Intelligent evaluation method and system for soil quality driven by multispectral images

Abstract

This application relates to the field of soil analysis technology and discloses a multispectral image-driven intelligent soil quality assessment method and system. By introducing a working condition parameter recording and explicit modeling mechanism, it effectively optimizes the consistency problem faced by multispectral remote sensing data in long-term application and cross-regional promotion. Traditional soil quality remote sensing assessment methods often use single-temporal data for modeling, neglecting the systematic impact of changes in observation conditions on spectral data, resulting in a significant performance degradation of the model when applied in different times and regions. This method eliminates external interference factors through spectral correction and state decoupling techniques, enhances the stability of feature expression through multi-scale feature construction constrained by mechanisms, balances the model's universality and specificity through a hierarchical modeling architecture, ensures performance for cross-regional application and long-term operation through regional adaptation and drift detection techniques, and maintains the model's continuous effectiveness through a continuous learning mechanism.

Description

Technical Field

[0001] This application relates to the field of soil analysis technology, and in particular to a multispectral image-driven intelligent evaluation method and system for soil quality. Background Technology

[0002] In the field of agricultural remote sensing and intelligent soil monitoring, multispectral image-driven soil quality assessment technology has become a research hotspot. This technology acquires reflectance spectral data of the soil surface in multiple bands using multispectral sensors mounted on UAVs or satellites, and combines this with machine learning algorithms to establish a mapping model between spectral characteristics and soil physicochemical properties, enabling rapid prediction of quality indicators such as organic matter, nutrient content, and pH value. Existing methods typically employ a supervised learning framework, relying on collected soil samples for model training, and are applied to the spatial distribution assessment of soil quality in the same area.

[0003] Existing multispectral soil quality assessment methods face a serious regional transfer dilemma. Due to the inherent differences in soil types, parent materials, and climatic conditions across different regions, models trained in the source region are difficult to directly apply to the target region. Current transfer learning methods assume that the source and target domains share similar feature distributions, but the differences in the physicochemical mechanisms of soil spectral responses invalidate this assumption. In particular, for special soil types such as saline-alkali soils and red soils, simple feature alignment strategies cannot capture their unique spectral response patterns, leading to a significant decline in model performance and even negative transfer when applied across regions. Summary of the Invention

[0004] This application proposes a multispectral image-driven intelligent soil quality evaluation method and system to address the problem that models trained in the source region are difficult to directly apply to the target region due to the inherent differences in soil types, parent materials, and climate conditions in different regions.

[0005] To achieve the above objectives, this application adopts the following technical solution: a multispectral image-driven intelligent evaluation method for soil quality, the specific steps of which are as follows:

[0006] S1. Data Acquisition: Acquire multispectral remote sensing images of the target area and corresponding soil sample information, and simultaneously record imaging time, illumination conditions, sensor operating status and observation geometric parameters;

[0007] S2. Spectral Correction: Based on the imaging time, illumination conditions, and observation geometric parameters, radiometric correction and illumination normalization are performed on the multispectral remote sensing image to obtain surface reflectance data at a uniform scale.

[0008] S3. State Decoupling: Identify and eliminate non-bare soil areas and abnormal reflection areas, construct soil water content state factors and surface structure state factors, and decouple soil spectral information.

[0009] S4. Feature Construction: Extract multi-scale spectral features based on the decoupled soil spectral data and form stable feature expressions related to soil physicochemical properties;

[0010] S5. Hierarchical modeling: Based on the stable features, a prediction model is constructed that includes a general feature expression layer and a soil type-specific expression layer to achieve adaptive fusion of different expression results;

[0011] S6. Regional Adaptation: Use multispectral data of the target region to constrain and optimize the prediction model, reducing the performance degradation of the model in cross-regional applications.

[0012] S7. Drift Detection: Continuously detect changes in the distribution of input features and dynamically adjust the model parameter update strategy based on the detection results;

[0013] S8. Continuous updates: The model is incrementally updated based on newly acquired multispectral data to maintain the long-term stability of soil quality assessment results.

[0014] Furthermore, in step S1, the specific steps for data acquisition are as follows:

[0015] S1.1. Use a remote sensing device equipped with a multispectral sensor to acquire multispectral image data of the target area, select representative locations within the image coverage area according to preset spatial distribution rules to collect soil samples, and conduct laboratory analysis on the soil samples to obtain organic matter content, nutrient content and pH values.

[0016] S1.2 When acquiring multispectral image data, simultaneously record the imaging time, solar altitude angle, cloud cover, sensor exposure parameters, operating temperature, flight altitude, and imaging angle, and establish and store the correlation between the operating parameters and the corresponding image data.

[0017] Furthermore, in step S2, the specific steps for spectral correction are as follows:

[0018] S2.1 Read the sensor parameters and imaging time recorded in step S1.2, convert the digital signal value of the multispectral image into radiance value according to the radiometric calibration parameters, eliminate sensor nonlinear error, perform atmospheric correction in combination with atmospheric state parameters, eliminate atmospheric scattering and absorption interference, and obtain the true surface reflectance data.

[0019] S2.2 Extract the solar elevation angle, imaging angle and surface slope information recorded in step S1.2, establish an illumination geometry model, normalize the reflectivity data after radiation correction, and unify the reflectivity data of different imaging times and observation angles to standard observation conditions to obtain surface reflectivity data of a uniform scale.

[0020] Furthermore, in step S3, the specific steps for state decoupling are as follows:

[0021] S3.1 Read the surface reflectance data from step S2, calculate the reflectance ratio or difference between the red band and the near-infrared band, identify the vegetation coverage area based on the vegetation index threshold, identify water bodies, shadows and areas with strong reflectance abnormalities through reflectance statistical analysis, remove the identified non-bare soil areas and abnormal areas, and retain the bare soil spectral data.

[0022] S3.2. Select short-wave infrared reflectance to calculate water state factor, select texture parameters or multi-angle data to calculate structural state factor, establish a relationship model of the influence of state factor on reflectance, decompose reflectance into soil intrinsic reflectance component and state reflectance component, and subtract state component to obtain soil intrinsic spectral information.

[0023] Furthermore, in step S4, the specific steps for feature construction are as follows:

[0024] S4.1 Read the intrinsic spectral information of the soil from step S3, calculate the ratio characteristics, difference characteristics and normalized combination characteristics of reflectance in each band, differentiate the spectral curve to obtain the first derivative characteristics and second derivative characteristics, calculate the gray-level co-occurrence matrix parameters under different spatial windows to obtain texture characteristics, and statistically analyze the similarity of adjacent pixels to obtain spatial clustering characteristics.

[0025] S4.2 Based on the spectral response mechanism of soil organic matter, salinity and nutrients, the features are classified into darkening feature group, salinization feature group and absorption feature group respectively. The correlation coefficient between the features and physicochemical properties is calculated on the source region and target region samples. Features with stable and high correlation coefficients are selected and combined to form a stable feature expression.

[0026] Furthermore, in step S5, the specific steps of layered modeling are as follows:

[0027] S5.1 Read the stable feature expression from step S4, collect the spectral features and physicochemical properties of soil samples from the source area, construct a general feature expression layer for all samples to establish the mapping relationship between spectra and physicochemical properties, divide the samples into saline-alkali soil, red soil and black soil sample groups according to soil type, and construct soil type-specific expression layers to establish the mapping relationship for each type.

[0028] S5.2 Extract stable features of soil samples from the target area and input them into the stratified prediction model. Calculate the feature value distribution of the sample in each feature group, determine the similarity between the sample and each soil type and normalize it into weight coefficients. Input the sample features into each expression layer to obtain the output results. Based on the weight coefficients, perform weighted summation of the outputs of each layer to obtain the soil quality prediction results.

[0029] Furthermore, in step S6, the specific steps for region adaptation are as follows:

[0030] S6.1. Obtain multispectral data of the target area and extract stable feature expressions according to steps S2 to S4. Calculate the distance between the sample features of the target area and the feature distribution center of the source area. Mark the samples with a distance exceeding the threshold as high deviation samples. Input the samples into the hierarchical prediction model to obtain the prediction results and calculate the uncertainty value. Establish a scoring function based on the degree of deviation and uncertainty and select high-scoring samples as suitable samples.

[0031] S6.2. Adjust the parameters of the hierarchical prediction model using the adapted samples, keeping the general expression layer parameters unchanged, and only adjusting the specific expression layer parameters and fusion weight coefficients corresponding to the target region. Introduce a parameter change penalty term in the loss function to limit the deviation. Verify the prediction error on the adapted samples and source region samples. Stop adjusting when the accuracy requirements are met to obtain the adapted model.

[0032] Furthermore, in step S7, the specific steps for drift detection are as follows:

[0033] S7.1. Acquire multispectral data of the target area at preset time intervals and extract stable feature expressions according to steps S2 to S4. Establish a sliding time window to store sample feature statistics. Calculate the statistical distance between the feature distribution of the current window and the historical benchmark window as the offset. Extract the current working condition parameters and calculate the numerical difference with the historical working condition. When the offset or difference exceeds the threshold, it is determined that working condition drift has occurred and the intensity level is recorded.

[0034] S7.2 Read the drift intensity level recorded in step S7.1, determine the parameter update range according to the intensity, adjust only the fusion weight coefficient when there is a slight drift, adjust the parameters and weight coefficient of the specific expression layer when there is a moderate drift, and adjust the parameters and weight coefficient of the general expression layer and the specific expression layer when there is a severe drift. Analyze the direction of change of the working condition parameters to generate parameter adjustment constraints, and pass the update strategy and constraints to step S8.

[0035] Furthermore, in step S8, the specific steps are continuously updated as follows:

[0036] S8.1. Based on the drift time detected in step S7.1 or the preset update cycle, new multispectral data is acquired in the target area and soil samples are collected to obtain physicochemical property parameters. Stable feature expressions are obtained by processing according to steps S2 to S4. The distribution density and coverage of the new samples are calculated. Samples from different feature areas, different working conditions, and different soil types are selected to form an incremental training sample set. Representative samples are selected from the historical sample database to form a memory sample set.

[0037] S8.2 Read the update strategy and adjustment constraints passed in step S7.2, combine the incremental training sample set and the memory sample set to form a mixed training sample set, select the model level and parameters to be adjusted according to the update strategy, calculate the prediction error of incremental samples and memory samples in the objective function, introduce directional constraints and magnitude penalty terms, verify the prediction accuracy after optimization, and replace the model parameters to complete the update when the requirements are met.

[0038] A multispectral image-driven intelligent soil quality evaluation system includes a data acquisition module, a spectral correction module, a state decoupling module, a feature construction module, a hierarchical modeling module, a region adaptation module, a drift detection module, and a continuous update module.

[0039] The data acquisition module is used to acquire multispectral remote sensing images and soil sample information of the target area, and simultaneously record imaging time, illumination conditions, sensor working status and observation geometric parameters;

[0040] The spectral correction module is used to perform radiometric correction and illumination normalization on multispectral images based on operating parameters to obtain surface reflectance data at a uniform scale.

[0041] The state decoupling module is used to identify and eliminate non-bare soil areas, construct water-bearing state factors and surface structure state factors, and extract intrinsic soil spectral information;

[0042] The feature construction module is used to extract multi-scale spectral features, group features according to the response mechanism of soil physicochemical properties, and form a stable feature expression.

[0043] The hierarchical modeling module is used to construct a prediction model that includes a general expression layer and a type-specific expression layer, enabling adaptive fusion of results from different expression layers;

[0044] The region adaptation module is used to evaluate the sample quality of the target region and adjust the model parameters through constraint optimization to reduce the performance degradation of cross-regional applications;

[0045] The drift detection module is used to monitor changes in feature distribution and operating parameters, determine the drift intensity level, and dynamically formulate model update strategies.

[0046] The continuous update module is used for incremental learning based on new data, and combines the memory sample mechanism to update model parameters to maintain the long-term stability of evaluation results.

[0047] The beneficial effects of this invention are as follows:

[0048] This application provides a multispectral image-driven intelligent soil quality assessment method and system. By introducing a working condition parameter recording and explicit modeling mechanism, it effectively optimizes the consistency problem faced by multispectral remote sensing data in long-term application and cross-regional promotion. Traditional soil quality remote sensing assessment methods often use single-temporal data for modeling, neglecting the systematic impact of changes in observation conditions on spectral data, resulting in a significant performance degradation of the model when applied in different times and regions. This method eliminates external interference factors through spectral correction and state decoupling techniques, enhances the stability of feature expression through multi-scale feature construction constrained by mechanisms, balances the model's universality and specificity through a hierarchical modeling architecture, ensures performance for cross-regional application and long-term operation through regional adaptation and drift detection techniques, and maintains the model's continuous effectiveness through a continuous learning mechanism. The overall technical solution forms a complete closed loop from data acquisition, standardization processing, feature extraction, model construction to dynamic maintenance, enabling the intelligent soil quality assessment system to adapt to differences in soil types in different regions, changes in environmental conditions at different times, and data distribution drift during long-term operation, thereby improving the accuracy, stability, and reliability of the assessment results. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0050] Figure 1 This is a diagram illustrating the method steps.

[0051] Figure 2 This is a system flowchart. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1: Please refer to Figure 1 A multispectral image-driven intelligent soil quality assessment method is proposed, with the following specific steps:

[0054] S1. Data Acquisition: Acquire multispectral remote sensing images of the target area and corresponding soil sample information, and simultaneously record imaging time, illumination conditions, sensor operating status and observation geometric parameters;

[0055] S2. Spectral Correction: Based on imaging time, illumination conditions, and observation geometric parameters, radiometric correction and illumination normalization are performed on multispectral remote sensing images to obtain surface reflectance data at a uniform scale.

[0056] S3. State Decoupling: Identify and eliminate non-bare soil areas and abnormal reflection areas, construct soil water content state factors and surface structure state factors, and decouple soil spectral information.

[0057] S4. Feature Construction: Extract multi-scale spectral features based on the decoupled soil spectral data and form stable feature expressions related to soil physicochemical properties;

[0058] S5. Hierarchical Modeling: Based on stable features, a prediction model is constructed that includes a general feature expression layer and a soil type-specific expression layer to achieve adaptive fusion of different expression results;

[0059] S6. Regional Adaptation: Use multispectral data of the target region to constrain and optimize the prediction model, reducing the performance degradation of the model in cross-regional applications.

[0060] S7. Drift Detection: Continuously detect changes in the distribution of input features and dynamically adjust the model parameter update strategy based on the detection results;

[0061] S8. Continuous updates: The model is incrementally updated based on newly acquired multispectral data to maintain the long-term stability of soil quality assessment results.

[0062] In this embodiment, step S1 simultaneously records operational information such as imaging time, illumination conditions, sensor operating status, and observation geometric parameters while acquiring multispectral remote sensing images, laying the foundation for subsequent standardized processing of spectral data. Compared to the traditional method of only acquiring remote sensing images, this step fully preserves external condition information affecting the spectral response, making multispectral data acquired at different times and from different devices traceable and comparable. This solves the consistency problem in multi-source data fusion analysis from the source, providing reliable data support for constructing a soil quality assessment model applicable across time and space.

[0063] Step S2 performs radiometric correction and illumination geometric normalization on the multispectral images based on the recorded operating parameters, effectively eliminating the systematic interference of sensor response differences, atmospheric scattering and absorption, and changes in illumination geometry on the spectral data. By converting image data under different observation conditions into surface reflectance at a uniform scale, this step achieves a standardized representation of multi-temporal remote sensing data, significantly reducing spectral fluctuations caused by differences in observation conditions. This creates the prerequisite for subsequent extraction of stable soil characteristics and avoids misinterpreting changes in observation conditions as changes in soil quality.

[0064] Step S3 identifies and eliminates non-bare soil areas and anomalous reflectance areas, and decouples spectral information by constructing water content state factors and surface structure state factors, thus separating the instantaneous state of the soil from its inherent properties. Traditional methods often directly incorporate the influence of state factors such as changes in soil moisture content and differences in surface roughness into the soil quality assessment results, causing the assessment results to be interfered with by short-term environmental fluctuations. This step extracts spectral information reflecting the intrinsic physicochemical properties of the soil through decoupling processing, making the assessment results more focused on the intrinsic quality characteristics of the soil and improving the stability and interpretability of the assessment results.

[0065] Step S4 extracts multi-scale spectral features based on decoupled soil spectral data and forms grouped stable feature expressions according to the spectral response mechanism of soil physicochemical properties. By dividing the features into darkening feature group, salinization feature group, and absorption feature group, this step establishes a clear correspondence between features and soil quality indicators, enhancing the physical meaning of feature expressions. Simultaneously, stability screening retains features that are stable across different regions and soil types, providing a robust feature foundation for constructing cross-regional applicable prediction models and avoiding the problems of lack of specificity and regional adaptability in feature selection in traditional methods.

[0066] Step S5 constructs a hierarchical prediction model structure comprising a general feature expression layer and a soil type-specific expression layer, and achieves adaptive fusion of results from different expression layers. The general expression layer captures the common spectral response patterns of soils in different regions, while the specific expression layer learns the unique response characteristics of each soil type. The two are dynamically fused through weight coefficients. This hierarchical architecture effectively solves the contradiction between generality and specificity that traditional single models struggle to balance. It allows the model to utilize the common knowledge accumulated from a large-scale dataset while accurately adapting to the individual characteristics of specific soil types, significantly improving the model's prediction accuracy and generalization ability across different soil type regions.

[0067] Step S6 filters suitable samples for the target region through credibility assessment and adjusts model parameters using constrained optimization, effectively reducing performance degradation in cross-regional applications. Traditional transfer methods often directly use all data from the target region for model adjustment, which can easily introduce noisy samples leading to negative transfer. This step identifies high-quality suitable samples through feature deviation analysis and prediction uncertainty assessment, while employing parameter freezing strategies and regularization constraints to limit the adjustment range. This achieves effective model adaptation to the target region while preserving the knowledge accumulated in the source region, resulting in stable adaptation even with a limited sample size.

[0068] Step S7 continuously monitors changes in feature distribution and operating parameters to promptly detect operating condition drift during system operation and dynamically adjusts the model parameter update strategy based on the drift intensity. In long-term application, factors such as changes in environmental conditions, equipment aging, and seasonal changes can cause the input data distribution to gradually shift. Traditional fixed models cannot cope with this gradual drift, leading to a continuous decline in prediction accuracy. This step establishes a drift perception and response mechanism, formulating update strategies based on drift severity levels. This avoids model oscillations caused by over-updates and prevents performance degradation due to drift accumulation, ensuring the model's continued effectiveness in complex and changing environments.

[0069] Step S8 achieves long-term stable maintenance of model performance by incrementally updating based on new data and preventing the model from forgetting historical conditions through a memory sample mechanism. Traditional methods often focus only on fitting new data during model updates, neglecting the preservation of historical data, causing the model to lose its predictive ability for old conditions while adapting to new ones. This step trains the model using a mixture of new and memory samples, combined with an update strategy and parameter constraints passed on by drift detection. This allows the model to absorb new features from new data while retaining effective knowledge from historical data, forming a complete continuous learning loop and ensuring the long-term reliable operation of the soil quality assessment system.

[0070] This method effectively optimizes the consistency issues faced by multispectral remote sensing data in long-term application and cross-regional promotion by introducing a working condition parameter recording and explicit modeling mechanism. Traditional soil quality remote sensing assessment methods often use single-temporal data for modeling, neglecting the systematic impact of changes in observation conditions on spectral data, leading to a significant performance degradation of the model when applied in different times and regions. This method eliminates external interference factors through spectral correction and state decoupling techniques, enhances the stability of feature expression through multi-scale feature construction constrained by mechanisms, balances the model's universality and specificity through a hierarchical modeling architecture, ensures performance for cross-regional application and long-term operation through regional adaptation and drift detection techniques, and maintains the model's continuous effectiveness through a continuous learning mechanism. The overall technical solution forms a complete closed loop from data acquisition, standardization processing, feature extraction, model construction to dynamic maintenance, enabling the intelligent soil quality assessment system to adapt to differences in soil types in different regions, changes in environmental conditions at different times, and data distribution drift during long-term operation, thereby improving the accuracy, stability, and reliability of the assessment results.

[0071] Example 2: Please refer to Figure 1 In step S2, the specific steps for spectral correction are as follows:

[0072] S2.1 Read the sensor parameters and imaging time recorded in step S1.2, convert the digital signal value of the multispectral image into radiance value according to the radiometric calibration parameters, eliminate sensor nonlinear error, perform atmospheric correction in combination with atmospheric state parameters, eliminate atmospheric scattering and absorption interference, and obtain the true surface reflectance data.

[0073] S2.2 Extract the solar elevation angle, imaging angle and surface slope information recorded in step S1.2, establish an illumination geometry model, normalize the reflectivity data after radiation correction, and unify the reflectivity data of different imaging times and observation angles to standard observation conditions to obtain surface reflectivity data of a uniform scale.

[0074] In this embodiment: Step S1.1 involves acquiring multispectral image data of the target area using a remote sensing device equipped with a multispectral sensor, and collecting soil samples at representative locations within the image coverage area according to a preset spatial distribution rule, thus establishing a spatial correspondence between the remote sensing image data and the measured soil physicochemical properties. Laboratory analysis of the soil samples yields organic matter content, nutrient content, and pH values, providing reliable labeled data for the subsequent construction of a mapping model between spectral characteristics and soil quality indicators. This step ensures accurate spatial matching between remote sensing observation data and ground-based ground data, laying a data foundation for establishing a stable soil quality prediction model.

[0075] Step S1.2 involves simultaneously recording operational parameters such as imaging time, solar altitude angle, cloud cover, sensor exposure parameters, operating temperature, flight altitude, and imaging angle while acquiring multispectral image data. These operational parameters are then linked to the corresponding image data and stored, thus fully preserving information about external conditions affecting the quality of the spectral data. The recording of these operational parameters enables targeted correction processing of spectral data under different observation conditions, providing a necessary basis for eliminating the systematic impact of changes in observation conditions on the spectral response. This ensures the comparability and traceability of remote sensing data acquired across multiple time periods and devices.

[0076] Compared to traditional data acquisition methods that only collect remote sensing images and soil samples, step S1 significantly improves the standardization and completeness of data acquisition by simultaneously recording complete operating condition parameter information. Traditional methods often neglect to record observation conditions during the data acquisition stage, resulting in systematic biases in data acquired from different batches due to differences in observation conditions. Subsequent processing cannot accurately trace and correct these biases, affecting the model's generalization ability. This step, by establishing a correlation storage mechanism between image data and operating condition parameters, provides complete input information for subsequent standardization processes such as radiometric correction and illumination normalization. This enables multispectral data acquired at different times, with different equipment, and under different environmental conditions to be converted into comparable data under a unified standard.

[0077] By establishing the spatial correspondence in step S1.1 and recording the working condition information in step S1.2, the overall data acquisition process achieves the unification of remote sensing observation data, ground truth data, and observation condition information, providing a high-quality training dataset for constructing a soil quality assessment model applicable across time and space. This data acquisition method fundamentally solves the data quality problems caused by inconsistent data sources and unclear observation conditions in traditional methods, creating favorable conditions for the implementation of subsequent steps and improving the reliability and practicality of the overall method.

[0078] Example 3: Please refer to Figure 1 In step S3, the specific steps of state decoupling are as follows:

[0079] S3.1 Read the surface reflectance data from step S2, calculate the reflectance ratio or difference between the red band and the near-infrared band, identify the vegetation coverage area based on the vegetation index threshold, identify water bodies, shadows and areas with strong reflectance abnormalities through reflectance statistical analysis, remove the identified non-bare soil areas and abnormal areas, and retain the bare soil spectral data.

[0080] S3.2. Select short-wave infrared reflectance to calculate water state factor, select texture parameters or multi-angle data to calculate structural state factor, establish a relationship model of the influence of state factor on reflectance, decompose reflectance into soil intrinsic reflectance component and state reflectance component, and subtract state component to obtain soil intrinsic spectral information.

[0081] In this embodiment: Step S2.1 reads the sensor parameters and imaging time recorded in step S1.2, and converts the digital signal values ​​of the multispectral image into radiance values ​​according to the radiometric calibration parameters. This eliminates the nonlinear error of the sensor's own response and ensures the comparability of data acquired by different devices. Furthermore, atmospheric correction is performed using atmospheric state parameters, effectively eliminating the interference of atmospheric molecular scattering and aerosol absorption on the spectral signal, and obtaining reflectance data that reflects the true reflectance characteristics of the land surface. This step completes the conversion from the raw sensor signal to the true reflectance of the land surface, providing a reliable physical quantity basis for the subsequent extraction of stable soil spectral characteristics.

[0082] Step S2.2 establishes a quantitative model describing the influence of illumination geometry on reflectivity by extracting the solar elevation angle, imaging angle, and surface slope information recorded in step S1.2. By normalizing the radiometrically corrected reflectivity data, reflectivity data acquired at different imaging times and observation angles are unified to the equivalent reflectivity under standard observation conditions, eliminating the influence of solar position variations and observation geometry differences on spectral data. This step standardizes multi-temporal remote sensing data in terms of illumination geometry, making data acquired at different times directly comparable and providing consistency assurance for cross-temporal soil quality monitoring.

[0083] The formula for calculating normalized reflectance is as follows:

[0084]

[0085] R is the normalized reflectance; A is the observed reflectance after radiometric correction; B is the cosine of the solar zenith angle under standard observation conditions; C is the cosine of the sensor-observed zenith angle under standard observation conditions; D is the cosine of the solar zenith angle under actual observation conditions; E is the cosine of the sensor-observed zenith angle under actual observation conditions; F is a very small positive number to prevent the denominator from being zero, with a value of 0.0001.

[0086] Compared to traditional spectral data processing methods that only perform simple radiometric calibration, step S2 significantly improves the standardization level of multispectral data through systematic radiometric correction and illumination geometry normalization. Traditional methods often ignore the complex influence of atmospheric conditions and illumination geometry on spectral data, directly using raw reflectance data for analysis. This leads to systematic biases between data acquired at different times or from different devices, causing the model to misinterpret spectral changes caused by these external factors as true changes in soil quality during training. This step introduces a correction mechanism driven by operating parameters, quantitatively correcting the three main sources of interference: sensor characteristics, atmospheric conditions, and illumination geometry, transforming the raw observation data into standardized data that reflects the essential characteristics of the land surface.

[0087] Through radiometric correction in step S2.1 and geometric normalization in step S2.2, the overall spectral correction process comprehensively compensates for changes in observation conditions, enabling subsequent feature extraction and model construction to be based on more consistent input data. This systematic correction approach effectively reduces data fluctuations caused by differences in observation conditions, improving the stability and repeatability of extracted features. This step creates a good data quality foundation for subsequent state decoupling and feature construction, enabling the model to more accurately capture the true variation patterns of soil quality and reducing the impact of external interference factors on the evaluation results.

[0088] Example 4: Please refer to Figure 1 In step S3, the specific steps of state decoupling are as follows:

[0089] S3.1 Read the surface reflectance data from step S2, calculate the reflectance ratio or difference between the red band and the near-infrared band, identify the vegetation coverage area based on the vegetation index threshold, identify water bodies, shadows and areas with strong reflectance abnormalities through reflectance statistical analysis, remove the identified non-bare soil areas and abnormal areas, and retain the bare soil spectral data.

[0090] S3.2. Select short-wave infrared reflectance to calculate water state factor, select texture parameters or multi-angle data to calculate structural state factor, establish a relationship model of the influence of state factor on reflectance, decompose reflectance into soil intrinsic reflectance component and state reflectance component, and subtract state component to obtain soil intrinsic spectral information.

[0091] In this embodiment: Step S3.2 calculates the water content state factor by selecting short-wave infrared reflectance and calculates the structural state factor by selecting texture parameters or multi-angle data, establishing a quantitative relationship model of the influence of the state factor on reflectance. By decomposing reflectance into intrinsic soil reflectance components and state reflectance components, and subtracting the influence of the state components, this step achieves effective separation between the instantaneous state and intrinsic properties of the soil. This decoupling process obtains intrinsic spectral information reflecting the inherent physicochemical properties of the soil, eliminates short-term interference of soil water content fluctuations and surface structure changes on spectral data, and improves the stability of feature expression.

[0092] Compared to traditional methods that directly use corrected reflectance data for feature extraction, this step significantly improves the purity and stability of soil spectral information by introducing interference region identification and state factor decoupling mechanisms. Traditional methods often mix spectral information from non-target features such as vegetation and water bodies into the analysis process, and directly incorporate the influence of transient state factors such as soil moisture content and surface roughness into the soil quality assessment results, leading to interference from various non-essential factors. This step systematically processes interference factors from both spatial and physical perspectives through precise identification and removal in step S3.1 and state decoupling processing in step S3.2.

[0093] The intrinsic spectral information of soil obtained through state decoupling focuses more on the intrinsic quality characteristics of the soil. This allows the subsequently extracted spectral features to stably reflect the true state of the soil's physicochemical properties, such as organic matter and nutrients, without being affected by short-term factors such as temporary increases in water content after rainfall or changes in surface structure after tillage. This approach improves the temporal stability and spatial comparability of feature expression, laying a reliable feature foundation for constructing soil quality prediction models applicable across time phases and regions, and making the evaluation results more scientific and practical.

[0094] Example 5: Please refer to Figure 1 In step S4, the specific steps for feature construction are as follows:

[0095] S4.1 Read the intrinsic spectral information of the soil from step S3, calculate the ratio characteristics, difference characteristics and normalized combination characteristics of reflectance in each band, differentiate the spectral curve to obtain the first derivative characteristics and second derivative characteristics, calculate the gray-level co-occurrence matrix parameters under different spatial windows to obtain texture characteristics, and statistically analyze the similarity of adjacent pixels to obtain spatial clustering characteristics.

[0096] S4.2 Based on the spectral response mechanism of soil organic matter, salinity and nutrients, the features are classified into darkening feature group, salinization feature group and absorption feature group respectively. The correlation coefficient between the features and physicochemical properties is calculated on the source region and target region samples. Features with stable and high correlation coefficients are selected and combined to form a stable feature expression.

[0097] In this embodiment: Step S4.1 captures the relative relationship information between different bands by calculating the ratio characteristics, difference characteristics, and normalized combination characteristics of reflectance for each band. First-order and second-order derivative characteristics are obtained by differentiating the spectral curves, extracting the rate of change and curvature information of the spectral curves. Texture features are obtained by calculating the gray-level co-occurrence matrix parameters under different spatial windows, characterizing the spatial variation characteristics of the soil surface. Spatial clustering features are obtained by statistically analyzing the similarity of adjacent pixels, describing the spatial distribution pattern of soil types. This step constructs a multi-scale feature set from both spectral and spatial dimensions, providing a rich feature foundation for comprehensively characterizing soil quality information.

[0098] Step S4.2 establishes a physical correlation between features and soil physicochemical properties by classifying multi-scale features into darkening, salinization, and absorption feature groups based on the spectral response mechanisms of soil organic matter, salinity, and nutrients. By calculating the correlation coefficients between features and physicochemical properties on samples from both the source and target regions, and selecting features with stable and high correlation coefficients for combination, this step ensures that the selected features maintain stable response relationships under different regional conditions. This feature selection method, based on mechanistic constraints and stability verification, forms stable feature expressions with clear physical meaning and cross-regional applicability.

[0099] The formula for calculating the feature stability score is as follows:

[0100]

[0101] S is the stability score of the feature; G is the correlation coefficient between the feature and the physicochemical properties in the source region samples; H is the correlation coefficient between the feature and the physicochemical properties in the target region samples; I is a very small positive number to prevent the denominator from being zero, with a value of 0.001.

[0102] Compared to traditional methods that rely on empirical band combinations or blindly select features using statistical methods, this step significantly improves the comprehensiveness and stability of feature representation by introducing multi-scale feature extraction and mechanistic constraint-based selection mechanisms. Traditional methods often focus only on a few classic vegetation or soil indices, lacking in-depth mining of spectral information. Furthermore, the feature selection process lacks physical mechanism guidance, resulting in poor stability of selected features when applied in different regions. This step, through the multi-dimensional feature set constructed in step S4.1, covers information from multiple levels, including band operations, derivative changes, and spatial texture. Step S4.2, based on the spectral response mechanism of soil physicochemical properties, performs feature grouping and cross-regional stability verification, ensuring the physical interpretability and applicability of the features.

[0103] By employing a mechanistic-constrained feature construction approach, the stable feature representations formed in this step not only accurately reflect the essential attributes of soil quality but also maintain consistent response relationships across different soil types and regional conditions. This provides reliable input variables for subsequently constructing predictive models with generalization capabilities. This feature construction strategy enhances the interpretability and robustness of the model, making the evaluation results more scientific and credible, while reducing the risk of performance degradation due to feature failure when the model is applied across regions.

[0104] Example 6: Please refer to Figure 1 In step S5, the specific steps for layered modeling are as follows:

[0105] S5.1 Read the stable feature expression from step S4, collect the spectral features and physicochemical properties of soil samples from the source area, construct a general feature expression layer for all samples to establish the mapping relationship between spectra and physicochemical properties, divide the samples into saline-alkali soil, red soil and black soil sample groups according to soil type, and construct soil type-specific expression layers to establish the mapping relationship for each type.

[0106] S5.2 Extract stable features of soil samples from the target area and input them into the stratified prediction model. Calculate the feature value distribution of the sample in each feature group, determine the similarity between the sample and each soil type and normalize it into weight coefficients. Input the sample features into each expression layer to obtain the output results. Based on the weight coefficients, perform weighted summation of the outputs of each layer to obtain the soil quality prediction results.

[0107] In this embodiment: Step S5.1 involves collecting the spectral characteristics and physicochemical properties of soil samples from the source region, constructing a general feature expression layer using all samples to establish a mapping relationship between spectra and physicochemical properties, and learning the common spectral response patterns of soils from different regions. Simultaneously, the samples are divided into saline-alkali soil, red soil, and black soil sample groups according to soil type, and soil type-specific expression layers are constructed to establish mapping relationships for each type, thereby specifically learning the unique spectral response characteristics of each soil type. This hierarchical modeling structure achieves an organic combination of general and specialized knowledge, making the model both widely applicable and able to accurately capture the individual characteristics of different soil types.

[0108] Step S5.2 extracts stable features from soil samples in the target area and inputs them into a stratified prediction model. It calculates the feature value distribution of the sample in each feature group, determines the similarity between the sample and each soil type, and normalizes it into weight coefficients, thus achieving automatic identification of the soil type to which the target sample belongs. By inputting sample features into each expression layer to obtain output results, and then weighting and summing the outputs of each layer according to the weight coefficients, this step achieves dynamic fusion of prediction results from the general expression layer and the specific expression layer. This allows the prediction results to adaptively adjust the contribution ratio of each expression layer according to the actual type characteristics of the sample.

[0109] The formula for calculating the final prediction result is as follows:

[0110]

[0111] In the formula, the weighting coefficient L k The calculation formula is as follows:

[0112]

[0113] Feature distance N k The calculation formula is:

[0114]

[0115] Y represents the final prediction result; J represents the weight coefficient of the general layer; K represents the output of the general feature expression layer; For the k-th type of special layer weight coefficient; Output for the k-th soil type specific expression layer; This represents the distance between the sample characteristics and the distribution of the characteristics of the kth soil type. P represents the distance between the sample feature and the general feature distribution; P₱ represents the feature value of the sample in feature group j; Let be the mean value of the k-th soil type in feature group j; is the standard deviation of the k-th soil type in characteristic group j; U is a very small positive number to prevent the denominator from being zero, and takes the value 0.0001;

[0116] Compared to traditional methods that use a single model or multiple independent models for soil quality prediction, this step significantly improves the applicability and prediction accuracy of the model by introducing a hierarchical modeling architecture and an adaptive fusion mechanism. Traditional single-model methods often struggle to account for the unique characteristics of different soil types. A model that performs well on one soil type will show a significant performance drop when applied to other types. While multiple independent model methods are highly targeted, they lack the utilization of common knowledge and are prone to prediction instability on type boundary samples. This step achieves the collaborative expression of common and unique knowledge through the hierarchical structure constructed in step S5.1, and dynamically adjusts the weights of each expression layer based on sample characteristics through the adaptive fusion mechanism in step S5.2.

[0117] By employing a hierarchical modeling and adaptive fusion approach, this step enables the prediction model to automatically match the optimal prediction strategy based on the actual characteristics of the samples to be evaluated. This approach leverages the generalized patterns accumulated from a large number of samples to enhance the model's generalization ability, while simultaneously using a specific expression layer to accurately capture the unique patterns of different soil types, thereby improving prediction accuracy. This modeling method effectively addresses the shortcomings of traditional methods in adapting to soil type diversity, providing a model foundation that is both universal and targeted for cross-regional soil quality assessment.

[0118] Example 7: Please refer to Figure 1 In step S6, the specific steps for region adaptation are as follows:

[0119] S6.1. Obtain multispectral data of the target area and extract stable feature expressions according to steps S2 to S4. Calculate the distance between the sample features of the target area and the feature distribution center of the source area. Mark the samples with a distance exceeding the threshold as high deviation samples. Input the samples into the hierarchical prediction model to obtain the prediction results and calculate the uncertainty value. Establish a scoring function based on the degree of deviation and uncertainty and select high-scoring samples as suitable samples.

[0120] S6.2. Adjust the parameters of the hierarchical prediction model using the adapted samples, keeping the general expression layer parameters unchanged, and only adjusting the specific expression layer parameters and fusion weight coefficients corresponding to the target region. Introduce a parameter change penalty term in the loss function to limit the deviation. Verify the prediction error on the adapted samples and source region samples. Stop adjusting when the accuracy requirements are met to obtain the adapted model.

[0121] In this embodiment: Step S6.1 acquires multispectral data of the target region and extracts stable feature representations according to steps S2 to S4, calculating the distance between the feature distribution centers of the target region and the source region, thereby achieving a quantitative assessment of the distribution differences between the target region and source region samples. By marking samples with distances exceeding a threshold as high-deviation samples and inputting these samples into a hierarchical prediction model to calculate the uncertainty value of the prediction results, this step establishes a dual evaluation mechanism that comprehensively considers both the degree of feature deviation and prediction reliability. By establishing a scoring function based on the degree of deviation and uncertainty and selecting high-scoring samples as suitable samples, it ensures that the samples used for model adjustment have high quality and representativeness.

[0122] Step S6.2 achieves precise local optimization of the model by adjusting the parameters of the hierarchical prediction model using adapted samples, keeping the general expression layer parameters unchanged and adjusting only the specific expression layer parameters and fusion weight coefficients corresponding to the target region. By introducing a penalty term for parameter changes into the loss function to limit the deviation, this step avoids the negative transfer phenomenon of the model losing knowledge of the source region due to overfitting to the target region. By simultaneously verifying the prediction error on both adapted and source region samples and stopping the adjustment when the accuracy requirements are met, this step ensures that the adapted model maintains good performance in both the target and source regions.

[0123] Compared to traditional transfer learning methods that directly use all data from the target region for model retraining or simple parameter fine-tuning, this step significantly improves the effectiveness and stability of cross-regional model adaptation by introducing sample quality assessment and constraint optimization mechanisms. Traditional methods often ignore the quality differences of samples in the target region, using all samples indiscriminately for model adjustment, which easily introduces noisy and outlier samples, leading to a decline in model performance. Furthermore, the lack of effective constraints on the parameter adjustment range can easily result in overfitting to the target region data while forgetting knowledge of the source region. This step identifies high-quality adaptation samples through the dual assessment mechanism established in step S6.1, and achieves controlled optimization through the parameter freezing strategy and regularization constraints designed in step S6.2.

[0124] By employing a combined strategy of sample selection and constraint optimization, this step ensures that the model adaptation process effectively absorbs the feature distribution information of the target region while preserving the common knowledge accumulated in the source region. This adaptation method achieves stable performance improvements even with a limited sample size in the target region, effectively reducing performance degradation in cross-regional applications. This step lays the technical foundation for the widespread application of soil quality assessment methods, enabling models trained in the source region to quickly adapt to new target regions while maintaining reliable assessment accuracy.

[0125] Example 8: Please refer to Figure 1 In step S7, the specific steps for drift detection are as follows:

[0126] S7.1. Acquire multispectral data of the target area at preset time intervals and extract stable feature expressions according to steps S2 to S4. Establish a sliding time window to store sample feature statistics. Calculate the statistical distance between the feature distribution of the current window and the historical benchmark window as the offset. Extract the current working condition parameters and calculate the numerical difference with the historical working condition. When the offset or difference exceeds the threshold, it is determined that working condition drift has occurred and the intensity level is recorded.

[0127] S7.2 Read the drift intensity level recorded in step S7.1, determine the parameter update range according to the intensity, adjust only the fusion weight coefficient when there is a slight drift, adjust the parameters and weight coefficient of the specific expression layer when there is a moderate drift, and adjust the parameters and weight coefficient of the general expression layer and the specific expression layer when there is a severe drift. Analyze the direction of change of the working condition parameters to generate parameter adjustment constraints, and pass the update strategy and constraints to step S8.

[0128] In this embodiment: Step S7.1 establishes a continuous monitoring mechanism for changes in the distribution of input data by acquiring multispectral data of the target area at preset time intervals and extracting stable feature expressions according to steps S2 to S4. By establishing a sliding time window to store sample feature statistics and calculating the statistical distance between the feature distribution of the current window and the historical benchmark window as the offset, this step achieves quantitative detection of data distribution drift. At the same time, by extracting the current operating condition parameters and calculating the numerical difference with the historical operating conditions, this step establishes the ability to analyze the correlation between feature drift and operating condition changes. When the offset or difference exceeds the threshold, the operating condition drift is promptly determined and the intensity level is recorded, providing an accurate basis for subsequent dynamic adjustment of the model update strategy.

[0129] Step S7.2 reads the drift intensity level recorded in step S7.1 and determines the parameter update range based on the intensity, establishing a graded response model update strategy. For mild drift, only the fusion weight coefficients are adjusted; for moderate drift, the parameters and weight coefficients of the specific expression layer are adjusted; and for severe drift, the parameters and weight coefficients of the general expression layer and specific expression layer are adjusted. This graded strategy avoids over-updates caused by minor fluctuations while ensuring a sufficient response to significant drifts. By analyzing the direction of parameter changes under operating conditions to generate parameter adjustment constraints, this step provides a clear optimization direction for the subsequent continuous learning process, making model updates more efficient and stable.

[0130] Compared to traditional maintenance methods that use fixed models or periodic full retraining, this step significantly improves the model's adaptability and stability in long-term operation by introducing a drift detection mechanism and a dynamic update strategy. Traditional fixed-model methods cannot cope with gradual changes in data distribution caused by environmental conditions and equipment aging, leading to a continuous decline in prediction accuracy over time. While periodic full retraining can update the model, it lacks specificity, has high computational costs, and can easily destroy historically accumulated knowledge. This step uses the continuous monitoring mechanism established in step S7.1 to perceive changes in data distribution in real time, and uses the hierarchical response strategy designed in step S7.2 to accurately determine the update range based on the degree of drift.

[0131] By combining drift detection with dynamic strategy adjustment, this step transforms the model maintenance process from passive, periodic updates to proactive, on-demand responses, improving both update efficiency and effectiveness. This maintenance method ensures the model maintains a consistent match with actual data distribution over long-term application, effectively preventing model performance degradation due to environmental changes. This step provides technical support for the long-term reliable operation of the soil quality assessment system, enabling it to adapt to complex and ever-changing real-world application environments and maintain stable assessment accuracy.

[0132] Example 9: Please refer to Figure 1 In step S8, the specific steps are continuously updated as follows:

[0133] S8.1. Based on the drift time detected in step S7.1 or the preset update cycle, new multispectral data is acquired in the target area and soil samples are collected to obtain physicochemical property parameters. Stable feature expressions are obtained by processing according to steps S2 to S4. The distribution density and coverage of the new samples are calculated. Samples from different feature areas, different working conditions, and different soil types are selected to form an incremental training sample set. Representative samples are selected from the historical sample database to form a memory sample set.

[0134] S8.2 Read the update strategy and adjustment constraints passed in step S7.2, combine the incremental training sample set and the memory sample set to form a mixed training sample set, select the model level and parameters to be adjusted according to the update strategy, calculate the prediction error of incremental samples and memory samples in the objective function, introduce directional constraints and magnitude penalty terms, verify the prediction accuracy after optimization, and replace the model parameters to complete the update when the requirements are met.

[0135] In this embodiment: Step S8.1 acquires new multispectral data in the target area and collects soil samples to obtain physicochemical property parameters based on the drift time detected in step S7.1 or a preset update cycle, establishing a demand-triggered incremental data acquisition mechanism. By processing the data according to steps S2 to S4 to obtain stable feature expressions and calculate the distribution density and coverage of the newly added samples, this step achieves the evaluation of the quality and representativeness of the new data. By selecting samples from different feature regions, different working conditions, and different soil types to form an incremental training sample set, and simultaneously selecting representative samples from the historical sample database to form a memory sample set, this step ensures that the model update can both learn new features and retain historical knowledge.

[0136] Step S8.2 reads the update strategy and adjustment constraints passed in step S7.2, and combines the incremental training sample set and the memory sample set into a mixed training sample set, thus achieving collaborative training of new and old knowledge. By selecting the model level and parameters to be adjusted according to the update strategy, and calculating the prediction error of the incremental samples and memory samples in the objective function, this step establishes a dual-objective optimization mechanism. By introducing directional constraints and magnitude penalty terms to limit parameter changes, and replacing the model parameters after verifying that the prediction accuracy meets the requirements after optimization, this step ensures the controllability and effectiveness of model updates, and achieves continuous maintenance of model performance.

[0137] Compared to traditional model update methods that rely on full data retraining or simple incremental learning, this step significantly improves the effectiveness and stability of continuous model learning by introducing a memory sample mechanism and a constraint optimization strategy. While traditional full data retraining methods can fully utilize new data, they are computationally expensive and time-consuming, making them unsuitable for real-time updates. Simple incremental learning methods, on the other hand, are prone to model forgetting historical data, leading to catastrophic forgetting. This step achieves balanced sampling of new and old data through the dual-sample set construction mechanism established in step S8.1, and the constraint optimization scheme designed in step S8.2 preserves historical knowledge while absorbing new knowledge.

[0138] By employing a combined strategy of incremental updates and memory retention, this step enables the model to continuously adapt to changes in data distribution over long-term operation without losing its predictive ability for historical conditions. This update method ensures both computational efficiency and maintains the stability of model performance, forming a complete continuous learning loop from drift detection and strategy formulation to parameter updates. This step provides key technical support for the long-term stable operation of the soil quality assessment system, enabling the system to maintain reliable assessment accuracy in complex and ever-changing real-world application scenarios, and ensuring the long-term validity and practical value of the assessment results.

[0139] Please see Figure 2A multispectral image-driven intelligent soil quality evaluation system includes a data acquisition module, a spectral correction module, a state decoupling module, a feature construction module, a hierarchical modeling module, a regional adaptation module, a drift detection module, and a continuous update module.

[0140] The data acquisition module is used to acquire multispectral remote sensing images and soil sample information of the target area, and simultaneously record imaging time, illumination conditions, sensor working status and observation geometric parameters;

[0141] The spectral correction module is used to perform radiometric correction and illumination normalization on multispectral images based on operating parameters to obtain surface reflectance data at a uniform scale.

[0142] The state decoupling module is used to identify and eliminate non-bare soil areas, construct water-bearing state factors and surface structure state factors, and extract intrinsic soil spectral information;

[0143] The feature construction module is used to extract multi-scale spectral features, group features according to the response mechanism of soil physicochemical properties, and form a stable feature expression.

[0144] The hierarchical modeling module is used to construct a prediction model that includes a general expression layer and a type-specific expression layer, enabling adaptive fusion of results from different expression layers;

[0145] The region adaptation module is used to evaluate the sample quality of the target region and adjust the model parameters through constraint optimization to reduce the performance degradation of cross-regional applications;

[0146] The drift detection module is used to monitor changes in feature distribution and operating parameters, determine the drift intensity level, and dynamically formulate model update strategies.

[0147] The continuous update module is used for incremental learning based on new data, and combines the memory sample mechanism to update model parameters to maintain the long-term stability of evaluation results.

[0148] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multispectral image-driven intelligent soil quality evaluation method, characterized in that, The specific steps are as follows: S1. Data Acquisition: Acquire multispectral remote sensing images of the target area and corresponding soil sample information, and simultaneously record imaging time, illumination conditions, sensor operating status and observation geometric parameters; S2. Spectral Correction: Based on the imaging time, illumination conditions, and observation geometric parameters, radiometric correction and illumination normalization are performed on the multispectral remote sensing image to obtain surface reflectance data at a uniform scale. S3. State Decoupling: Identify and eliminate non-bare soil areas and abnormal reflection areas, construct soil water content state factors and surface structure state factors, and decouple soil spectral information. S4. Feature Construction: Extract multi-scale spectral features based on the decoupled soil spectral data and form stable feature expressions related to soil physicochemical properties; S5. Hierarchical modeling: Based on the stable features, a prediction model is constructed that includes a general feature expression layer and a soil type-specific expression layer to achieve adaptive fusion of different expression results; S6. Regional Adaptation: Use multispectral data of the target region to constrain and optimize the prediction model, reducing the performance degradation of the model in cross-regional applications. S7. Drift Detection: Continuously detect changes in the distribution of input features and dynamically adjust the model parameter update strategy based on the detection results; S8. Continuous updates: The model is incrementally updated based on newly acquired multispectral data to maintain the long-term stability of soil quality assessment results.

2. The multispectral image-driven intelligent soil quality evaluation method according to claim 1, characterized in that, In step S1, the specific steps for data acquisition are as follows: S1.

1. Use a remote sensing device equipped with a multispectral sensor to acquire multispectral image data of the target area, select representative locations within the image coverage area according to preset spatial distribution rules to collect soil samples, and conduct laboratory analysis on the soil samples to obtain organic matter content, nutrient content and pH values. S1.2 When acquiring multispectral image data, simultaneously record the imaging time, solar altitude angle, cloud cover, sensor exposure parameters, operating temperature, flight altitude, and imaging angle, and establish and store the correlation between the operating parameters and the corresponding image data.

3. The xxxx according to claim 1, characterized in that, In step S2, the specific steps for spectral correction are as follows: S2.1 Read the sensor parameters and imaging time recorded in step S1.2, convert the digital signal value of the multispectral image into radiance value according to the radiometric calibration parameters, eliminate sensor nonlinear error, perform atmospheric correction in combination with atmospheric state parameters, eliminate atmospheric scattering and absorption interference, and obtain the true surface reflectance data. S2.2 Extract the solar elevation angle, imaging angle and surface slope information recorded in step S1.2, establish an illumination geometry model, normalize the reflectivity data after radiation correction, and unify the reflectivity data of different imaging times and observation angles to standard observation conditions to obtain surface reflectivity data of a uniform scale.

4. The multispectral image-driven intelligent soil quality evaluation method according to claim 1, characterized in that, In step S3, the specific steps of state decoupling are as follows: S3.1 Read the surface reflectance data from step S2, calculate the reflectance ratio or difference between the red band and the near-infrared band, identify the vegetation coverage area based on the vegetation index threshold, identify water bodies, shadows and areas with strong reflectance abnormalities through reflectance statistical analysis, remove the identified non-bare soil areas and abnormal areas, and retain the bare soil spectral data. S3.

2. Select short-wave infrared reflectance to calculate water state factor, select texture parameters or multi-angle data to calculate structural state factor, establish a relationship model of the influence of state factor on reflectance, decompose reflectance into soil intrinsic reflectance component and state reflectance component, and subtract state component to obtain soil intrinsic spectral information.

5. The multispectral image-driven intelligent soil quality evaluation method according to claim 1, characterized in that, In step S4, the specific steps for feature construction are as follows: S4.1 Read the intrinsic spectral information of the soil from step S3, calculate the ratio characteristics, difference characteristics and normalized combination characteristics of reflectance in each band, differentiate the spectral curve to obtain the first derivative characteristics and second derivative characteristics, calculate the gray-level co-occurrence matrix parameters under different spatial windows to obtain texture characteristics, and statistically analyze the similarity of adjacent pixels to obtain spatial clustering characteristics. S4.2 Based on the spectral response mechanism of soil organic matter, salinity and nutrients, the features are classified into darkening feature group, salinization feature group and absorption feature group respectively. The correlation coefficient between the features and physicochemical properties is calculated on the source region and target region samples. Features with stable and high correlation coefficients are selected and combined to form a stable feature expression.

6. The multispectral image-driven intelligent soil quality evaluation method according to claim 1, characterized in that, In step S5, the specific steps of layered modeling are as follows: S5.1 Read the stable feature expression from step S4, collect the spectral features and physicochemical properties of soil samples from the source area, construct a general feature expression layer for all samples to establish the mapping relationship between spectra and physicochemical properties, divide the samples into saline-alkali soil, red soil and black soil sample groups according to soil type, and construct soil type-specific expression layers to establish the mapping relationship for each type. S5.2 Extract stable features of soil samples from the target area and input them into the stratified prediction model. Calculate the feature value distribution of the sample in each feature group, determine the similarity between the sample and each soil type and normalize it into weight coefficients. Input the sample features into each expression layer to obtain the output results. Based on the weight coefficients, perform weighted summation of the outputs of each layer to obtain the soil quality prediction results.

7. The multispectral image-driven intelligent soil quality evaluation method according to claim 1, characterized in that, In step S6, the specific steps for region adaptation are as follows: S6.

1. Obtain multispectral data of the target area and extract stable feature expressions according to steps S2 to S4. Calculate the distance between the sample features of the target area and the feature distribution center of the source area. Mark the samples with a distance exceeding the threshold as high deviation samples. Input the samples into the hierarchical prediction model to obtain the prediction results and calculate the uncertainty value. Establish a scoring function based on the degree of deviation and uncertainty and select high-scoring samples as suitable samples. S6.

2. Adjust the parameters of the hierarchical prediction model using the adapted samples, keeping the general expression layer parameters unchanged, and only adjusting the specific expression layer parameters and fusion weight coefficients corresponding to the target region. Introduce a parameter change penalty term in the loss function to limit the deviation. Verify the prediction error on the adapted samples and source region samples. Stop adjusting when the accuracy requirements are met to obtain the adapted model.

8. The multispectral image-driven intelligent soil quality evaluation method according to claim 1, characterized in that, In step S7, the specific steps for drift detection are as follows: S7.

1. Acquire multispectral data of the target area at preset time intervals and extract stable feature expressions according to steps S2 to S4. Establish a sliding time window to store sample feature statistics. Calculate the statistical distance between the feature distribution of the current window and the historical benchmark window as the offset. Extract the current working condition parameters and calculate the numerical difference with the historical working condition. When the offset or difference exceeds the threshold, it is determined that working condition drift has occurred and the intensity level is recorded. S7.2 Read the drift intensity level recorded in step S7.1, determine the parameter update range according to the intensity, adjust only the fusion weight coefficient when there is a slight drift, adjust the parameters and weight coefficient of the specific expression layer when there is a moderate drift, and adjust the parameters and weight coefficient of the general expression layer and the specific expression layer when there is a severe drift. Analyze the direction of change of the working condition parameters to generate parameter adjustment constraints, and pass the update strategy and constraints to step S8.

9. The multispectral image-driven intelligent soil quality evaluation method according to claim 1, characterized in that, In step S8, the specific steps are continuously updated as follows: S8.

1. Based on the drift time detected in step S7.1 or the preset update cycle, new multispectral data is acquired in the target area and soil samples are collected to obtain physicochemical property parameters. Stable feature expressions are obtained by processing according to steps S2 to S4. The distribution density and coverage of the new samples are calculated. Samples from different feature areas, different working conditions, and different soil types are selected to form an incremental training sample set. Representative samples are selected from the historical sample database to form a memory sample set. S8.2 Read the update strategy and adjustment constraints passed in step S7.2, combine the incremental training sample set and the memory sample set to form a mixed training sample set, select the model level and parameters to be adjusted according to the update strategy, calculate the prediction error of incremental samples and memory samples in the objective function, introduce directional constraints and magnitude penalty terms, verify the prediction accuracy after optimization, and replace the model parameters to complete the update when the requirements are met.

10. A multispectral image-driven intelligent soil quality evaluation system, characterized in that, The multispectral image-driven intelligent soil quality evaluation system is used to execute the multispectral image-driven intelligent soil quality evaluation method according to any one of claims 1 to 9. The system includes a data acquisition module, a spectral correction module, a state decoupling module, a feature construction module, a hierarchical modeling module, a region adaptation module, a drift detection module, and a continuous update module. The data acquisition module is used to acquire multispectral remote sensing images and soil sample information of the target area, and simultaneously record imaging time, illumination conditions, sensor working status and observation geometric parameters; The spectral correction module is used to perform radiometric correction and illumination normalization on multispectral images based on operating parameters to obtain surface reflectance data at a uniform scale. The state decoupling module is used to identify and eliminate non-bare soil areas, construct water-bearing state factors and surface structure state factors, and extract intrinsic soil spectral information; The feature construction module is used to extract multi-scale spectral features, group features according to the response mechanism of soil physicochemical properties, and form a stable feature expression. The hierarchical modeling module is used to construct a prediction model that includes a general expression layer and a type-specific expression layer, enabling adaptive fusion of results from different expression layers; The region adaptation module is used to evaluate the sample quality of the target region and adjust the model parameters through constraint optimization to reduce the performance degradation of cross-regional applications; The drift detection module is used to monitor changes in feature distribution and operating parameters, determine the drift intensity level, and dynamically formulate model update strategies. The continuous update module is used for incremental learning based on new data, and combines the memory sample mechanism to update model parameters to maintain the long-term stability of evaluation results.

Images