Saline-containing water body extraction method based on combination of SNDI and deep learning

Abstract

The invention discloses a salt-containing water body extraction method based on combination of SNDI and deep learning. The training method comprises the following steps: acquiring remote sensing image data; an SNDI index is calculated; carrying out region screening by utilizing an SNDI index to obtain a preliminary mask; performing correction and optimization to obtain a secondary mask; at least superposing the secondary mask and the remote sensing image data to serve as training data, and marking the training data to obtain a training set; and iteratively updating the multi-channel deep learning network by using the training set to obtain a salt-containing water body extraction model. According to the method, the preliminary mask used for prompting the salt lake water body area is obtained through preliminary screening and is corrected, training of the deep learning model is carried out through combination of the corrected mask and the remote sensing image data, and the obtained model can automatically optimize and adjust the threshold value according to different image features. The method is suitable for remote sensing images under different seasons, illumination and pollution conditions, and the possibility of misclassification is reduced.

Description

Technical Field

[0001] The present invention relates to the field of geographic remote sensing technology, and particularly to a method for extracting saline water bodies based on the combination of SNDI and deep learning. Background Art

[0002] In remote sensing images, the extraction of water bodies usually relies on water indices (such as NDWI, MNDWI, AWEI, NDVI). These traditional index models use the spectral feature differences of different bands to distinguish water bodies and non-water body areas. However, the reflection characteristics of saline water bodies in the visible and near-infrared bands are quite different from those of general fresh water, and the extraction accuracy of traditional water indices is relatively low.

[0003] For example, NDWI (Normalized Difference Water Index): uses the difference between the green band and the near-infrared band to detect water bodies. MNDWI (Modified Normalized Difference Water Index): is based on the green band and the short-wave infrared band, suitable for water body detection in the built environment, and suppresses interference from buildings, soil, etc. AWEI (Automated Water Extraction Index): is used for automatic water body extraction and is suitable for complex scenes with a large amount of shadow interference.

[0004] In addition, the principle of a relatively special water index SNDI is based on the spectral reflection characteristics of high-salinity water bodies in the blue band and the near-infrared (NIR) band. High-salinity water bodies show lower reflectivity in the near-infrared band (NIR), while the reflectivity in the blue band is slightly higher than that of ordinary fresh water. Therefore, SNDI magnifies the contrast between high-salinity water bodies and other surface cover types by calculating the ratio of these two bands.

[0005] Although SNDI is effective in extracting saline water bodies, its ability to identify and adapt to complex surfaces is limited. For example, there may be dry salt crusts and saline-alkali lands around high-salinity water bodies. The spectral characteristics of these areas are similar to those of water bodies, and SNDI may misjudge them as saline water bodies, resulting in non-water body noise in the extraction results; SNDI is mainly based on the spectral characteristics of the blue band and the near-infrared band and is easily affected by lighting conditions, shadows, and atmospheric conditions. Especially under high-reflection or thin-cloud conditions, the risk of misjudgment increases: The threshold selection of SNDI has a great impact on the results, and the threshold needs to be adjusted under different regions and different imaging conditions to adapt to different saline water body environments. If the threshold is set improperly, it is easy to cause missed detection or false detection. In summary, to improve the accuracy, it is usually necessary to combine with other remote sensing technologies to make up for its limitations, but no effective solution has emerged in the current technical solutions. Summary of the Invention

[0006] In view of the deficiencies of the prior art, the purpose of the present invention is to provide a method for extracting saline water bodies based on the combination of SNDI and deep learning.

[0007] To achieve the foregoing invention purpose, the technical solutions adopted by the present invention include:

[0008] In a first aspect, the present invention provides a training method for a saline water body extraction model based on the combination of SNDI and deep learning, which includes:

[0009] Obtain remote sensing image data, where the remote sensing image data covers the blue light band and the near-infrared band of a selected area;

[0010] Calculate the SNDI index based on the remote sensing image data;

[0011] Use the SNDI index for regional screening to obtain a preliminary mask indicating saline water bodies;

[0012] Correct and optimize the preliminary mask to obtain a secondary mask;

[0013] At least use the secondary mask and the remote sensing image data superimposed as training data, and mark the training data to obtain a training set;

[0014] Use the training set to iteratively update a multi-channel deep learning network to obtain a saline water body extraction model.

[0015] In a second aspect, the present invention further provides a method for extracting saline water bodies based on the combination of SNDI and deep learning, which includes:

[0016] Provide a saline water body extraction model trained by the above training method;

[0017] Input the remote sensing image data of the target area into the saline water body extraction model to obtain the distribution result of the saline water bodies in the target area, where the remote sensing image data of the target area covers the blue light band and the near-infrared band of the target area.

[0018] In a third aspect, the present invention further provides a training system for a saline water body extraction model based on the combination of SNDI and deep learning, which includes:

[0019] An image acquisition module for acquiring remote sensing image data, where the remote sensing image data covers the blue light band and the near-infrared band of a selected area;

[0020] An index calculation module for calculating the SNDI index based on the remote sensing image data;

[0021] A region screening module for screening regions using the SNDI index to obtain a preliminary mask indicating saline water bodies;

[0022] A secondary optimization module for correcting and optimizing the preliminary mask to obtain a secondary mask;

[0023] An image overlay module for using at least the secondary mask and the remote sensing image data as training data after overlaying, and marking the training data to obtain a training set;

[0024] An iterative update module for iteratively updating a multi-channel deep learning network using the training set to obtain a saline water body extraction model.

[0025] Fourthly, the present invention also provides a readable storage medium, in which a computer program is stored, and when the computer program is run, the steps of the above training method are executed.

[0026] Based on the above technical solutions, compared with the prior art, the beneficial effects of the present invention at least include:

[0027] The present invention obtains a preliminary mask for prompting the saline lake water body region through the preliminary screening of SNDI and corrects the preliminary mask, and uses the combination of the corrected mask and the remote sensing image data to train a deep learning model. The obtained model can automatically optimize and adjust the threshold according to different image features, making it applicable to remote sensing images under different seasons, lighting, and pollution conditions, and reducing the possibility of misclassification.

[0028] In addition, the technical solution adopted by the present invention can automatically denoise, extract local features through a convolutional neural network, and has strong robustness to the interference generated by high-reflection regions (such as bare land and saline-alkali land), thereby reducing the influence of interference and making the extraction of saline water bodies more stable.

[0029] The fine segmentation ability of this technical solution can perform excellently in image edge extraction, achieve more precise boundary recognition, can effectively extract the complex and irregular boundaries of saline water bodies, and is suitable for application scenarios with high requirements for spatial resolution. This technology can combine multiple multi-spectral remote sensing data, which makes the extraction result not dependent on a single sensor, enhancing the diversity of data sources and the stability of the result.

[0030] This comprehensive method combining SNDI and deep learning not only ensures the accuracy of segmentation but also has significant advantages in automation, adaptability, and efficiency, and is very suitable for the extraction application of saline water bodies such as salt lakes and saline lakes.

[0031] The above description is only an overview of the technical solution of the present invention. In order to enable those skilled in the art to more clearly understand the technical means of the present application and implement it in accordance with the content of the specification, the following describes the preferred embodiments of the present invention as follows. Detailed implementation manners

[0032] The principle of SNDI is based on the spectral reflection characteristics of high-salinity water bodies in the blue band and the near-infrared (NIR) band. High-salinity water bodies exhibit lower reflectance in the near-infrared band (NIR), while the reflectance in the blue band is slightly higher than that of ordinary fresh water. Therefore, SNDI calculates the ratio of these two bands to amplify the contrast between high-salinity water bodies and other land cover types.

[0033] Although the current SNDI index evaluation method is applicable to the extraction of saline water bodies, there are also some limitations and disadvantages:

[0034] Sensitivity to salt crusts and saline-alkali lands: Dry salt crusts and saline-alkali lands may exist around high-salinity water bodies. The spectral characteristics of these areas are similar to those of water bodies, and SNDI may misclassify them as water bodies, resulting in non-water body noise in the extraction results.

[0035] Dependence on specific bands and affected by imaging conditions: SNDI is mainly based on the spectral characteristics of the blue band and the near-infrared band, and is easily affected by lighting conditions, shadows, and atmospheric conditions. Especially under high-reflection or thin cloud conditions, the risk of misclassification increases.

[0036] Sensitivity of threshold setting: The threshold selection of SNDI has a great impact on the results. Thresholds need to be adjusted under different regions and imaging conditions to adapt to different saline water body environments. If the threshold is set improperly, it is easy to cause missed detections or false detections.

[0037] Limitations of spatio-temporal changes: The scope, salinity, and salt concentration of high-salinity water bodies such as salt lakes will change with seasons and climate. Therefore, the extraction results of SNDI in a single time phase may not accurately reflect the actual dynamic changes of water bodies. It is recommended to combine with multi-temporal analysis to improve the accuracy.

[0038] Limitations in supporting deep learning models: Relying solely on the SNDI index cannot fully capture the texture and complex features in water bodies. Therefore, in more complex scenarios (such as the junction between salt crusts and water bodies or saline-alkali lands covered with vegetation), the single-index method of SNDI may not perform as well as the combination with deep learning models.

[0039] In view of the deficiencies in the prior art, the inventors of this case have proposed the technical solution of the present invention through long-term research and a large number of practices. The following will further explain the technical solution, its implementation process, principle, etc.

[0040] In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0041] Moreover, relational terms such as "first" and "second" are only used to distinguish one component or method step with the same name from another, and do not necessarily require or imply any such actual relationship or order between these components or method steps.

[0042] An embodiment of the present invention provides a training method for a saline water body extraction model based on the combination of SNDI and deep learning, which includes the following steps:

[0043] 1. Obtain remote sensing image data, where the remote sensing image data covers the blue light band and the near-infrared band of a selected area;

[0044] 2. Calculate the SNDI index based on the remote sensing image data;

[0045] 3. Use the SNDI index for region screening to obtain a preliminary mask indicating the saline water body;

[0046] 4. Correct and optimize the preliminary mask to obtain a secondary mask;

[0047] 5. At least use the secondary mask and the remote sensing image data superimposed as training data, and mark the training data to obtain a training set;

[0048] 6. Use the training set to iteratively update a multi-channel deep learning network to obtain a saline water body extraction model.

[0049] As some typical application examples, the SNDI (Salinity Normalized Difference Index) index used in the above technical solution is a spectral index constructed based on remote sensing data and is usually used to distinguish saline areas (water bodies or saline-alkali lands) from non-saline areas. Its general form of expression is:

[0050]

[0051] Where: Banda: the band sensitive to salt (select the short-wave infrared SWIR band 1.57 - 1.65 μm); Bandb: the reference band with obvious contrast to salt (select the near-infrared NIR band 0.85 - 0.88 μm or other bands).

[0052] In the above step 3, through this index, the saline area usually exhibits numerical characteristics on the SNDI image that are significantly different from those of the non-saline area. The specific screening method and steps are as follows:

[0053] (1) Remote sensing image preprocessing

[0054] First, select suitable remote sensing data, such as Landsat-8, Sentinel-2, MODIS, etc.; secondly, perform radiometric calibration, geometric correction, and atmospheric correction on the remote sensing image to obtain reflectance data with physical significance.

[0055] (2) Calculate the SNDI index image

[0056] Calculate the index value of each pixel according to the SNDI formula to generate an SNDI index map with clear spatial distribution:

[0057]

[0058] Calculate the index value of each pixel through GIS or remote sensing processing software (such as ENVI, ArcGIS, Google Earth Engine, or Python code implementation) to obtain the index distribution map.

[0059] (3) Set the threshold and conduct preliminary screening

[0060] Set the threshold for the SNDI index image using statistical analysis, histogram analysis, and experience. Extract the histogram and frequency distribution of the SNDI of the pixels in the study area, select sample points in typical saline areas, analyze the SNDI value range of these sample areas, and determine a reasonable threshold according to the inflection point and significant demarcation position of the SNDI value frequency distribution curve.

[0061] Generally, SNDI>0 indicates the salt enrichment area, and being close to 1 indicates high-salinity water bodies or saline-alkali areas; in this patent, according to the actual situation of the saline water body, SNDI>0.2 is set as the preliminary extraction threshold (after empirical analysis, when SNDI>0.2, the salinity of the saline water body>1 g / L).

[0062]

[0063] (4) Generate a preliminary mask image

[0064] According to the set threshold, obtain a preliminary saline water body mask through pixel binarization; this mask is presented in binary form (1 and 0), where 1 represents the saline area (water body candidate area) and 0 represents the non-saline area.

[0065] In the above step 4, the specific steps for correcting and optimizing the preliminary mask to obtain the secondary mask include:

[0066] For some land areas, they are misidentified as saline water bodies due to abnormal reflectivity; the edge areas of saline water bodies, vegetation interference, or mixed pixels result in non-identification; there may be errors such as isolated small areas or speckle noise in the mask map, and these errors are eliminated through deep learning methods and optimization methods. The following are the specific steps in some embodiments:

[0067] (1) Prepare the deep learning dataset

[0068] Perform preliminary annotation on the remote sensing image and the preliminary mask, and obtain the true annotation (Ground Truth) through the method of manual correction. Use remote sensing images (multi-band images): such as multiple bands of Landsat-8 or Sentinel-2 to annotate the mask (binary image): the saline water body area is marked as 1, and the background is marked as 0. Ensure that the data volume of the training set accounts for about 70-80%, the validation set accounts for about 10-15%, and the test set accounts for about 10-15%.

[0069] (2) Data augmentation and preprocessing before training

[0070] Augment the data through the method of manual intervention, and appropriately perform image flipping (left and right, up and down), rotation transformation (random rotation of 90°, 180°, 270°), scale transformation (random scaling), cropping and stitching (cropping small images to adapt to the network input), and standard normalization (normalizing the data to between 0 and 1) to improve the generalization ability of the model.

[0071] (3) Build a semantic segmentation model (U-Net)

[0072] U-Net is a classic fully convolutional neural network (FCN), which is widely used in remote sensing image semantic segmentation tasks. The core structure of U-Net consists of two parts: an encoder and a decoder, with a skip connection mechanism in the middle, which can effectively retain multi-scale spatial information and is suitable for high-precision remote sensing image segmentation.

[0073] In the first step, define the model input layer. The dimension of the input remote sensing image is: (Batch_size, Channels, Height, Width). For multi-spectral remote sensing data: Channels is the number of bands. For example, the common band combination of Landsat-8 images (such as: 4 bands or more).

[0074] The example code for this step is as follows:

[0075]

[0076]

[0077] The second step is to construct the encoder (downsampling path). The encoder is responsible for gradually downsampling to extract high-level semantic features. Each step includes a convolutional module (DoubleConv) and a pooling layer (MaxPooling).

[0078] The example code (U-Net encoder) for this step is as follows:

[0079]

[0080] The third step: Construct the decoder (upsampling path). The decoder gradually upsamples the feature map. The upsampling (Upsampling) and skip connection (Skip Connection) restore the spatial details of the image.

[0081] The example code (U-Net decoder) for this step is as follows:

[0082]

[0083] The fourth step: Define the model output layer. The output is a mask map, and the number of channels is the number of classes (here it is binary classification: saline water body / background, and the number of channels is 1). The Sigmoid activation function is used to output a probability map between 0 and 1.

[0084] The example code for this step is as follows:

[0085]

[0086]

[0087] The fifth step: Instantiate the U-Net model, complete instantiation (input 4 bands, output 1 mask channel):

[0088] model = UNet(in_channels = 4, out_channels = 1)

[0089] The sixth step: Define the loss function and optimizer:

[0090] criterion = nn.BCELoss() # Binary cross-entropy loss

[0091] optimizer = torch.optim.Adam(model.parameters(), lr = 1e-4)

[0092] (4) Model training process

[0093] The Adam optimizer is used to optimize the loss function that combines the Dice loss and the cross-entropy loss. The learning rate strategy is set to 0.0001 - 0.001, and the learning rate decay strategy is CosineAnnealingLR. The training parameters are set as follows: Epoch: 30 - 100, Batch Size: 4 - 32. During training, the model performance is evaluated on the validation set regularly, and the Dice coefficient is monitored.

[0094] The model is verified and evaluated. Using the validation set and the test set, the evaluation is carried out using the Dice coefficient. According to the evaluation results, the hyperparameters and the model structure are continuously adjusted to ensure that the Dice coefficient > 0.80.

[0095] (5) Model prediction (generating the secondary mask)

[0096] The trained model is used for inference. The remotely sensed image corresponding to the preliminary mask is input into the model, and the model outputs a probability image (binary mask, where 1 represents the saline water body area and 0 represents the non-saline water body area). The final secondary mask is generated according to the threshold (0.2).

[0097] The remaining steps 5 and 6 can be implemented in a general way.

[0098] The principle of the above technical solution is that by combining the Salinity-Normalized Difference Index (SNDI) with the deep learning model, the high-salinity water body feature extraction ability of SNDI can be fully utilized, and the performance of the deep learning model in saline water body recognition can be improved.

[0099] Before training the deep learning model, the SNDI is used to extract the initial high-salinity water body mask to help the model focus on the water body area and reduce the interference of other areas. Then, through manual correction or combination with geographical information labels, more accurate training data is generated to improve the recognition accuracy of the model for targets such as salt lakes and saline lakes.

[0100] In some embodiments, the training method specifically includes:

[0101] After preprocessing the remotely sensed image data, the SNDI index is calculated. The preprocessing includes radiometric correction, atmospheric correction, and extraction of the region of interest. The calculation of the SNDI index is carried out in the region of interest.

[0102] In some embodiments, the calculation method of the SNDI index can be expressed as:

[0103]

[0104] Wherein, SNDI represents the SNDI index, Blue represents the response value in the blue light band, and NIR represents the response value in the near-infrared band.

[0105] In some embodiments, the area screening is performed in a fixed threshold manner, and specifically includes the following process:

[0106] Determine the SNDI index corresponding to any pixel using a fixed threshold, and mark the pixel as saline water body and non-saline water body;

[0107] Generate the preliminary mask based on the marks corresponding to the multiple pixels.

[0108] In some embodiments, the process of correcting and optimizing the preliminary mask may specifically include:

[0109] Perform denoising and morphological processing on the preliminary mask;

[0110] Correct the misjudged area in the preliminary mask by manual adjustment.

[0111] Based on the above correction method, the preferred solution adopted in the embodiments of the present invention can also be implemented in the following manner:

[0112] Extract the misjudged area of the saline water body extraction model, where the misjudged area refers to the difference area between the predicted mask and the true mask and the area with a confidence level lower than the confidence threshold;

[0113] Perform manual correction and data augmentation on the misjudged area to form a supplementary data set, and merge the supplementary data set with the training set to obtain an incremental data set;

[0114] Use the incremental data set to perform secondary iterative training on the saline water body extraction model, and the learning rate of the secondary iterative training is lower than that of the first iterative update.

[0115] When the above technical solution is specifically implemented, the model can improve the accuracy through the strategy of Active Learning. Specifically, it is to use the misjudged areas in the mask after model prediction (that is, the areas where the model makes mistakes or has low confidence), and use these areas as the key for annotation and training to further improve the accuracy of the deep learning model. This method is applicable to areas where there are differences or large uncertainties between the model predicted mask and the true mask. Focus on the samples that are most difficult for the model to learn, accelerate the model learning speed, and significantly improve the mask prediction accuracy. The specific steps are as follows:

[0116] The first step: Misjudged area identification and extraction, compare the predicted mask output by the model with the true mask (or manually corrected mask) to find the areas with large differences or low confidence:

[0117] Pixel-level error identification: Error area = |Predicted mask - True mask|

[0118] The misjudgment area includes: pixels misjudged by the model (false positives, false negatives), and uncertain areas where the prediction probability is close to 0.5.

[0119] Second step: Manually refine the annotation and expert correction of the misjudgment area. For the selected misjudgment area, perform manual correction and expert-guided modification, and carefully correct the annotation mask of the misjudgment area of the model to improve the accuracy and reliability of the annotation.

[0120] Third step: Construct an incremental training set mainly composed of the misjudgment area. Use the corrected misjudgment area as key samples and add them to the training set:

[0121] Incremental training data set = Original training data + Corrected misjudgment area

[0122] Select 10% - 20% of the corrected data and add it to the original training set.

[0123] Fourth step: Data augmentation. Perform data augmentation on the annotation data of the misjudgment area. Through steps such as flipping, rotating, and scaling, further expand the diversity of training samples, which can better improve the generalization ability and avoid overfitting.

[0124] Fifth step: Model retraining (Fine-tuning). Retrain or fine-tune the model with the incremental training set containing the misjudgment area. Use a relatively small learning rate of 0.00001 - 0.0001 to prevent drastic changes in the model's weights, and guide the model to accurately learn the features of the misjudgment area through the loss function (Dice).

[0125] Sixth step: Iterative refinement. Continuously perform the above process through multiple rounds of iteration. In each round of iteration, re-evaluate the model, continuously discover new misjudgment areas, continuously refine the training set, and continuously fine-tune the model to significantly improve the final mask prediction accuracy.

[0126] In some embodiments, the multi-channel deep learning network includes any one of U-Net and SegNet.

[0127] In some embodiments, the loss functions used in the iterative update include the cross-entropy loss function and the Dice loss function.

[0128] In some embodiments, the optimizers used in the iterative update include the Adam optimizer.

[0129] Based on the above training method, the second aspect of the embodiments of the present invention further provides a saline water body extraction method combining SNDI with deep learning, which includes the following steps:

[0130] Provide a saline water body extraction model trained by using the training method provided in any of the above embodiments;

[0131] Input the remote sensing image data of the target area into the saline water body extraction model to obtain the distribution result of the saline water body in the target area, where the remote sensing image data of the target area covers the blue band and the near-infrared band of the target area.

[0132] As some typical implementation cases of the above technical solutions, the main steps of the training method and the saline water body extraction method provided by the present invention may include:

[0133] 1. Data preparation and preprocessing

[0134] Collect remote sensing image data: Obtain multi-spectral remote sensing images of the target area, such as satellite data of Landsat, Sentinel-2, etc., covering the bands required by SNDI such as blue and near-infrared (NIR).

[0135] Radiometric correction and atmospheric correction: Preprocess the image data to eliminate the atmospheric influence and improve the image quality, laying a foundation for calculating SNDI and generating a mask.

[0136] Radiometric calibration, converting digital number (DN) to radiance: Convert the digital number (DN) in the image to the radiance value, and this conversion requires the calibration coefficient of the sensor. The general formula is:

[0137] L = G×DN + B

[0138] Where L is the radiance, G is the gain coefficient, and B is the offset.

[0139] Further convert the radiance value to the surface reflectance, which usually requires considering parameters such as the solar altitude angle and solar irradiance.

[0140] Extract ROI (region of interest): Determine the salt lake and the surrounding non-water areas, and mark the areas that the model needs to focus on learning.

[0141] 2. Calculate SNDI and generate a preliminary mask

[0142] Calculate SNDI: Use the blue band and the near-infrared (NIR) band to calculate the SNDI index to highlight high-salinity water bodies. The formula is as follows:

[0143] SNDI = (Blue - NIR)(Blue + NIR)SNDI

[0144] =\frac{(Blue - NIR)}{(Blue + NIR)}SNDI

[0145] =(Blue + NIR)(Blue - NIR)

[0146] Set a threshold to generate a preliminary mask: Set an appropriate threshold according to the SNDI value (for example, SNDI > 0 indicates high-salinity water bodies), and generate a preliminary mask; Mark the high-salinity water bodies in the form of a binary preliminary mask, and retain this mask as the preliminary identification result of high-salinity water bodies during further optimization.

[0147] The setting of the threshold here is determined by using machine learning methods. First, divide the pixels in the image into two categories: "saline water bodies" and "non-water bodies". Using the labeled samples, find the most suitable SNDI threshold through a machine learning model, or it can be done by using statistical analysis and manual correction. The obtained threshold is generally the fixed threshold used for preliminary screening.

[0148] 3. Optimize the mask and verify manually

[0149] Denoising and morphological processing: Perform denoising and morphological processing (such as erosion, dilation) on the preliminary mask to remove small-area noise points and incomplete boundaries, making the mask smoother and more continuous.

[0150] In the denoising process, the median filtering method is used to replace the pixel value with the median of the neighboring pixels, which is suitable for removing salt-and-pepper noise and retaining edge information. First, select the window size, usually an odd-sized square window, such as 3x3, 5x5, etc. The larger the window, the more obvious the smoothing effect, but the greater the risk of losing details. Second, select the center of the window as the current pixel, take out all the pixel values within the window to form a list. Third, sort the pixel values in the list by size and take the value at the middle position as the median. Fourth, replace the current pixel value with the calculated median, slide the window to the next pixel, and repeat the above steps.

[0151] Suppose a 3x3 median filter is applied to a window of an image, and the pixel values within the window are as follows:

[0152] 10, 100, 10

[0153] 100, 255, 100

[0154] 10, 100, 10

[0155] Sort the pixel values: [10, 10, 10, 10, 100, 100, 100, 100, 255]

[0156] Take the median value: 100, and replace the pixel value at the center of the window with the median value 100.

[0157] Manual verification and correction: Compare the preliminary mask with the real image, manually check the misclassified areas, exclude non-water body noises, and ensure accurate marking of the saline water body areas.

[0158] 4. Combine the mask and the original image to generate optimized training data

[0159] Overlay the mask: Overlay the optimized mask with the multispectral image to provide multi-channel information for each pixel, including RGB, NIR bands, and the SNDI mask channel.

[0160] Generate training labels: Mark the salt lake area as the high-salinity water body category, and mark other areas as the background or non-water body category, and generate training labels for the deep learning model accordingly.

[0161] Based on the existing geographic information data (such as topographic maps, land use maps, geographic databases released by the government), align the geographic information with the remote sensing image, and use GIS tools to map the geographic information onto the image. Generate a binary label map.

[0162] Data augmentation: Enhance and expand the training data by means of rotation, translation, color change, etc. to improve the generalization ability of the model.

[0163] Spectral enhancement: Randomly adjust the brightness, contrast, and saturation to simulate the lighting conditions in different environments and improve the adaptability of the model.

[0164] 5. Train the deep learning model

[0165] Design a multi-channel deep learning network: Select a network structure that supports multiple input channels (such as U-Net, SegNet, etc.), and use the original image bands and the SNDI mask as input feature channels.

[0166] Normalize the data, normalize the image pixel values to the range of 0-1 or -1 to 1 to facilitate faster convergence during training; crop or scale the image to the input size required by U-Net (usually 256x256 or 512x512, etc.). Divide the dataset into a training set, a validation set, and a test set. For example, 70% is used as the training set, 20% is used as the validation set, and 10% is used as the test set.

[0167] Build a U-Net network. The encoder uses a series of downsampling layers (usually convolutional layers + max pooling layers) to extract high-dimensional features of the image. The decoder uses a series of upsampling layers to map the high-dimensional features back to the image resolution for reconstructing the segmentation boundary; passing low-level features from the encoder layers to the corresponding decoder layers helps to recover the details of the segmentation boundary.

[0168] In the binary classification task, Binary Cross-Entropy Loss is selected. For the segmentation task, Dice Loss can also be chosen to increase the robustness to imbalanced data.

[0169] Adam is usually used as the optimizer, and hyperparameters such as the learning rate are set.

[0170] Train the model: Use the optimized training dataset to train the model, enabling the model to learn SNDI features and enhancing the ability to identify saline water bodies.

[0171] The initial learning rate is set to 0.001 and dynamically adjusted using a learning rate scheduler. The batch size is selected according to the memory size, commonly 4, 8, or 16. The number of iterations is generally set to 100 to 300 epochs, depending on the data volume and convergence situation.

[0172] For each batch of data passing through the U-Net model, a segmentation prediction output is generated; the error between the prediction result and the true label is calculated using the set loss function; the gradient is calculated through backpropagation, and the network parameters are updated using the optimizer, enabling the model to gradually learn more accurate segmentation boundaries; after each epoch, the loss and segmentation accuracy are evaluated on the validation set to monitor whether the model is overfitting.

[0173] Model optimization: Adjust hyperparameters such as the learning rate and batch size through the loss curve and accuracy curve on the validation set; if the validation loss no longer decreases over multiple epochs, training can be stopped early to prevent overfitting; if overfitting is detected, the types of data augmentation transformations can be increased to improve the generalization ability of the model.

[0174] 6. Validation and Accuracy Evaluation

[0175] Evaluate the model using the test set: Use the data that has not participated in training as the test set to evaluate the precision and recall rate of the model, with particular attention to the extraction effect of saline water bodies.

[0176] Evaluate the model performance on the test set, calculate segmentation metrics such as IoU, Dice coefficient, and accuracy; analyze the image samples that perform poorly in the test set to understand the deficiencies of the model, such as misclassified regions or complex structures that cannot be recognized; compare the model results with the preliminary SNDI mask to observe the improvements in accuracy and details of the model.

[0177] Model Saving and Deployment: Save the trained model in a format convenient for deployment (such as.h5, SavedModel, etc.); use model quantization technology to optimize the model for deployment on the server.

[0178] The key innovations of the above exemplary technical solutions are mainly reflected in the following aspects:

[0179] 1. Utilize SNDI to enhance the ability to identify saline water bodies

[0180] By processing specific band combinations of remote sensing images, SNDI enhances the difference between saline water bodies and the surrounding environment, making the features of saline water bodies more prominent. Combining with the exponential features generated by SNDI can effectively improve the sensitivity and recognition rate of deep learning models for saline water bodies.

[0181] 2. Fusion of deep learning models and traditional indices

[0182] Use the SNDI index as a feature input into the deep learning model, combined with the learning ability of the convolutional neural network (U-Net) for high-dimensional features, so as to better distinguish saline water bodies from other types of water bodies and ground objects in terms of spatial and spectral features. This way of fusing deep learning and indices improves the deficiencies of traditional methods that only use threshold methods or deep learning alone.

[0183] 3. Automated and accurate water body classification and segmentation

[0184] The deep learning model can automatically learn the texture, spectrum and other features of saline water bodies, and perform fine classification and boundary segmentation in combination with the enhanced features of SNDI. Compared with traditional methods, the method combining SNDI and deep learning can extract the boundaries of irregular saline water bodies more accurately, especially in areas with complex backgrounds.

[0185] 4. Dynamic threshold adjustment and self-adaptive enhancement

[0186] Combined with the adaptive characteristics of deep learning, the threshold of SNDI can be adjusted according to different environmental conditions and spectral changes, solving the problem of decreased accuracy of traditional fixed threshold methods in different scenarios, and making the method more adaptable in different salt lake environments.

[0187] 5. Multi-source data fusion supports more complex scenario applications

[0188] This method allows the fusion of data from different satellite sensors (such as Landsat, Sentinel-2, etc.), enabling the model to process data with different spectral resolutions. Through the deep learning model, the common features of saline water bodies can be extracted from data with different resolutions, improving the generalization ability of this method.

[0189] The third aspect of the embodiments of the present invention further provides a training system for a saline water body extraction model based on the combination of SNDI and deep learning, which includes:

[0190] An influence acquisition module, configured to acquire remote sensing image data, where the remote sensing image data covers the blue light band and the near-infrared band of a selected area;

[0191] An index calculation module, configured to calculate the SNDI index based on the remote sensing image data;

[0192] A region screening module, configured to perform region screening using the SNDI index to obtain a preliminary mask indicating a saline water body;

[0193] A secondary optimization module, configured to correct and optimize the preliminary mask to obtain a secondary mask;

[0194] An image superposition module, configured to at least superpose the secondary mask and the remote sensing image data as training data, and mark the training data to obtain a training set;

[0195] An iterative update module, configured to iteratively update a multi-channel deep learning network using the training set to obtain a saline water body extraction model.

[0196] Correspondingly, the embodiments of the present invention further provide a readable storage medium, in which a computer program is stored, and when the computer program is run, the steps of the training method provided in any of the above embodiments are executed.

[0197] The technical solution of the present invention will be further described in detail below through several embodiments. However, the selected embodiments are only used to illustrate the present invention and do not limit the scope of the present invention.

[0198] Embodiment 1

[0199] In this embodiment, a method combining SNDI and deep learning is applied to monitor the dynamic changes of the saline lake water body by analyzing remote sensing images of different periods.

[0200] 1. Data collection and preprocessing

[0201] Select data sources with good spatial resolution and spectral resolution (such as Sentinel-2 or Landsat series).

[0202] Perform radiometric and atmospheric corrections and image normalization.

[0203] 2. SNDI calculation and dynamic threshold adjustment

[0204] SNDI index calculation: Based on the selected bands, calculate the SNDI index image for each time point to ensure the comparability of time series data. The SNDI calculation formula is as follows:

[0205]

[0206] Perform dynamic threshold adjustment: Dynamically adjust the threshold according to the statistical characteristics of each time series image.

[0207] 3. Training data generation and model initialization

[0208] Perform preliminary threshold segmentation on the SNDI of each time point to generate a binary preliminary mask, and perform correction and optimization.

[0209] Combine historical images and pre-annotated data, and use data augmentation to generate diverse training data.

[0210] 4. Deep learning model construction and training

[0211] Select the spatio-temporal convolutional network (U-Net) as the model, and the input includes multi-temporal image bands and the corresponding SNDI index.

[0212] Train the model, use the cross-entropy loss function and the Adam optimizer for learning, and dynamically adjust the learning rate and hyperparameters.

[0213] 5. Extraction of the temporal variation of saline water bodies

[0214] After the model training is completed, use the model to predict the distribution of saline water bodies at each time point, generate a binary preliminary mask image to identify the boundaries of saline water bodies; compare the water body masks of different periods to identify the expansion or contraction areas of saline water bodies.

[0215] Denoise the extracted temporal water body mask and verify the results.

[0216] Example 2

[0217] In view of the problem of misjudgment areas in the results of saline water bodies extracted by the SNDI preliminary mask in this example, an active learning optimization method based on the misjudgment areas of the mask is proposed to significantly improve the accuracy of the deep learning model. The specific implementation steps are as follows:

[0218] 1: Initial model training and identification of misjudgment areas

[0219] Using the preliminary mask generated by combining remote sensing images (such as Sentinel-2 satellite images) with the calculated SNDI index as training data, a U-Net semantic segmentation network is constructed to train the initial model: the input data size is 256×256, and the images use the blue and near-infrared bands; the model optimizer selects the Adam optimizer, and the initial learning rate is 0.0001; the loss function selects the Dice loss function.

[0220] Apply the trained model to the test set to obtain the initial prediction mask, and identify the misjudgment areas after comparing with the actual ground truth mask: the misjudgment areas are defined as:

[0221] Misjudgment area pixel = |Model prediction mask - Actual ground truth mask|

[0222] Save the misjudgment areas in binary form, where the area with a value of 1 is the misjudgment area.

[0223] 2: Manual refinement and annotation of misjudgment areas

[0224] Select the misjudgment areas identified by the model, and conduct fine annotation and correction of the misjudgment areas by remote sensing image interpretation experts: correction method: if the model-predicted pixel in the misjudgment area is 1 but should be 0 in reality, it is manually corrected to 0; if the model-predicted pixel is 0 but should be 1 in reality, it is manually corrected to 1. Add the misjudgment area mask after manual fine annotation to the incremental training set to form a new training sample.

[0225] 3: Data augmentation for incremental training sample data

[0226] Perform data augmentation on the misjudgment area data to increase the data volume and data diversity. Use horizontal and vertical flipping; random rotation (90°, 180°, 270°); scale change (scaling ±10%); after enhancement, the number of misjudgment area samples increases to more than 3 times the original sample number, forming an incremental training data set.

[0227] 4: Model retraining (incremental training)

[0228] Use the incremental training data set containing misjudgment areas to train the original model, and fine-tune the hyperparameters as shown in the following table:

[0229] Parameter Parameter value Model structure U-Net Optimizer Adam Fine-tuning learning rate le-5 Batch Size 8 Number of training iterations 20 epochs Loss function Dice Loss

[0230] During the fine-tuning process, calculate the Dice metric on the validation set after each epoch to monitor the accuracy change.

[0231] 5: Repeat the training iteration optimization of misjudgment areas

[0232] The fine-tuned model performs mask prediction on the new test set to identify misjudged regions again; repeat the above (Step 2) to (Step 4) to finely correct the misjudged regions and perform incremental training again; iterate and optimize 2 - 3 times until the misjudged regions are significantly reduced and the mask prediction accuracy of the model reaches a satisfactory standard (such as Dice coefficient > 0.85).

[0233] Analysis of implementation effect

[0234] Through the method of the above embodiment, the Dice coefficient of the initial model obtained in the first training is about 0.75; after introducing active training and iteration of misjudged regions, the Dice coefficient of the model is increased to above 0.85; the number of misjudged regions is significantly reduced (reduced by about 40% - 60%), the mask boundary is significantly improved, and the final mask prediction accuracy is significantly improved.

[0235] Specific effect examples:

[0236] Epoch Dice coefficient (validation set) Ratio of misjudged area (%) Initial training 0.75 25％ First iteration 0.82 15％ Second iteration 0.86 8％

[0237] It shows that the method of this embodiment can significantly improve the mask prediction accuracy and reduce the proportion of misjudged regions; the advantage of this method is based on the idea of active learning, and uses the misjudged regions of the mask for targeted training; through the fine annotation of misjudged regions and the method of incremental data enhancement, the annotation cost is reduced and the data utilization efficiency is improved; the model is optimized iteratively, and the recognition accuracy of the model for complex regions is greatly improved.

[0238] Comparative example 1

[0239] This comparative example is generally the same as Example 1, the main difference is that the step of correcting and optimizing the preliminary mask is omitted, and the preliminary mask obtained by using a fixed threshold is directly superimposed with the remote sensing image data as the training data.

[0240] Test example

[0241] Use the method of Example 1: combining SNDI and deep learning; Example 2: the method of active learning optimization based on mask misjudged regions; Comparative example 1: the method of directly obtaining the preliminary mask by using a fixed threshold; to extract saline water bodies in Gahai Salt Lake respectively.

[0242] Example 1 uses the SNDI index combined with deep learning (U-Net model). First, calculate the SNDI index for the remote sensing image to obtain the preliminary saline water body mask; then, perform mask refinement correction through the U-Net model. The Dice loss function, Adam optimizer, and learning rate of 0.0001 are used during the training process, and the iteration is 50 times.

[0243] The simulation results show that the mask edge of this method is accurate and continuous, the Dice coefficient reaches 0.88, and the saline water body is accurately identified.

[0244] In Example 2, an active learning optimization method based on mask misjudgment regions was used. For the misjudgment regions found in the initial training of Example 1, active learning optimization was carried out. First, the misjudgment regions were identified. After being corrected by manual annotation, incremental training data was formed, and retraining was carried out using an active learning strategy.

[0245] The simulation results show that the Dice coefficient is further increased to 0.91, the proportion of misjudgment regions is reduced from the initial 25% to 6%, the boundary is more accurate, and the extraction accuracy of saline water bodies is significantly improved.

[0246] In Comparative Example 1, a method of directly obtaining a preliminary mask using a fixed threshold was adopted. In this comparative example, only a fixed threshold (SNDI = 0.2) was used to binarize the remote sensing image to directly generate a preliminary mask without further optimization.

[0247] The simulation results show that the Dice coefficient of this method is only 0.72, the mask edge is blurred, there is more noise, and the proportion of obvious mis-extraction and missed-extraction regions is relatively large.

[0248] After comparing the advantages and disadvantages of the three methods, it is concluded that:

[0249] Advantages of Example 1: The mask has a high accuracy and can accurately identify the saline water body region. Disadvantages: There may be a certain proportion of misjudgment regions in the initial training stage. Advantages of Example 2: The mask has the highest accuracy, very few misjudgment regions, and significantly improves the overall accuracy. Disadvantages: Additional human and computational costs are required for active learning optimization. Advantages of Comparative Example 1: Simple and fast operation, no need for complex model training. Disadvantages: Low mask accuracy, serious misjudgment, not applicable to scenarios with high-precision requirements.

[0250] Through comprehensive analysis, it can be seen that on the basis of Example 1, Example 2 has a significant advantage in accuracy (Dice = 0.91) by introducing active learning optimization, and is more suitable for actual scenarios with high requirements for the extraction accuracy of saline water bodies.

[0251] Based on the above examples and comparative examples, it can be clear that in the embodiments of the present invention, a preliminary mask for prompting the saline lake water body region is obtained through the preliminary screening of SNDI and the preliminary mask is corrected, and the corrected mask is combined with remote sensing image data for the training of a deep learning model. The obtained model can automatically optimize and adjust the threshold according to different image features, making it applicable to remote sensing images under different seasons, lighting, and pollution conditions, and reducing the possibility of misclassification.

[0252] In addition, the technical solution adopted in the embodiments of the present invention can automatically reduce noise, extract local features through a convolutional neural network, and has strong robustness to the interference generated by high-reflection regions (such as bare land and saline-alkali land), thereby reducing the influence of interference and making the extraction of saline water bodies more stable.

[0253] The fine segmentation ability of this technical solution can perform excellently in image edge extraction, achieve more refined boundary recognition, effectively extract the complex and irregular boundaries of saline water bodies, and is suitable for application scenarios with high spatial resolution requirements. This technology can combine various multispectral remote sensing data, which makes the extraction results not rely on a single sensor, enhancing the diversity of data sources and the stability of the results.

[0254] This comprehensive method combining SNDI and deep learning not only ensures the accuracy of segmentation but also has significant advantages in automation, adaptability, and efficiency, and is very suitable for the extraction applications of saline water bodies such as salt lakes and saline lakes.

[0255] It should be understood that the above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly, and should not be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims

1. A training method for a saline water body extraction model based on the combination of SNDI and deep learning, characterized in that, Including: Obtain remote sensing image data, where the remote sensing image data covers the blue light band and the near-infrared band of a selected area; Calculate the SNDI index based on the remote sensing image data; Use the SNDI index for regional screening to obtain a preliminary mask indicating saline water bodies; Correct and optimize the preliminary mask to obtain a secondary mask; At least use the secondary mask and the remote sensing image data superimposed as training data, and label the training data to obtain a training set; Use the training set to iteratively update a multi-channel deep learning network to obtain a saline water body extraction model.

2. The training method according to claim 1, wherein Specifically including: Perform preprocessing on the remote sensing image data and then calculate the SNDI index. The preprocessing includes radiometric calibration, atmospheric correction, and extraction of the region of interest. The calculation of the SNDI index is performed in the region of interest.

3. The training method according to claim 1, characterized in that, The calculation method of the SNDI index is expressed as: where SNDI represents the SNDI index, Blue represents the response value of the blue light band, and NIR represents the response value of the near-infrared band.

4. The training method according to claim 1, characterized in that The regional screening is carried out in a fixed threshold manner, specifically including: Use a fixed threshold to determine the SNDI index corresponding to any pixel, and label the pixel as a saline water body and a non-saline water body; Generate the preliminary mask with the labels corresponding to multiple pixels.

5. The training method according to claim 1, wherein The process of correcting and optimizing the preliminary mask specifically includes: Perform denoising and morphological processing on the preliminary mask; Manually adjust the misjudged area in the preliminary mask.

6. The training method according to claim 5, wherein Also including: Extract the misjudged area of the saline water body extraction model, where the misjudged area refers to the difference area between the predicted mask and the true mask and the area with a confidence level lower than the confidence threshold; Perform manual correction and data augmentation on the misjudged area to form a supplementary data set, and merge the supplementary data set with the training set to obtain an incremental data set; Use the incremental data set to perform secondary iterative training on the saline water body extraction model, and the learning rate of the secondary iterative training is lower than the learning rate of the previous iterative update.

7. The training method according to claim 1, wherein The multi-channel deep learning network includes any one of U-Net and SegNet; And / or, the loss function used in the iterative update includes the cross-entropy loss function and the Dice loss function; And / or, the optimizer used in the iterative update includes the Adam optimizer.

8. A method for extracting saline water bodies based on the combination of SNDI and deep learning, characterized in that, Including: Provide a saline water body extraction model trained by using the training method described in any one of claims 1-7; Input the remote sensing image data of the target area into the saline water body extraction model to obtain the distribution result of the saline water bodies in the target area, where the remote sensing image data of the target area covers the blue light band and the near-infrared band of the target area.

9. A training system for a saline water body extraction model based on the combination of SNDI and deep learning, characterized in that, Including: An acquisition module for acquiring remote sensing image data, where the remote sensing image data covers the blue light band and the near-infrared band of a selected area; An index calculation module for calculating the SNDI index based on the remote sensing image data; A regional screening module for using the SNDI index for regional screening to obtain a preliminary mask indicating saline water bodies; The secondary optimization module is used to correct and optimize the preliminary mask to obtain a secondary mask; The image superposition module is used to superpose at least the secondary mask and the remote sensing image data as training data, and mark the training data to obtain a training set; The iterative update module is used to iteratively update the multi-channel deep learning network by using the training set to obtain a saline water body extraction model.

10. A readable storage medium, characterized in that, A computer program is stored in the readable storage medium, and when the computer program is run, it executes the steps of the training method according to any one of claims 1-7.

Images