Machine learning-based surface matrix parameter hyperspectral data inversion method and system

Abstract

The invention relates to the technical field of remote sensing data processing and earth surface parameter inversion, and discloses an earth surface matrix parameter hyperspectral data inversion method and system based on machine learning. Comprising the following steps: constructing a multi-source heterogeneous hyperspectral data set; performing feature screening on the preprocessed hyperspectral data set based on an adaptive band selection algorithm, constructing a dynamic weight matrix by calculating mutual information entropy and inter-class distance measurement between spectral bands to realize intelligent screening of key feature bands, and combining spectral derivative conversion and spectral index calculation to generate an enhanced feature vector; and a multi-task transfer learning neural network model is constructed, and an output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture. And performing preprocessing and feature enhancement operation which is the same as that of the training data on the hyperspectral image data of the to-be-inverted region, inputting the trained neural network model, and outputting a surface matrix parameter inversion result.

Description

Technical Field

[0001] This application relates to the fields of remote sensing data processing and surface parameter inversion technology, and in particular to a method and system for inverting hyperspectral data of surface matrix parameters based on machine learning. Background Technology

[0002] Surface matrix parameters (such as soil organic matter content, mineral component concentration, and water content) are core data for the material composition and environmental evolution of the Earth's surface system. Hyperspectral remote sensing, with its advantages of continuous bands and high spectral resolution, has become an important technical means for rapidly acquiring such parameters. However, existing inversion methods suffer from the following technical bottlenecks: 1. Traditional empirical models rely on prior assumptions: Methods based on physical models or statistical regression (such as multiple linear regression and partial least squares regression) usually assume that there is a linear relationship between spectral characteristics and parameters, which is difficult to adapt to the nonlinear spectral response of complex surface matrices and is sensitive to noise.

[0003] 2. Fixed band selection method is inefficient: Traditional band selection is mostly based on correlation analysis or expert experience, ignoring the higher-order correlation and spatial context information between spectral bands, resulting in feature redundancy or loss of key information, and limiting the inversion accuracy.

[0004] 3. Isolated processing of multi-parameter inversion: Existing methods mostly model single parameters independently, without utilizing the inherent correlation between parameters. This leads to problems such as repeated model training and resource waste, and makes it difficult to capture the optimization potential of multi-parameter collaborative inversion.

[0005] 4. High cost of sample labeling and insufficient generalization ability: It is time-consuming and laborious to obtain high-quality experimental samples. When traditional models are applied across regions, the generalization performance decreases due to environmental differences (such as atmospheric conditions and terrain undulations). It is urgent to introduce techniques such as transfer learning to improve the adaptability of the model. Summary of the Invention

[0006] This application provides a method and system for inverting hyperspectral data of surface matrix parameters based on machine learning, aiming to overcome the problems of insufficient spatiotemporal coverage and poor real-time performance in traditional water quality monitoring that relies on manual sampling and laboratory analysis; although monitoring networks based on single-point sensors can achieve real-time data acquisition, they are limited by the monitoring range and cannot meet the dynamic assessment needs of large-scale water areas.

[0007] In a first aspect, this application provides a machine learning-based method for inverting hyperspectral data of land surface matrix parameters, including: A multi-source heterogeneous hyperspectral dataset is constructed, which includes hyperspectral reflectance data of surface matrix samples in the target area and the corresponding measured parameter true values. The hyperspectral dataset is preprocessed with radiometric calibration, atmospheric correction and spectral denoising, and spatial location information and environmental parameters of surface matrix samples are collected. The preprocessed hyperspectral dataset is used for feature selection based on an adaptive band selection algorithm. By calculating the mutual information entropy between spectral bands and the inter-class distance metric, a dynamic weight matrix is ​​constructed to achieve intelligent selection of key feature bands. Enhanced feature vectors are generated by combining spectral derivative transformation and spectral index calculation. A multi-task transfer learning neural network model is constructed, which includes a spectral feature encoding layer, a spatial context fusion layer, and a parameter inversion output layer. The spectral feature encoding layer uses an attention mechanism to enhance the feature representation of key bands, the spatial context fusion layer extracts neighborhood spatial correlation features through a convolutional neural network, and the output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture. Perform the same preprocessing and feature enhancement operations on the hyperspectral image data of the inversion area as on the training data, input the trained neural network model, and output the inversion results of surface matrix parameters, which include at least one parameter among soil organic matter content, mineral component concentration, and water content.

[0008] In some embodiments, before performing the same preprocessing and feature enhancement operations on the hyperspectral image data of the region to be inverted as on the training data, the method further includes: training the neural network model using a hybrid loss function, wherein the hybrid loss function includes a regression loss term based on mean squared error and a feature preservation loss term based on structural similarity; during training, an adversarial example enhancement strategy is introduced to improve the robustness of the model; and the model's generalization ability is optimized by transferring prior knowledge of the same origin region through domain adaptation techniques.

[0009] In some embodiments, the method further includes: embedding a model uncertainty quantification module in the neural network model, calculating the confidence evaluation value of the inversion parameters using a Bayesian neural network or a Monte Carlo method, wherein the confidence evaluation value is based on the variance or entropy value of the predicted distribution and is used to screen low-confidence samples for active learning and iterative optimization of the model.

[0010] In some embodiments, the collection of spatial location information and environmental parameters of surface matrix samples includes: obtaining the geographic coordinates of the surface matrix samples through a GNSS positioning device, and collecting environmental parameters of the sample points. The environmental parameters include surface temperature, air humidity, vegetation coverage, terrain slope and aspect. The vegetation coverage is calculated by combining UAV remote sensing imagery with pixel binarization, and the terrain slope and aspect are obtained based on DEM data through spatial analysis algorithms.

[0011] In some embodiments, the intelligent screening of key feature bands by constructing a dynamic weight matrix through calculating the mutual information entropy and inter-class distance metrics between spectral bands includes: calculating the mutual information entropy between each band and the true value of the surface matrix parameter to quantify the correlation between the band and the target parameter; using Mahalanobis distance to calculate the inter-class distance of different categories of samples on each band to construct a band discrimination capability matrix; fusing mutual information entropy and inter-class distance through an adaptive weight formula; and screening key feature bands according to the dynamic weight matrix.

[0012] In some embodiments, the step of combining spectral derivative transformation and spectral index calculation to generate an enhanced feature vector includes: performing a first- or second-order spectral derivative transformation on the selected key feature bands to eliminate baseline drift and low-frequency noise, thereby obtaining a derivative spectral matrix; calculating a targeted spectral index based on the key bands, wherein the spectral index includes, but is not limited to, the normalized vegetation index, the soil-regulated vegetation index, and the mineral absorption characteristic index; and concatenating the original spectral reflectance, derivative spectral values, and spectral index values ​​into an enhanced feature vector.

[0013] In some embodiments, the spectral feature encoding layer employs a multi-head self-attention mechanism to enhance the feature representation of key bands by calculating the attention weights of features in different bands; the spatial context fusion layer employs a residual neural network structure with dilated convolutions to expand the receptive field and extract multi-scale neighborhood spatial correlation features without reducing spatial resolution; the parameter inversion output layer sets up a multi-task shared feature extraction layer and an independent parameter regression branch, the shared layer employs a fully connected network to fuse spectral and spatial features, and each branch achieves the collaborative inversion of soil organic matter content, mineral component concentration, and water content through adaptive weight allocation.

[0014] In some embodiments, the multiple methods further include: introducing a self-supervised pre-training module between the spectral feature encoding layer and the spatial context fusion layer, using a contrastive learning method to pre-train unlabeled hyperspectral data, and improving the model's general representation ability of spectral-spatial features by maximizing the feature similarity of different augmented views of the same sample, wherein the augmented views include random band masks, spectral noise perturbations, and spatial clipping operations.

[0015] In some embodiments, the hyperspectral image data of the region to be inverted undergoes the same preprocessing and feature enhancement operations as the training data, including: performing radiometric calibration to correct sensor response bias, atmospheric correction based on the MODTRAN model to eliminate the influence of aerosols and water vapor, and spectral denoising using wavelet transform or median filtering; performing band selection based on the key band index determined during the training phase, and generating enhanced feature vectors according to the same derivative transformation method and spectral index calculation formula to ensure that the feature space of the input model is consistent with the training data.

[0016] Secondly, this application provides a machine learning-based hyperspectral data inversion system for surface matrix parameters, the system comprising: The data construction unit is used to construct a multi-source heterogeneous hyperspectral dataset, which includes hyperspectral reflectance data of surface matrix samples in the target area and the corresponding measured parameter true values. The hyperspectral dataset is preprocessed with radiometric calibration, atmospheric correction and spectral denoising, and spatial location information and environmental parameters of surface matrix samples are collected. The feature selection unit is used to perform feature selection on the preprocessed hyperspectral dataset based on the adaptive band selection algorithm. By calculating the mutual information entropy between spectral bands and the inter-class distance metric, a dynamic weight matrix is ​​constructed to achieve intelligent selection of key feature bands. The enhanced feature vector is generated by combining spectral derivative transformation and spectral index calculation. The model building unit is used to construct a multi-task transfer learning neural network model, which includes a spectral feature encoding layer, a spatial context fusion layer, and a parameter inversion output layer. The spectral feature encoding layer uses an attention mechanism to enhance the feature representation of key bands, the spatial context fusion layer extracts neighborhood spatial correlation features through a convolutional neural network, and the output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture.

[0017] The inversion output unit is used to perform the same preprocessing and feature enhancement operations on the hyperspectral image data of the area to be inverted as on the training data. It inputs the trained neural network model and outputs the inversion results of surface matrix parameters, which include at least one parameter among soil organic matter content, mineral component concentration, and water content.

[0018] Thirdly, embodiments of this application provide a computer device, the computer device including a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and, when executing the computer program, implement the method provided in any embodiment of this application.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the method provided in any embodiment of this application.

[0020] This invention employs an adaptive band selection algorithm to screen key feature bands, combining spectral derivative transformation and spectral index to generate enhanced feature vectors. This effectively suppresses noise interference and strengthens the spectral response characteristics of target parameters, resulting in improved inversion accuracy compared to traditional fixed-band methods. A multi-task transfer learning neural network utilizes the correlation between parameters to share feature representations, reducing the cost of repetitive training. Simultaneously, domain adaptation technology transfers prior knowledge from the same source region, eliminating the need for extensive relabeling of samples when the model is applied across regions, thus reducing generalization error. A hybrid loss function combines regression loss and feature preservation loss to avoid overfitting. Adversarial example augmentation strategies enhance the model's resistance to noise and anomalous samples, ensuring the stability of inversion results in complex environments. A standardized workflow is established from data preprocessing and feature augmentation to model training and inversion output, supporting efficient inversion of multiple parameters such as soil organic matter content, mineral component concentration, and water content, significantly improving the engineering application capabilities of remote sensing inversion of surface matrix parameters.

[0021] In summary, this invention, through the deep integration of machine learning and hyperspectral remote sensing technology, systematically solves the shortcomings of traditional inversion methods in terms of feature selection, model generalization, and multi-parameter coordination, and has significant technological advancements and practical application value.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic flowchart illustrating the steps of a machine learning-based hyperspectral data inversion method for surface matrix parameters provided in an embodiment of this application. Figure 2 This is a schematic block diagram of the structure of a machine learning-based hyperspectral data inversion system for land surface matrix parameters provided in one embodiment of this application; Figure 3 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0028] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0029] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0030] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0031] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0032] Surface matrix parameters (such as soil organic matter content, mineral component concentration, and water content) are core data for the material composition and environmental evolution of the Earth's surface system. Hyperspectral remote sensing, with its advantages of continuous bands and high spectral resolution, has become an important technical means for rapidly acquiring such parameters. However, existing inversion methods suffer from the following technical bottlenecks: 1. Traditional empirical models rely on prior assumptions: Methods based on physical models or statistical regression (such as multiple linear regression and partial least squares regression) usually assume that there is a linear relationship between spectral characteristics and parameters, which is difficult to adapt to the nonlinear spectral response of complex surface matrices and is sensitive to noise.

[0033] 2. Fixed band selection method is inefficient: Traditional band selection is mostly based on correlation analysis or expert experience, ignoring the higher-order correlation and spatial context information between spectral bands, resulting in feature redundancy or loss of key information, and limiting the inversion accuracy.

[0034] 3. Isolated processing of multi-parameter inversion: Existing methods mostly model single parameters independently, without utilizing the inherent correlation between parameters. This leads to problems such as repeated model training and resource waste, and makes it difficult to capture the optimization potential of multi-parameter collaborative inversion.

[0035] 4. High cost of sample labeling and insufficient generalization ability: It is time-consuming and laborious to obtain high-quality experimental samples. When traditional models are applied across regions, the generalization performance decreases due to environmental differences (such as atmospheric conditions and terrain undulations). It is urgent to introduce techniques such as transfer learning to improve the adaptability of the model.

[0036] To solve the above problem, please refer to Figure 1 This application provides a machine learning-based method for inverting hyperspectral data of land surface matrix parameters. The method is applied to computer equipment.

[0037] like Figure 1 As shown, the provided method includes steps S101 to S104. The computer device can be a handheld terminal, a laptop computer, a wearable device, or a robot, etc. This is used to implement steps S101 to S104 and their corresponding embodiments.

[0038] Step S101. Construct a multi-source heterogeneous hyperspectral dataset, which includes hyperspectral reflectance data of surface matrix samples in the target area and the corresponding measured parameter true values. Perform radiometric calibration, atmospheric correction and spectral denoising preprocessing on the hyperspectral dataset, and collect spatial location information and environmental parameters of surface matrix samples.

[0039] Specifically, by constructing a multi-source dataset containing hyperspectral reflectance data of surface matrix samples, true values ​​of measured parameters, spatial location information, and environmental parameters, and by improving data quality through preprocessing, a reliable input is provided for subsequent modeling.

[0040] Hyperspectral reflectance data are collected from spectral data of samples in the target area by ground-based spectrometers (such as ASD FieldSpec) or UAV / satellite payloads (such as Hyperion), covering the visible-near-infrared-short-wave infrared band (400-2500nm), with a sampling interval ≤10 nm.

[0041] The true values ​​of the measured parameters are obtained by simultaneously collecting parameters such as soil organic matter content (potassium dichromate oxidation method), mineral component concentration (X-ray fluorescence spectrometer), and moisture content (drying method) to ensure high accuracy of sample labels.

[0042] Spatial location and environmental parameters: The sample coordinates are recorded by GPS (accuracy ≤10 m), and environmental variables such as terrain parameters (DEM data), meteorological data (temperature, humidity), and vegetation coverage (estimated by UAV RGB imagery) are collected simultaneously for subsequent model calibration.

[0043] Radiometric calibration involves converting the raw DN value to absolute reflectance and correcting it using the instrument's built-in calibration coefficients or a reference whiteboard. The formula is: ρ(λ) = (L(λ) * π * d 2 L(λ) / (E(λ)*cosθ); where L(λ) is the instrument observation radiance, E(λ) is the incident irradiance, d is the Earth-Sun distance, and θ is the solar zenith angle.

[0044] Atmospheric correction uses physical models (such as FLAASH and ACOLITE) to remove the effects of atmospheric scattering and absorption, and restore the true reflectance of the surface. Input parameters include aerosol thickness and water vapor content (based on MODIS / AIRS remote sensing inversion or ground measurement).

[0045] For high-frequency noise, spectral denoising uses Savitzky-Golay filtering (window size 5-11, polynomial order 2-3) or wavelet transform for smoothing, while preserving the morphological characteristics of the spectral curve.

[0046] Step S102. Based on the adaptive band selection algorithm, feature filtering is performed on the preprocessed hyperspectral dataset. By calculating the mutual information entropy between spectral bands and the inter-class distance metric, a dynamic weight matrix is ​​constructed to achieve intelligent filtering of key feature bands. Enhanced feature vectors are generated by combining spectral derivative transformation and spectral index calculation.

[0047] Specifically, a dynamic weighting algorithm that integrates mutual information entropy and inter-class distance is used to select key bands, and nonlinear characteristics are enhanced by combining spectral derivatives and exponential calculations, thus solving the problems of redundancy and information loss in traditional fixed band selection.

[0048] The construction of the dynamic weight matrix includes: Mutual Information Entropy (MI): This measures the correlation between band Bi and the target parameter Y, and the formula is: ; Where p(b,y) is the joint probability distribution, and p(b) and p(y) are the marginal probability distributions. The higher the MI value, the stronger the correlation between the band and the parameter.

[0049] Fisher ratio (inter-class distance): measures the ability of a band to distinguish between samples of different classes. The formula is: ; Where μ1 and μ2 are the means of the two classes of samples, and σ1 2 ,σ2 2The variance is represented by a larger value, indicating stronger discriminative power. Dynamic weight fusion generates a band weight matrix by standardizing and weighting the MI and Fisher values ​​(the weight coefficients can be optimized through grid search), and then selecting the top 30% of bands by weight as key features.

[0050] Feature enhancement is achieved through spectral derivative transformation: calculating the first derivative (R′(λ) = (R(λ + Δλ)). R(λ The second derivative (Δλ) / 2Δλ) and the second derivative are used to eliminate baseline drift and highlight the differences in absorption peak position and slope.

[0051] Spectral index calculation: Indices are designed based on parameter characteristics, such as the normalized water index for water content and the ratio index (e.g., R650 / R700) for organic matter, generating an enhanced feature vector (dimension ≤ 50) containing the original bands, derivatives, and indices.

[0052] Step S103. Construct a multi-task transfer learning neural network model, which includes a spectral feature encoding layer, a spatial context fusion layer, and a parameter inversion output layer. The spectral feature encoding layer uses an attention mechanism to enhance the feature representation of key bands, the spatial context fusion layer extracts neighborhood spatial correlation features through a convolutional neural network, and the output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture.

[0053] Specifically, a three-layer neural network architecture is designed, which strengthens key bands through an attention mechanism, extracts spatial context through CNN, achieves parameter collaborative inversion through multi-task learning, and enhances cross-regional generalization ability by combining transfer learning.

[0054] The spectral feature coding layer includes: Input layer: receiving the enhanced feature vector (dimension D), normalized (x′= x After μ / σ), input. Channel attention module: calculates the band dimension weights, the formula is: α=Sigmoid(W2*ReLU(W1*GlobalAvgPool(X))); where X is the input feature, W1 and W2 are the parameters of the fully connected layer, and the output is a weighted feature to enhance the expression of key bands.

[0055] The fully connected layer extracts nonlinear spectral features through 2-3 hidden layers (128-256 neurons, ReLU activation function).

[0056] The spatial context fusion layer uses a convolutional neural network (CNN): it divides the hyperspectral image into k×k pixel patches (e.g., 5×5), inputs them into a 2D convolutional layer (3×3 kernel, stride 1, padding 1), and after batch normalization and ReLU activation, it extracts the spatial correlation of the neighborhood (e.g., spatial heterogeneity of soil texture).

[0057] Feature fusion combines spectral coding features and spatial features through splicing or addition operations, and then inputs them into subsequent layers.

[0058] The multi-task output layer avoids redundant training by sharing underlying features: the parameters of the spectral coding layer and the spatial fusion layer are shared. Independent output heads: each parameter (organic matter, mineral concentration, water content) corresponds to an independent fully connected layer (1 neuron, linear activation function), the loss function is the mean squared error (MSE), and the total loss function is the weighted sum of multiple tasks.

[0059] Transfer learning strategies include: Pre-training: Training the base model on large-scale data in the source region (such as the North China Plain) to learn the general spectrum-parameter mapping relationship. Fine-tuning: Inputting a small number of samples into the target region (such as the Southwest Hills), freezing the first 2 / 3 of the network parameters, and updating only the last 1 / 3 of the layers and the output layer to adapt to terrain and atmospheric differences and improve generalization ability.

[0060] Step S104. Perform the same preprocessing and feature enhancement operations on the hyperspectral image data of the area to be inverted as on the training data, input the trained neural network model, and output the surface matrix parameter inversion results, wherein the inversion results include at least one parameter among soil organic matter content, mineral component concentration, and water content.

[0061] By performing the same preprocessing and feature enhancement on the inverted region data as on the training data, the trained model is input and multi-parameter inversion results are output.

[0062] Data preprocessing is performed simultaneously by performing radiometric calibration, atmospheric correction, and spectral denoising on a pixel-by-pixel basis on the hyperspectral image of the target area, ensuring complete consistency with the training data preprocessing workflow and avoiding errors introduced by processing differences. Pixel spectral curves are extracted, and dynamic weights are calculated, bands are selected, and enhanced feature vectors are generated according to the S102 method (using the optimal band index and exponential formula determined during the training phase).

[0063] Model inference and output are achieved by inputting single-pixel features or spatial patches into a neural network, outputting continuous values ​​for each parameter (e.g., organic matter content 0-10%, water content 0-50%). Spatial interpolation (e.g., Kriging) is performed on the inversion results, or a raster map is directly generated, and spatial location information is overlaid. The final output is a thematic map of surface matrix attributes containing multiple parameters, supporting visualization and analysis on a GIS platform.

[0064] In some embodiments, before performing the same preprocessing and feature enhancement operations on the hyperspectral image data of the region to be inverted as on the training data, the method further includes: training the neural network model using a hybrid loss function, wherein the hybrid loss function includes a regression loss term based on mean squared error and a feature preservation loss term based on structural similarity; during training, an adversarial example enhancement strategy is introduced to improve the robustness of the model; and the model's generalization ability is optimized by transferring prior knowledge of the same origin region through domain adaptation techniques.

[0065] By fusing regression loss and feature preservation loss to optimize model fitting ability, combining adversarial examples to enhance robustness, and utilizing domain adaptation techniques to transfer knowledge from the same region, the problem of cross-domain generalization is solved.

[0066] The hybrid loss function design includes: a regression loss term that uses mean squared error (MSE) to measure the numerical difference between the predicted and true values; a feature preservation loss term that uses structural similarity (SSIM) to measure the structural similarity between the predicted and true features; high-level features are extracted through a pre-trained VGG network; and Gram matrix similarity is calculated.

[0067] Adversarial example enhancement employs the Fast Gradient Sign Method (FGSM) to generate adversarial examples, which are then mixed with the original examples during training to improve the model's robustness to noise and spectral variations.

[0068] Domain adaptation techniques employ deep adaptive neural networks (DANNs), adding a domain discriminator after the feature encoding layer. They achieve feature alignment between the source domain (e.g., a known region) and the target domain (e.g., a new region) through gradient inversion (GRL) layers, and promote the model's learning of domain-invariant general features through domain classification cross-entropy loss calculation.

[0069] In some embodiments, the method further includes: embedding a model uncertainty quantification module in the neural network model, calculating the confidence evaluation value of the inversion parameters using a Bayesian neural network or a Monte Carlo method, wherein the confidence evaluation value is based on the variance or entropy value of the predicted distribution and is used to screen low-confidence samples for active learning and iterative optimization of the model.

[0070] The confidence level of the inversion results is evaluated by embedding an uncertainty quantification module, and the predicted distribution is calculated by Bayesian inference or Monte Carlo sampling. Low-confidence samples are screened to iteratively optimize the model.

[0071] Uncertainty quantification methods include: Bayesian Neural Networks (BNNs): Introducing the probability distribution (such as mean and variance) of weights into fully connected layers, approximating the posterior distribution through variational inference, sampling weights to generate multiple prediction results during prediction, and using variance to characterize uncertainty.

[0072] Monte Carlo dropout (MC Dropout) maintains the activation of the dropout layer during training and samples multiple times during prediction (e.g., T=50 times), evaluating uncertainty by the entropy value of the prediction results.

[0073] The active learning process involves setting a confidence threshold (such as variance > 0.1 or entropy > 0.8), filtering low-confidence samples, manually labeling them, adding them to the training set, retraining the model, and iterating repeatedly until the accuracy stabilizes.

[0074] In some embodiments, the collection of spatial location information and environmental parameters of surface matrix samples includes: obtaining the geographic coordinates of the surface matrix samples through a GNSS positioning device, and collecting environmental parameters of the sample points. The environmental parameters include surface temperature, air humidity, vegetation coverage, terrain slope and aspect. The vegetation coverage is calculated by combining UAV remote sensing imagery with pixel binarization, and the terrain slope and aspect are obtained based on DEM data through spatial analysis algorithms.

[0075] By acquiring precise geographic coordinates through GNSS and combining them with UAV remote sensing and DEM data to collect multi-dimensional environmental parameters, spatial and environmental context information is provided for the model.

[0076] Geographic coordinates are collected using high-precision GNSS equipment (such as Trimble R10, with a positioning accuracy of ≤2cm) to record the latitude and longitude (WGS84 coordinate system) and altitude of sample points, ensuring that the spatial location is accurately matched with the pixels of the remote sensing image (error ≤1 pixel).

[0077] Environmental parameters were acquired, including: surface temperature and air humidity, measured in real time using a portable weather station (such as Vaisala WXT530) with accuracies of ±0.5℃ and ±2%RH, respectively.

[0078] Vegetation coverage includes: acquiring RGB images (resolution ≤10cm) by drone, converting them to HSV color space, segmenting vegetation areas, and calculating coverage using the pixel binarization method.

[0079] Topographic slope and aspect were determined by downloading 30m resolution DEM data (such as SRTM) and resampling it to image resolution; then, using ArcGIS / SAGA's topographic analysis tools, slope (unit: degrees) and aspect (unit: degrees, 0-360°) were calculated based on differential geometry methods.

[0080] In some embodiments, the intelligent screening of key feature bands by constructing a dynamic weight matrix through calculating the mutual information entropy and inter-class distance metrics between spectral bands includes: calculating the mutual information entropy between each band and the true value of the surface matrix parameter to quantify the correlation between the band and the target parameter; using Mahalanobis distance to calculate the inter-class distance of different categories of samples on each band to construct a band discrimination capability matrix; fusing mutual information entropy and inter-class distance through an adaptive weight formula; and screening key feature bands according to the dynamic weight matrix.

[0081] By quantifying band-parameter correlation through mutual information, measuring class separability through Mahalanobis distance, and adaptively fusing to construct a weight matrix, the feature redundancy problem of high-dimensional spectra can be solved.

[0082] Mutual information entropy calculation involves dividing the spectral values ​​of each band into 10 equally spaced intervals, statistically analyzing the joint probability matrix, and then calculating the mutual information.

[0083] Mahalanobis distance, an inter-class metric, is used to calculate the inter-class Mahalanobis distance for binary classification problems (such as high / low organic matter content). In multi-class classification, the average inter-class Mahalanobis distance is used.

[0084] Normalized MI and Mahalanobis distance (0-1 interval) are used, and an adaptive weighting formula is employed to select the top 40% of bands as key features.

[0085] In some embodiments, the step of combining spectral derivative transformation and spectral index calculation to generate an enhanced feature vector includes: performing a first- or second-order spectral derivative transformation on the selected key feature bands to eliminate baseline drift and low-frequency noise, thereby obtaining a derivative spectral matrix; calculating a targeted spectral index based on the key bands, wherein the spectral index includes, but is not limited to, the normalized vegetation index, the soil-regulated vegetation index, and the mineral absorption characteristic index; and concatenating the original spectral reflectance, derivative spectral values, and spectral index values ​​into an enhanced feature vector.

[0086] By removing baseline noise through derivative transformation of key bands and combining it with targeted spectral index to enhance parameter-sensitive features, multi-dimensional enhanced features are constructed.

[0087] The spectral derivative transformation includes: First derivative: using the symmetric finite difference method, with a window size of 5 bands. Second derivative: further calculating the derivative of the first derivative to highlight the changes in the curvature of the absorption peaks.

[0088] Spectral index calculation is performed using the normalized difference vegetation index (NDVI = (NIR)). R) / (NIR+R), monitoring the impact of vegetation cover on soil parameters); Soil-modified vegetation index (SAVI=(1+L)(NIR) R) / (NIR+R+L) (L=0.5, to reduce soil background interference, NIR is the value in the near-infrared band, and R is the value in the infrared band); Mineral Absorption Characteristic Index (MAI): Targeting mineral absorption peaks (such as Fe at 900nm). 2+ (Absorption), calculate R850 / R950 to characterize mineral concentration.

[0089] Feature stitching generates an enhancement vector with a dimension of 3K+M by stitching together the original reflectance (key bands), first derivative, second derivative, and spectral index in band order (K ​​is the number of key bands, and M is the number of indices).

[0090] In some embodiments, the spectral feature encoding layer employs a multi-head self-attention mechanism to enhance the feature representation of key bands by calculating the attention weights of features in different bands; the spatial context fusion layer employs a residual neural network structure with dilated convolutions to expand the receptive field and extract multi-scale neighborhood spatial correlation features without reducing spatial resolution; the parameter inversion output layer sets up a multi-task shared feature extraction layer and an independent parameter regression branch, the shared layer employs a fully connected network to fuse spectral and spatial features, and each branch achieves the collaborative inversion of soil organic matter content, mineral component concentration, and water content through adaptive weight allocation.

[0091] Multi-head self-attention is introduced to enhance inter-band correlation, dilated convolution expands the spatial receptive field, and multi-task branching enables parameter collaborative inversion, thereby improving feature representation capabilities.

[0092] The spectral feature encoding layer (multi-head self-attention) maps key band features into a Query, Key, Value matrix (dimension dk), performs parallel multi-head attention computation, and concatenates the outputs of each head before generating enhanced features through a fully connected layer, highlighting higher-order correlations between bands (such as the synergistic response of organic matter in 600-700nm and 1400-1900nm).

[0093] The spatial context fusion layer (dilated convolutional residual network) uses 3 convolutional layers with dilation rate (e.g., [1, 2, 4]) to expand the receptive field without increasing the parameters (e.g., 3×3 dilated convolution + rate=2 is equivalent to a 7×7 receptive field); residual connections avoid gradient vanishing.

[0094] The multi-task output layer includes: a shared feature layer that fuses spectral and spatial features through two fully connected layers (256 neurons); and independent regression branches: each parameter corresponds to one fully connected layer, with branch weights adaptively allocated, where the observation noise variance of the task is estimated through prior knowledge or data.

[0095] In some embodiments, the multiple methods further include: introducing a self-supervised pre-training module between the spectral feature encoding layer and the spatial context fusion layer, using a contrastive learning method to pre-train unlabeled hyperspectral data, and improving the model's general representation ability of spectral-spatial features by maximizing the feature similarity of different augmented views of the same sample, wherein the augmented views include random band masks, spectral noise perturbations, and spatial clipping operations.

[0096] By using unlabeled data to learn pre-trained models through contrastive learning, the general representation ability of spectral-spatial features can be enhanced, alleviating the problem of insufficient labeled samples.

[0097] Enhanced view generation includes: random band masking: randomly occluding 10-30% of the bands (simulating sensor failure or noise); spectral noise perturbation: adding Gaussian noise (mean 0, standard deviation 0.05) or salt-and-pepper noise (probability 0.1); spatial cropping: randomly cropping 80% of the region from the image patch while preserving the local context.

[0098] The contrastive learning objective function employs NT-Xent loss, which maximizes the feature cosine similarity between two augmented views of the same sample.

[0099] The pre-training process first trains the contrastive learning module on unlabeled data for 100 epochs, and then fine-tunes the entire network using labeled data to improve the model's ability to extract features from unseen samples.

[0100] In some embodiments, the hyperspectral image data of the region to be inverted undergoes the same preprocessing and feature enhancement operations as the training data, including: performing radiometric calibration to correct sensor response bias, atmospheric correction based on the MODTRAN model to eliminate the influence of aerosols and water vapor, and spectral denoising using wavelet transform or median filtering; performing band selection based on the key band index determined during the training phase, and generating enhanced feature vectors according to the same derivative transformation method and spectral index calculation formula to ensure that the feature space of the input model is consistent with the training data.

[0101] By employing a rigorous data processing workflow that synchronizes training and inversion phases, we ensure consistency in the input feature space and avoid model bias caused by differences in processing.

[0102] Radiometric calibration: Using the sensor’s built-in laboratory calibration coefficients, the DN value is converted into reflectance. Atmospheric correction is based on the MODTRAN model. Input parameters include aerosol type (rural / urban), water vapor content (from ERA5 reanalysis data), and solar zenith angle. Output is surface reflectance.

[0103] Spectral denoising selects wavelet transform (suitable for periodic noise) or median filtering (suitable for salt-and-pepper noise) based on the noise level, and the window size is optimized by the signal-to-noise ratio (e.g., use 5-point median filtering when SNR>20dB).

[0104] Band selection uses the optimal band index determined during the training phase (such as the band numbers of the top 30% of the weights) to directly extract the corresponding bands from the inversion data. The derivative and exponent calculations strictly reuse the transformation parameters from training (such as the size of the derivative window and the position of the band in the exponent formula). For example, a 5-point window is used to calculate the first derivative during training, and this must be kept consistent during inversion to avoid changes in the feature dimension.

[0105] Embodiments of this application also provide a machine learning-based hyperspectral data inversion system for surface matrix parameters. This machine learning-based hyperspectral data inversion system is used to execute the aforementioned machine learning-based hyperspectral data inversion method for surface matrix parameters. The machine learning-based hyperspectral data inversion system for surface matrix parameters can be configured on a server or terminal.

[0106] The server can be a standalone server, a server cluster, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal can be an electronic device such as a mobile phone, tablet, laptop, desktop computer, user digital assistant, or wearable device.

[0107] like Figure 2 As shown, the hyperspectral data inversion system for surface matrix parameters 200 based on machine learning includes: Data construction unit 201 is used to construct a multi-source heterogeneous hyperspectral dataset. The hyperspectral dataset includes hyperspectral reflectance data of surface matrix samples in the target area and the corresponding measured parameter true values. The hyperspectral dataset is preprocessed with radiometric calibration, atmospheric correction and spectral denoising, and spatial location information and environmental parameters of surface matrix samples are collected. The feature selection unit 202 is used to perform feature selection on the preprocessed hyperspectral dataset based on the adaptive band selection algorithm. By calculating the mutual information entropy between spectral bands and the inter-class distance metric, a dynamic weight matrix is ​​constructed to achieve intelligent selection of key feature bands. The enhanced feature vector is generated by combining spectral derivative transformation and spectral index calculation. Model building unit 203 is used to build a multi-task transfer learning neural network model, which includes a spectral feature encoding layer, a spatial context fusion layer and a parameter inversion output layer. The spectral feature encoding layer uses an attention mechanism to enhance the feature representation of key bands, the spatial context fusion layer extracts neighborhood spatial correlation features through a convolutional neural network, and the output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture.

[0108] The inversion output unit 204 is used to perform the same preprocessing and feature enhancement operations on the hyperspectral image data of the area to be inverted as on the training data. It inputs the trained neural network model and outputs the inversion results of surface matrix parameters, which include at least one parameter among soil organic matter content, mineral component concentration, and water content.

[0109] In some embodiments, before performing the same preprocessing and feature enhancement operations on the hyperspectral image data of the region to be inverted as on the training data, the method further includes: training the neural network model using a hybrid loss function, wherein the hybrid loss function includes a regression loss term based on mean squared error and a feature preservation loss term based on structural similarity; during training, an adversarial example enhancement strategy is introduced to improve the robustness of the model; and the model's generalization ability is optimized by transferring prior knowledge of the same origin region through domain adaptation techniques.

[0110] In some embodiments, the method further includes: embedding a model uncertainty quantification module in the neural network model, calculating the confidence evaluation value of the inversion parameters using a Bayesian neural network or a Monte Carlo method, wherein the confidence evaluation value is based on the variance or entropy value of the predicted distribution and is used to screen low-confidence samples for active learning and iterative optimization of the model.

[0111] In some embodiments, the collection of spatial location information and environmental parameters of surface matrix samples includes: obtaining the geographic coordinates of the surface matrix samples through a GNSS positioning device, and collecting environmental parameters of the sample points. The environmental parameters include surface temperature, air humidity, vegetation coverage, terrain slope and aspect. The vegetation coverage is calculated by combining UAV remote sensing imagery with pixel binarization, and the terrain slope and aspect are obtained based on DEM data through spatial analysis algorithms.

[0112] In some embodiments, the intelligent screening of key feature bands by constructing a dynamic weight matrix through calculating the mutual information entropy and inter-class distance metrics between spectral bands includes: calculating the mutual information entropy between each band and the true value of the surface matrix parameter to quantify the correlation between the band and the target parameter; using Mahalanobis distance to calculate the inter-class distance of different categories of samples on each band to construct a band discrimination capability matrix; fusing mutual information entropy and inter-class distance through an adaptive weight formula; and screening key feature bands according to the dynamic weight matrix.

[0113] In some embodiments, the step of combining spectral derivative transformation and spectral index calculation to generate an enhanced feature vector includes: performing a first- or second-order spectral derivative transformation on the selected key feature bands to eliminate baseline drift and low-frequency noise, thereby obtaining a derivative spectral matrix; calculating a targeted spectral index based on the key bands, wherein the spectral index includes, but is not limited to, the normalized vegetation index, the soil-regulated vegetation index, and the mineral absorption characteristic index; and concatenating the original spectral reflectance, derivative spectral values, and spectral index values ​​into an enhanced feature vector.

[0114] In some embodiments, the spectral feature encoding layer employs a multi-head self-attention mechanism to enhance the feature representation of key bands by calculating the attention weights of features in different bands; the spatial context fusion layer employs a residual neural network structure with dilated convolutions to expand the receptive field and extract multi-scale neighborhood spatial correlation features without reducing spatial resolution; the parameter inversion output layer sets up a multi-task shared feature extraction layer and an independent parameter regression branch, the shared layer employs a fully connected network to fuse spectral and spatial features, and each branch achieves the collaborative inversion of soil organic matter content, mineral component concentration, and water content through adaptive weight allocation.

[0115] In some embodiments, the multiple methods further include: introducing a self-supervised pre-training module between the spectral feature encoding layer and the spatial context fusion layer, using a contrastive learning method to pre-train unlabeled hyperspectral data, and improving the model's general representation ability of spectral-spatial features by maximizing the feature similarity of different augmented views of the same sample, wherein the augmented views include random band masks, spectral noise perturbations, and spatial clipping operations.

[0116] In some embodiments, the hyperspectral image data of the region to be inverted undergoes the same preprocessing and feature enhancement operations as the training data, including: performing radiometric calibration to correct sensor response bias, atmospheric correction based on the MODTRAN model to eliminate the influence of aerosols and water vapor, and spectral denoising using wavelet transform or median filtering; performing band selection based on the key band index determined during the training phase, and generating enhanced feature vectors according to the same derivative transformation method and spectral index calculation formula to ensure that the feature space of the input model is consistent with the training data.

[0117] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the machine learning-based hyperspectral data inversion system for land surface matrix parameters and each module described above can be referred to the corresponding process in the aforementioned embodiments of the machine learning-based hyperspectral data inversion method for land surface matrix parameters, and will not be repeated here.

[0118] The aforementioned machine learning-based hyperspectral data inversion system for surface matrix parameters can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the computer device shown.

[0119] Please see Figure 3 , Figure 3 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0120] The storage medium may store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform an embodiment of any machine learning-based hyperspectral data inversion method for surface matrix parameters.

[0121] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0122] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When executed by a processor, the computer program enables the processor to perform any machine learning-based hyperspectral data inversion system method for surface matrix parameters.

[0123] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0124] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0125] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: A multi-source heterogeneous hyperspectral dataset is constructed, which includes hyperspectral reflectance data of surface matrix samples in the target area and the corresponding measured parameter true values. The hyperspectral dataset is preprocessed with radiometric calibration, atmospheric correction and spectral denoising, and spatial location information and environmental parameters of surface matrix samples are collected. The preprocessed hyperspectral dataset is used for feature selection based on an adaptive band selection algorithm. By calculating the mutual information entropy between spectral bands and the inter-class distance metric, a dynamic weight matrix is ​​constructed to achieve intelligent selection of key feature bands. Enhanced feature vectors are generated by combining spectral derivative transformation and spectral index calculation. A multi-task transfer learning neural network model is constructed, which includes a spectral feature encoding layer, a spatial context fusion layer, and a parameter inversion output layer. The spectral feature encoding layer uses an attention mechanism to enhance the feature representation of key bands, the spatial context fusion layer extracts neighborhood spatial correlation features through a convolutional neural network, and the output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture. Perform the same preprocessing and feature enhancement operations on the hyperspectral image data of the inversion area as on the training data, input the trained neural network model, and output the inversion results of surface matrix parameters, which include at least one parameter among soil organic matter content, mineral component concentration, and water content.

[0126] In some embodiments, before performing the same preprocessing and feature enhancement operations on the hyperspectral image data of the region to be inverted as on the training data, the method further includes: training the neural network model using a hybrid loss function, wherein the hybrid loss function includes a regression loss term based on mean squared error and a feature preservation loss term based on structural similarity; during training, an adversarial example enhancement strategy is introduced to improve the robustness of the model; and the model's generalization ability is optimized by transferring prior knowledge of the same origin region through domain adaptation techniques.

[0127] In some embodiments, the method further includes: embedding a model uncertainty quantification module in the neural network model, calculating the confidence evaluation value of the inversion parameters using a Bayesian neural network or a Monte Carlo method, wherein the confidence evaluation value is based on the variance or entropy value of the predicted distribution and is used to screen low-confidence samples for active learning and iterative optimization of the model.

[0128] In some embodiments, the collection of spatial location information and environmental parameters of surface matrix samples includes: obtaining the geographic coordinates of the surface matrix samples through a GNSS positioning device, and collecting environmental parameters of the sample points. The environmental parameters include surface temperature, air humidity, vegetation coverage, terrain slope and aspect. The vegetation coverage is calculated by combining UAV remote sensing imagery with pixel binarization, and the terrain slope and aspect are obtained based on DEM data through spatial analysis algorithms.

[0129] In some embodiments, the intelligent screening of key feature bands by constructing a dynamic weight matrix through calculating the mutual information entropy and inter-class distance metrics between spectral bands includes: calculating the mutual information entropy between each band and the true value of the surface matrix parameter to quantify the correlation between the band and the target parameter; using Mahalanobis distance to calculate the inter-class distance of different categories of samples on each band to construct a band discrimination capability matrix; fusing mutual information entropy and inter-class distance through an adaptive weight formula; and screening key feature bands according to the dynamic weight matrix.

[0130] In some embodiments, the step of combining spectral derivative transformation and spectral index calculation to generate an enhanced feature vector includes: performing a first- or second-order spectral derivative transformation on the selected key feature bands to eliminate baseline drift and low-frequency noise, thereby obtaining a derivative spectral matrix; calculating a targeted spectral index based on the key bands, wherein the spectral index includes, but is not limited to, the normalized vegetation index, the soil-regulated vegetation index, and the mineral absorption characteristic index; and concatenating the original spectral reflectance, derivative spectral values, and spectral index values ​​into an enhanced feature vector.

[0131] In some embodiments, the spectral feature encoding layer employs a multi-head self-attention mechanism to enhance the feature representation of key bands by calculating the attention weights of features in different bands; the spatial context fusion layer employs a residual neural network structure with dilated convolutions to expand the receptive field and extract multi-scale neighborhood spatial correlation features without reducing spatial resolution; the parameter inversion output layer sets up a multi-task shared feature extraction layer and an independent parameter regression branch, the shared layer employs a fully connected network to fuse spectral and spatial features, and each branch achieves the collaborative inversion of soil organic matter content, mineral component concentration, and water content through adaptive weight allocation.

[0132] In some embodiments, the multiple methods further include: introducing a self-supervised pre-training module between the spectral feature encoding layer and the spatial context fusion layer, using a contrastive learning method to pre-train unlabeled hyperspectral data, and improving the model's general representation ability of spectral-spatial features by maximizing the feature similarity of different augmented views of the same sample, wherein the augmented views include random band masks, spectral noise perturbations, and spatial clipping operations.

[0133] In some embodiments, the hyperspectral image data of the region to be inverted undergoes the same preprocessing and feature enhancement operations as the training data, including: performing radiometric calibration to correct sensor response bias, atmospheric correction based on the MODTRAN model to eliminate the influence of aerosols and water vapor, and spectral denoising using wavelet transform or median filtering; performing band selection based on the key band index determined during the training phase, and generating enhanced feature vectors according to the same derivative transformation method and spectral index calculation formula to ensure that the feature space of the input model is consistent with the training data.

[0134] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the processor described above can be referred to the corresponding process in the method embodiments described above, and will not be repeated here.

[0135] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement the steps of the hyperspectral data inversion method for surface matrix parameters based on machine learning provided in the above embodiments of this application.

[0136] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0137] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for inverting surface matrix parameters from hyperspectral data based on machine learning, characterized in that, include: A multi-source heterogeneous hyperspectral dataset is constructed, which includes hyperspectral reflectance data of surface matrix samples in the target area and the corresponding measured parameter true values. The hyperspectral dataset is preprocessed with radiometric calibration, atmospheric correction and spectral denoising, and spatial location information and environmental parameters of surface matrix samples are collected. The preprocessed hyperspectral dataset is used for feature selection based on an adaptive band selection algorithm. By calculating the mutual information entropy between spectral bands and the inter-class distance metric, a dynamic weight matrix is ​​constructed to achieve intelligent selection of key feature bands. Enhanced feature vectors are generated by combining spectral derivative transformation and spectral index calculation. A multi-task transfer learning neural network model is constructed, which includes a spectral feature encoding layer, a spatial context fusion layer, and a parameter inversion output layer. The spectral feature encoding layer uses an attention mechanism to enhance the feature representation of key bands, the spatial context fusion layer extracts neighborhood spatial correlation features through a convolutional neural network, and the output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture. Perform the same preprocessing and feature enhancement operations on the hyperspectral image data of the inversion area as on the training data, input the trained neural network model, and output the inversion results of surface matrix parameters, which include at least one parameter among soil organic matter content, mineral component concentration, and water content.

2. The method according to claim 1, characterized in that, Before performing the same preprocessing and feature enhancement operations on the hyperspectral image data of the region to be inverted as on the training data, the process further includes: The neural network model is trained using a hybrid loss function, which includes a regression loss term based on mean squared error and a feature preservation loss term based on structural similarity. During training, an adversarial example enhancement strategy is introduced to improve the robustness of the model, and the model's generalization ability is optimized by transferring prior knowledge of the same region through domain adaptation technology.

3. The method according to claim 2, characterized in that, The method further includes: A model uncertainty quantification module is embedded in the neural network model to calculate the confidence evaluation value of the inversion parameters using a Bayesian neural network or Monte Carlo method. The confidence evaluation value is based on the variance or entropy value of the predicted distribution and is used to screen low-confidence samples for active learning and iterative optimization of the model.

4. The method according to claim 1, characterized in that, The spatial location information and environmental parameters of the collected surface matrix samples include: The geographic coordinates of the surface matrix sample are obtained through GNSS positioning equipment, and environmental parameters of the sample point are collected. The environmental parameters include surface temperature, air humidity, vegetation coverage, terrain slope and aspect. The vegetation coverage is calculated by combining UAV remote sensing imagery with pixel binarization, and the terrain slope and aspect are obtained based on DEM data through spatial analysis algorithms.

5. The method according to claim 1, characterized in that, The method of constructing a dynamic weight matrix by calculating the mutual information entropy between spectral bands and the inter-class distance metric to achieve intelligent screening of key feature bands includes: Calculate the mutual information entropy between each band and the true value of the surface matrix parameters to quantify the correlation between the band and the target parameters; Mahalanobis distance was used to calculate the inter-class distance of samples of different categories in each band, and a band discrimination capability matrix was constructed. The mutual information entropy and inter-class distance are fused using an adaptive weighting formula; key feature bands are selected based on the dynamic weighting matrix.

6. The method according to claim 1, characterized in that, The method of combining spectral derivative transformation and spectral index calculation to generate enhanced feature vectors includes: The first or second order spectral derivative transformation is performed on the selected key feature bands to eliminate baseline drift and low-frequency noise, and the derivative spectral matrix is ​​obtained. Targeted spectral indices are calculated based on key bands. These indices include, but are not limited to, normalized vegetation index, soil-regulated vegetation index, and mineral absorption characteristic index. The original spectral reflectance, derivative spectral values, and spectral index values ​​are then concatenated into an enhanced feature vector.

7. The method according to claim 1, characterized in that, The spectral feature encoding layer adopts a multi-head self-attention mechanism, which strengthens the feature representation of key bands by calculating the attention weights of features in different bands; the spatial context fusion layer adopts a residual neural network structure with dilated convolution, which expands the receptive field and extracts multi-scale neighborhood spatial correlation features without reducing spatial resolution. The parameter inversion output layer is configured with a multi-task shared feature extraction layer and an independent parameter regression branch. The shared layer uses a fully connected network to fuse spectral and spatial features. Each branch achieves the collaborative inversion of soil organic matter content, mineral component concentration, and water content through adaptive weight allocation.

8. The method according to claim 7, characterized in that, The multiple methods also include: A self-supervised pre-training module is introduced between the spectral feature encoding layer and the spatial context fusion layer. The contrastive learning method is used to pre-train the model on unlabeled hyperspectral data. By maximizing the feature similarity of different augmented views of the same sample, the model’s ability to represent spectral-spatial features in a general way is improved. The augmented views include random band masks, spectral noise perturbations, and spatial clipping operations.

9. The method according to claim 1, characterized in that, The hyperspectral image data of the region to be inverted undergoes the same preprocessing and feature enhancement operations as the training data, including: The hyperspectral image data were sequentially subjected to radiometric calibration to correct sensor response deviation, atmospheric correction based on the MODTRAN model to eliminate the influence of aerosols and water vapor, and spectral denoising was performed using wavelet transform or median filtering. Bands are selected based on the key band index determined during the training phase, and enhanced feature vectors are generated using the same derivative transformation method and spectral index calculation formula to ensure that the feature space of the input model is consistent with the training data.

10. A machine learning-based hyperspectral data inversion system for surface matrix parameters, characterized in that, include: The data construction unit is used to construct a multi-source heterogeneous hyperspectral dataset, which includes hyperspectral reflectance data of surface matrix samples in the target area and the corresponding measured parameter true values. The hyperspectral dataset is preprocessed with radiometric calibration, atmospheric correction and spectral denoising, and spatial location information and environmental parameters of surface matrix samples are collected. The feature selection unit is used to perform feature selection on the preprocessed hyperspectral dataset based on the adaptive band selection algorithm. By calculating the mutual information entropy between spectral bands and the inter-class distance metric, a dynamic weight matrix is ​​constructed to achieve intelligent selection of key feature bands. The enhanced feature vector is generated by combining spectral derivative transformation and spectral index calculation. The model building unit is used to construct a multi-task transfer learning neural network model, which includes a spectral feature encoding layer, a spatial context fusion layer, and a parameter inversion output layer. The spectral feature encoding layer uses an attention mechanism to enhance the feature representation of key bands, the spatial context fusion layer extracts neighborhood spatial correlation features through a convolutional neural network, and the output layer realizes multi-parameter collaborative inversion based on a multi-task learning architecture. The inversion output unit is used to perform the same preprocessing and feature enhancement operations on the hyperspectral image data of the area to be inverted as on the training data. It inputs the trained neural network model and outputs the inversion results of surface matrix parameters, which include at least one parameter among soil organic matter content, mineral component concentration, and water content.

Images