indicates medium fertility, 0.6-0.8 indicates high fertility, and >0.8 indicates very high fertility. Combining the cascade forest model with the soil fertility classification rules creates a soil fertility classification model. This combination allows the model to take into account the advantages of machine learning while incorporating domain expert knowledge, improving the scientific nature and interpretability of the classification.

[0071] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0072] The soil data feature vectors of all citrus orchard sampling points were input into the soil fertility classification model, and the soil fertility of each sampling point was classified into five levels to generate a soil fertility grade dataset of the sampling points.

[0073] Based on the soil fertility grade dataset and geographic coordinate information of the sampling points, a spatial distribution layer of the sampling points was established. Each sampling point was classified and marked according to five fertility grades: very low, low, medium, high, and very high, to obtain the fertility grade distribution of discrete sampling points.

[0074] The spatial autocorrelation analysis of the fertility grade distribution of discrete sampling points was conducted, the spatial autocorrelation index of the fertility grade of the sampling points was calculated, and the semivariogram model was constructed to obtain the spatial structure parameters of soil fertility.

[0075] Based on the spatial structure parameters of soil fertility, an improved Kriging interpolation algorithm was used to spatially predict the soil fertility level in unsampled areas. The interpolation results of the spherical model, exponential model, and Gaussian model were adaptively weighted and fused to obtain a continuous fertility distribution field.

[0076] The continuous fertility distribution field is rasterized, the grid unit size is set to 5 meters × 5 meters, and the corresponding soil fertility grade attribute is assigned to each grid unit to construct the rasterized soil fertility distribution data;

[0077] The rasterized soil fertility distribution data was rendered using the geographic information system, and a five-color grading display scheme was designed. Auxiliary layers were added according to the terrain characteristics of the citrus orchard and the management block division to generate a soil fertility distribution map with geographic reference information.

[0078] Specifically, inputting soil data feature vectors from all citrus orchard sampling points into the soil fertility classification model is the initial step in generating a soil fertility distribution map. Specifically, the 12-dimensional fused feature vectors for each sampling point obtained in the previous step are sequentially input into the trained cascade forest model. Through a four-layer cascade structure, a probability distribution is ultimately generated for each sampling point, indicating its belonging to one of the five fertility classes. The model outputs a probability distribution that reflects the likelihood of a sample belonging to each fertility class. For example, a probability distribution of [0.05, 0.15, 0.65, 0.10, 0.05] for a sampling point indicates that it has the highest probability of belonging to "medium" fertility. Based on the principle of maximum probability, a fertility class label is assigned to each sampling point, selecting the class with the highest probability as the fertility class for that point. This generates a dataset containing fertility class information for all sampling points. A spatial distribution layer for the sampling points is created based on the soil fertility class dataset and the geographic coordinates of the sampling points. This step involves associating the fertility class information of the sampling points with their geographic coordinates, displaying the distribution of the sampling points in geographic space. In the implementation, sampling points are first divided into five categories based on fertility level and marked with different colors: very low fertility (red), low fertility (orange), medium fertility (yellow), high fertility (light green), and very high fertility (dark green). Each sampling point is then marked and assigned a corresponding color in geographic space based on its latitude and longitude coordinates, forming a point layer containing both spatial location and attribute information. At this point, fertility distribution exists only at discrete sampling point locations, forming a fertility level distribution for each discrete sampling point.

[0079] Spatial autocorrelation analysis of the fertility grade distribution at discrete sampling points was performed to assess the clustering and variability of the spatial distribution of soil fertility. Spatial autocorrelation refers to the spatial interdependence of observations within the study area and is typically expressed using the Moran's I index. The Moran's I index ranges from -1 to 1, with positive values indicating positive correlation (clustering of similar values), negative values indicating negative correlation (clustering of dissimilar values), and 0 indicating random distribution. The calculation requires the construction of a spatial weight matrix to define the adjacency between sampling points, typically using the inverse of the inter-point distance as weight. The results of the spatial autocorrelation analysis are used to guide subsequent spatial interpolation. For example, a high degree of positive correlation indicates strong spatial continuity in the soil fertility distribution, making it suitable for spatial interpolation. Furthermore, a semivariogram model is constructed to characterize the spatial structure of soil fertility. The semivariogram describes how the spatial correlation between sampling points changes with distance, typically increasing initially and then stabilizing with increasing distance. According to the distribution characteristics of the actual data, an appropriate theoretical model (spherical model, exponential model or Gaussian model) is selected to fit the experimental semivariogram to obtain parameters reflecting the spatial structure of soil fertility, including the nugget value (reflecting the degree of random variation), the sill value (reflecting the degree of total variation) and the range (reflecting the influence range of spatial correlation).

[0080] Based on the spatial structure parameters of soil fertility, an improved kriging interpolation algorithm was used to spatially predict soil fertility levels in unsampled areas. Kriging interpolation is an optimal linear unbiased estimation method based on spatial statistics. Its core concept is to calculate the optimal weight coefficients for the estimated points based on the spatial distribution and correlation of known sample points. Traditional kriging interpolation uses only a single model, while the improved kriging interpolation combines the advantages of multiple models. In the specific implementation, a spherical model, an exponential model, and a Gaussian model were used for kriging interpolation. Each model has its own characteristics: the spherical model is suitable for describing phenomena with rapid changes over medium distances, the exponential model is suitable for describing phenomena with relatively gradual changes, and the Gaussian model is suitable for describing phenomena with very smooth changes. The interpolation results of the three models were adaptively weighted and fused. The weight coefficients were determined based on the cross-validation error of each model, with models with lower cross-validation errors being given higher weights. This multi-model fusion strategy improved interpolation accuracy and produced a continuous fertility distribution field covering the entire citrus orchard.

[0081] The purpose of rasterizing the continuous fertility distribution field is to achieve the conversion from continuous interpolation results to discrete management units. The rasterization process divides the study area into regular grids, and each grid cell represents an independent plot. The grid cell size is set to 5 meters × 5 meters. This scale can not only ensure the fine expression of the spatial variation characteristics of soil fertility, but also adapt to the scale of agricultural production management. During the rasterization process, the study area is divided into uniform grids according to the specifications of 5 meters × 5 meters. The fertility level interpolation calculation is performed on the center point of each grid cell, and the calculation result is used as the attribute value of the grid cell. For special cases, such as when the grid cell coincides with the sampling point, the measured fertility level of the sampling point is directly used as the attribute value of the unit. Through rasterization, a regular grid data structure containing spatial position and fertility level attributes is formed, which is convenient for subsequent analysis and management.

[0082] The rasterized soil fertility distribution data is rendered using a geographic information system to generate an intuitive soil fertility distribution map. The rendering process includes designing a color scheme, adding legends and coordinate information, etc. The color scheme uses a five-color grading display: red represents extremely low fertility areas, orange represents low fertility areas, yellow represents medium fertility areas, light green represents high fertility areas, and dark green represents extremely high fertility areas. The color transition intuitively reflects the changes in fertility levels. At the same time, auxiliary layers are added according to the terrain characteristics of the citrus orchard and the division of management blocks, such as contour layers to display terrain undulations, road and water system layers to facilitate positioning and navigation, and management block boundary layers to facilitate precise fertilization management. The final generated soil fertility distribution map contains geographic reference information, such as coordinate grids, scales, and compasses, which are convenient for field comparison.

[0083] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0084] A soil fertility diagnosis rule base was constructed based on the soil fertility distribution map. Expert knowledge was encoded into a structured diagnostic rule set that included soil nutrient deficiency symptoms, nutrient excess symptoms, and nutrient balance relationships. A soil fertility case library containing typical diagnostic cases was also established to obtain a soil fertility diagnosis knowledge base.

[0085] Conduct rule-based reasoning analysis on regional data in the soil fertility distribution map, and use the diagnostic rules in the soil fertility diagnosis knowledge base to perform forward chain reasoning to identify symptoms of soil nutrient deficiency or imbalance, thereby obtaining regional soil fertility problem diagnosis results.

[0086] The regional soil fertility problem diagnosis results were matched with typical cases in the soil fertility diagnosis knowledge base. The similarity between the new and old cases was evaluated using an improved Euclidean distance calculation method. The diagnostic information of the most similar case was extracted as an auxiliary reference to obtain a comprehensive diagnosis result.

[0087] Based on the comprehensive diagnosis results and the target citrus yield, the fertilizer requirements for each region are calculated based on the principle of nutrient balance. The fertilizer calculation formula takes into account the nutrient requirements for the target yield, the effective nutrient content of the soil, the nutrient content of the fertilizer, and the utilization rate to generate basic fertilizer application data.

[0088] Basic fertilizer application data is modified based on soil type, fertilizer form, and climatic conditions. The soil-fertilizer-crop system coupling model is used to adjust the dynamic process of nutrient supply, calculate the appropriate fertilization time and method for each region, and obtain a precise fertilization plan.

[0089] The precise fertilization plan is spatially mapped according to the management zones, and a zoned fertilization layer including the type, amount, time and method of fertilizer is constructed. A differentiated fertilization recommendation map is generated through geographic information system rendering processing.

[0090] Specifically, constructing a soil fertility diagnosis rule base based on soil fertility distribution maps first requires converting expert knowledge about citrus orchard soil fertility into a computer-processable form. The soil fertility diagnosis rule base utilizes an "if-then" rule set, encompassing three categories of rules: soil nutrient deficiency symptoms, nutrient excess symptoms, and nutrient balance relationships. Nutrient deficiency symptom rules describe the manifestations of various nutrient deficiencies, such as "If soil pH < 5.5 and available phosphorus < 10 mg / kg, then diagnose acidic phosphorus deficiency." Nutrient excess symptom rules describe the manifestations of nutrient excess, such as "If available potassium > 250 mg / kg and the calcium-magnesium ratio < 3:1, then diagnose potassium excess and calcium-magnesium imbalance." Nutrient balance relationship rules describe the interactions between nutrients, such as "If the nitrogen-phosphorus ratio > 15:1, then diagnose nitrogen-phosphorus imbalance." Each rule is assigned a credibility coefficient, indicating its reliability, ranging from 0 to 1. A typical case library was also established to collect and organize existing soil fertility diagnosis cases. Each case includes three pieces of information: soil physical and chemical indicators, symptom descriptions, and diagnostic results. Cases are categorized and stored by fertility level and problem type for quick retrieval. The rule library and case library together constitute the soil fertility diagnosis knowledge base, providing knowledge support for subsequent diagnosis.

[0091] A forward chain reasoning method is used to analyze the regional data in the soil fertility distribution map using rule-based reasoning. This method involves applying inference rules from known facts to draw conclusions. The specific steps include data preparation, rule matching, conflict resolution, and result output. In the data preparation phase, the soil physical and chemical data and fertility levels for each region in the soil fertility distribution map are input into the inference system as known facts. In the rule matching phase, the data for each region is matched against the rules in the diagnostic rule library, selecting all rules whose conditional parts (if parts) match the data for the current region. The conflict resolution phase handles the simultaneous application of multiple rules, applying a comprehensive scoring method based on rule priority and credibility to select the rule with the highest score for execution. In the result output phase, the conclusion part (then part) of the matching rule is used as the diagnosis result for the regional soil fertility problem. Forward chain reasoning triggers rule execution from the input data until no new rules can be triggered, forming an inference chain and outputting a regional soil fertility problem diagnosis result, including the problem type, symptom description, and severity.

[0092] The regional soil fertility problem diagnosis results were matched against typical cases in the soil fertility diagnosis knowledge base. An improved Euclidean distance calculation method was used to assess the similarity between the new and old cases. While traditional Euclidean distance calculation simply treats all features equally, the improved Euclidean distance method incorporates feature weights, assigning higher weights to important features. The calculation process includes four steps: feature selection, data normalization, weight determination, and distance calculation. Feature selection extracts key features from the diagnosis results, including soil physical and chemical indicators and problem symptom characteristics. Data normalization normalizes features of different dimensions so that they are distributed within the same range. Weight determination assigns weights to each feature based on the information gain ratio. Distance calculation uses the weighted Euclidean distance formula to calculate the distance between the current region and each case in the case base, with smaller distances indicating higher similarity. Cases are sorted by similarity from high to low, and the top N most similar cases are selected. Diagnostic information from these cases is extracted and combined with the results of rule-based reasoning. A comprehensive diagnostic result, including a problem description, cause analysis, and treatment recommendations, is obtained through evidence synthesis.

[0093] Based on the comprehensive diagnostic results and the target citrus yield, the fertilizer requirements for each region are calculated based on the principle of nutrient balance. The nutrient balance principle states that the nutrients required for crop growth should be met through a combination of soil supply and exogenous fertilization. The calculation process takes into account the target yield nutrient requirement, the available soil nutrient content, the fertilizer nutrient content, and the utilization rate. The target yield nutrient requirement refers to the amount of nitrogen, phosphorus, potassium, and other nutrients required to produce a unit of citrus output, determined by consulting the citrus nutrient requirement parameter table; the available soil nutrient content is the amount of nutrients available for crop absorption and utilization, as determined through soil testing; the fertilizer nutrient content refers to the amount of active ingredients in the applied fertilizer; and the utilization rate factor refers to the proportion of applied nutrients absorbed and utilized by the crop, which is influenced by soil properties, climatic conditions, and fertilization method. According to the nutrient balance principle, fertilizer application calculation requires first determining the target yield, then consulting the unit yield nutrient requirement parameters, calculating the total requirement, subtracting the soil supply, and then converting the final fertilizer application amount based on the fertilizer utilization rate. Separate calculations are made for the three main nutrients, nitrogen, phosphorus, and potassium, to generate baseline fertilization data, including regional nitrogen, phosphorus, and potassium fertilizer usage. These baseline fertilization data are then adjusted based on soil type, fertilizer form, and climatic conditions. A coupled soil-fertilizer-crop system model is used to adjust the dynamics of nutrient supply. Soil type adjustment accounts for the impact of different soil textures on nutrient fixation and release. For example, sandy soils require more frequent fertilization and less fertilization per application, while clay soils require the opposite. Fertilizer form adjustment considers the characteristics of different fertilizer types. For example, quick-acting fertilizers release nutrients quickly but for a short duration, while slow-release fertilizers release nutrients slowly but have a long-lasting effect. Climate adjustment considers the effects of temperature, rainfall, and other factors on nutrient transformation and leaching. For example, during the rainy season, nitrogen fertilizer application may need to be reduced or slow-release nitrogen fertilizers may be used instead. The coupled soil-fertilizer-crop system model is a mathematical model that describes the interactions among these three factors. It comprises three submodules: soil nutrient dynamics, fertilizer transformation dynamics, and crop uptake dynamics. The model inputs include base fertilizer application rates, soil physical and chemical properties, fertilizer characteristics, and climate data. By simulating the migration and transformation of nutrients in the soil and crop absorption, it outputs dynamic curves of nutrient availability at different stages. Based on these dynamic curves and the nutrient requirements of citrus fruits at each growth stage, the optimal time for fertilization is determined. Annual fertilization is divided into four phases: early growth, flower bud differentiation, fruit expansion, and post-harvest recovery. Fertilizer is allocated in different proportions to each phase based on nutrient requirements. Furthermore, the optimal fertilization method is determined based on the soil characteristics of each region, including broadcast fertilization, hole fertilization, furrow fertilization, rhizosphere injection, and foliar spraying, to form a precise fertilization plan.

[0094] Precision fertilization plans are spatially mapped according to management zones, creating a zoned fertilization layer that includes fertilizer type, amount, time, and method. The spatial mapping process begins by identifying management zones. Management zones are spatial units defined based on soil fertility distribution, terrain characteristics, and cultivation management needs. These zones are generally divided according to fertility level, terrain slope, and natural boundaries such as roads and water systems. Each management zone is assigned corresponding fertilization plan attributes, including fertilizer type, amount, time, and method. Fertilizer type specifies the type of fertilizer appropriate for each area, such as nitrogen, phosphate, potash, compound, and organic fertilizers; amount specifies the specific amount of each fertilizer, measured in kilograms per hectare; time specifies the specific dates or growth period for fertilization at each stage; and method specifies the most appropriate application technique. This attribute information is associated with the spatial zones to form a zoned fertilization layer. This is then rendered using a geographic information system (GIS) with an intuitive symbology and color scheme to generate a differentiated fertilization recommendation map. The recommended map is in the form of a base map overlay. The base map is a topographic map of the citrus orchard, with fertilizer zoning layers and fertilizer information labels superimposed to facilitate intuitive understanding and implementation by managers.

[0095] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0096] The basic fertilizer application data were corrected for soil type. A soil correction coefficient matrix was constructed based on the differences in nutrient fixation capacity of different soil textures. The nitrogen, phosphorus, and potassium nutrient availability conversion rates of sandy soils, loamy soils, and clay soils were incorporated into the calculation to obtain the soil type-corrected fertilizer application data.

[0097] The fertilizer application data after soil type correction were corrected for fertilizer form. A fertilizer form correction model was established based on the nutrient release characteristics of different fertilizer types. The dynamic changes in nutrient availability of quick-acting fertilizers, slow-release fertilizers, and organic fertilizers were considered to obtain fertilizer application data after fertilizer form correction.

[0098] The fertilizer application data after fertilizer form correction is corrected for climate conditions. A relationship model between meteorological data and nutrient availability is constructed. The risk of nutrient leaching loss is quantitatively assessed using rainfall, temperature, and humidity to obtain the fertilizer application data after climate condition correction.

[0099] The climatically corrected fertilizer application data were input into the soil-fertilizer-crop system coupling model. Based on the material balance principle and discrete element method, the migration and transformation process of nutrients in the soil was simulated. The dynamic curves of nutrient absorption by citrus at different stages were calculated to obtain the dynamic data of nutrient supply.

[0100] Based on the dynamic data of nutrient supply and the characteristics of nutrient demand during the key growth periods of citrus, the optimal fertilization time window was determined using a time series analysis method. The early growth period, flower bud differentiation period, fruit expansion period, and post-harvest recovery period were divided into different fertilization stages, and a phased fertilization schedule was obtained.

[0101] Based on the phased fertilization time plan and the physical and chemical properties of the soil in each area, a combination of fertilization methods including broadcasting, hole application, furrow application, rhizosphere injection and foliar spraying was designed. The optimal fertilization method code was assigned to each area. The revised fertilization amount data, phased fertilization time plan and fertilization method code were integrated to obtain a precise fertilization plan.

[0102] Specifically, correcting basic fertilizer application data for soil type first requires constructing a soil correction coefficient matrix based on the differences in nutrient fixation capacity among different soil textures. Soil texture refers to the ratio of sand, silt, and clay in the soil, which directly affects the soil's nutrient fixation capacity and availability. To construct the soil correction coefficient matrix, the third aliquot of the geotagged soil sample is first analyzed for soil particle composition. Soil texture type is determined based on the sand, silt, and clay content. Soil texture is then classified into three categories using the soil texture triangle diagram: sandy soil, loamy soil, and clay soil. The fixation characteristics and available conversion rates of the three major nutrients, nitrogen, phosphorus, and potassium, were then analyzed for each soil type. Sandy soils have low nutrient fixation capacity and high leaching losses, with available conversion rates of approximately 0.6, 0.4, and 0.7 for nitrogen, phosphorus, and potassium, respectively. Loamy soils have moderate nutrient fixation capacity, with available conversion rates of approximately 0.5, 0.3, and 0.6 for nitrogen, phosphorus, and potassium, respectively. Clayey soils have strong nutrient fixation capacity and low leaching losses, with available conversion rates of approximately 0.4, 0.2, and 0.5 for nitrogen, phosphorus, and potassium, respectively. These conversion rate data were organized into a 3×3 soil correction coefficient matrix, with rows representing soil type and columns representing nutrient types. Fertilizer rate correction was calculated by dividing the base fertilizer rate by the corresponding available conversion rate. For example, when applying nitrogen fertilizer to sandy soil, the base fertilizer rate should be divided by 0.6 to obtain the soil type-corrected fertilizer rate data.

[0103] Fertilizer form correction is a secondary correction performed on soil type-corrected fertilizer application data based on the nutrient release characteristics of different fertilizer types. Fertilizer form correction begins with the establishment of a fertilizer form correction model, which describes the dynamic nutrient release process of different fertilizer types. Fast-acting fertilizers such as urea, ammonium sulfate, and superphosphate release nutrients rapidly but for a short period of time, with a steeply declining release curve. Slow-release fertilizers such as sulfur-coated urea and polyammonium phosphate release nutrients more slowly but for a long period of time, with a gently declining release curve. Organic fertilizers such as farmyard manure and commercial organic fertilizers release nutrients most slowly but have the longest-lasting effects, with a release curve that remains approximately constant. Based on these characteristics, a nutrient release rate function is established to describe the nutrient release amount of fertilizers at different time points. During the correction process, for quick-acting fertilizers, due to their rapid release, which can easily lead to nutrient waste, it is necessary to increase the number of applications and reduce the single dosage. The correction factor is 0.8-0.9. For slow-release fertilizers, due to their steady release, they are suitable for single-application, and the correction factor is 1.0-1.1. For organic fertilizers, due to their long-term effects, the base dosage needs to be increased, and the correction factor is 1.2-1.5. Multiply the correction factor by the soil type-corrected fertilizer application rate to obtain the fertilizer form-corrected fertilizer application rate data.

[0104] To apply climate-corrected fertilizer application rate data after fertilizer form correction, a model linking meteorological data with nutrient availability must be constructed. Climate influences nutrient availability primarily through rainfall, temperature, and humidity. Rainfall influences nutrient leaching losses, particularly water-soluble nitrogen; temperature influences microbial activity and nutrient conversion rates; and humidity influences soil aeration and microbial activity. This model analyzes historical meteorological data and nutrient residue testing data to establish a correlation function between meteorological factors and nutrient loss rates. First, historical and forecasted meteorological data for the citrus orchard area were collected, including monthly average rainfall, temperature, and humidity. A nutrient leaching risk index (nutrient leaching risk index) was then calculated for each month. This index is a weighted combination of rainfall, temperature, and humidity. Finally, a climate correction factor was determined based on the leaching risk index, with higher correction factors for high-risk months and lower factors for low-risk months. The climate correction factor was multiplied by the fertilizer form-corrected application rate to obtain climate-corrected fertilizer application data, thus enabling adaptation to seasonal climate changes.

[0105] Fertilizer application data corrected for climate conditions are input into a coupled soil-fertilizer-crop system model. The model simulates nutrient migration and transformation in the soil based on the material balance principle and discrete element method. The coupled soil-fertilizer-crop system model is a multi-component, multi-process dynamic simulation system consisting of three submodules: soil, fertilizer, and crop. The soil submodule describes the movement of soil water and nutrients; the fertilizer submodule describes the dynamic release of nutrients from different fertilizers; and the crop submodule describes the absorption and utilization of nutrients by citrus. The material balance principle states that the input, output, and transformation of each component in the system must conform to the law of conservation of mass. The discrete element method discretizes continuous space and time into finite elements, and the global solution is obtained by solving the mass transfer equation for each element. The model inputs include corrected fertilizer application data, soil physical and chemical properties, fertilizer properties, and meteorological data. Numerical simulations are used to calculate the dynamic changes in soil nutrient concentrations and nutrient uptake rates by citrus at different times. The model outputs dynamic nutrient supply data, which describe the matching between soil nutrient supply capacity and citrus nutrient demand at each point in the growing season.

[0106] Based on nutrient supply dynamics data and the nutrient requirements during key citrus growth periods, a time series analysis method was used to determine the optimal fertilization window. This method first smoothed the nutrient supply and demand dynamics data to eliminate the effects of short-term fluctuations. The difference between the two curves was then calculated to represent the difference in nutrient supply and demand. Finally, peak-valley analysis was used to identify the time points with the largest difference. These points represent periods of minimal nutrient supply and represent the optimal fertilization window. Based on the characteristics of the citrus growth cycle, the annual fertilization period is divided into four key fertilization phases: the early growth phase (before spring shoot bud break), the flower bud differentiation phase (from flower bud differentiation to flowering), the fruit expansion phase (from young fruit expansion to fruit maturity), and the post-harvest recovery phase (from fruit harvest to dormancy). Nutrient requirements vary for each phase: nitrogen is the primary nutrient during the early growth phase, phosphorus is the primary nutrient during the flower bud differentiation phase, potassium is the primary nutrient during the fruit expansion phase, and a balanced nutrient balance is maintained during the post-harvest recovery phase. Based on the characteristics and duration of each phase, the total fertilizer application rate is rationally allocated to develop a phased fertilization schedule. Based on the phased fertilization schedule and the physical and chemical properties of each region's soil, a fertilization method combination was designed, including broadcasting, hole application, furrow application, rhizosphere injection, and foliar spraying. Different fertilization methods are suitable for different soil conditions and fertilizer types. Broadcasting is suitable for large areas with flat terrain, offering ease of use but lower utilization rates. Hole application is suitable for young trees and sparsely planted orchards, providing concentrated nutrient supply but requiring more work. Furrow application is suitable for mature trees and inter-row plantings, balancing efficiency and utilization rates. Rhizosphere injection is suitable for fast-acting nutrient supplementation, offering precision and efficiency but higher costs. Foliar spraying is suitable for trace elements and emergency nutrient supplementation, offering a quick response but poor sustainability. For each region, the suitability of various fertilization methods was comprehensively evaluated based on soil characteristics, tree age distribution, and management conditions. The optimal fertilization method combination was selected and assigned a corresponding method code. Finally, the revised fertilizer application rate data, phased fertilization schedule, and fertilization method code were integrated to form a precise fertilization plan, which includes comprehensive information such as region number, soil type, fertilizer type, phased fertilizer rate, application time, and fertilization method.

[0107] The above describes the citrus orchard soil fertility query method based on artificial intelligence in the embodiment of the present application. The following describes the citrus orchard soil fertility query system based on artificial intelligence in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the artificial intelligence-based citrus orchard soil fertility query system includes:

[0108] Detection module 201 is used to collect surface and deep soil samples in the citrus orchard, perform physical and chemical index testing and on-site rapid testing on the soil samples, and use a spatiotemporal dual interpolation algorithm to fill missing values to obtain a standardized multimodal soil dataset;

[0109] Extraction module 202, for extracting physical and chemical characteristics and biological activity characteristics from the standardized multimodal soil dataset, and fusing multimodal features using a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism to obtain a fused feature vector;

[0110] An input module 203 is configured to input the fused feature vector into a cascade forest model having a four-layer cascade structure, perform five-level classification based on the soil fertility comprehensive index, and obtain a soil fertility classification model;

[0111] A classification module 204 is configured to classify the soil fertility levels of the citrus orchard sampling points according to the soil fertility classification model and generate a soil fertility distribution map;

[0112] The diagnosis module 205 is used to perform intelligent soil fertility diagnosis based on the soil fertility distribution map, calculate the amount of fertilizer to be applied according to the nutrient balance principle, and generate a differentiated fertilization recommendation map.

[0113] Through the collaborative cooperation of the above components, surface and deep soil samples were collected in the citrus orchard for physical and chemical index detection and on-site rapid detection, and the spatiotemporal dual interpolation algorithm was used to fill in the missing values to obtain a standardized multimodal soil dataset. The physical and chemical characteristics and biological activity characteristics of the soil data were extracted, and a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism were used for multimodal feature fusion. The fused feature vector was input into a cascade forest model containing a 4-layer cascade structure for soil fertility classification, and a soil fertility distribution map was generated and intelligent diagnosis was performed. Finally, the fertilizer amount was calculated to generate a differentiated fertilization recommendation map. This effectively solved the technical problems of limited sampling points, long detection cycle, fixed fertilization scheme, and single fertility assessment in traditional citrus orchard soil fertility management, and realized accurate soil fertility assessment and differentiated fertilization recommendation based on multi-source data. Specifically, the beneficial effects of this scheme are reflected in: First, the hierarchical hexagonal sampling strategy and multimodal data collection method were adopted to expand the sampling coverage, improve the representativeness and timeliness of soil data, and solve the problem of insufficient spatial representation caused by limited sampling points in traditional methods; secondly, through multimodal feature fusion The strategy integrates soil physical and chemical characteristics with biological activity characteristics to comprehensively characterize soil fertility, overcoming the one-sidedness of relying solely on physical and chemical indicators to assess fertility. Third, the Cascade Forest Algorithm (CFAL) is applied as an artificial intelligence model to replace traditional soil classification methods. The CFAL's multi-level cascade structure can extract deep features layer by layer, while the dual-type forest combination design enhances the model's generalization ability and robustness to outliers. These algorithmic features directly improve the accuracy and adaptability of soil fertility classification, especially in citrus orchards with limited sample sizes. Fourth, a continuous soil fertility distribution map is generated based on spatial interpolation technology, which intuitively displays the spatial variation characteristics of fertility and provides a decision-making basis for precise management. Fifth, the two artificial intelligence technologies of rule reasoning and case reasoning are integrated for intelligent soil fertility diagnosis. This not only utilizes expert knowledge but also takes into account historical experience, resulting in more comprehensive and accurate diagnostic results. Finally, through multi-factor correction and soil-fertilizer-crop system coupling model to optimize fertilization plans, a precise match of fertilizer type, amount, timing, and method is achieved, significantly improving fertilizer utilization efficiency and reducing environmental pollution risks.

[0114] above Figure 2 The artificial intelligence-based citrus orchard soil fertility query system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The artificial intelligence-based citrus orchard soil fertility query device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0115] Figure 3This is a schematic diagram of the structure of an AI-based citrus orchard soil fertility query device provided by an embodiment of the present invention. This AI-based citrus orchard soil fertility query device 300 may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions and operations within the AI-based citrus orchard soil fertility query device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, allowing the AI-based citrus orchard soil fertility query device 300 to execute the series of instructions and operations stored in the storage medium 330 to implement the steps of the AI-based citrus orchard soil fertility query method described above.

[0116] The artificial intelligence-based citrus orchard soil fertility query device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the artificial intelligence-based citrus orchard soil fertility query equipment shown does not constitute a limitation of the artificial intelligence-based citrus orchard soil fertility query equipment provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0117] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the artificial intelligence-based citrus orchard soil fertility query method.

[0118] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0119] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling an artificial intelligence-based citrus orchard soil fertility query device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc., various media that can store program code.

[0120] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A citrus orchard soil fertility query method based on artificial intelligence, characterized in that: The method comprises: Surface and deep soil samples were collected from citrus orchards, and physical and chemical index tests and on-site rapid tests were performed on the soil samples. A spatiotemporal dual interpolation algorithm was used to fill missing values to obtain a standardized multimodal soil dataset. Extracting physical and chemical characteristics and biological activity features from the standardized multimodal soil dataset, and fusing multimodal features using a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism to obtain a fused feature vector; The fused feature vector is input into a cascade forest model with a 4-layer cascade structure, and a five-level classification is performed based on the soil fertility comprehensive index to obtain a soil fertility classification model, including: A cascade forest model framework with a 4-layer cascade structure is constructed. Two random forest classifiers and two completely random tree forest classifiers are set in each cascade structure. The number of decision trees of the random forest classifier is 100, and the number of decision trees of the completely random tree forest classifier is 100, to obtain an initial cascade forest structure; a five-fold cross-validation partition is performed on the fusion feature vector, the data set is divided into a training set and a validation set, the training samples and the test samples are divided according to a ratio of 8:2, and stratified sampling is used to maintain the distribution consistency of various samples to obtain a model training data set; the model training data set is input into the first layer of the initial cascade forest structure, each forest independently processes the input features, and generates a category probability distribution by voting to obtain a first-layer feature representation; the first-layer feature representation is compared with the original fusion feature vector The rows are spliced as the input of the second-level cascade, and the Gini impurity of each node is calculated through random feature selection and decision tree splitting threshold optimization. The optimal splitting point is selected to build a decision tree to obtain the second-level feature representation; the second-level feature representation is spliced with the original fusion feature vector and input into the third and fourth cascades in turn. The hierarchical soil fertility features are extracted through the multi-layer cascade structure, and the last layer outputs the probability distribution of five categories. The final classification result is determined by majority voting to obtain the trained cascade forest model; based on the trained cascade forest model and the soil fertility comprehensive index, the soil fertility level is divided into five levels: very low, low, medium, high, and very high. A soil fertility level division rule is constructed, and the cascade forest model and the soil fertility level division rule are combined to form a soil fertility classification model; Classifying the soil fertility levels of the citrus orchard sampling points according to the soil fertility classification model to generate a soil fertility distribution map; Based on the soil fertility distribution map, intelligent soil fertility diagnosis is performed, the amount of fertilizer to be applied is calculated according to the nutrient balance principle, and a differentiated fertilization recommendation map is generated.

2. The artificial intelligence-based soil fertility query method for a citrus orchard according to claim 1, characterized in that: The surface and deep soil samples were collected in the citrus orchard, and the soil samples were tested for physical and chemical indicators and rapid on-site testing. The missing values were filled using a spatiotemporal dual interpolation algorithm to obtain a standardized multimodal soil dataset, including: An irregular sampling network was divided in the citrus orchard according to the stratified hexagonal sampling strategy. Soil samples from the surface layer (0-20 cm) and the depth (20-40 cm) were collected at each sampling node. The precise geographic coordinates of the sampling points were recorded to obtain geo-tagged soil samples. The geographically marked soil sample was separated into three equal parts, and the first equal part was sent to the laboratory for testing of conventional physical and chemical indicators such as pH value, organic matter content, total nitrogen, available phosphorus, available potassium, trace element content, and cation exchange capacity to obtain a full physical and chemical indicator data set; A second portion of the geographically marked soil sample is rapidly scanned using a four-channel near-infrared spectrometer to collect spectral reflectance data within a wavelength range of 400-2500 nm, and the spectral reflectance data are converted into rapid soil physical and chemical parameters using a pre-built spectral model to obtain a spectral conversion parameter set; performing soil particle composition analysis and structural property determination on the third aliquot of the geographically marked soil sample to construct a soil physical property database for use in subsequent fertilization program design and correction coefficient determination, thereby obtaining a soil physical property data set; A multi-parameter sensor array with self-calibration function is arranged at the sampling node, and the sensor data collection time interval is set to 4 hours and the collection period is 30 days. The data is transmitted to the edge computing unit through a low-power wide area network to obtain a time-series continuous monitoring data stream; Performing data quality assessment on the physical and chemical full index dataset, the spectral conversion parameter set, the soil physical property dataset, and the time-series continuous monitoring data stream, identifying multidimensional outliers using a modified Mahalanobis distance algorithm, and filling missing values and outliers using a dual interpolation algorithm combining tensor decomposition with spatiotemporal autoregression to obtain a cleaned multi-source dataset; The cleaned multi-source dataset was normalized using an adaptive weighted Z-score normalization method. The weight coefficient was determined by the objective function of minimizing the variance within each parameter group and maximizing the variance between groups. The standardized indicators were reconstructed according to four dimensions: nutrient status, acid-base balance, physical structure, and biological activity, to obtain a structured standardized multimodal soil dataset.

3. The artificial intelligence-based soil fertility query method for a citrus orchard according to claim 1, characterized in that: The physical and chemical characteristics and biological activity characteristics of the standardized multimodal soil dataset are extracted, and a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism are used to fuse the multimodal features to obtain a fused feature vector, including: Calculating the soil fertility index based on the physical and chemical parameters in the standardized multimodal soil dataset, and weighting the organic matter index, total nitrogen index, available phosphorus index, available potassium index, and pH suitability index according to different weights to obtain a physical and chemical characteristic matrix; performing nonlinear data transformation on the soil dehydrogenase activity and microbial carbon-nitrogen ratio in the standardized multimodal soil dataset to construct a nine-dimensional feature vector representing the dynamic characteristics of the soil microbial community and obtain a biological activity feature matrix; The physicochemical feature matrix and the biological activity feature matrix are subjected to intra-modal feature dimensionality reduction, respectively, and the main features of each modality are extracted using the kernel principal component analysis method. The number of principal components is selected by maximizing the variance contribution rate to obtain a set of intra-modal features after dimensionality reduction; Calculating a feature weight coefficient for the modal feature set after dimensionality reduction, using the information gain method to measure the contribution of each feature to the soil fertility classification, and dividing the information gain rate of each feature by the sum of the information gain rates of all features to obtain a feature weight vector; An inter-modal feature fusion matrix is constructed based on the feature weight vectors, and the features of different modalities are mapped to a shared latent space using an adaptive tensor decomposition technique. The optimal mapping parameters are determined by a joint optimization of minimizing the reconstruction error and maximizing the inter-modal mutual information to obtain a multi-modal interactive feature representation. The multimodal interactive feature representation is cascaded and spliced with the original dimensionality reduction features. The nonlinear interaction relationship between features is captured through a high-order feature cross network. A residual connection structure is designed to avoid gradient problems during training, and a twelve-dimensional fusion feature vector containing comprehensive information on four aspects: soil nutrient status, acid-base balance, nutrient supply capacity, and microbial activity is constructed.

4. The artificial intelligence-based soil fertility query method for a citrus orchard according to claim 1, characterized in that: The step of classifying the soil fertility levels of the citrus orchard sampling points according to the soil fertility classification model to generate a soil fertility distribution map includes: Input the soil data feature vectors of all citrus orchard sampling points into the soil fertility classification model, perform five-level classification judgment on the soil fertility of each sampling point, and generate a sampling point soil fertility grade dataset; Based on the soil fertility grade dataset and geographic coordinate information of the sampling points, a spatial distribution layer of the sampling points is established, and each sampling point is classified and marked according to five fertility grades: very low, low, medium, high, and very high, to obtain the fertility grade distribution of discrete sampling points; Performing spatial autocorrelation analysis on the fertility grade distribution of the discrete sampling points, calculating the spatial autocorrelation index of the fertility grade of the sampling points, constructing a semivariogram model, and obtaining soil fertility spatial structure parameters; Based on the soil fertility spatial structure parameters, an improved Kriging interpolation algorithm is used to spatially predict the soil fertility level of the unsampled area, and the interpolation results of the spherical model, exponential model and Gaussian model are adaptively weighted fused to obtain a continuous fertility distribution field; The continuous fertility distribution field is rasterized, the grid unit size is set to 5 meters × 5 meters, and each grid unit is assigned a corresponding soil fertility grade attribute to construct rasterized soil fertility distribution data; The rasterized soil fertility distribution data is rendered using a geographic information system, a five-color hierarchical display scheme is designed, auxiliary layers are added according to the terrain characteristics of the citrus orchard and the management block division, and a soil fertility distribution map with geographic reference information is generated.

5. The artificial intelligence-based soil fertility query method for a citrus orchard according to claim 1, characterized in that: The intelligent soil fertility diagnosis is performed based on the soil fertility distribution map, the amount of fertilizer is calculated according to the nutrient balance principle, and a differentiated fertilization recommendation map is generated, including: A soil fertility diagnosis rule base is constructed based on the soil fertility distribution map, expert knowledge is encoded into a structured diagnosis rule set including soil nutrient deficiency symptoms, nutrient excess symptoms, and nutrient balance relationships, and a soil fertility case library including typical diagnosis cases is established to obtain a soil fertility diagnosis knowledge base; Performing rule reasoning analysis on the regional data in the soil fertility distribution map, performing forward chain reasoning using the diagnostic rules in the soil fertility diagnosis knowledge base, identifying symptoms of soil nutrient deficiency or imbalance, and obtaining regional soil fertility problem diagnosis results; The regional soil fertility problem diagnosis results are matched with typical cases in the soil fertility diagnosis knowledge base for similarity, an improved Euclidean distance calculation method is used to evaluate the similarity between the new and old cases, and the diagnostic information of the most similar case is extracted as an auxiliary reference to obtain a comprehensive diagnostic result; Based on the comprehensive diagnosis results and the target citrus yield, the fertilizer requirement for each region is calculated based on the nutrient balance principle. The fertilizer amount calculation formula takes into account the nutrient requirement for the target yield, the effective nutrient content of the soil, the nutrient content of the fertilizer, and the utilization rate to generate basic fertilizer amount data; The basic fertilizer application data is modified according to soil type, fertilizer form, and climatic conditions. The soil-fertilizer-crop system coupling model is used to adjust the dynamic process of nutrient supply, calculate the appropriate fertilization time and method for each region, and obtain a precise fertilization plan; The precise fertilization plan is spatially mapped according to the management zones, and a zoned fertilization layer including the type of fertilizer, amount of fertilizer, time of fertilizer application and method of fertilizer application is constructed, and a differentiated fertilization recommendation map is generated through geographic information system rendering processing.

6. The artificial intelligence-based soil fertility query method for a citrus orchard according to claim 5, characterized in that: The basic fertilization amount data is corrected according to soil type, fertilizer form, and climatic conditions and environmental factors, and the soil-fertilizer-crop system coupling model is applied to adjust the dynamic process of nutrient supply, calculate the appropriate fertilization time and fertilization method for each region, and obtain a precise fertilization plan, including: The basic fertilizer application rate data was corrected for soil type. A soil correction coefficient matrix was constructed based on the differences in nutrient fixation capacity of different soil textures. The nitrogen, phosphorus, and potassium nutrient availability conversion rates of sandy soils, loamy soils, and clay soils were added to the calculation to obtain the soil type-corrected fertilizer application rate data. Performing fertilizer form correction on the soil type-corrected fertilizer application data, establishing a fertilizer form correction model based on the nutrient release characteristics of different fertilizer types, taking into account the dynamic changes in nutrient availability of quick-acting fertilizers, slow-release fertilizers, and organic fertilizers, to obtain fertilizer form-corrected fertilizer application data; The fertilizer application amount data after the fertilizer form correction is corrected for climate conditions, a relationship model between meteorological data and nutrient availability is constructed, and the risk of nutrient leaching loss is quantitatively assessed through rainfall, temperature, and humidity to obtain the fertilizer application amount data after climate condition correction; The fertilization amount data corrected by the climate conditions are input into the soil-fertilizer-crop system coupling model. Based on the material balance principle and discrete element method, the migration and transformation process of nutrients in the soil is simulated, and the dynamic curve of nutrient absorption by citrus in different periods is calculated to obtain the dynamic data of nutrient supply; Based on the dynamic data of nutrient supply and the characteristics of nutrient demand during the key growth period of citrus, a time series analysis method is used to determine the optimal fertilization time window, and the early growth period, flower bud differentiation period, fruit expansion period and post-harvest recovery period are divided into different fertilization stages to obtain a staged fertilization time plan; Based on the phased fertilization time plan and the physical and chemical properties of the soil in each area, a combination of fertilization methods including broadcasting, hole application, furrow application, rhizosphere injection and foliar spraying was designed, and an optimal fertilization method code was assigned to each area. The revised fertilization amount data, the phased fertilization time plan and the fertilization method code were integrated to obtain a precise fertilization plan.

7. A citrus orchard soil fertility query system based on artificial intelligence, characterized in that: For implementing the artificial intelligence-based citrus orchard soil fertility query method according to any one of claims 1 to 6, the artificial intelligence-based citrus orchard soil fertility query system comprises: A detection module is used to collect surface and deep soil samples in citrus orchards, perform physical and chemical index testing and on-site rapid testing on the soil samples, and use a spatiotemporal dual interpolation algorithm to fill missing values to obtain a standardized multimodal soil dataset; an extraction module for extracting physical and chemical characteristics and biological activity characteristics from the standardized multimodal soil dataset, and fusing multimodal features using a hierarchical feature fusion strategy and an adaptive feature weight allocation mechanism to obtain a fused feature vector; An input module is used to input the fused feature vector into a cascade forest model containing a 4-layer cascade structure, perform five-level classification based on the comprehensive soil fertility index, and obtain a soil fertility classification model, including: constructing a cascade forest model framework containing a 4-layer cascade structure, setting 2 random forest classifiers and 2 completely random tree forest classifiers in each cascade structure, wherein the number of decision trees of the random forest classifier is 100, and the number of decision trees of the completely random tree forest classifier is 100, to obtain an initial cascade forest structure; performing a five-fold cross-validation division on the fused feature vector, dividing the data set into a training set and a validation set, dividing the training samples and the test samples in a ratio of 8:2, and using stratified sampling to maintain the consistency of the distribution of various samples to obtain a model training data set; inputting the model training data set into the first layer of the initial cascade forest structure, each forest independently processes the input features, and generates category probabilities by voting. distribution to obtain the first-layer feature representation; the first-layer feature representation is spliced with the original fused feature vector as the input of the second-layer cascade, and the Gini impurity of each node is calculated through random feature selection and decision tree splitting threshold optimization, and the optimal splitting point is selected to build a decision tree to obtain the second-layer feature representation; the second-layer feature representation is spliced with the original fused feature vector, and the results are input into the third and fourth cascades in turn, and hierarchical soil fertility features are extracted through a multi-layer cascade structure. The last layer outputs the probability distribution of five categories, and the final classification result is determined by majority voting to obtain a trained cascade forest model; based on the trained cascade forest model and the soil fertility comprehensive index, the soil fertility level is divided into five levels: extremely low, low, medium, high, and extremely high, and a soil fertility level division rule is constructed. The cascade forest model and the soil fertility level division rule are combined to form a soil fertility classification model; A classification module, configured to classify the soil fertility levels of the citrus orchard sampling points according to the soil fertility classification model and generate a soil fertility distribution map; The diagnosis module is used to perform intelligent soil fertility diagnosis based on the soil fertility distribution map, calculate the amount of fertilizer to be applied according to the nutrient balance principle, and generate a differentiated fertilization recommendation map.

8. A citrus orchard soil fertility query device based on artificial intelligence, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and the processor implements the artificial intelligence-based citrus orchard soil fertility query method described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor executes the artificial intelligence-based soil fertility query method for a citrus orchard as claimed in any one of claims 1 to 6.

Images