Remote sensing image-based ground vegetation leaf area index remote sensing inversion method

Abstract

The invention discloses a ground vegetation leaf area index remote sensing inversion method based on a remote sensing image. The method comprises the following steps: data acquisition and preprocessing; performing multi-scale space-time non-local filtering fusion; dynamic feature extraction; carrying out transfer learning fine tuning; time sequence deep learning integration; and model output and post-processing. According to the invention, through multi-source space-time fusion and dynamic feature distribution, vegetation LAI inversion with high resolution and high continuity is realized; the transfer learning and the time sequence deep network enhance the adaptability of the model to a new region and time sequence change; the stability and reliability of large-scale application are guaranteed through full-process automation and uncertainty evaluation, and the model is obviously superior to an existing single-source or static model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image processing and vegetation parameter inversion, and particularly relates to a ground vegetation leaf area index remote sensing inversion method based on multi-source remote sensing images. BACKGROUND

[0002] Leaf area index (LAI) is an important parameter for studying vegetation canopy structure and ecological processes, and is widely used in climate simulation, agricultural monitoring, ecological environment assessment and other fields. Existing LAI inversion methods mainly rely on a single remote sensing data source or are based on empirical models, which are difficult to balance high spatial resolution and temporal and spatial continuity, and are prone to distortion under complex terrain and cloud cover conditions. Synthetic aperture radar (SAR), thermal infrared, visible and near-infrared multi-source remote sensing data have strong complementarity, but traditional fusion methods cannot fully exploit multi-source spatio-temporal information, and the model generalization ability is insufficient. With the development of deep learning and transfer learning technology, introducing it into LAI inversion can significantly improve the accuracy and stability, but so far there is no systematic global inversion framework, and it cannot dynamically adapt to the differences in multi-source data features of different vegetation types and terrain conditions. SUMMARY

[0003] In view of the problems of precision limitation, temporal and spatial discontinuity and insufficient model generalization ability of single source data in the prior art, the present application provides a vegetation leaf area index remote sensing inversion method combining multi-source remote sensing data fusion, transfer learning fine-tuning and time series deep learning, to realize global high-precision and automated LAI inversion.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0005] A ground vegetation leaf area index remote sensing inversion method based on remote sensing images, comprising the following steps:

[0006] (1) Data acquisition and preprocessing: acquiring multi-source remote sensing image data, including optical visible light band image, near-infrared band image, thermal infrared image and synthetic aperture radar image, and respectively performing radiation correction, atmospheric correction, geometric registration and cloud, cloud shadow and noise detection and removal processing on the multi-source images;

[0007] (2) Multi-scale spatio-temporal non-local filtering fusion: performing multi-scale pyramid decomposition on the preprocessed multi-source images, extracting image block features at different spatial scales and different time phases by constructing a spatio-temporal non-local filtering operator, and performing weighted fusion to obtain high-resolution, high-continuity fused images;

[0008] (3) Dynamic feature extraction: based on the fused images, spectral features, texture features and structural features are extracted in turn, and the importance of the features under different vegetation types and terrain environments is preliminarily evaluated to generate a feature vector set;

[0009] (4) Transfer learning fine-tuning: Based on the random forest or XGBoost model, transfer learning is used to fine-tune the base model by using a small number of ground-measured LAI samples from different ecological zones to dynamically update the model parameters and improve the model's generalization ability.

[0010] (5) Time series deep learning integration: Construct an SSA (singular spectrum analysis) module to decompose time series features, use a bidirectional time series convolution network BiTCN to extract long-term and short-term time series changes, and combine an attention mechanism to enhance key time series information, output the leaf area index prediction value.

[0011] (6) Model output and post-processing: The outputs of the transfer learning model and the time series deep learning model are combined through a weighted fusion strategy to generate the final LAI inversion results, and spatial continuity and temporal consistency verification is performed to output the global ground vegetation leaf area index distribution map.

[0012] Preferably, the cloud, cloud shadow and noise detection and removal process in step (1) includes a multi-level detection process combining spectral index discrimination based on threshold method, morphological filtering and deep learning semantic segmentation, and morphological dilation and spectral interpolation filling are performed on the detected cloud layer area to ensure the integrity and spatiotemporal continuity of the fused image.

[0013] Preferably, the multi-scale pyramid decomposition in step (2) uses a combination of Gaussian pyramid and Laplacian pyramid to decompose multi-layer images of different spatial resolutions, and based on an improved spatio-temporal non-local filtering algorithm, the similarity weight matrix between image blocks is calculated to jointly weight and fuse multi-source and multi-temporal features in the spatio-temporal dimension, to improve the leaf area index inversion accuracy in complex terrain areas.

[0014] Preferably, the texture feature extraction in step (3) uses a multi-scale texture analysis method combining gray level co-occurrence matrix, local binary pattern and wavelet transform, and dynamically adjusts the texture feature extraction window size according to vegetation type and slope information to realize adaptive extraction of features in different ecological zones such as alpine, arid, humid and tropical rainforest.

[0015] Preferably, in step (4), the transfer learning fine-tuning stage, first pre-train the base model on global typical ecological zone remote sensing-ground measurement paired samples, then use a small amount of target area field measurement data for incremental fine-tuning, and optimize the hyperparameters through Bayesian optimization or genetic algorithm to ensure the adaptability and stability of the model in the target area.

[0016] Preferably, the SSA module in step (5) first decomposes the time series features to obtain principal components and noise components; the BiTCN module includes a bidirectional convolution layer and a residual connection to capture temporal changes at different time scales; and the Attention mechanism learns the weights of key moments within a time window to respond to extreme weather or seasonal mutations in a timely manner, thereby improving the spatiotemporal continuity of LAI inversion.

[0017] Preferably, the model output is subjected to a post-processing stage, based on a dynamic feature weight distribution mechanism, to use a reinforcement learning framework or a meta-learning algorithm to adjust the weights of spectral, texture, structural and temporal features in the final fusion in real time, so as to adapt to the differences in the influence of different vegetation types, terrain undulations and seasonal changes on leaf area index.

[0018] Preferably, in step (6), the model output result is subjected to spatial kriging interpolation and temporal sequence smoothing filtering to generate a global leaf area index distribution map, which is cross-validated with ground observation station data and other remote sensing inversion results, and output includes MAE, RMSE and R 2 , and other precision evaluation index reports.

[0019] Preferably, the method further comprises a model uncertainty evaluation step: based on Monte Carlo dropout or an ensemble learning framework, the transfer learning and temporal deep network are randomly sampled multiple times to calculate the confidence interval and uncertainty distribution map of the leaf area index inversion result, to assist subsequent decision-making and risk assessment.

[0020] Preferably, the method can be integrated into a cloud-based remote sensing data processing platform to parallelly process multi-source image data through a distributed computing cluster, to realize efficient and automated LAI inversion of global different ecological zones, different resolutions and multi-temporal remote sensing images, and provide an API interface to support downstream applications such as climate models, agricultural monitoring and ecological assessment.

[0021] Compared with the prior art, the present application has the following advantages:

[0022] The present application realizes high-resolution and high-continuity vegetation LAI inversion through multi-source spatiotemporal fusion and dynamic feature distribution; transfer learning and temporal deep network enhance the adaptability of the model to new regions and temporal changes; and full-process automation and uncertainty evaluation ensure the stability and reliability of large-scale applications, which is significantly better than existing single-source or static models. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a schematic diagram of the overall process of the present application;

[0024] Figure 2 is an overall performance output precision report;

[0025] Figure 3 to assess results by season;

[0026] Figure 4 to assess performance for different vegetation types;

[0027] Figure 5 to assess quality control indicators;

[0028] Figure 6 to assess uncertainty. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0030] As shown in Figure 1 a remote sensing image-based ground vegetation leaf area index remote sensing inversion method, comprising the following steps:

[0031] (1) Data acquisition and preprocessing: acquiring multi-source remote sensing image data, including optical visible light band image, near-infrared band image, thermal infrared image and synthetic aperture radar image, and respectively performing radiation correction, atmospheric correction, geometric registration and cloud, cloud shadow and noise detection and removal processing on the multi-source image;

[0032] The cloud, cloud shadow and noise detection and removal processing in step (1) includes a multi-level detection process combining spectral index discrimination based on threshold method, morphological filtering and deep learning semantic segmentation, and morphological dilation and spectral interpolation filling are performed on the detected cloud layer area to ensure the integrity and spatiotemporal continuity of the fused image.

[0033] (2) Multi-scale spatiotemporal non-local filtering fusion: multi-scale pyramid decomposition is performed on the preprocessed multi-source image, and through constructing a spatiotemporal non-local filtering operator, image block features are extracted and weighted fused under different spatial scales and different time phases to obtain high-resolution and high-continuity fused image;

[0034] In step (2), the multi-scale pyramid decomposition adopts a method combining Gaussian pyramid and Laplacian pyramid, and multi-layer decomposition is performed on different spatial resolution images, and based on the improved spatiotemporal non-local filtering algorithm, the similarity weight matrix between image blocks is calculated to jointly weight and fuse multi-source multi-temporal features in the spatiotemporal dimension, so as to improve the leaf area index inversion accuracy in complex terrain areas.

[0035] (3) Dynamic feature extraction: Based on the fused image, spectral features, texture features, and structural features are extracted in turn, and the importance of the features under different vegetation types and terrain environments is preliminarily evaluated to generate a feature vector set;

[0036] In step (3), the texture feature extraction is achieved by a multi-scale texture analysis method combining a gray level co-occurrence matrix, a local binary pattern, and a wavelet transform, and the texture feature extraction window size is dynamically adjusted according to the vegetation type and slope information to realize adaptive extraction of features in different ecological regions such as high-cold, high-drought, humid, and tropical rainforest.

[0037] (4) Transfer learning fine-tuning: Based on a random forest or XGBoost model, transfer learning is performed on the base model by using a small amount of ground-measured LAI samples in different ecological regions to dynamically update the model parameters to improve the generalization ability of the model when applied in new regions.

[0038] In step (4), the transfer learning fine-tuning stage, the base model is first pre-trained on the remote sensing-ground measurement paired samples in global typical ecological regions, then incrementally fine-tuned using a small amount of field measurement data in the target region, and the hyperparameters are optimized through Bayesian optimization or genetic algorithm to ensure the adaptability and stability of the model in the target region.

[0039] (5) Time series deep learning integration: An SSA (singular spectrum analysis) module is constructed to decompose the time series features, a bidirectional time series convolution network BiTCN is used to extract long-term and short-term time series changes, and an Attention mechanism is used to enhance key time series information to output the leaf area index prediction value.

[0040] In step (5), the SSA module first decomposes the time series features to obtain principal components and noise components; the BiTCN module includes a bidirectional convolution layer and a residual connection to capture time series changes at different time scales; the Attention mechanism learns the weights of key moments within a time window to respond to extreme weather or seasonal mutations in time, thereby improving the spatiotemporal continuity of LAI inversion.

[0041] In the model output and post-processing stage, based on a dynamic feature weight distribution mechanism, a reinforcement learning framework or a meta-learning algorithm is used to adjust the weights of spectral, texture, structural, and time series features in the final fusion in real time to adapt to the differences in the influence of different vegetation types, terrain undulations, and seasonal changes on the leaf area index.

[0042] (6) Model output and post-processing: The outputs of the transfer learning model and the time series deep learning model are fused by a weighted fusion strategy to generate the final LAI inversion result, and the spatial continuity and temporal consistency are verified to output a global ground vegetation leaf area index distribution map.

[0043] In the step (6), the model output result is generated into a global leaf area index distribution map after spatial kriging interpolation and time series smoothing filtering, and is cross-verified with ground observation station data and other remote sensing inversion results, and outputs MAE, RMSE and R 2 and other precision evaluation indexes.

[0044] The method further comprises a model uncertainty evaluation step: based on Monte Carlo dropout or an integrated learning framework, the transfer learning and time series deep network are randomly sampled multiple times, the confidence interval and uncertainty distribution map of the leaf area index inversion result are calculated, to assist subsequent decision-making and risk assessment.

[0045] The method can be integrated into a cloud remote sensing data processing platform, and through a distributed computing cluster, multi-source image data is processed in parallel to realize efficient and automated LAI inversion of global different ecological zones, different resolutions and multi-temporal remote sensing images, and an API interface is provided to support downstream applications such as climate models, agricultural monitoring and ecological assessment.

[0046] Specific implementation steps

[0047] 1. Multi-source remote sensing data preprocessing and fusion

[0048] Obtain optical (10m, 30m), SAR (100m), thermal infrared and other multi-source remote sensing images;

[0049] Respectively perform radiation and atmospheric correction, geometric registration;

[0050] Multi-level cloud, shadow and noise detection and removal based on threshold method, morphological filtering and deep learning semantic segmentation;

[0051] Construct a Gaussian-Laplacian pyramid, calculate the inter-block similarity weight through a spatio-temporal non-local filtering operator, and weightedly fuse on different pyramid levels and time phases to obtain a fine-grained fused image.

[0052] 2. Dynamic feature extraction

[0053] Extract spectral, texture (gray level co-occurrence matrix, LBP, wavelet transform) and structural features from the fused image respectively;

[0054] According to the vegetation type, slope and terrain conditions, dynamically adjust the texture window and feature selection threshold to generate a joint feature vector set.

[0055] 3. Transfer learning fine-tuning

[0056] Pre-train a random forest or XGBoost base model on global typical ecological zones (alpine, arid, humid, tropical) paired remote sensing-ground measured LAI samples;

[0057] Incremental fine-tuning is performed using a small amount of field measurement data from the target area, and hyperparameters are searched through Bayesian optimization or genetic algorithms.

[0058] 4. Time Series Deep Learning Integration

[0059] Construct the SSA module to perform principal component decomposition on the original time series features and remove noise;

[0060] BiTCN is used to capture bidirectional temporal dependencies and combined with residual connections;

[0061] Introducing the Attention mechanism to output temporal feature representation according to the weights of key moments within the time window;

[0062] The SSA-BiTCN-Attention output is weightedly integrated with the transfer learning model results to generate preliminary LAI predictions.

[0063] 5. Dynamic feature weight allocation and post-processing

[0064] In the fusion stage, the weights of spectral, texture, structural and temporal features are adjusted in real time based on meta-learning or reinforcement learning frameworks;

[0065] Perform spatial Kriging interpolation and temporal smoothing filtering on the preliminary prediction results;

[0066] Cross-validate with ground station observations and other inversion results to form the final global LAI distribution map and output MAE, RMSE, R 2 Equal precision report;

[0067] Use Monte Carlo dropout or ensemble learning to perform uncertainty assessment and generate confidence intervals and uncertainty distributions.

[0068] 6. System integration and automation

[0069] The method is deployed on a cloud-based distributed computing platform to achieve automated processing of multi-temporal and multi-source images;

[0070] Provides standard API interfaces to support downstream applications such as climate models, agricultural monitoring, and ecological assessment.

[0071] Specific calculation algorithm

[0072] 1. Multi-source image preprocessing and fusion algorithm

[0073] (1) Radiation and atmospheric correction

[0074] Optical image correction uses the FLAASH model to calculate the path radiation and ground reflectivity of the input visible light / near-infrared image based on the surface atmospheric parameters (aerosol type, water vapor content):

[0075] L sensor (λ)=T(λ)L ground (λ)+L path (λ)

[0076] After the inverse solution, the surface reflectivity p(λ) is obtained:

[0077]

[0078] SAR radiation correction for Sentinel-1 SAR image, Gamma-MAP algorithm is applied to remove scatterer noise first, and then according to the terrain and transmission loss correction, the surface scattering coefficient σ is calculated 0 .

[0079] (2) Cloud, shadow and noise detection and filling

[0080] Multi-stage detection process

[0081] Threshold discrimination: calculate the normalized difference cloud index (NDCI) and shadow index (NDVI-NDWI combination) to preliminarily segment;

[0082] Morphological filtering: open and close operation is performed on the threshold result to remove small block artifacts;

[0083] Deep learning semantic segmentation: U-Net network is used for fine segmentation of cloud / shadow boundary, and the loss function adopts Dice+Cross-Entropy hybrid.

[0084] Interpolation filling: based on the adjacent time same position pixels and spatial similar pixels, the detected cloud / shadow pixels are filled by weighted interpolation:

[0085]

[0086] wherein, the weight w i =exp(-α||I(x,y,t)-I(x i ,y i ,t i )||).

[0087] (3) Multi-scale pyramid fusion

[0088] Pyramid decomposition: Gaussian pyramid and Laplacian pyramid are constructed for each source image I:

[0089] G0=I,G l+1 =Downsample(G l ),L l =G l -Upsample(G l+1 )

[0090] Temporal and spatial non-local filtering fusion on each layer Laplacian sub-band The similarity weight is calculated in the sliding block Ω p

[0091]

[0092] Then cross-layer and cross-phase weighted fusion:

[0093]

[0094] Finally, the fused image is reconstructed layer by layer:

[0095] 2. Dynamic feature extraction and adaptive selection

[0096] (1) Spectral index calculation

[0097] For the fused image Calculate the commonly used index:

[0098]

[0099] (2) Texture and structure features

[0100] Gray level co-occurrence matrix (GLCM): Calculate the matrix statistics such as energy, contrast, entropy, etc. in the window;

[0101] Wavelet transform: Perform two-dimensional wavelet decomposition on the image, and extract the high-frequency sub-band energy as the texture index;

[0102] LBP and SIFT: Extract local binary pattern histogram and key point descriptor at multiple scales.

[0103] (3) Feature adaptive selection

[0104] According to the pre-classified vegetation ecological types (alpine, arid, etc.), use the information gain ratio IG(R|F j ) and principal component analysis (PCA) to select the top K most discriminative features, forming a K-dimensional feature vector x.

[0105] 3. Transfer learning fine-tuning algorithm

[0106] (1) Pre-training of basic model

[0107] Take XGBoost as an example: loss function Where Ω is the regularization term, and the pre-training set is the paired samples of typical ecological zones around the world.

[0108] (2) Incremental fine-tuning

[0109] Target area small sample set ​Based on the basic model parameters, continue to iteratively minimize:

[0110]

[0111] At the same time, Bayesian optimization (Gaussian process modeling hyperparameter space) is used to automatically search for key hyperparameters such as learning rate and tree depth to ensure optimal performance on the validation set.

[0112] 4. Time Series Deep Learning Integration Algorithm

[0113] (1) SSA singular spectrum analysis

[0114] Construct the time series matrix X∈R L×M , perform singular value decomposition: Retain the first r principal components and reconstruct the denoised sequence

[0115] (2) Bidirectional Temporal Convolutional Network (BiTCN) - Network Layer Definition: Forward and Backward Convolution

[0116]

[0117] and merged into Add residual connections to stabilize training.

[0118] (3) Temporal Attention Mechanism

[0119] Calculating attention weights

[0120]

[0121] Get the moment weighted output c t =∑ i α t,i h i

[0122] (4) Integration and fusion

[0123] The output of BiTCN-Attention with fine-tuning XGBoost output Linear fusion:

[0124]

[0125] Where σ is the Sigmoid function, a and b are learnable parameters, and the fusion weights are adaptively adjusted according to different regions.

[0126] 5. Post-processing and uncertainty assessment

[0127] (1) Spatial Kriging interpolation

[0128]

[0129] (2) Time smoothing filter Cubic Exponential Smoothing:

[0130] S t = a y t + (1-a) (S t-1 + T t-1 ), T t = b (S t - S t-1 ) + (1-b) T t-1 .

[0131] (3) Uncertainty evaluation using Monte Carlo Dropout, multiple forward predictions Calculate the mean and variance as the confidence interval:

[0132]

[0133] Algorithm flow implementation

[0134] Step 1: Data acquisition and preprocessing

[0135] 1.1 Optical images (Sentinel-2 10m, Landsat-8 30m), Sentinel-1 SAR (VV / VH), MODIS thermal infrared, etc. Align by collection date;

[0136] 1.2 Radiometric and atmospheric correction: Apply FLAASH-based optical correction algorithm and Gamma-MAP-based SAR correction respectively;

[0137] 1.3 Geometric registration: Take SRTM30m DEM as the reference, use sub-pixel registration method based on phase cross-correlation and feature point matching to achieve error <0.5 pixels for each source image;

[0138] 1.4 Cloud and shadow detection: Use the cascade of normalized difference cloud index (NDCI), morphological filtering and U-Net semantic segmentation network to achieve detection accuracy of over 95%; The detected cloud / shadow area is filled by image similarity weighted interpolation to ensure time series continuity.

[0139] Step 2: Multi-scale pyramid fusion

[0140] 2.1 Construct a three-layer pyramid: Gaussian pyramid down-sampling and Laplacian pyramid decomposition for each source image;

[0141] 2.2 Spatio-temporal non-local filtering: Extract N x N pixel blocks by block at each pyramid level, calculate the weight matrix W through block similarity, and fuse temporal and multi-source features;

[0142] 2.3 Reconstructing the fused image: layer by layer up-fusion, output the final fine-grained fused image, which has a spatial resolution close to the finest image and takes into account multi-temporal information.

[0143] Step 3: Dynamic feature extraction and initial selection

[0144] 3.1 Spectral features: Calculate NDVI, EVI, SAVI, and more than ten indexes;

[0145] 3.2 Texture features: In the fused image, select the gray level co-occurrence matrix (GLCM) contrast, correlation, energy, entropy, and wavelet high-frequency sub-band energy;

[0146] 3.3 Structural features: Extract canopy structure information based on multi-scale SIFT features and local binary pattern (LBP);

[0147] 3.4 Feature initial selection and adaptation: Based on vegetation classification (alpine, arid, humid, tropical), use information gain and principal component analysis (PCA) to remove redundant features, and generate K-dimensional feature vectors.

[0148] Step 4: Fine-tuning of transfer learning and hyperparameter optimization

[0149] 4.1 In the paired samples of global typical ecological zones (more than ten sites), use XGBoost as the basis, and use 10,000 ground measured LAI samples for pre-training;

[0150] 4.2 Select 200-500 field measurement samples in the target area, fine-tune the model parameters using incremental learning, and search for tree depth, learning rate, and other hyperparameters through Bayesian optimization (Gaussian Process);

[0151] 4.3 The MAE of the fine-tuned model in the target area is less than 0.2, and R 2 >0.85.

[0152] Step 5: Time series deep learning integration

[0153] 5.1 SSA decomposition: Perform singular spectrum decomposition on the K-dimensional time series feature sequence, and select the first M principal components to reconstruct the denoised sequence;

[0154] 5.2 BiTCN architecture: Build a bidirectional time series convolution layer (kernel size = 3, dilation = 1, 2, 4), and introduce residual skip connection;

[0155] 5.3 Attention mechanism: After the BiTCN output, connect a time attention layer to dynamically allocate feature weights at each time;

[0156] 5.4 Output concatenation: The output of SSA–BiTCN–Attention is combined with the prediction result of the fine-tuned XGBoost model to obtain the final LAI prediction value through linear weighting.

[0157] Step 6: Post-processing and evaluation

[0158] 6.1 Spatial interpolation: Perform Kriging interpolation on the initial prediction to fill local holes;

[0159] 6.2 Time smoothing: Apply triple exponential smoothing filter to remove abnormal fluctuations;

[0160] 6.3 Cross-validation: Compare with ground stations and third-party LAI products (MODISLAI), and output MAE, RMSE, R 2 Report;

[0161] 6.4 Uncertainty assessment: Based on Monte Carlo dropout, multiple random predictions are made and confidence intervals are calculated.

[0162] Example 1

[0163] Complete implementation from remote sensing data to LAI inversion

[0164] In this case, we used a certain area (100 km2, including plains, hills, and low mountains) as the experimental object. The goal was to invert the vegetation LAI value in the study area by integrating multi-source remote sensing images, dynamically extracting features, and combining XGBoost and time series deep learning models.

[0165] 1. Data acquisition and preprocessing

[0166] 1.1 Data Selection

[0167] Select the following multi-source, time-series remote sensing data:

[0168] Optical data: Sentinel-2MSI (10m resolution, bands: red, NIR, blue).

[0169] Radar data: Sentinel-1 SAR (30m resolution, VV / VH polarization).

[0170] Land surface temperature: MODIS LST product (1000m resolution, twice daily observations).

[0171] 1.2 Data Correction and Alignment

[0172] Optical image radiation and atmospheric correction:

[0173] Method: Use ENVI FLAASH tool.

[0174] Input parameters: Sensor type: Sentinel-2 MSI; Atmospheric model: Mid-latitude summer; Aerosol model: Continental; Output: Surface reflectance image.

[0175] SAR radiometric correction:

[0176] Method: GMT SAR software was used.

[0177] Input parameters: Image type: SLC. Denoising filter: Gamma-MAP filter. Output: Backscatter coefficient (σ0).

[0178] Data geometric co-registration:

[0179] Registration reference data: SRTM DEM (30 m).

[0180] Accuracy: Sub-pixel registration (error < 0.5 pixels).

[0181] Cloud / shadow detection and filling:

[0182] Method: Fine-tuning of the U-Net model.

[0183] Detection accuracy: > 95%.

[0184] Interpolation method: Interpolation using NDVI and MODIS LST data from adjacent time periods.

[0185] 2. Multi-scale pyramid fusion

[0186] Temporal-spectral fusion of the above images with different resolutions:

[0187] Construction of the image resolution pyramid:

[0188] Levels: 3 levels (pyramid settings: optical, radar, temperature).

[0189] Method: Gaussian blur and Laplace decomposition were used.

[0190] Fusion formula:

[0191]

[0192] where the weight w s (x,y) is used to balance the contribution of different information sources. Dynamic weights are assigned based on the rate of feature change (high weight for optical images).

[0193] The output image resolution is the maximum resolution of Sentinel-2 (10 m), and the time coverage is monthly data for the year 2022.

[0194] 3. Dynamic feature extraction and selection

[0195] The following features are extracted in the fused image:

[0196] Spectral features

[0197] Compute basic vegetation indices, including NDVI, EVI and SAVI:

[0198]

[0199] Parameter settings:

[0200] RED: Sentinel-2 band B4 (665 nm).

[0201] NIR: Sentinel-2 band B8 (842 nm).

[0202] Texture features

[0203] Use the gray level co-occurrence matrix (GLCM) to extract texture statistics:

[0204] Energy E:

[0205]

[0206] Contrast C:

[0207]

[0208] Parameter settings:

[0209] Window size: 9x9 pixels.

[0210] Radar features, radar image VV and VH polarization ratio calculation:

[0211]

[0212] Temperature dynamic features, extract the daily variation range of MODIS LST (day minus night value).

[0213] Final output x = [NDVI, EVI, C, E, RPol, LST], forming the input features for our training model.

[0214] 4. Model training and inversion

[0215] 4.1 XGBoost model pre-training

[0216] Select global typical ecological zone data sets (alpine meadow, humid forest, etc.), construct a sample library (10000 sample points).

[0217] Model parameter settings:

[0218] Tree depth (D): 6.

[0219] Learning rate (η): 0.1.

[0220] Subsample ratio (γ): 0.8.

[0221] Objective loss function:

[0222]

[0223] 4.2 Incremental model fine-tuning

[0224] Incremental sample acquisition: Collect 200 distributed samples from the target area.

[0225] Sample label: Ground-truth LAI value.

[0226] Input incremental fine-tuning hyperparameter search:

[0227] Search range:

[0228] Tree depth (D): 5-10.

[0229] Learning rate (η): 0.01, 0.05, 0.1, 0.2.

[0230] Subsample ratio (γ): 0.6, 0.7, 0.8, 1.0.

[0231] Optimization method: Use Bayesian optimization (based on Gaussian process) to automatically explore the optimal hyperparameter combination to minimize the RMSE of the validation set.

[0232] Training output:

[0233] Extract the trained XGBoost model as the LAI basic inversion model.

[0234] Output MAE and R 2 : MAE: 0.15, R 2 : 0.90.

[0235] 5. Time series deep learning integration

[0236] To improve the accuracy of time series prediction, a deep learning module is added to model the apparent time series signal. 5.1 Dynamic denoising: SSA singular spectrum analysis

[0237] Construct a time series matrix:

[0238]

[0239] where x1, x2, …, x n represent the eigenvalues in the time series.

[0240] Singular value decomposition (SVD):

[0241] Singular value decomposition is performed on matrix X:

[0242]

[0243] The first r principal components are retained, and the components corresponding to small singular values are ignored, to reconstruct the denoised time series matrix:

[0244]

[0245] The output denoised reconstructed signal is used for subsequent modeling.

[0246] 5.2 BiTCN model construction

[0247] Network architecture:

[0248] Input components: denoised time series feature vectors.

[0249] Time series convolution (bidirectional): a bidirectional time convolution network (BiTCN) is used, and its output formula is:

[0250]

[0251] Merging output:

[0252]

[0253] Attention mechanism:

[0254] Dynamic weight allocation is performed on the time series:

[0255]

[0256] Weighted output:

[0257]

[0258] Output:

[0259] The model is mapped to the LAI value through a fully connected layer.

[0260] The loss function is MSE:

[0261]

[0262] 6. Fusion and final inversion

[0263] Fusion is realized based on the outputs of XGBoost and time series deep learning models:

[0264] Fusion formula: weight adaptive adjustment model contribution:

[0265]

[0266] The weight calculation is based on the MAE value (w1+w2=1):

[0267]

[0268] Verification results:

[0269] Three evaluation metrics are calculated on the test set:

[0270] Mean square error (MSE): 0.12.

[0271] Mean absolute error (MAE): 0.11.

[0272] Coefficient of determination (R 2 ):0.92.

[0273] 7. Post-processing and verification

[0274] 7.1 Spatial interpolation Since some pixels are missing, the Kriging interpolation method is used to fill the empty areas:

[0275]

[0276] Among them, the weights satisfy the Kriging conditions:

[0277]

[0278] Time smoothing uses a cubic exponential smoothing filter to further smooth the time series data:

[0279] S t =αy t +(1-α)(S t-1 +T t-1 )

[0280] Among them, the trend item is updated as follows:

[0281] T t =β(S t -S t-1 )+(1-β)T t-1

[0282] Uncertainty assessment uses Monte Carlo Dropout to detect prediction uncertainty:

[0283] Multiple forward sampling.

[0284] Output the mean and standard deviation to form a confidence interval:

[0285]

[0286] Results:

[0287] like Figure 2Overall performance output precision report is shown;

[0288] As Figure 3 Seasonal evaluation results are shown;

[0289] As Figure 4 Different vegetation type performance is shown;

[0290] As Figure 5 Quality control indicators are shown;

[0291] As Figure 6 Uncertainty evaluation is shown.

[0292] Overall evaluation:

[0293] The following advantages can be obtained from several sets of data tables:

[0294] High precision: test set R 2 reached 0.92;

[0295] Good stability: balanced performance in each season and vegetation type.

[0296] In this example, by integrating optical, radar and thermal infrared images, multi-scale fusion technology processes data, extracts spatio-temporal features, and combines XGBoost and bidirectional deep learning model, the high-precision inversion of vegetation LAI value in the study area is successfully realized. The main features of this method include dynamic feature selection, multi-source information fusion, regional adaptive model optimization, and rigorous uncertainty evaluation measures, which provides a complete technical route for similar tasks.

[0297] It can be understood that the present application is described through some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to these features and embodiments without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, these features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of the present application.

Claims

1. A remote sensing inversion method for ground vegetation leaf area index based on remote sensing images, characterized in that: The following steps are involved: (1) Data acquisition and preprocessing: Acquire multi-source remote sensing image data, including optical visible light band images, near infrared band images, thermal infrared images, and synthetic aperture radar images, and perform radiometric correction, atmospheric correction, geometric registration, and cloud, cloud shadow, and noise detection and removal on the multi-source images; (2) Multi-scale spatiotemporal non-local filtering fusion: The pre-processed multi-source images are subjected to multi-scale pyramid decomposition. By constructing a spatiotemporal non-local filtering operator, the image block features are extracted at different spatial scales and different phases and weighted fusion is performed to obtain a high-resolution and high-continuity fused image. (3) Dynamic feature extraction: Based on the fused image, spectral features, texture features, and structural features are extracted in sequence, and the importance of features under different vegetation types and terrain environments is preliminarily evaluated to generate a feature vector set; (4) Transfer learning fine-tuning: Using random forest or XGBoost as the basic model, transfer learning is performed on the basic model using a small number of ground-measured LAI samples from different ecological zones, and the model parameters are dynamically updated to improve the generalization ability of the model when applied to new areas; (5) Time series deep learning integration: Construct an SSA (singular spectrum analysis) module to decompose time series features, use a bidirectional temporal convolutional network (BiTCN) to extract long-term and short-term time series changes, and combine the attention mechanism to enhance key time series information and output the leaf area index prediction value; (6) Model output and post-processing: The outputs of the transfer learning model and the time series deep learning model are used to generate the final LAI inversion results through a weighted fusion strategy, and the spatial continuity and temporal consistency are verified to output a global ground vegetation leaf area index distribution map.

2. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: The cloud, cloud shadow and noise detection and removal process in step (1) includes a multi-stage detection process combining spectral index discrimination based on threshold method, morphological filtering and deep learning semantic segmentation, and morphological expansion and spectral interpolation filling are performed on the detected cloud area to ensure the integrity and spatiotemporal continuity of the fused image, which includes 1. Radiation and atmospheric correction Optical image correction uses the FLAASH model to calculate the path radiation and ground reflectivity of the input visible light / near-infrared image based on the surface atmospheric parameters (aerosol type, water vapor content): L sensor (λ)=T(λ)L ground (λ)+L path (l) After inverse solution, we get the surface reflectivity ρ(λ): SAR radiation correction: For Sentinel-1SAR images, the Gamma-MAP algorithm is applied to remove scatterer noise first, and then the surface scattering coefficient σ is calculated based on terrain and transmission loss correction. 0 ; 2. Cloud, shadow and noise detection and filling Multi-level detection process Threshold discrimination: Calculate the normalized difference cloud index (NDCI) and shadow index (NDVI-NDWI combination) for preliminary segmentation; Morphological filtering: perform opening and closing operations on the threshold results to remove small artifacts; Deep learning semantic segmentation: The U-Net network performs fine segmentation of cloud / shadow boundaries, and the loss function adopts a hybrid of Dice and Cross-Entropy. Interpolation fills the detected cloud / shadow pixels based on weighted interpolation of adjacent pixels at the same location and spatially similar pixels: Among them, the weight w i =exp(-α||I(x,y,t)-I(x i ,y i ,t i )||).

3. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: In the step (2), the multi-scale pyramid decomposition adopts a method combining Gaussian pyramid and Laplacian pyramid to perform multi-layer decomposition on images of different spatial resolutions, and based on the improved spatiotemporal non-local filtering algorithm, by calculating the similarity weight matrix between image blocks, the multi-source and multi-temporal features are jointly weighted and fused in the spatiotemporal dimension to improve the leaf area index inversion accuracy in complex terrain areas. Multi-scale pyramid fusion Pyramid decomposition constructs Gaussian pyramid and Laplacian pyramid for each source image I: G0=I,G l+1 =Downsample(G l ),L l =G l -Upsample(G l+1 ) Spatiotemporal non-local filtering fusion for each layer of Laplacian subband (source s, time phase t), in the sliding block Ω p Calculate similarity weight internally: Then cross-layer and cross-temporal weighted fusion: Finally, the fused image is reconstructed layer by layer:

4. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: In the step (3), the texture feature extraction is carried out by a multi-scale texture analysis method combining gray-level co-occurrence matrix, local binary pattern and wavelet transform, and the texture feature extraction window size is dynamically adjusted according to the vegetation type and slope information to achieve feature adaptive extraction of different ecological zones such as high cold, high drought, humid and tropical rain forests. Dynamic feature extraction and adaptive selection 1. Spectral index calculation for fused images Calculate common exponents:

2. Texture and structural characteristics Gray Level Co-occurrence Matrix (GLCM): Calculates matrix statistics such as energy, contrast, and entropy within the window; Wavelet transform: Perform two-dimensional wavelet decomposition on the image and extract high-frequency sub-band energy as texture index; LBP and SIFT: Extract local binary pattern histograms and key point descriptors at multiple scales.

3. Feature Adaptive Selection According to the pre-classified vegetation ecological type (alpine, drought, etc.), the information gain ratio IG (R|F j ) and principal component analysis (PCA) together to screen the top K most discriminative features to form a K-dimensional feature vector x.

5. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: In the transfer learning fine-tuning phase of step (4), the basic model is first pre-trained on remote sensing-ground measurement paired samples of typical ecological regions around the world, and then incremental fine-tuned using a small amount of field measurement data from the target area, and hyperparameters are optimized through Bayesian optimization or genetic algorithm to ensure the adaptability and stability of the model in the target area. Transfer learning fine-tuning algorithm 1. Basic model pre-training Taking XGBoost as an example: loss function Where Ω is the regularization term, and the pre-training set is paired samples of typical ecological zones around the world.

2. Incremental fine-tuning Small sample set in target area Based on the basic model parameters, continue to iteratively minimize: At the same time, Bayesian optimization (Gaussian process modeling hyperparameter space) is used to automatically search for key hyperparameters such as learning rate and tree depth to ensure the optimal performance of the evidence set.

6. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: In step (5), the SSA module first decomposes the time series features to obtain the principal components and noise components; the BiTCN module includes a bidirectional convolution layer and a residual connection to capture temporal changes at different time scales; the Attention mechanism achieves timely response to extreme weather or seasonal mutations by learning the weights of key moments within the time window, thereby improving the spatiotemporal continuity of LAI inversion; Time Series Deep Learning Ensemble Algorithm 1. SSA Singular Spectrum Analysis Construct the time series matrix X∈R L×M , perform singular value decomposition: Retain the first r principal components and reconstruct the denoised sequence 2. Bidirectional Temporal Convolutional Network (BiTCN) Network layer definition: forward and backward convolution and merged into Add residual connections to stabilize training.

3. Temporal Attention Mechanism Calculating attention weights Get the moment weighted output c t =∑ i α t,i h i 4. Integration and fusion The output of BiTCN-Attention with fine-tuning XGBoost output Linear fusion: ω=σ(aMAE ML -bMAE DL ), Where σ is the Sigmoid function, a and b are learnable parameters, and the fusion weights are adaptively adjusted according to different regions.

7. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: In the model output and post-processing stage, based on a dynamic feature weight allocation mechanism, a reinforcement learning framework or a meta-learning algorithm is used to adjust the weights of spectral, texture, structural and temporal features in the final fusion in real time to adapt to the differences in the impact of different vegetation types, terrain undulations and seasonal changes on leaf area index.

8. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: In step (6), the model output results are subjected to spatial kriging interpolation and time series smoothing filtering to generate a global leaf area index distribution map, which is cross-validated with ground observation station data and other remote sensing inversion results. The output includes MAE, RMSE and R 2 And other precision evaluation index reports, specific 1. Spatial Kriging Interpolation 2. Time smoothing filter triple exponential smoothing: S t =αy t +(1-α)(S t-1 +Tt -1 ),T t =β(S t -s t-1 )+(1-β)T t-1 。 3. Uncertainty assessment uses Monte Carlo Dropout and multiple forward predictions Compute the mean and variance as confidence intervals:

9. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: The method further includes a model uncertainty assessment step: based on a Monte Carlo dropout or ensemble learning framework, multiple random samplings are performed on the transfer learning and time series deep network to calculate the confidence interval and uncertainty distribution map of the leaf area index inversion results to assist subsequent decision-making and risk assessment.

10. The method for remote sensing inversion of ground vegetation leaf area index based on remote sensing images according to claim 1, characterized in that: The method can be integrated into a cloud-based remote sensing data processing platform. By processing multi-source image data in parallel through distributed computing clusters, it can achieve efficient and automated LAI inversion for remote sensing images of different ecological zones, different resolutions, and multiple temporal phases around the world. It also provides an API interface to support downstream applications such as climate modeling, agricultural monitoring, and ecological assessment.

Images