Multi-source fusion hyperspectral soil nutrient remote sensing inversion system
The multi-source fusion hyperspectral soil nutrient remote sensing inversion system solves the problems of inversion accuracy and coverage in complex scenarios using single hyperspectral remote sensing, achieving efficient and accurate soil nutrient monitoring and improving inversion accuracy and coverage in complex terrain.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA GEOLOGICAL SURVEY CHANGSHA NATURAL RESOURCES COMPREHENSIVE SURVEY CENT
- Filing Date
- 2026-03-16
- Publication Date
- 2026-05-08
AI Technical Summary
Existing single hyperspectral remote sensing methods for detecting soil nutrients neglect the coupling relationship between nutrients in complex scenarios such as mining area reclamation and saline-alkali land, resulting in poor inversion accuracy and significant interference from factors such as soil particle size and water content, as well as insufficient coverage.
A multi-source fusion hyperspectral soil nutrient remote sensing inversion system was adopted, including a satellite remote sensing platform, a near-ground airborne platform, and ground sensing equipment. Sensitive bands were selected through the CARS-IRIV algorithm, and data source weights were dynamically allocated by combining an attention model. Partial least squares regression, random forest, and CatBoost multi-objective regression models were used to generate multi-dimensional fusion feature vectors. The CNN-Transformer deep learning model was integrated to generate spatial distribution maps of soil nutrients and fertilizer prescription maps.
It improves the accuracy of soil nutrient inversion in complex terrain, increases coverage by more than 40%, improves the accuracy of organic carbon inversion in saline-alkali land and humid areas by 20%-30%, significantly enhances noise resistance, increases computational efficiency by 50%, shortens inversion time by 33%, and improves model robustness and adaptability.
Smart Images
Figure CN121994722A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of agricultural remote sensing and soil science technology, specifically to a multi-source fusion hyperspectral soil nutrient remote sensing inversion system. Background Technology
[0002] In soil nutrient monitoring, hyperspectral remote sensing can capture characteristic absorption peaks of nutrients such as organic matter, nitrogen, phosphorus, and potassium in the soil by acquiring hundreds to thousands of continuous narrow-band spectral data. For example, organic matter exhibits low reflectivity in the visible light band due to carbon bond vibrations, while nitrogen shows significant absorption characteristics in the near-infrared band. Traditional chemical detection requires destructive sampling and is costly, while hyperspectral technology, by constructing spectral-nutrient relationship models, enables large-scale, real-time monitoring, significantly improving the efficiency of agricultural resource management.
[0003] Traditional soil nutrient detection methods rely on laboratory chemical analysis, which suffers from low efficiency, high cost, and insufficient spatial coverage. Meanwhile, single-mode hyperspectral remote sensing inversion is significantly affected by factors such as soil particle size and moisture content, and faces bottlenecks such as data redundancy and poor environmental adaptability. Existing single-mode hyperspectral remote sensing methods for soil nutrient detection ignore the coupling relationships between nutrients, resulting in poor inversion accuracy in complex scenarios such as mining area reclamation and saline-alkali land. Summary of the Invention
[0004] The main objective of this invention is to provide a multi-source fusion hyperspectral soil nutrient remote sensing inversion system, which aims to solve the problem that existing single hyperspectral remote sensing methods for detecting soil nutrients ignore the coupling relationship between nutrients and have poor inversion accuracy in complex scenarios such as mining area reclamation and saline-alkali land.
[0005] The technical solution proposed in this invention is as follows: A multi-source fusion hyperspectral soil nutrient remote sensing inversion system includes: Data acquisition module: including satellite remote sensing platform, near-ground airborne platform, and ground sensing equipment, used to acquire hyperspectral data, multispectral data, ground sampling data, and meteorological data; Preprocessing module: used to perform SG smoothing algorithm, multivariate scattering correction and first-order differential transformation on hyperspectral data, realize spatiotemporal registration of multi-source data through geographic information system, and interpolate first-resolution satellite data into second-resolution raster data; Feature fusion module: used to extract preset quantity sensitive bands that are strongly correlated with soil organic carbon and total nitrogen using the CARS-IRIV algorithm, and dynamically allocate data source weights using an attention model to generate multi-dimensional fused feature vectors; Inversion Model Module: Used to integrate partial least squares regression, random forest basic models, CatBoost multi-objective regression model, and CNN-Transformer deep learning model; Decision output module: used to verify the inversion accuracy through independent sample testing, generate a first-resolution spatial distribution map of soil nutrients and a variable fertilization prescription map based on the geographic information system, and display them visually.
[0006] Preferably, the CARS-IRIV algorithm used in the feature fusion module includes: an iterative screening process based on a genetic algorithm, generating a correlation heatmap by calculating the Pearson correlation coefficient between spectral bands and nutrient content, screening sensitive bands whose absolute values of correlation coefficients with soil organic carbon and total nitrogen are greater than or equal to preset values, marking the screened sensitive bands as a subset of spectral features, and using the subset of spectral features as the core input parameter for feature fusion.
[0007] Preferably, the inversion model module is also used to: calculate the contribution of hyperspectral, topographic, and meteorological data to the inversion results based on the attention model through cosine similarity, generate a dynamic weight matrix, prioritize the input weight of hyperspectral features, and adaptively adjust the weight of auxiliary data according to soil type and topographic conditions.
[0008] Preferably, the CatBoost multi-objective regression model includes: a parallel optimization mechanism with dual loss functions to simultaneously minimize the root mean square error of soil organic carbon and total nitrogen predictions; the CatBoost multi-objective regression model has a built-in category feature processing unit to automatically encode non-numerical variables such as soil texture and land use type.
[0009] Preferably, the CNN-Transformer deep learning model includes: a convolutional layer for extracting local features from spectral data; a self-attention layer for modeling long-distance dependencies between different bands; and an enhanced feature sampling module for reducing redundant information through sliding window dimensionality reduction.
[0010] Preferably, the decision output module includes: a nutrient distribution map rendering unit, used to overlay topographic contour lines and land use type base maps; and a variable fertilization prescription map generation unit, used to generate suggested fertilization data per acre by combining the crop's fertilizer requirement model.
[0011] Preferably, the satellite remote sensing platform includes Gaofen-5, Zhuhai-1, and Sentinel-2 satellites; Gaofen-5 and Zhuhai-1 satellites are used to collect hyperspectral data, covering soil organic matter and total nitrogen sensitive bands; Sentinel-2 satellite is used to simultaneously collect multispectral data, which is used to provide vegetation index and terrain roughness auxiliary features. The near-ground airborne platform includes a drone equipped with a portable spectrometer; the drone is used to: perform high-resolution spectral scanning of field-level areas to supplement satellite data in observation blind spots in complex terrain; The ground sensing equipment includes a ground spectrometer, a global navigation satellite receiving terminal, and a meteorological sensor; the ground spectrometer is used to collect laboratory spectra of soil and simultaneously measure the true values of soil organic matter and total nitrogen, while the global navigation satellite receiving terminal acquires the coordinates of the sampling points; the meteorological sensor is used to collect temperature, humidity, and light intensity in real time.
[0012] Preferably, the preprocessing module is used to remove high-frequency noise from hyperspectral data using the SG smoothing algorithm, eliminate baseline drift caused by soil particle size differences by combining the multivariate scattering correction algorithm, and enhance the characteristics of weak absorption peaks through first-order differential transformation. It is also used for geocoding based on geographic information systems, unifying hyperspectral data, multispectral data, ground sampling data, and meteorological data into raster data of second resolution through kriging interpolation, thereby generating aligned multi-source datasets.
[0013] Preferably, the multidimensional fusion feature vector includes: spectral reflectance, texture features, and meteorological parameters, and the multidimensional fusion feature is used to construct a multidimensional input space; The integrated partial least squares regression algorithm is used for: preliminary modeling of linear relationships and screening of principal components; the random forest basic model is used for: handling nonlinear relationships and verifying the effectiveness of spectral feature subsets by ranking features by importance. The CatBoost multi-objective regression model is used to: simultaneously optimize using dual loss functions, improve the accuracy of collaborative inversion by utilizing the coupling relationship between soil organic carbon and total nitrogen, and automatically encode soil texture and land use type during the iteration process.
[0014] Preferably, the convolutional layer is used to: extract local spectral features; the self-attention layer is used to: model cross-band dependencies and capture long-distance spectral-nutrient correlations; and the enhanced feature sampling module is used to: reduce redundant information by using a sliding window for dimensionality reduction.
[0015] The above technical solution can achieve the following beneficial effects: The multi-source fusion hyperspectral soil nutrient remote sensing inversion system proposed in this invention achieves multi-scale soil nutrient monitoring from the regional level to the field level through a three-layer architecture of satellite remote sensing platform (remote sensing satellite), near-ground airborne platform (UAV), and ground sensing equipment. It solves the problem of blind spots in complex terrain observation of single satellite data, improving coverage by more than 40%. At the same time, it collects meteorological data and topographic data simultaneously. Through a dynamic weighting mechanism, it automatically weakens the interference of environmental factors such as water and light on spectral signals, improving the soil organic carbon inversion accuracy in saline-alkali land and humid areas by 20%-30% compared with single hyperspectral data, greatly improving the inversion accuracy and significantly enhancing the noise resistance. Attached Figure Description
[0016] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of an embodiment of a multi-source fusion hyperspectral soil nutrient remote sensing inversion system proposed in this invention. Detailed Implementation
[0018] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.
[0019] This invention proposes a multi-source fusion hyperspectral soil nutrient remote sensing inversion system.
[0020] As attached Figure 1 As shown, in one embodiment of the multi-source fusion hyperspectral soil nutrient remote sensing inversion system proposed in this invention, the multi-source fusion hyperspectral soil nutrient remote sensing inversion system includes: Data acquisition module: including satellite remote sensing platform, near-ground airborne platform, and ground sensing equipment, used to acquire hyperspectral data (e.g., 400-2500nm), multispectral data (acquired by Sentinel-2 satellite), ground sampling data, and meteorological data.
[0021] Preprocessing module: Used to perform SG smoothing algorithm, multiplicative scatter correction (MSC) and first-order differential transformation on hyperspectral data, realize spatiotemporal registration of multi-source data through geographic information system (GIS), and interpolate satellite data at first resolution (e.g. 30m) into raster data at second resolution (e.g. 10m).
[0022] Specifically, the SG smoothing algorithm is a digital signal processing method based on local polynomial fitting. Its core function is to effectively filter out noise and smooth data while preserving the signal trend. It is especially suitable for processing sequences with regular fluctuations but containing noise.
[0023] Feature fusion module: Used to extract a preset number (e.g., 20) of sensitive bands that are strongly correlated with soil organic carbon (SOC) and total nitrogen (TN) using the CARS-IRIV algorithm, and dynamically allocate data source weights using an attention model to generate a multi-dimensional fused feature vector (e.g., 106×98).
[0024] Inversion model module: It is used to integrate the basic models of Partial Least Squares Regression (PLSR), Random Forest (RF), CatBoost multi-objective regression model, and CNN-Transformer deep learning model.
[0025] Specifically, ensemble partial least squares regression is a combination of partial least squares regression and ensemble learning. The core idea is to build multiple different PLSR models and integrate their prediction results through voting, averaging, and other methods to obtain a more robust regression model with stronger generalization ability.
[0026] Specifically, the CNN-Transformer deep learning model is a hybrid deep learning architecture that combines the local feature extraction capability of convolutional neural networks (CNNs) with the global dependency modeling capability of Transformers. Its core advantage is that it retains the efficient capture of local spatial / temporal features of CNNs while modeling long-distance dependencies through the self-attention mechanism of Transformers. It is particularly suitable for processing spatiotemporal sequence data (such as GNSS positioning data, remote sensing images, speech signals, text + image multimodal data, etc.).
[0027] Decision output module: used to verify the inversion accuracy through independent sample testing, generate a first-resolution (30m) spatial distribution map of soil nutrients and a variable fertilization prescription map based on the geographic information system, and display them visually, such as supporting web and mobile visualization.
[0028] The multi-source fusion hyperspectral soil nutrient remote sensing inversion system proposed in this invention achieves multi-scale soil nutrient monitoring from the regional level to the field level through a three-layer architecture of satellite remote sensing platform (remote sensing satellite), near-ground airborne platform (UAV), and ground sensing equipment. It solves the problem of blind spots in complex terrain observation of single satellite data, improving coverage by more than 40%. At the same time, it collects meteorological data and topographic data simultaneously. Through a dynamic weighting mechanism, it automatically weakens the interference of environmental factors such as water and light on spectral signals, improving the soil organic carbon inversion accuracy in saline-alkali land and humid areas by 20%-30% compared with single hyperspectral data, greatly improving the inversion accuracy and significantly enhancing the noise resistance.
[0029] In this scheme, the CARS-IRIV algorithm extracts 20 core bands strongly correlated with SOC and TN through iterative screening using genetic algorithms and correlation heatmap analysis, removes more than 80% of redundant bands, reduces the feature dimension from 200+ dimensions to below 50 dimensions, improves data processing efficiency by 50%, avoids the negative impact of the "curse of dimensionality" on the model, and based on the dynamic weight allocation mechanism of the attention model, it perceives the contribution of different data sources in real time. For example, in cloudy weather, it automatically reduces the weight of satellite spectral data and increases the weight of UAV data, which improves the inversion stability under complex meteorological conditions by 25% and significantly enhances the robustness of the model.
[0030] In this scheme, the CatBoost model uses a dual loss function to simultaneously optimize the prediction accuracy of SOC and TN, and utilizes their strong correlation to reduce the information loss of the single-objective model. In the reclaimed soil of the mining area, the inversion determination coefficient R² of TN increased from 0.80 to 0.95, and the root mean square error of prediction decreased from 0.12 g / kg to 0.07 g / kg, with an accuracy improvement of 18.7%. The CNN-Transformer architecture captures local spectral features through convolutional layers and models long-distance dependencies across bands through self-attention layers, improving the adaptability to different types of soil such as black soil and red soil by 30%. In particular, in the differentiation of nitrogen and phosphorus nutrients with similar spectral features, the recognition accuracy is improved by 22% compared with the traditional model. The enhanced feature sampling module (RFS) reduces the dimensionality of 1000 satellite images from 48 hours to 32 hours through sliding window dimensionality reduction, improving the computational efficiency by 33%, and supporting real-time processing and rapid inversion of large-scale regional data.
[0031] Furthermore, the CARS-IRIV algorithm used in the feature fusion module includes: an iterative screening process based on a genetic algorithm, which generates a correlation heatmap by calculating the Pearson correlation coefficient between spectral bands and nutrient content, screening sensitive bands whose absolute values of correlation coefficients with soil organic carbon and total nitrogen are greater than or equal to preset values (e.g., 0.7), marking the screened sensitive bands as a subset of spectral features, and using the subset of spectral features as the core input parameter for feature fusion.
[0032] Meanwhile, the aforementioned inversion model module is also used to: calculate the contribution of hyperspectral, topographic, and meteorological data to the inversion results based on the attention model and cosine similarity, generate a dynamic weight matrix, prioritize the input weight of hyperspectral features, and adaptively adjust the weight of auxiliary data according to soil type and topographic conditions.
[0033] The inversion model module is also used to: adaptively adjust the weight of auxiliary data according to soil type and topographic conditions. Hyperspectral data, which directly reflects the vibrational information of material molecules, has a default weight of ≥60%. Topographic data and meteorological data are dynamically adjusted through cosine similarity. For example, the humidity weight in humid areas is increased to 20% to suppress water interference.
[0034] Specifically, the CatBoost multi-objective regression model includes: a parallel optimization mechanism with dual loss functions to simultaneously minimize the root mean square error (RMSE) of soil organic carbon and total nitrogen predictions; the CatBoost multi-objective regression model has a built-in category feature processing unit that automatically encodes non-numerical variables such as soil texture and land use type, thereby improving the model's generalization ability in complex scenarios.
[0035] In addition, the CNN-Transformer deep learning model includes: convolutional layers for extracting local features from spectral data; self-attention layers for modeling long-distance dependencies between different bands; and enhanced feature sampling module (RFS) for reducing redundant information through sliding window dimensionality reduction, which improves computational efficiency by more than 30% compared to traditional models.
[0036] Meanwhile, the decision output module includes: a nutrient distribution map rendering unit, which uses the GB / T41475-2022 standard color scale to overlay topographic contour lines and land use type base maps; and a variable fertilization prescription map generation unit, which combines the fertilizer requirement model of crops (such as corn / wheat crops) to generate suggested data on fertilizer application per acre, and supports integration with the ISO11783 agricultural machinery data interface.
[0037] Specifically, the satellite remote sensing platform includes Gaofen-5, Zhuhai-1, and Sentinel-2 satellites; Gaofen-5 and Zhuhai-1 satellites are used to collect hyperspectral data, covering sensitive bands for soil organic matter (SOC) and total nitrogen (TN); Sentinel-2 satellite is used to simultaneously collect multispectral data, which is used to provide vegetation index and terrain roughness auxiliary features.
[0038] The near-ground airborne platform includes a drone equipped with a portable spectrometer; the drone is used to: perform high-resolution spectral scanning of field-level areas to supplement satellite data in observation blind spots in complex terrain; The ground sensing equipment includes a ground spectrometer, a global navigation satellite receiver terminal, and a meteorological sensor. The ground spectrometer is used to collect laboratory spectra of soil and simultaneously determine the true values of soil organic carbon (potassium dichromate oxidation method) and total nitrogen (Kjeldahl nitrogen determination method). The global navigation satellite receiver terminal (GNSS receiver) acquires the coordinates of the sampling points. The meteorological sensor is used to collect temperature, humidity, and light intensity in real time for subsequent data correction and model input.
[0039] Meanwhile, the preprocessing module is used to remove high-frequency noise from hyperspectral data using the SG smoothing algorithm, eliminate baseline drift caused by soil particle size differences by combining the multivariate scattering correction algorithm, and enhance the characteristics of weak absorption peaks through first-order differential transformation.
[0040] The preprocessing module is also used for geocoding based on the geographic information system, unifying hyperspectral data, multispectral data, ground sampling data, and meteorological data into raster data of the second resolution through Kriging interpolation, thereby generating an aligned multi-source dataset.
[0041] In addition, the multidimensional fusion feature vector includes spectral reflectance, texture features, and meteorological parameters, and the multidimensional fusion features are used to construct a multidimensional input space.
[0042] The integrated partial least squares regression algorithm is used for: preliminary modeling of linear relationships and screening of principal components; the random forest basic model is used for: handling nonlinear relationships and verifying the effectiveness of spectral feature subsets by ranking features by importance.
[0043] The CatBoost multi-objective regression model is used to: simultaneously optimize using dual loss functions, improve the accuracy of collaborative inversion by utilizing the coupling relationship between soil organic carbon and total nitrogen, automatically encode soil texture and land use type during the iteration process to avoid artificial feature engineering, and improve the TN inversion determination coefficient (R²) by 15% compared with the single-objective model in reclaimed soils in mining areas.
[0044] Meanwhile, the convolutional layer is used to extract local spectral features; the self-attention layer is used to model cross-band dependencies and capture long-distance spectral-nutrient correlations ignored by traditional models; and the enhanced feature sampling module is used to reduce redundant information by using a sliding window to reduce the training time of 1000 images from 48 hours to 32 hours.
[0045] In addition, this scheme uses independent sample testing, and the verification indicators must meet the following requirements: R² ≥ 0.96 and RMSE ≤ 0.56 g / kg for soil organic carbon; R² ≥ 0.95 and RMSE ≤ 0.07 g / kg for total nitrogen; if the accuracy decreases in special scenarios such as mining areas, regional parameter adjustments will be automatically triggered.
[0046] Nutrient spatial distribution map is based on GIS rendering. Soil organic carbon is rendered in warm colors, and adjacent color scales have a ΔE ≥ 5 to ensure differentiation. Topographic contour lines and land use vector maps are overlaid. Combined with crop fertilizer requirement models, zoning suggestions are made based on the total nitrogen inversion results: reduce nitrogen fertilizer by 10%-20% in high-content areas and increase nitrogen fertilizer by 5%-10% in low-content areas. A fertilizer instruction file (variable fertilization treatment map) conforming to ISO11783 standard is generated to directly drive the operation of variable fertilizer applicators.
[0047] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0048] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A multi-source fusion hyperspectral soil nutrient remote sensing inversion system, characterized in that, include: Data acquisition module: including satellite remote sensing platform, near-ground airborne platform, and ground sensing equipment, used to acquire hyperspectral data, multispectral data, ground sampling data, and meteorological data; Preprocessing module: used to perform SG smoothing algorithm, multivariate scattering correction and first-order differential transformation on hyperspectral data, realize spatiotemporal registration of multi-source data through geographic information system, and interpolate first-resolution satellite data into second-resolution raster data; Feature fusion module: used to extract preset quantity sensitive bands that are strongly correlated with soil organic carbon and total nitrogen using the CARS-IRIV algorithm, and dynamically allocate data source weights using an attention model to generate a multi-dimensional fused feature vector; Inversion Model Module: Used to integrate partial least squares regression, random forest basic models, CatBoost multi-objective regression model, and CNN-Transformer deep learning model; Decision output module: used to verify the inversion accuracy through independent sample testing, generate a first-resolution spatial distribution map of soil nutrients and a variable fertilization prescription map based on the geographic information system, and display them visually.
2. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1, characterized in that, The CARS-IRIV algorithm used in the feature fusion module includes: an iterative screening process based on a genetic algorithm, which generates a correlation heatmap by calculating the Pearson correlation coefficient between spectral bands and nutrient content, screening sensitive bands whose absolute values of correlation coefficients with soil organic carbon and total nitrogen are greater than or equal to preset values, marking the screened sensitive bands as a subset of spectral features, and using the subset of spectral features as the core input parameter for feature fusion.
3. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1, characterized in that, The inversion model module is also used to: calculate the contribution of hyperspectral, topographic, and meteorological data to the inversion results based on the attention model and cosine similarity, generate a dynamic weight matrix, prioritize the input weight of hyperspectral features, and adaptively adjust the weight of auxiliary data according to soil type and topographic conditions.
4. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1, characterized in that, The CatBoost multi-objective regression model includes: a parallel optimization mechanism with dual loss functions to simultaneously minimize the root mean square error of soil organic carbon and total nitrogen predictions; the CatBoost multi-objective regression model has a built-in category feature processing unit to automatically encode non-numerical variables such as soil texture and land use type.
5. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1, characterized in that, The CNN-Transformer deep learning model includes: a convolutional layer for extracting local features from spectral data; a self-attention layer for modeling long-distance dependencies between different bands; and an enhanced feature sampling module for reducing redundant information through sliding window dimensionality reduction.
6. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1, characterized in that, The decision output module includes: a nutrient distribution map rendering unit, used to overlay topographic contour lines and land use type base maps; and a variable fertilization prescription map generation unit, used to combine crop fertilizer requirement models to generate suggested fertilization data per acre.
7. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1, characterized in that, The satellite remote sensing platform includes Gaofen-5, Zhuhai-1, and Sentinel-2 satellites; Gaofen-5 and Zhuhai-1 satellites are used to collect hyperspectral data, covering soil organic matter and total nitrogen sensitive bands; Sentinel-2 satellite is used to simultaneously collect multispectral data, which is used to provide vegetation index and terrain roughness auxiliary features. The near-ground airborne platform includes a drone equipped with a portable spectrometer; the drone is used to: perform high-resolution spectral scanning of field-level areas to supplement satellite data in observation blind spots in complex terrain; The ground sensing equipment includes a ground spectrometer, a global navigation satellite receiving terminal, and a meteorological sensor; the ground spectrometer is used to collect laboratory spectra of soil and simultaneously measure the true values of soil organic matter and total nitrogen, while the global navigation satellite receiving terminal acquires the coordinates of the sampling points; the meteorological sensor is used to collect temperature, humidity, and light intensity in real time.
8. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 7, characterized in that, The preprocessing module is used to remove high-frequency noise from hyperspectral data using the SG smoothing algorithm, eliminate baseline drift caused by soil particle size differences by combining the multivariate scattering correction algorithm, and enhance the characteristics of weak absorption peaks through first-order differential transformation. It is also used for geocoding based on geographic information systems, unifying hyperspectral data, multispectral data, ground sampling data, and meteorological data into raster data of second resolution through kriging interpolation, thereby generating aligned multi-source datasets.
9. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 2, characterized in that, The multidimensional fusion feature vector includes: spectral reflectance, texture features, and meteorological parameters. The multidimensional fusion feature is used to construct a multidimensional input space. The integrated partial least squares regression algorithm is used for: preliminary modeling of linear relationships and screening of principal components; the random forest basic model is used for: handling nonlinear relationships and verifying the effectiveness of spectral feature subsets by ranking features by importance. The CatBoost multi-objective regression model is used to: simultaneously optimize using dual loss functions, improve the accuracy of collaborative inversion by utilizing the coupling relationship between soil organic carbon and total nitrogen, and automatically encode soil texture and land use type during the iteration process.
10. A multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 5, characterized in that, The convolutional layer is used to extract local spectral features; the self-attention layer is used to model cross-band dependencies and capture long-distance spectral-nutrient correlations; and the enhanced feature sampling module is used to reduce redundant information by using a sliding window to reduce dimensionality.