Search

CN-121982439-A - VOD remote sensing data downscaling model construction method based on physical constraint

CN121982439ACN 121982439 ACN121982439 ACN 121982439ACN-121982439-A

Abstract

The application belongs to the field of artificial intelligence and remote sensing science and technology, and particularly discloses a physical constraint-based VOD remote sensing data downscaling model construction method, which comprises the steps of firstly collecting SMAP-IB VOD and SM data and multisource driving variables; preprocessing, generating a daily NDVI by adopting a FSDAF algorithm, calculating physical parameters based on RTM, simulating a transmission process of a microwave signal, constructing a PIDNet model, training the model through a joint loss function, and generating 1 km high-resolution VOD data. The method solves the problems that the spatial resolution of the existing VOD product is low (9-36 km) and the downscaling result lacks physical consistency, and the generated 1 km VOD has fine spatial details and physical credibility, can support the scenes of field-scale agricultural moisture management, community-scale ecological hydrologic simulation and the like, and provides high-quality data support for global carbon circulation assessment and drought disaster monitoring.

Inventors

  • Cui Dunyue
  • ZHANG XIANG
  • GU XIHUI
  • CHEN NENGCHENG

Assignees

  • 湖北珞珈实验室
  • 中国地质大学（武汉）

Dates

Publication Date
20260505
Application Date
20251127

Claims (10)

  1. 1. The VOD remote sensing data downscaling model construction method based on physical constraint is characterized by comprising the following steps of: S10, acquiring various remote sensing data, including VOD and soil humidity data inverted by an SMAP-IB algorithm, a driving variable data set and radiation transmission model auxiliary data, and carrying out reprojection, resampling, clipping and space-time registration on all the data so as to unify the time and space resolution of the data; S20, generating day-by-day seamless NDVI data with high spatial resolution by using a flexible space-time data fusion algorithm based on the driving variable data set preprocessed in the step S10; S30, calculating physical parameters based on the radiation transmission model auxiliary data and the soil humidity data which are preprocessed in the step S10 by using a radiation transmission model, wherein the physical parameters comprise soil effective temperature, soil dielectric constant, soil reflectivity and simulated brightness temperature, and simulating a microwave signal transmission process; And S40, constructing and training a physical constraint deep learning model, wherein the input of the model comprises the driving variable data set and the VOD data which are preprocessed in the step S10 and the physical parameters calculated in the step S30, outputting a VOD predicted value, and the training of the model uses a joint loss function which comprises data loss and physical loss, wherein the data loss is used for constraining the difference between the VOD predicted value and a real VOD value, and the physical loss is used for constraining the difference between the brightness temperature and the theoretical brightness temperature which are simulated by the radiation transmission model based on the predicted VOD.
  2. 2. The method for constructing a physical constraint-based VOD remote sensing data downscaling model according to claim 1, wherein step S20 is specifically: based on the driving variable dataset preprocessed in the step S10, a flexible space-time data fusion method is adopted, and 1 km day-by-day seamless NDVI data is generated by taking 0.05 degree day-by-day NDVI and 1 km month-by-month NDVI as inputs and fusing time prediction and space prediction.
  3. 3. The method for constructing a physical constraint-based VOD remote sensing data downscaling model according to claim 2, wherein the time prediction The calculation formula is as follows: In the formula, The time predicted value at the time of t 2 is represented, Low spatial resolution image pixel representing time t 1 and time t 2 The sum of the high-resolution pixel reflectivity differences corresponding to the belonging ground object class c; The number of the high spatial resolution image pixels corresponding to the ground object category c in the low spatial resolution image pixels is represented, Is the total number of the ground object categories, M represents the number of high spatial resolution pixels belonging to feature class c; the spatial prediction The calculation formula is as follows: In the formula, Is a thin plate spline function.
  4. 4. The method for constructing a physical constraint-based VOD remote sensing data downscaling model according to claim 1 or 2, wherein in step S20, the flexible spatio-temporal data fusion algorithm is FSDAF algorithm, and the FSDAF algorithm is implemented by: Taking high spatial resolution NDVI data at a known time t 1 and low spatial resolution NDVI data at a predicted time t 2 as inputs, and establishing a conversion model of the same spatial pixel change at different time through unsupervised cluster analysis; Acquiring a time variation residual error of a pixel level through time sequence difference analysis; Performing spatial interpolation operation by using a thin plate spline function, and calculating a spatial residual by combining spatial characteristic information of the image; optimizing residual allocation by adopting a self-adaptive weighting strategy; and (3) realizing the predictive reconstruction of the high-resolution image at the time t 2 through spatial neighborhood correlation analysis.
  5. 5. The method for constructing a physical constraint-based VOD remote sensing data downscaling model according to claim 4, wherein the predictive reconstruction calculation formula of the FSDAF algorithm is as follows: In the formula, Representing the reflectivity value of the pixel of the high spatial resolution image at the moment t 2 to be predicted; Representing the reflectivity value of the pixel of the high spatial resolution image at the time t 1 ; Represents weight parameter, k represents kth high spatial resolution image pixel, n represents similar pixel number of the same window, weight The calculation method is that , wherein, W is the window size, Representing the total temporal and spatial variation values after the allocation of the residual.
  6. 6. The method for constructing a physical constraint-based VOD remote sensing data downscaling model according to claim 1, wherein in step S30, the radiation transmission model is a τ - ω radiation transmission model, and a core formula of the τ - ω radiation transmission model is as follows: In the formula, Representing the effective temperature of the soil; Representing vegetation temperature; representing the reflectivity of the soil; For sensor incident angle; The dielectric constant of the soil is obtained by utilizing a Dobson model estimation; is the vegetation attenuation coefficient, i.e., VOD.
  7. 7. The method for constructing a physical constraint-based downscaling model of VOD remote sensing data according to claim 1, wherein in step S40, the physical constraint deep learning model is based on a long-short-term memory network LSTM, the input layer includes 5 types 36 km of driving variable data, 36 km SMAP-IB VOD data and 4 types 36 km of physical parameters, the output layer is a VOD predicted value, the LSTM includes an input layer, a 2-layer hidden layer and an output layer, and each hidden layer includes 6 neurons.
  8. 8. The VOD remote sensing data downscaling model construction method based on physical constraint according to claim 1, wherein in step S40, the data Loss and the physical Loss are both SmoothL a Loss function, and the joint Loss function is a sum of the data Loss and the physical Loss; the data loss function expression is: the physical damage function expression is: In the formula, Default to 1.0; Predicting the output VOD for the physical constraint deep learning model, VOD for SMAP-IB products; Is that The theoretical brightness temperature predicted value obtained by forward calculation of the radiation transmission model, And (3) forward calculating theoretical brightness temperature true values for VOD (video on demand) and soil humidity data of the SMAP-IB through a radiation transmission model.
  9. 9. The VOD remote sensing data downscaling model construction method based on physical constraint according to claim 1, wherein in step S10, the driving variable data set includes a digital elevation model DEM, a surface roughness, a normalized vegetation index NDVI, a surface temperature LST and a soil moisture SM; Wherein, the DEM is ASTER GDEM with 30m resolution, and the surface roughness is calculated by the formula Calculation to generate 1 km surface roughness data The standard deviation of the NDVI is 1 km DEM meshes, the NDVI is obtained by fusing 1 km month-by-month data of the national earth system science data center and 0.05-degree day-by-day data of the NASA MCD19A3CMG, and the LST and SM are 1 km day-by-day resolution, and are derived from the national Qinghai-Tibet plateau science data center.
  10. 10. The VOD remote sensing data downscaling model construction method based on physical constraints according to claim 1, wherein 500 m NASA MCD12Q1 products are selected as land use type data, 250 m global soil database is selected as soil texture data, and ERA5 analysis data sets are selected as first layer soil temperature, third layer soil temperature and surface temperature in the radiation transmission model auxiliary data.

Description

VOD remote sensing data downscaling model construction method based on physical constraint Technical Field The application belongs to the field of artificial intelligence and remote sensing science and technology, and particularly relates to a physical constraint-based VOD remote sensing data downscaling model construction method. Background The vegetation optical thickness (Vegetation Optical Depth, VOD) is a key parameter for representing the moisture content and biomass of vegetation, and the effective monitoring of the vegetation state is realized by quantifying the attenuation degree of microwave signals in the vegetation canopy. The method has irreplaceable functions in the fields of global carbon circulation evaluation, extreme climate event ecological response monitoring such as drought, biomass estimation and the like, and is used as a microwave remote sensing inversion parameter to make up for the limitation of optical vegetation index dependence on clear sky conditions. At present, main stream VOD products are generated by means of inversion of microwave satellite platforms such as AMSR-E/AMSR2, ASCAT, SMOS, SMAP and the like, but are limited by an observation mechanism and sensor performance, the spatial resolution is generally low (mostly 9-36 km), and application requirements such as fine-scale ecological process analysis, field hydrologic simulation, agricultural climate monitoring and the like are difficult to meet. Although satellites such as sentinel-1 may provide high resolution data, the contradiction of temporal resolution and spatial coverage limits their ability to continuously monitor. The existing VOD downscaling research is in a starting stage and has obvious defects. Traditional statistical or machine learning methods deviate from the physical mechanism of microwave and vegetation interaction, resulting in a lack of physical consistency in the results. Therefore, there is a need to develop a high-resolution VOD generation method that combines the advantages of physical mechanism and data driving, and solves the problems of insufficient spatial resolution and low physical reliability of the existing product. Disclosure of Invention Aiming at the defects of the prior art, the application aims to provide a VOD remote sensing data downscaling model construction method based on physical constraint, which realizes accurate generation of 1 km high-resolution VOD data by integrating a physical supervision mechanism of a Radiation Transmission Model (RTM) and data learning capability of a long-short-term memory network (LSTM) and provides technical support for dynamic monitoring of fine-scale vegetation. In order to achieve the above object, in a first aspect, the present application provides a method for constructing a scale-down model of VOD remote sensing data based on physical constraints, comprising the following steps: S10, acquiring various remote sensing data, including VOD and soil humidity data inverted by an SMAP-IB algorithm, a driving variable data set and radiation transmission model auxiliary data, and carrying out reprojection, resampling, clipping and space-time registration on all the data so as to unify the time and space resolution of the data; S20, generating day-by-day seamless NDVI data with high spatial resolution by using a flexible space-time data fusion algorithm based on the driving variable data set preprocessed in the step S10; S30, calculating physical parameters based on the radiation transmission model auxiliary data and the soil humidity data which are preprocessed in the step S10 by using a radiation transmission model, wherein the physical parameters comprise soil effective temperature, soil dielectric constant, soil reflectivity and simulated brightness temperature, and simulating a microwave signal transmission process; And S40, constructing and training a physical constraint deep learning model, wherein the input of the model comprises the driving variable data set and the VOD data which are preprocessed in the step S10 and the physical parameters calculated in the step S30, outputting a VOD predicted value, and the training of the model uses a joint loss function which comprises data loss and physical loss, wherein the data loss is used for constraining the difference between the VOD predicted value and a real VOD value, and the physical loss is used for constraining the difference between the brightness temperature and the theoretical brightness temperature which are simulated by the radiation transmission model based on the predicted VOD. As a further preferred, step S20 is specifically: based on the driving variable dataset preprocessed in the step S10, a flexible space-time data fusion method is adopted, and 1 km day-by-day seamless NDVI data is generated by taking 0.05 degree day-by-day NDVI and 1 km month-by-month NDVI as inputs and fusing time prediction and space prediction. As a further preference, the temporal predictionThe calculation form