A training method for ignoring cloud interference in remote sensing deep learning parameter inversion

Abstract

The application discloses a training method for ignoring cloud interference in remote sensing deep learning parameter inversion, and relates to the field of geospatial artificial intelligence. The method comprises the following steps: performing cloud identification on an original remote sensing image to generate cloud mask data, and rasterizing ground measured data into label data matched with the remote sensing image in space; marking the cloud-shielded area in the label data by using the cloud mask data to obtain cloud mask label data; constructing a remote sensing deep learning parameter inversion model, inputting the original remote sensing image for prediction, and outputting a parameter prediction map; calculating the loss of cloud interference between the parameter prediction map and the cloud mask label data based on a cloud mask loss function; and optimizing the model parameters through gradient back propagation. The method solves the technical problem that the existing training method cannot effectively deal with the cloud interference problem, thereby limiting the inversion accuracy, and effectively deals with the interference problem of cloud shielding on model training, thereby improving the inversion accuracy.

Description

Technical Field

[0001] This application relates to the field of geospatial artificial intelligence (GeoAI), and in particular to a training method that ignores cloud interference in remote sensing deep learning parameter inversion. Background Technology

[0002] Remote sensing deep learning parameter inversion refers to the process of using deep learning technology to infer physical, chemical, or biological parameters of the Earth's surface or atmosphere from remote sensing data. It is a key research direction in geospatial artificial intelligence (GeoAI) and has significant application value in environmental monitoring, agricultural monitoring, and hydrological monitoring. Currently, the main methods to address cloud interference in remote sensing deep learning parameter inversion are to first convert image data into tabular data and remove cloud-containing data, or to transform continuous value prediction regression tasks into discrete value prediction image segmentation tasks. However, existing methods, due to the disruption of image data structure leading to limited feature extraction, or the loss of numerical accuracy caused by converting continuous values ​​to discrete values, cannot effectively handle cloud interference, thus limiting the accuracy and efficiency of remote sensing deep learning parameter inversion.

[0003] Currently, in related technologies, training methods that ignore cloud interference in remote sensing deep learning parameter inversion have the technical problem of being unable to effectively handle cloud interference, resulting in limited inversion accuracy. Summary of the Invention

[0004] This application provides a training method for remote sensing deep learning parameter inversion that ignores cloud interference. The method involves generating cloud masks by identifying clouds in remote sensing images, rasterizing ground-based measured data into matching label data, using cloud masks to mark cloud-occluded areas in the label data, improving the image segmentation network to construct a parameter inversion model, inputting remote sensing images for prediction, and outputting a parameter prediction map. The mean squared error loss function is improved to a cloud mask ignoring loss function. The cloud interference ignoring loss between the prediction map and the label data is calculated. Based on this loss, model parameters are optimized through gradient backpropagation to complete model training. These techniques solve the problem of existing remote sensing deep learning parameter inversion training methods that fail to effectively handle cloud interference, leading to limited inversion accuracy. This method effectively addresses the interference of cloud occlusion on model training, improving inversion accuracy.

[0005] This application provides a training method for ignoring cloud interference in remote sensing deep learning parameter inversion, comprising: performing cloud identification on the original remote sensing image to generate cloud mask data; rasterizing the ground measured data into label data spatially matched with the remote sensing image; using the cloud mask data to mark the cloud-occluded areas in the label data to obtain cloud mask label data; improving the pre-trained image segmentation network to construct an image-based remote sensing deep learning parameter inversion model; inputting the original remote sensing image into the remote sensing deep learning parameter inversion model for prediction and outputting a parameter prediction map; improving the mean square error loss function to obtain a cloud mask ignoring loss function; calculating the cloud interference ignoring loss between the parameter prediction map and the cloud mask label data based on the cloud mask ignoring loss function; and optimizing the model parameters through gradient backpropagation based on the cloud interference ignoring loss to complete the training of the remote sensing deep learning parameter inversion model.

[0006] In a possible implementation, cloud identification is performed on the original remote sensing image to generate cloud mask data, and the following processing is performed: determining whether the original remote sensing image has a data quality assessment band; if so, cloud pixels are identified based on the cloud information in the data quality assessment band to obtain cloud mask data; if not, cloud identification is performed on the original remote sensing image through a segmentation model to obtain cloud mask data.

[0007] In a possible implementation, the ground-measured data is rasterized into label data that spatially matches the remote sensing image, and the following processing is performed: the ground-measured data is converted into vector spatial data through coordinate resolution and data format conversion; geospatial interpolation is performed on the vector spatial data to obtain continuous field data; the raster grid is determined with the spatial extent of the original remote sensing image as a constraint, and the raster cell is determined with the resolution of the original remote sensing image as a constraint; the continuous field data is rasterized using the raster grid and the raster cells to obtain label data.

[0008] In a possible implementation, the pre-trained image segmentation network is improved to construct an image-based remote sensing deep learning parameter inversion model, and the following processing is performed: obtaining the pre-trained image segmentation network; deleting the terminal network based on the pre-trained image segmentation network to obtain a pruned image segmentation network; and setting a preset output convolutional layer, which is connected to the pruned image segmentation network to obtain the remote sensing deep learning parameter inversion model.

[0009] In a possible implementation, the mean squared error loss function is improved to obtain a cloud mask ignoring loss function, and the following processing is performed: obtain the mean squared error loss function; modify the mean squared error loss function into a piecewise function to obtain the cloud mask ignoring loss function.

[0010] In a possible implementation, the following processing is performed: the expression for the cloud mask neglect loss function is as follows:

[0011]

[0012] Wherein, CM-MSE represents the cloud masking neglect loss function, H represents the total number of rows in the original remote sensing image, W represents the total number of columns in the original remote sensing image, h represents the row number of the current pixel, w represents the column number of the current pixel, and y h,w This represents the parameter label value at position (h, w) in the cloud mask label data. The parameter prediction value at position (h, w) in the parameter prediction graph represents the model prediction parameter value, and the ignore_index represents the value marked as the cloud-occluded area in the cloud mask label data.

[0013] In a possible implementation, based on the ignored cloud interference loss, the model parameters are optimized through gradient backpropagation to complete the training of the remote sensing deep learning parameter inversion model, and the following processing is performed: the gradient of the ignored cloud interference loss with respect to the parameters of the remote sensing deep learning parameter inversion model is calculated through the backpropagation algorithm; the parameters of the remote sensing deep learning parameter inversion model are updated through the optimizer using the gradient, and iterative training is performed until the remote sensing deep learning parameter inversion model converges.

[0014] In a possible implementation, after training the remote sensing deep learning parameter inversion model is completed, the following processing is performed: the trained remote sensing deep learning parameter inversion model is deployed to the remote sensing data processing system, the remote sensing image to be inverted is input, and the parameter inversion results ignoring cloud interference are output through the remote sensing deep learning parameter inversion model. The parameter inversion results include: surface temperature, leaf area index, soil moisture or water quality parameters.

[0015] This application proposes a training method for remote sensing deep learning parameter inversion that ignores cloud interference. First, cloud identification is performed on the original remote sensing image to generate cloud mask data. Ground-measured data is rasterized into label data spatially matched to the remote sensing image. Then, the cloud-occluded areas in the label data are marked using the cloud mask data, resulting in cloud mask label data. Next, a pre-trained image segmentation network is improved to construct an image-based remote sensing deep learning parameter inversion model. The original remote sensing image is input into the remote sensing deep learning parameter inversion model for prediction, outputting a parameter prediction map. The mean squared error loss function is then improved to obtain a cloud mask ignoring loss function. Based on this cloud mask ignoring loss function, the cloud interference ignoring loss between the parameter prediction map and the cloud mask label data is calculated. Finally, based on the cloud interference ignoring loss, the model parameters are optimized through gradient backpropagation to complete the training of the remote sensing deep learning parameter inversion model. This method effectively addresses the interference of cloud occlusion on model training and improves inversion accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the method according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating a training method for ignoring cloud interference in remote sensing deep learning parameter inversion, provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram illustrating the cloud mask ignoring loss function in a training method for ignoring cloud interference in remote sensing deep learning parameter inversion provided in this embodiment of the application. Detailed Implementation

[0019] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] In the following description, references to "some embodiments" describe a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0022] This application provides a training method for remote sensing deep learning parameter inversion that ignores cloud interference, such as... Figure 1 As shown, the method includes:

[0023] Step S100: Perform cloud recognition on the original remote sensing image to generate cloud mask data, and rasterize the ground measurement data into label data that spatially matches the remote sensing image.

[0024] Specifically, raw remote sensing imagery refers to unprocessed or only minimally corrected image data acquired by satellites, aircraft, or other remote sensing platforms, recording electromagnetic wave reflection or radiation information of targets on the Earth's surface. Cloud mask data is obtained by directly identifying clouds from raw remote sensing images or by using large-scale visual models. Cloud mask data is data that marks cloud areas and can be represented by binary values ​​(e.g., 1 for cloud areas and 0 for non-cloud areas).

[0025] Ground-based measured data refers to data obtained on the ground through methods such as field measurement, sampling, and observation. This data reflects the actual state and attributes of the Earth's surface, including vegetation data, soil data, water body data, meteorological data, land use, and land cover data. Geographic Information System (GIS) tools or programming libraries (such as Python's GDAL) are used to convert ground-based measured data into raster data that matches the spatial resolution of remote sensing imagery.

[0026] In one possible implementation, cloud identification is performed on the original remote sensing image to generate cloud mask data. Step S100 further includes step S110, determining whether the original remote sensing image has a data quality assessment band. Specifically, a remote sensing data processing tool (such as GDAL or other remote sensing data processing libraries) is used to read the metadata of the original remote sensing image file and check whether a data quality assessment band (QA band) exists in the metadata. The QA band has a specific name or identifier and is part of the remote sensing image, generated by the satellite sensor during data acquisition. These bands contain information about the data quality of each pixel, such as cloud markers, shadow markers, snow / ice markers, etc.

[0027] Step S120: If yes, based on the cloud information in the data quality assessment band, identify cloud pixels to obtain cloud mask data. Specifically, if the original remote sensing image has a QA band, then according to the definition of the QA band, parse the cloud layer marker information of each pixel to generate cloud mask data.

[0028] Step S130: If not, perform cloud recognition on the original remote sensing image using a segmentation model to obtain cloud mask data. Specifically, if the original remote sensing image does not have a QA band, use a pre-trained segmentation model to perform cloud recognition on the original remote sensing image, and generate cloud mask data based on the model's output. The segmentation model can be selected as the Segment Anything Model (SAM). SAM breaks through segmentation boundaries and has powerful zero-shot generalization capabilities, enabling segmentation tasks for unknown objects and unfamiliar scenes.

[0029] In one possible implementation, the ground-measured data is rasterized into label data that spatially matches the remote sensing image. Step S100 further includes step S140, converting the ground-measured data into vector spatial data through coordinate resolution and data format conversion. Specifically, the geographic coordinate information (such as longitude and latitude) in the ground-measured data is parsed. Geographic Information System (GIS) tools or programming libraries (such as GeoPandas in Python) are used to convert the ground-measured data into a vector spatial data format (such as GeoJSON or Shapefile).

[0030] Step S150: Perform geospatial interpolation on the vector spatial data to obtain continuous field data. Specifically, select an appropriate interpolation method based on the data characteristics, such as inverse distance weighted interpolation or Kriging interpolation. Use GIS tools or programming libraries (such as Python's SciPy or PyKrige) to perform interpolation processing, converting discrete vector data into continuous field data.

[0031] Step S160: Determine the raster grid based on the spatial extent of the original remote sensing image and the raster cells based on the resolution of the original remote sensing image. Specifically, determine the boundaries of the raster grid based on the spatial extent of the original remote sensing image. Determine the size of the raster cells based on the resolution of the original remote sensing image. For example, use GDAL to read the spatial extent and resolution of the original remote sensing image and use NumPy to create the raster grid.

[0032] Step S170 involves rasterizing the continuous field data using the raster grid and raster cells to obtain label data. Specifically, GIS tools or programming libraries (such as GDAL) are used to rasterize the continuous field data, converting it into a raster format to ensure its spatial resolution matches the original remote sensing image. The rasterized label data is then saved. The rasterized label data maintains consistency with the original remote sensing image in terms of spatial resolution and number of rows and columns. Through a series of geospatial processing steps, it is ensured that the ground-measured data can accurately match the original remote sensing image, providing high-quality label data for subsequent remote sensing data processing and analysis.

[0033] Step S200: Use the cloud mask data to mark the cloud-obscured areas in the tag data to obtain cloud mask tag data.

[0034] Specifically, cloud mask data is fused pixel-by-pixel with rasterized label data. Cloud-obscured areas are ignored; that is, the label data is masked using the cloud mask data, assigning the value `ignore_index` to the label data corresponding to locations in the original remote sensing image where cloud cover exists, resulting in rasterized label data after the cloud mask. For example, using NumPy for pixel-by-pixel fusion, for areas marked as 1 in the cloud mask data, the corresponding location in the label data is marked as `ignore_index`.

[0035] Step S300: Improve the pre-trained image segmentation network, construct an image-based remote sensing deep learning parameter inversion model, input the original remote sensing image into the remote sensing deep learning parameter inversion model for prediction, and output a parameter prediction map.

[0036] Specifically, a pre-trained image segmentation network is a deep learning model used to segment images into different regions or objects, such as U-Net and DeepLab. The architecture of the pre-trained image segmentation network is adjusted to adapt to the characteristics of remote sensing data. The improved network is then used to perform parameter inversion on the original remote sensing image, outputting a parameter prediction map.

[0037] In one possible implementation, the pre-trained image segmentation network is improved to construct an image-based remote sensing deep learning parameter inversion model. Step S300 further includes step S310, obtaining the pre-trained image segmentation network. Specifically, a suitable pre-trained image segmentation network, such as U-Net or DeepLab, is selected according to the task requirements. The weights of the pre-trained model are loaded using a deep learning framework (such as TensorFlow or PyTorch).

[0038] Step S320: Based on the pre-trained image segmentation network, remove the terminal networks to obtain a pruned image segmentation network. Specifically, find the last Softmax and Argmax layers in the pre-trained image segmentation network, remove these layers from the model, retain the previous network structure, and construct the pruned image segmentation network based on the removed network. For example, PyTorch can be used to remove the last Softmax and Argmax layers. The Softmax layer is a commonly used activation function layer for the output layer of a neural network, especially in image segmentation problems. Its function is to transform a vector into a probability distribution, where the value of each element represents the predicted probability of that category. The Argmax layer is a commonly used layer in deep learning models, mainly used to determine the index with the highest probability in the neural network output, used to select the most likely category from the output of the segmentation model.

[0039] Step S330: A preset output convolutional layer is established, which is connected to the image segmentation network to obtain the remote sensing deep learning parameter inversion model. Specifically, a 1×1 convolutional layer with one output channel is added to the network after the image segmentation is removed. This layer generates the final parameter inversion result, which is single-band floating-point prediction data with the same row and column number as the original input remote sensing image. The 1×1 convolutional layer is then connected to the network after the image segmentation is removed, completing the construction of the remote sensing deep learning parameter inversion model.

[0040] Step S400: Improve the mean squared error loss function to obtain the cloud mask ignoring loss function. Based on the cloud mask ignoring loss function, calculate the ignoring cloud interference loss between the parameter prediction map and the cloud mask label data.

[0041] Specifically, the Mean Squared Error (MSE) loss function is one of the commonly used loss functions in regression problems, used to measure the average squared error between predicted and actual values. An improvement to the MSE loss function is made by ignoring the error in cloud-covered areas and only calculating the error in non-cloudy areas, resulting in the cloud mask ignoring loss function. When using the cloud mask ignoring loss function to calculate the loss between the parameter prediction map and the cloud mask label data, pixels in cloud-covered areas are skipped, resulting in the cloud interference ignoring loss.

[0042] like Figure 2 As shown, in one possible implementation, the mean squared error loss function is improved to obtain a cloud mask ignoring loss function. Step S400 further includes step S410, obtaining the mean squared error loss function. Specifically, standard MSE functions are defined using Python and TensorFlow or PyTorch, which forms the basis for the improvement.

[0043] Step S420: Modify the mean squared error loss function into a piecewise function to obtain the cloud mask ignoring loss function. The expression of the cloud mask ignoring loss function is as follows:

[0044]

[0045] Wherein, CM-MSE represents the cloud masking neglect loss function, H represents the total number of rows in the original remote sensing image, W represents the total number of columns in the original remote sensing image, h represents the row number of the current pixel, w represents the column number of the current pixel, and y h,w This represents the parameter label value at position (h, w) in the cloud mask label data. The parameter prediction value at position (h, w) in the parameter prediction graph represents the model prediction parameter value, and the ignore_index represents the value marked as the cloud-occluded area in the cloud mask label data.

[0046] Specifically, the standard MSE function is modified into a piecewise function that ignores cloud-occluded areas when calculating the loss. That is, a conditional check is added to the loss function: if the current pixel belongs to a cloud-occluded area, the loss for that pixel is not calculated. The calculation method for the cloud mask ignoring loss function is as follows: First, it iterates through every pixel of the cloud mask label data. For each pixel in the original remote sensing image, it calculates the true value y at its position (h, w). h,w Compared with model predictions The loss function calculates the squared difference between the true and predicted values ​​of a pixel. For each pixel location, it checks whether the pixel location is marked as cloud-covered (i.e., whether its value equals a specific `ignore_index`). If the pixel location is not cloud-covered, the squared difference between the true and predicted values ​​is calculated and summed. If the pixel location is cloud-covered, the corresponding predicted data value in the parameter prediction map is assigned the `ignore_index`. This ensures that the loss value for cloud-covered pixels remains constant at 0, and the gradient remains 0, thus not affecting the optimization of the remote sensing deep learning parameter inversion model parameters. This effectively ignores cloud interference during the training of the remote sensing deep learning parameter inversion model. Finally, the squared differences of all non-cloud-covered pixel locations are summed and divided by the total number of pixels in the original remote sensing image, H×W, to obtain the final loss value. The Cloud Mask Ignore Loss Function (CM-MSE) ignores cloud-covered pixels when calculating the mean squared error, allowing the model training to focus more on clear areas. This helps improve the model's prediction accuracy in cloudless areas while avoiding the negative impact of cloud-covered areas on model performance.

[0047] Step S500: Based on the ignored cloud interference loss, the model parameters are optimized through gradient backpropagation to complete the training of the remote sensing deep learning parameter inversion model.

[0048] Specifically, gradient backpropagation is an optimization algorithm that updates model parameters by calculating the gradient of the loss function ignored by the cloud mask, in order to minimize the loss. A suitable optimizer (such as Adam or SGD) is selected for parameter updates. During training, the gradient of the loss function ignored by the cloud mask is calculated, and the optimizer is used to update the model parameters, ultimately completing the training of the remote sensing deep learning parameter inversion model.

[0049] In one possible implementation, based on the cloud interference ignoring loss, the model parameters are optimized through gradient backpropagation to complete the training of the remote sensing deep learning parameter inversion model. Step S500 further includes step S510, calculating the gradient of the cloud interference ignoring loss with respect to the parameters of the remote sensing deep learning parameter inversion model using the backpropagation algorithm. Specifically, the input data (original remote sensing image) is forward-propagated through the remote sensing deep learning parameter inversion model to calculate the output (predicted parameter map). The cloud interference ignoring loss between the model output and the ground truth label is calculated using the Cloud Mask Ignoring Loss Function (CM-MSE). The gradient of the cloud interference ignoring loss with respect to the model parameters is calculated using a backpropagation algorithm, such as the automatic differentiation function of a deep learning framework (e.g., TensorFlow or PyTorch).

[0050] Step S520: Update the parameters of the remote sensing deep learning parameter inversion model using the gradient through an optimizer, and iterate the training until the remote sensing deep learning parameter inversion model converges. Specifically, use an optimizer (such as Adam or SGD) to update the parameters of the remote sensing deep learning parameter inversion model based on the calculated gradient to reduce loss. Repeat the steps of original remote sensing image input, parameter prediction map output, loss calculation ignoring cloud interference, gradient backpropagation, and model parameter update until the remote sensing deep learning parameter inversion model converges.

[0051] In one possible implementation, after training the remote sensing deep learning parameter inversion model is completed, the method further includes: deploying the trained remote sensing deep learning parameter inversion model to a remote sensing data processing system, inputting the remote sensing image to be inverted, and outputting parameter inversion results ignoring cloud interference through the remote sensing deep learning parameter inversion model, wherein the parameter inversion results include: surface temperature, leaf area index, soil moisture or water quality parameters.

[0052] Specifically, select a suitable deployment platform based on system requirements, such as a cloud server, local server, or edge device. If necessary, convert the model to a deployment-appropriate format, such as TensorFlow's SavedModel format or PyTorch's TorchScript. Configure the necessary software environment on the deployment platform, including deep learning frameworks, dependency libraries, and drivers. Deploy the trained remote sensing deep learning parameter inversion model to the remote sensing data processing system. Prepare the remote sensing image data to be inverted, and preprocess the images, including cropping, scaling, and normalization, to match the input format used during model training. Load the deployed remote sensing deep learning parameter inversion model into the remote sensing data processing system, feed the preprocessed input data into the model, execute the inference process, and obtain the inversion results from the model, including surface temperature, leaf area index, soil moisture, or water quality parameters. Analyze the model output and extract the required parameter values. Apply the inversion results to relevant fields, such as environmental monitoring and agricultural management. By deploying the trained remote sensing deep learning parameter inversion model into the remote sensing data processing system, parameters can be effectively inverted from new remote sensing images, ignoring cloud interference, thereby obtaining accurate parameters such as land surface temperature, leaf area index, soil moisture, or water quality. This not only improves the efficiency and accuracy of the inversion but can also be widely applied to various remote sensing application scenarios.

[0053] This application employs cloud identification to generate cloud masks from remote sensing images, rasterizes ground-measured data into matching label data, uses cloud masks to mark cloud-occluded areas in the label data, improves the image segmentation network to construct a parameter inversion model, inputs remote sensing images for prediction, outputs a parameter prediction map, improves the mean square error loss function to a cloud mask ignoring loss function, calculates the ignoring cloud interference loss between the prediction map and the label data, and optimizes model parameters through gradient backpropagation based on this loss to complete model training. These techniques solve the problem of limited inversion accuracy caused by the difficulty in effectively handling cloud interference in existing remote sensing deep learning parameter inversion training methods, achieving the technical effect of effectively handling the interference of cloud occlusion on model training and improving inversion accuracy.

[0054] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A training method for remote sensing deep learning parameter inversion that ignores cloud interference, characterized in that, The method includes: Cloud identification is performed on the original remote sensing images to generate cloud mask data, and the ground measurement data is rasterized into label data that spatially matches the remote sensing images. The cloud-obscured areas in the tag data are marked using the cloud mask data to obtain cloud mask tag data; The pre-trained image segmentation network is improved, and an image-based remote sensing deep learning parameter inversion model is constructed. The original remote sensing image is input into the remote sensing deep learning parameter inversion model for prediction, and the parameter prediction map is output. The mean squared error loss function is improved to obtain the cloud mask ignoring loss function. Based on the cloud mask ignoring loss function, the cloud interference ignoring loss between the parameter prediction map and the cloud mask label data is calculated. Based on the aforementioned neglect of cloud interference loss, the model parameters are optimized through gradient backpropagation to complete the training of the remote sensing deep learning parameter inversion model; The mean squared error loss function is improved to obtain the cloud mask neglect loss function, which includes: Obtain the mean squared error loss function; The mean squared error loss function is modified into a piecewise function to obtain the cloud mask ignoring loss function; The expression for the cloud mask neglect loss function is as follows: CM- ； Wherein, CM-MSE represents the cloud masking neglect loss function, H represents the total number of rows in the original remote sensing image, W represents the total number of columns in the original remote sensing image, h represents the row number of the current pixel, and w represents the column number of the current pixel. This represents the parameter label value at position (h, w) in the cloud mask label data. This represents the model prediction parameter value at position (h,w) in the parameter prediction graph, and ignore_index represents the value marked as a cloud-covered area in the cloud mask label data. Based on the aforementioned neglect of cloud interference loss, the model parameters are optimized through gradient backpropagation to complete the training of the remote sensing deep learning parameter inversion model, including: The gradient of the neglected cloud interference loss with respect to the parameters of the remote sensing deep learning parameter inversion model is calculated using the backpropagation algorithm; The parameters of the remote sensing deep learning parameter inversion model are updated by the optimizer using the gradient, and the training is iterated until the remote sensing deep learning parameter inversion model converges.

2. The training method for ignoring cloud interference in remote sensing deep learning parameter inversion as described in claim 1, characterized in that, Cloud identification is performed on the raw remote sensing imagery to generate cloud mask data, including: Determine whether the original remote sensing image has a data quality assessment band; If so, based on the cloud information in the data quality assessment band, identify cloud pixels and obtain cloud mask data; If not, cloud mask data is obtained by performing cloud identification on the original remote sensing image using a segmentation model.

3. The training method for ignoring cloud interference in remote sensing deep learning parameter inversion as described in claim 1, characterized in that, Ground-based measured data is rasterized into labeled data that spatially matches remote sensing imagery, including: The measured ground data is converted into vector space data through coordinate analysis and data format conversion; The vector spatial data is subjected to geospatial interpolation to obtain continuous field data; The raster grid is determined by the spatial extent of the original remote sensing image, and the raster cell is determined by the resolution of the original remote sensing image. The continuous field data is rasterized using the raster grid and the raster cells to obtain label data.

4. The training method for ignoring cloud interference in remote sensing deep learning parameter inversion as described in claim 1, characterized in that, Improvements are made to the pre-trained image segmentation network to construct an image-based remote sensing deep learning parameter inversion model, including: Obtain a pre-trained image segmentation network; By removing the terminal network from the pre-trained image segmentation network as the basic framework, a pruned image segmentation network is obtained. A preset output convolutional layer is connected to the image segmentation network to obtain a remote sensing deep learning parameter inversion model.

5. The training method for ignoring cloud interference in remote sensing deep learning parameter inversion as described in claim 1, characterized in that, After completing the training of the remote sensing deep learning parameter inversion model, the following steps are also included: The trained remote sensing deep learning parameter inversion model is deployed to the remote sensing data processing system. The remote sensing image to be inverted is input, and the remote sensing deep learning parameter inversion model outputs parameter inversion results ignoring cloud interference. The parameter inversion results include: surface temperature, leaf area index, soil moisture or water quality parameters.

Images