Soil moisture estimation method for irregular asynchronous multi-source remote sensing time series data

Abstract

The application belongs to the technical field of remote sensing application, and relates to a soil moisture estimation method for irregular asynchronous multi-source remote sensing time series data. The method comprises the following steps: acquiring multi-source remote sensing time series data, generating a binary time mask matrix, and extracting a soil moisture dynamic anomaly sequence; a sequence-to-sequence soil moisture estimation model is constructed to obtain pre-training model parameters; a dynamic mask-based mean square error loss is calculated, the model is fine-tuned, and a soil moisture anomaly field is output; a preset period mean field is reconstructed, and the soil moisture anomaly field is superimposed on the preset period mean field to obtain a spatiotemporally continuous high-resolution absolute soil moisture estimation result. The application solves the problem that high resolution and time continuity cannot be achieved simultaneously, enables the model to support missing data, eliminates the interference of systematic mean deviation on cross-region migration, significantly improves the cross-region migration capability of the model, and significantly improves the spatial reconstruction accuracy and generalization of soil moisture under sparse site conditions.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing application technology, specifically relating to a method for estimating soil moisture for irregular asynchronous multi-source remote sensing time-series data. Background Technology

[0002] Soil moisture is a core state variable in the land surface hydrological cycle, playing an irreplaceable role in agricultural drought monitoring, irrigation scheduling, water resource management, and climate change research. Obtaining continuous spatiotemporal soil moisture information at daily scales and 100-meter resolution is an urgent need in the fields of remote sensing and Earth system science. In recent years, with the operationalization of medium- and high-resolution satellites and the increasing maturity of global soil moisture products provided by passive microwave satellites, joint inversion of multi-source remote sensing data has become an important way to overcome the limitations of single-sensor capabilities. Data-driven methods based on deep learning have shown significant potential in this field, and by fusing multimodal remote sensing information such as optical, radar, and thermal infrared data, the estimation accuracy and spatiotemporal resolution of soil moisture can be improved to a certain extent.

[0003] However, existing technologies have the following drawbacks: (1) Downscaling methods rely on coarse-resolution satellite products as input, and inevitably inherit their systematic biases; although direct estimation methods have higher accuracy, they are constrained by the satellite revisit cycle and can only generate discrete time series, which cannot meet the needs of daily-scale continuous monitoring. (2) The revisit cycles of different satellite sensors vary, and the irregular loss of optical images caused by cloud and rain weather makes multi-source remote sensing data inherently asynchronous and incomplete in the time dimension. Existing methods generally use preprocessing methods such as time linear interpolation, image synthesis and missing data reconstruction to force multi-source time alignment, which not only introduces additional errors, but also causes a large amount of effective original observation information to be discarded and wasted. (3) Existing models are usually trained for specific regions. When applied to new regions, they require a large amount of local ground-based measurement data for retraining, making it difficult to achieve rapid deployment and promotion. (4) There is a systematic mean deviation between different sensors and stations. Existing methods are difficult to effectively separate dynamic changes from static background, resulting in poor consistency of cross-regional estimation.

[0004] Therefore, developing an estimation method that can directly process irregular asynchronous multi-source satellite observations, achieve diurnal temporal continuity, and is independent of coarse-resolution products during inference while possessing cross-regional migration capabilities is of significant scientific importance and application value. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a spatiotemporally continuous high-resolution soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data. This method addresses the technical problems of existing methods, such as the inability to simultaneously achieve high resolution and temporal continuity, the inability to natively process irregular asynchronous observations, limited cross-regional transferability, and the difficulty in eliminating mean bias. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data includes: Multi-source remote sensing time-series data were acquired and resampled to construct a unified time grid. Binary time mask matrices were generated for each of the multi-source remote sensing time-series data to characterize whether there were valid observations at the corresponding time. Soil moisture time-series data from ground-based measured stations in the target domain were collected, and the average value for each calendar month was calculated annually according to a preset period. Dynamic anomaly sequences of soil moisture were extracted. The multi-source remote sensing time-series data included radar image data, satellite optical image data, and meteorological reanalysis data. A sequence-to-sequence soil moisture estimation model was constructed. A proxy soil moisture product was used as a supervision label to pre-train the sequence-to-sequence soil moisture estimation model and obtain the pre-trained model parameters. The sequence-to-sequence soil moisture estimation model includes three encoders corresponding to different data sources and a decoder for fusing multiple encoded features. Load the pre-trained model parameters, divide the target area into training and validation sets based on ground measurement stations for spatial cross-validation, generate a dynamic mask based on the location of effective observations, calculate the mean square error loss based on the dynamic mask based on the error between the predicted result and the true value based on the location of effective observations, perform transfer fine-tuning on the sequence-to-sequence soil moisture estimation model using ground measurement station data in the target area, and output the soil moisture anomaly field in the target area using the transferred fine-tuned sequence-to-sequence soil moisture estimation model. Based on the preset periodic mean and external drift covariates, the preset periodic mean field of the target area is reconstructed using a spatial interpolation method. The soil moisture anomaly field is then superimposed with the preset periodic mean field to obtain the spatiotemporal continuous high-resolution absolute soil moisture estimation result of the target area.

[0006] In some optional embodiments, the length of the uniform time grid is 30-50 days; radar image data includes dual-polarization data, which is resampled to 10 meters after radiometric calibration and topographic radiometric correction; dual-polarization data includes VV polarization data and VH polarization data; satellite optical image data is resampled to 100 meters after atmospheric correction; meteorological reanalysis data is bilinearly interpolated from the original resolution of 0.1° to a 100-meter spatial grid.

[0007] In some optional embodiments, a value of 1 in the binary time mask indicates that there is a valid observation at the corresponding time, and a value of 0 in the binary time mask indicates that there is a missing observation at the corresponding time. The binary time mask is used to mask the encoder input, setting the input feature at the corresponding time with a value of 0 to zero before sending it to the encoder. When collecting the soil moisture time series of ground measurement stations in the target domain and calculating the average value by calendar month for each year, the average value is obtained by taking the average value of each calendar month for each year and then calculating the average value of each month.

[0008] In some optional embodiments, radar image data is used as the data source. A lightweight convolutional neural network is used to project local features, followed by a multi-head self-attention module.

[0009] In some optional embodiments, each encoder is used to extract temporal coding features from different data sources; the decoder fuses the temporal coding features from different data sources through a cross-attention mechanism, a feature splicing mechanism, and a weighted fusion mechanism, and outputs a soil moisture dynamic anomaly prediction sequence.

[0010] In some optional embodiments, when pre-training the sequence-to-sequence soil moisture estimation model, at least one of the mean absolute error loss function or the Huber loss function is used.

[0011] In some optional embodiments, the mean squared error loss of the dynamic mask is calculated only at the effective surface observation time step at the center of the time window; assuming Indicates the first time within the time window At each time step, the binary validity mask is: , These are the predicted soil moisture anomalies. To measure abnormal soil moisture values, If it is a loss, then: .

[0012] In some alternative embodiments, the sequence-to-sequence soil moisture estimation model is transferred and fine-tuned by using a full-parameter fine-tuning approach to update all parameters of the pre-trained model, or by using a low-rank adaptive fine-tuning approach to freeze the backbone parameters of the pre-trained model and update only the additional trainable parameters.

[0013] In some alternative embodiments, the external drift covariate is at least one of multidimensional embedding features, topographic factors, land use types, vegetation indices, and soil property indicators.

[0014] In some optional embodiments, multi-dimensional embedding features of the Earth base model are extracted at each site location, Pearson correlation coefficients of each dimension with the site's preset periodic mean soil moisture are calculated, and the effective number of drift dimensions and the family of variogram models are jointly selected through leave-one-out cross-validation. An upper limit is set for the effective number of drift dimensions, and the soil moisture anomaly field is superimposed with the preset periodic mean field to obtain the spatiotemporal continuous high-resolution absolute soil moisture estimation results for the target area.

[0015] The beneficial effects of this invention are: (1) This invention redefines soil moisture time series estimation as a sequence-to-sequence learning problem. By constructing a sequence-to-sequence soil moisture estimation model, it directly outputs a daily continuous time series, which solves the problem that high resolution and time continuity cannot be achieved simultaneously. No coarse resolution product is required as input during inference, thus avoiding the inherent systematic bias of downscaling methods. (2) This invention enables the model to natively support missing data at both the input and supervision ends by using the mean square error loss of binary time mask and dynamic mask, without the need for preprocessing and interpolation of input data, and makes full use of all effective original observation information; (3) This invention uses an anomaly decomposition strategy to split soil moisture estimation into two independent sub-tasks: dynamic anomaly prediction and static mean reconstruction, thereby eliminating the interference of systematic mean deviation on cross-regional migration. Combined with a two-stage training scheme of large-scale proxy pre-training and small sample migration fine-tuning, only a small amount of local ground measurement data is needed to complete high-quality regional adaptation, significantly improving cross-regional migration capability. (4) This invention uses the Earth basic model embedding vector as an external drift covariate, which can implicitly encode static environmental factors that affect the spatial distribution of soil moisture background values, such as vegetation cover, topographic relief, soil texture and land use. Compared with traditional auxiliary variables, it has a stronger spatial semantic expression ability and significantly improves the spatial reconstruction accuracy and generalization of the mean field under sparse site conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of a soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data provided in Embodiment 1 of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0018] Example 1 As an example, in order to solve the problems existing in the prior art, this embodiment provides a soil moisture estimation method for irregular asynchronous multi-source remote sensing time series data.

[0019] The implementation details of the method in this embodiment are described below. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0020] The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data in this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. (See attached...) Figure 1 As shown, the method provided in this embodiment includes steps 110-140.

[0021] Step 110: Acquire multi-source remote sensing time-series data and resample them to construct a unified time grid. Generate binary time mask matrices for each of the multi-source remote sensing time-series data to characterize whether there are valid observations at the corresponding time. Collect soil moisture time-series data from ground-based measured stations in the target domain and calculate the preset periodic mean value for each year according to calendar months. Extract dynamic anomaly sequences of soil moisture. The multi-source remote sensing time-series data includes radar image data, satellite optical image data, and meteorological reanalysis data.

[0022] In some optional embodiments, the length of the uniform time grid is 30-50 days; radar image data includes dual-polarization data, which is resampled to 10 meters after radiometric calibration and topographic radiometric correction; dual-polarization data includes VV polarization data and VH polarization data; satellite optical image data is resampled to 100 meters after atmospheric correction; meteorological reanalysis data is bilinearly interpolated from the original resolution of 0.1° to a 100-meter spatial grid.

[0023] In some optional embodiments, a value of 1 in the binary time mask indicates that there is a valid observation at the corresponding time, and a value of 0 in the binary time mask indicates that there is a missing observation at the corresponding time. The binary time mask is used to mask the encoder input, setting the input feature at the corresponding time with a value of 0 to zero before sending it to the encoder. When collecting the soil moisture time series of ground measurement stations in the target domain and calculating the average value by calendar month for each year, the average value is obtained by taking the average value of each calendar month for each year and then calculating the average value of each month.

[0024] For example: Sentinel-1 C-band SAR imagery of the study area, including VV and VH polarization data, was acquired and resampled to 10 meters after radiometric calibration and topographic radiometric correction. Sentinel-2 multispectral surface reflectance products were resampled to 100 meters. ERA5-Land daily meteorological data were bilinearly interpolated from 0.1° to a 100-meter spatial grid. A unified time grid was constructed with a 50-day time window and a daily step size. Binary time mask matrices were generated for Sentinel-1 and Sentinel-2 respectively; ERA5-Land did not require a mask due to its temporal continuity. For the measured time series of each ground station, the average of all valid observations for each calendar month was calculated for each year, and then the average of the twelve-month averages was taken to obtain the preset periodic average, i.e., the long-term average of the stations.

[0025] set up For the first The number of effective observations in a month For the mth month One valid soil moisture value, It is a dynamic abnormal sequence. This is the original measured time series. If the long-term average of the site is given, then: .

[0026] Subtracting the long-term mean of the stations from the original measured time series yields the dynamic anomaly sequence: This serves as the monitoring target during the fine-tuning phase.

[0027] Step 120: Construct a sequence-to-sequence soil moisture estimation model. Use proxy soil moisture products as supervision labels to pre-train the sequence-to-sequence soil moisture estimation model and obtain the pre-trained model parameters. The sequence-to-sequence soil moisture estimation model includes three encoders corresponding to different data sources and a decoder for fusing multiple encoded features.

[0028] In some optional embodiments, encoders whose data source is radar image data all employ a lightweight convolutional neural network to project local features followed by a multi-head self-attention module.

[0029] In some optional embodiments, each encoder is used to extract temporal coding features from different data sources; the decoder fuses the temporal coding features from different data sources through a cross-attention mechanism, a feature splicing mechanism, and a weighted fusion mechanism, and outputs a soil moisture dynamic anomaly prediction sequence.

[0030] Specifically, each data source is independently encoded by its corresponding encoder to obtain a temporal embedding sequence with a unified dimension. The decoder concatenates all encoder outputs along the sequence dimension to construct a unified multi-source context representation. At each decoding time step, the decoder uses the current hidden state as the query and calculates the attention distribution in the multi-source context through a multi-head cross-attention mechanism, adaptively extracting the cross-source and cross-time step features most relevant to the current prediction target. The distribution of attention weights implicitly reflects the differences in contributions from each data source at different prediction times. The fusion process is entirely data-driven and requires no manual weight specification. The hidden state after cross-attention fusion is mapped to the output layer via a feedforward network to generate a dynamic anomaly prediction sequence for soil moisture.

[0031] In some optional embodiments, when pre-training the sequence-to-sequence soil moisture estimation model, at least one of the mean absolute error loss function or the Huber loss function is used.

[0032] The Sentinel-1 encoder employs a lightweight convolutional neural network with an output embedding dimension of 256, followed by an 8-layer multi-head self-attention module with 8 attention heads. The feedforward network has a dimension of 512, and pre-layer normalization is used to ensure gradient stability. The input features at each time step are added to the temporal location encoding and then multiplied by the corresponding mask.

[0033] set up The input features at time t are , The time position encoding of the moment is The binary validity mask is , The input feature vector after time masking is ,but: .

[0034] The attention mask is derived from the binary temporal mask, assigning a negative infinity bias to the missing time steps. Let the attention mask matrix be... The query matrix of the attention mechanism is The key matrix of the attention mechanism is The value matrix of the attention mechanism is , The key matrix is ​​transpose, and the query dimension is... , The softmax activation function is used. Let the output feature matrix of the attention mechanism be: .

[0035] Sentinel-2 has a similar encoder structure to ERA5-Land but with a shallower layer count (2 encoder stack layers and 4 multi-head self-attention heads). The three encoders have independent parameters and output memory sequences separately. The multi-source fusion decoder contains 6 encoder stack layers. It uses a learnable estimation query matrix as initial input, passes through self-attention, and then sequentially performs cross-attention fusion with the memory sequences from the three encoders. The output is then processed through a linear projection layer. Soil moisture anomaly prediction sequences at 1 time step, targeting the surface layer (0-10 cm) and the rhizosphere layer (0-50 cm), respectively. Let the learnable estimation query matrix be... The maximum time window length supported by the model is The model has a unified embedding dimension of 1. ,but: .

[0036] Using SMAP Level 4 (SMS L4) soil moisture monitoring products as surrogate supervisory targets, pre-training was conducted over a large spatial domain in the North China Plain. The surrogate products served only as supervisory labels and were not input into the model. The pre-training loss function employed mean absolute error to enhance robustness against label noise introduced by resolution mismatch. The AdamW optimizer was used, with an initial learning rate parameter of [parameter missing]. The weight decay parameter is The batch size is 64. After training to convergence, all model weights are saved as initialization parameters for transfer fine-tuning.

[0037] Step 130: Load the pre-trained model parameters, divide the target area into training and validation sets based on ground measurement stations for spatial cross-validation, generate a dynamic mask based on the location of effective observations, calculate the mean squared error loss based on the dynamic mask based on the error between the predicted and true values ​​of the effective observation locations, perform transfer fine-tuning on the sequence-to-sequence soil moisture estimation model using ground measurement station data in the target area, and output the soil moisture anomaly field in the target area using the transferred fine-tuned sequence-to-sequence soil moisture estimation model.

[0038] In some optional embodiments, the mean squared error loss of the dynamic mask is calculated only at the effective surface observation time step at the center of the time window; assuming Indicates the first time within the time window At each time step, the binary validity mask is: , For predicted soil moisture anomalies, To measure abnormal soil moisture values, If it is a loss, then: .

[0039] Specifically, load the pre-trained model weights and execute the space on a site-by-site basis. Folded cross-validation (k=3 or 5) ensures that all data from the same site belong to the same fold, preventing information leakage caused by spatial autocorrelation. The fine-tuning loss function uses the dynamic mask mean square error, and the loss calculation is limited to the effective observation time step within the center of the time window (30 days). Let the dynamic mask mean square error loss during the fine-tuning stage be... , Indicates the first time within the time window Each time step For predicted soil moisture anomalies, To determine the measured soil moisture anomaly, then: .

[0040] In some alternative embodiments, when performing transfer fine-tuning on the sequence-to-sequence soil moisture estimation model, a full-parameter fine-tuning approach is used to update all parameters of the pre-trained model, or a low-rank adaptive fine-tuning approach is used to freeze the backbone parameters of the pre-trained model and only update the additional trainable parameters.

[0041] Full parameter fine-tuning updates all model parameters, suitable for scenarios with a relatively large sample size. Low-rank adaptive freezing of pre-trained weights injects only low-rank decomposition matrices with rank 8 into the query and value matrices of each attention layer, allowing for approximately 2% of the number of trainable parameters compared to full parameter fine-tuning, making it suitable for scenarios with a very small number of sites. Both methods can be completed on a personal computer within minutes, using unbiased root mean square error as the primary evaluation metric. Let the unbiased root mean square error be... , For the first Measured soil moisture values ​​at each time step. This is the long-term average of the site. These are the predicted soil moisture values ​​for the corresponding time steps. Let be the time-series mean of the predicted soil moisture values, then: .

[0042] Step 140: Based on the preset periodic mean and external drift covariates, the preset periodic mean field of the target area is reconstructed using a spatial interpolation method, and the soil moisture anomaly field is superimposed with the preset periodic mean field to obtain the spatiotemporal continuous high-resolution absolute soil moisture estimation result of the target area.

[0043] In some alternative embodiments, the external drift covariate is at least one of multidimensional embedding features, topographic factors, land use types, vegetation indices, and soil property indicators.

[0044] In some optional embodiments, multi-dimensional embedding features of the Earth base model are extracted at each site location, Pearson correlation coefficients of each dimension with the site's preset periodic mean soil moisture are calculated, and the effective number of drift dimensions and the family of variogram models are jointly selected through leave-one-out cross-validation. An upper limit is set for the effective number of drift dimensions, and the soil moisture anomaly field is superimposed with the preset periodic mean field to obtain the spatiotemporal continuous high-resolution absolute soil moisture estimation results for the target area.

[0045] Extract the 64-dimensional embedding vectors of the Alpha Earth baseline model at each site location (aggregated from the original 10-meter vector to 100-meter vectors). Calculate the Pearson correlation coefficients between each dimension and the mean soil moisture at each site. Use leave-one-out cross-validation to jointly select the optimal combination among the effective drift dimension and three families of variogram models (spherical, exponential, and Gaussian). Set an upper limit on the effective drift dimension to prevent overfitting. Let... For Kriging weights, For the selected Alpha Earth basic model, the drift covariates are... For regression coefficients, the external drift kriging estimator is... , To represent the measured mean soil moisture, the external drift kriging estimator is expressed as: .

[0046] Let the spatiotemporal continuous absolute soil moisture be... The soil moisture anomaly field is The long-term mean field is The final spatiotemporal continuous absolute soil moisture output is: .

[0047] This invention redefines soil moisture time-series estimation as a sequence-to-sequence learning problem and natively supports irregular asynchronous multi-source inputs at the model architecture level, thereby solving the problem of time continuity. Secondly, this invention decomposes absolute value estimation into two independent subtasks, dynamic anomaly prediction and static mean reconstruction, through anomaly decomposition, eliminating the interference of systematic mean deviation on model cross-regional migration, thereby solving the problem of cross-regional migration consistency.

[0048] At the input representation level, this invention uses a fixed-length time grid to uniformly represent all data sources and explicitly marks the validity of each data source at each time point with a binary time mask. This mask directly participates in the encoder attention calculation, masking missing moments. At the same time, a masking mechanism is also introduced at the output end, applying supervisory constraints only to valid measured moments. This achieves native tolerance for missing data on both the input and supervisory ends of the model, giving the model a natural tolerance for irregular observations at the architectural level, rather than relying on external completion algorithms. At the learning objective level, soil moisture is decomposed into dynamic anomaly components and station-specific long-term means. The model only learns to predict the former, thereby eliminating the interference of systematic mean deviations between different sensors and stations on cross-regional migration consistency. The mean field is reconstructed through an independent spatial interpolation step. The decoupled design of the two sub-tasks allows dynamic anomaly prediction and mean field reconstruction to be optimized independently without interference, and supports independent replacement and upgrading of their respective modules. At the training strategy level, a two-stage approach of surrogate pre-training and small-sample transfer fine-tuning is adopted. In the pre-training stage, a preliminary representation of surface dynamics is established using a large-scale coarse-resolution product anomaly sequence, enhancing the model's generalization ability to surface moisture processes in different climate zones. In the fine-tuning stage, the mean square error loss of dynamic masking is fully utilized to make full use of sparse ground-based measured records, achieving high-quality regional adaptation with very few local samples. At the mean field reconstruction level, the Earth base model embedding vector is introduced into the Kriging interpolation framework as an external drift covariate. This can implicitly encode static environmental factors that affect the spatial distribution of soil moisture background values, such as vegetation cover, topographic relief, soil texture, and land use. Compared with traditional auxiliary variables, it has a stronger spatial semantic expression ability, significantly improving the spatial reconstruction accuracy and generalization of the mean field under sparse site conditions.

[0049] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating soil moisture based on irregular asynchronous multi-source remote sensing time-series data, characterized in that, include: Multi-source remote sensing time-series data were acquired and resampled to construct a unified time grid. Binary time mask matrices were generated for each of the multi-source remote sensing time-series data to characterize whether there were valid observations at the corresponding time. Soil moisture time-series data from ground-based measured stations in the target domain were collected, and the average value for each calendar month was calculated annually according to a preset period. Dynamic anomaly sequences of soil moisture were extracted. The multi-source remote sensing time-series data included radar image data, satellite optical image data, and meteorological reanalysis data. A sequence-to-sequence soil moisture estimation model was constructed. A proxy soil moisture product was used as a supervision label to pre-train the sequence-to-sequence soil moisture estimation model and obtain the pre-trained model parameters. The sequence-to-sequence soil moisture estimation model includes three encoders corresponding to different data sources and a decoder for fusing multiple encoded features. Load the pre-trained model parameters, divide the target area into training and validation sets based on ground measurement stations for spatial cross-validation, generate a dynamic mask based on the location of effective observations, calculate the mean square error loss based on the dynamic mask based on the error between the predicted result and the true value based on the location of effective observations, perform transfer fine-tuning on the sequence-to-sequence soil moisture estimation model using ground measurement station data in the target area, and output the soil moisture anomaly field in the target area using the transferred fine-tuned sequence-to-sequence soil moisture estimation model. Based on the preset periodic mean and external drift covariates, the preset periodic mean field of the target area is reconstructed using a spatial interpolation method. The soil moisture anomaly field is then superimposed with the preset periodic mean field to obtain the spatiotemporal continuous high-resolution absolute soil moisture estimation result of the target area.

2. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, The uniform time grid length is set to 30-50 days; radar image data includes dual-polarization data, which is resampled to 10 meters after radiometric calibration and topographic radiometric correction; dual-polarization data includes VV polarization data and VH polarization data; satellite optical image data is resampled to 100 meters after atmospheric correction; meteorological reanalysis data is bilinearly interpolated from the original resolution of 0.1° to a 100-meter spatial grid.

3. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, A value of 1 in the binary time mask indicates that a valid observation exists at the corresponding time, while a value of 0 indicates that an observation is missing at the corresponding time. The binary time mask is used to mask the encoder input, setting the input features at the corresponding time with a value of 0 to zero before sending them to the encoder. When collecting the soil moisture time series of ground measurement stations in the target domain and calculating the average value by calendar month for each year, the average value is obtained by taking the average value of each calendar month for each year and then calculating the average value of each month.

4. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, The encoder, whose data source is radar image data, uses a lightweight convolutional neural network to project local features and then connects to a multi-head self-attention module.

5. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, Each encoder is used to extract temporal coding features from different data sources; The decoder fuses temporal coding features from different data sources through cross-attention, feature splicing, and weighted fusion mechanisms to output a predictive sequence of dynamic anomalies in soil moisture.

6. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, When pre-training the sequence-to-sequence soil moisture estimation model, at least one loss function is used, either the mean absolute error loss function or the Huber loss function.

7. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, The mean square error loss of the dynamic mask is calculated only at the effective surface observation time step in the center of the time window; let... Indicates the first time within the time window At each time step, the binary validity mask is: , These are the predicted soil moisture anomalies. To measure abnormal soil moisture values, If it is a loss, then: 。 8. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, When performing transfer fine-tuning on sequence-to-sequence soil moisture estimation models, either full-parameter fine-tuning is used to update all parameters of the pre-trained model, or low-rank adaptive fine-tuning is used to freeze the backbone parameters of the pre-trained model and only update the additional trainable parameters.

9. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, The external drift covariates are at least one of the following: multidimensional embedding features, topographic factors, land use type, vegetation index, and soil property index.

10. The soil moisture estimation method for irregular asynchronous multi-source remote sensing time-series data according to claim 1, characterized in that, Multidimensional embedding features of the Earth base model are extracted at each site location. Pearson correlation coefficients between each dimension and the site's preset periodic mean soil moisture are calculated. The effective number of drift dimensions and the family of variogram models are jointly selected through leave-one-out cross-validation. An upper limit for the effective number of drift dimensions is set. The soil moisture anomaly field is superimposed with the preset periodic mean field to obtain the spatiotemporal continuous high-resolution absolute soil moisture estimation results for the target area.

Images