CN-121994722-A - Multisource fuses hyperspectral soil nutrient remote sensing inversion system

CN121994722ACN 121994722 ACN121994722 ACN 121994722ACN-121994722-A

Abstract

The invention relates to the technical field of agricultural remote sensing and soil science, in particular to a multi-source fusion hyperspectral soil nutrient remote sensing inversion system, which realizes multi-scale soil nutrient monitoring from regional level to field level through a three-layer framework of a satellite remote sensing platform (remote sensing satellite), a near-ground space-based platform (unmanned aerial vehicle) and ground sensing equipment, solves the problem of observation blind areas of single satellite data in complex terrains, improves the coverage by more than 40%, synchronously acquires meteorological data and topographic data, automatically weakens the interference of environmental factors such as moisture, illumination and the like on the optical signals through a dynamic weighting mechanism, improves the inversion precision of organic carbon in soil in saline-alkali soil and wet areas by 20% -30% compared with single hyperspectral data, and greatly improves the anti-noise capability.

Inventors

ZOU PU
WANG KAIHUA
MO JIAWEI
HUANG CHAOYU
TANG YUANYUAN
CHEN FANGFANG
LV YUFENG
TANG YUKAI

Assignees

中国地质调查局长沙自然资源综合调查中心

Dates

Publication Date: 20260508
Application Date: 20260316

Claims (10)

1. A multi-source fusion hyperspectral soil nutrient remote sensing inversion system is characterized by comprising: the data acquisition module comprises a satellite remote sensing platform, a near-ground air-based platform and ground sensing equipment, and is used for acquiring hyperspectral data, multispectral data, ground sampling data and meteorological data; the preprocessing module is used for carrying out SG smoothing algorithm, multi-element scattering correction and first-order differential transformation on the hyperspectral data, realizing multi-source data space-time registration through a geographic information system and converting the first-resolution satellite data interpolation into second-resolution raster data; The feature fusion module is used for extracting preset quantity of sensitive wave bands related to soil organic carbon and total nitrogen intensity through a CARS-IRIV algorithm, dynamically distributing data source weights in combination with an attention model, and generating a multidimensional fusion feature vector; The inversion model module is used for integrating a partial least square regression model, a random forest basic model, a CatBoost multi-objective regression model and a CNN-transducer deep learning model; And the decision output module is used for verifying inversion accuracy through independent sample testing, generating a soil nutrient space distribution map with a first resolution and a variable fertilization prescription map based on the geographic information system, and carrying out visual display.
2. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1 is characterized in that the CARS-IRIV algorithm adopted by the characteristic fusion module comprises an iterative screening process based on a genetic algorithm, wherein a correlation thermodynamic diagram is generated by calculating pearson correlation coefficients of a spectrum band and nutrient content, sensitive bands with absolute values of the correlation coefficients of organic carbon and total nitrogen of soil being larger than or equal to a preset value are screened, the screened sensitive bands are marked as spectrum characteristic subsets, and the spectrum characteristic subsets are used as core input parameters of characteristic fusion.
3. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system of claim 1, wherein the inversion model module is further used for calculating contribution degree of hyperspectral, topographic and meteorological data to inversion results through cosine similarity based on the attention model, generating a dynamic weight matrix, preferentially reinforcing input weights of hyperspectral features, and adaptively adjusting auxiliary data weights according to soil types and topographic conditions.
4. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1 is characterized in that the CatBoost multi-target regression model comprises a double-loss function parallel optimization mechanism, a prediction root mean square error of soil organic carbon and total nitrogen is synchronously minimized, and a CatBoost multi-target regression model built-in type characteristic processing unit is used for automatically coding non-numerical variables of soil texture and land utilization type.
5. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1 is characterized in that the CNN-converter deep learning model comprises a convolution layer for extracting local features of spectrum data, a self-attention layer for modeling long-distance dependency relations among different wave bands, and an enhanced feature sampling module for reducing dimensions and redundant information through a sliding window.
6. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system is characterized in that the decision output module comprises a nutrient distribution map rendering unit and a variable fertilization prescription map generating unit, wherein the nutrient distribution map rendering unit is used for superposing a contour line of a terrain and a land utilization type base map, and the variable fertilization prescription map generating unit is used for generating fertilization amount proposal data per mu in combination with a fertilizer requirement model of crops.
7. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 1, wherein the satellite remote sensing platform comprises a high-score No. five satellite, a high-score No. one satellite, and a Sentinel-2 satellite, wherein the high-score No. five satellite and the high-score No. one satellite are used for collecting hyperspectral data and covering soil organic matters and a total nitrogen sensitive wave band, and the Sentinel-2 satellite is used for synchronously collecting multispectral data, wherein the multispectral data is used for providing vegetation indexes and topography roughness auxiliary features; The near-ground space-based platform comprises an unmanned aerial vehicle carrying a portable spectrometer, wherein the unmanned aerial vehicle is used for carrying out high-resolution spectrum scanning on a field-level region and supplementing satellite data in an observation blind area of complex terrain; The ground sensing equipment comprises a ground spectrometer, a global navigation satellite receiving terminal and a meteorological sensor, wherein the ground spectrometer is used for collecting a spectrum of a soil laboratory, synchronously measuring true value data of soil organic matters and total nitrogen, the global navigation satellite receiving terminal is used for obtaining coordinates of sampling points, and the meteorological sensor is used for collecting temperature, humidity and illumination intensity in real time.
8. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 7, wherein the preprocessing module is used for removing high-frequency noise of hyperspectral data by adopting an SG smoothing algorithm, eliminating baseline drift caused by soil granularity difference by combining a multi-element scattering correction algorithm, and enhancing weak absorption peak characteristics by first-order differential transformation; and the system is also used for unifying hyperspectral data, multispectral data, ground sampling data and meteorological data to grid data with a second resolution through a Kriging interpolation method based on the geocoding of the geographic information system, so as to generate an aligned multisource data set.
9. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system according to claim 2, wherein the multi-dimensional fusion feature vector comprises spectral reflectivity, texture features and meteorological parameters, and the multi-dimensional fusion features are used for constructing a multi-dimensional input space; The integrated partial least squares regression algorithm is used for initially modeling a linear relation and screening a principal component score, and the random forest basic model is used for processing a nonlinear relation and verifying the effectiveness of a spectrum feature subset through feature importance sequencing; The CatBoost multi-objective regression model is used for synchronously optimizing by adopting a double-loss function, improving the collaborative inversion precision by utilizing the coupling relation of soil organic carbon and total nitrogen, and automatically coding the soil texture and the land utilization type in the iterative process.
10. The multi-source fusion hyperspectral soil nutrient remote sensing inversion system as claimed in claim 5, wherein the convolution layer is used for extracting local spectral features, the self-attention layer is used for modeling cross-band dependence and capturing long-distance spectrum-nutrient correlation, and the enhanced feature sampling module is used for reducing dimension and redundant information through a sliding window.

Description

Multisource fuses hyperspectral soil nutrient remote sensing inversion system Technical Field The invention relates to the technical field of agricultural remote sensing and soil science, in particular to a multi-source fusion hyperspectral soil nutrient remote sensing inversion system. Background In the soil nutrient detection scheme, hyperspectral remote sensing can capture characteristic absorption peaks of nutrients such as organic matters, nitrogen, phosphorus, potassium and the like in soil by acquiring hundreds to thousands of continuous narrow-band spectral data. For example, organic matter exhibits low reflectivity due to carbon bond vibration in the visible light band, while nitrogen has a distinct absorption characteristic in the near infrared band. The traditional chemical detection needs destructive sampling and has high cost, and the hyperspectral technology can realize large-scale and real-time monitoring by constructing a spectrum-nutrient relation model, thereby remarkably improving the management efficiency of agricultural resources. The traditional soil nutrient detection scheme relies on laboratory chemical analysis, has the problems of low efficiency, high cost, insufficient space coverage and the like, and single hyperspectral remote sensing inversion is obviously interfered by factors such as soil granularity, water content and the like, and faces the bottlenecks such as data redundancy, poor environmental adaptability and the like. The existing single hyperspectral remote sensing soil nutrient detection scheme ignores the coupling relation among nutrients, and has poor inversion precision under complex scenes such as mining area reclamation, saline-alkali soil and the like. Disclosure of Invention The invention mainly aims to provide a multi-source fusion hyperspectral soil nutrient remote sensing inversion system, which aims to solve the problems that the existing single hyperspectral remote sensing soil nutrient detection scheme ignores the coupling relation between nutrients and has poor inversion precision in complex scenes such as mining area reclamation, saline-alkali soil and the like. The technical scheme provided by the invention is as follows: A multi-source fused hyperspectral soil nutrient remote sensing inversion system, comprising: the data acquisition module comprises a satellite remote sensing platform, a near-ground air-based platform and ground sensing equipment, and is used for acquiring hyperspectral data, multispectral data, ground sampling data and meteorological data; the preprocessing module is used for carrying out SG smoothing algorithm, multi-element scattering correction and first-order differential transformation on the hyperspectral data, realizing multi-source data space-time registration through a geographic information system and converting the first-resolution satellite data interpolation into second-resolution raster data; The feature fusion module is used for extracting preset quantity of sensitive wave bands related to soil organic carbon and total nitrogen intensity through a CARS-IRIV algorithm, dynamically distributing data source weights in combination with an attention model, and generating a multidimensional fusion feature vector; The inversion model module is used for integrating a partial least square regression model, a random forest basic model, a CatBoost multi-objective regression model and a CNN-transducer deep learning model; And the decision output module is used for verifying inversion accuracy through independent sample testing, generating a soil nutrient space distribution map with a first resolution and a variable fertilization prescription map based on the geographic information system, and carrying out visual display. Preferably, the CARS-IRIV algorithm adopted by the feature fusion module comprises an iterative screening process based on a genetic algorithm, wherein a correlation thermodynamic diagram is generated by calculating pearson correlation coefficients of a spectrum band and nutrient content, sensitive bands with absolute values of the correlation coefficients of organic carbon and total nitrogen of soil being larger than or equal to a preset value are screened, the screened sensitive bands are marked as spectrum feature subsets, and the spectrum feature subsets are used as core input parameters of feature fusion. Preferably, the inversion model module is further used for calculating contribution degrees of hyperspectral, topographic and meteorological data to inversion results based on the attention model, generating a dynamic weight matrix, preferentially strengthening input weights of hyperspectral features, and adaptively adjusting auxiliary data weights according to soil types and topographic conditions. Preferably, the CatBoost multi-objective regression model comprises a double-loss function parallel optimization mechanism for synchronously minimizing the prediction root mean square error of soil organic carbo