A method for detecting lunar south pole impact craters based on an improved TransUnet network

Abstract

To address the challenges of extracting impact craters from lunar polar regions due to lighting conditions, and the inability of Digital Elevation Models (DEMs) to extract small impact craters, this invention proposes AM-TransUNet+, a small impact crater extraction algorithm for the future lunar south pole region of interest. This algorithm is based on the Natural Language Processing Transformer model and the Image Segmentation Convolutional Network UNet+ model. Using imagery from the Lunar Reconnaissance Orbiter Narrow Angle Camera (LRO-NAC, 0.5m / pixel), the algorithm incorporates a convolutional block attention module and depthwise separable convolutions into the core TransUNet model, further accelerating model convergence and improving accuracy while significantly reducing the number of training parameters without compromising accuracy. The accuracy of the algorithm is validated using surface images of Mars and Mercury. Experimental results demonstrate that the AM-TransUNet+ algorithm exhibits better transferability and accuracy across different deep space exploration remote sensing data.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and specifically to a method for detecting lunar south pole impact craters based on an improved TransUnet network. Background Technology

[0002] Impact craters, as the most prominent landform on the lunar surface, are diverse in type and shape, exhibiting a ring-shaped depression structure of varying sizes and uneven aggregation. The study of impact craters can not only infer the relative geological age of the lunar surface, surface characteristics, and the presence of water ice, but also has applications in spacecraft positioning and navigation, lunar base site selection, and obstacle avoidance on the lunar surface.

[0003] In recent years, many countries around the world have formulated lunar exploration plans. The accurate and rapid identification of impact craters has always been a key focus of research in the field of deep space exploration. Many researchers have also proposed a series of lunar surface impact crater extraction algorithms (CDA), the development of which is shown in Figure 1. Based on the different data sources of these algorithms, they can be roughly divided into two categories: lunar surface images observed by optical sensors, and digital elevation models (DEMs) obtained through laser altimeter scanning. Among them, Silburt, based on DEM data from the Lunar Reconnaissance Orbiter (LRO) and the Kaguya lunar orbiter, used Keras and Tensorflow to build and train a U-Net segmentation model. It utilized skip connections to achieve information fusion between low-level and high-level features, and combined with the Povilaitis (crater diameter 5-20 km) and Head (crater diameter >20 km) databases, identified nearly 92% of impact craters. Wang Yiran et al. developed a DEM-based CDA for LOLA DEM to detect impact craters with 3D morphological features (such as edge height, internal slope, and depth). Their constructed impact crater dataset (LU1319373) contains over 1 million impact craters covering a distance of 1 km, with a detection rate of approximately 85%. Salih et al., based on Lunar Reconnaissance Orbiter Narrow Angle Camera (LRO NAC) images, used YOLOv3 to detect impact craters in six mid-latitude regions with varying illumination levels, and further studied the degradation characteristics of impact craters. Yang et al., based on LRO NAC images, used Domain Adaptive (DA) to effectively detect unlabeled ground truth samples and proposed a new network, CraterDANet, to extract impact craters from the Chang'e-4 landing area, establishing a new lunar impact crater dataset containing 20,000 impact craters.

[0004] Due to the high resolution of images observed by optical sensors, small impact craters can be detected. However, optical sensors are susceptible to illumination variations, making it difficult to classify craters through image processing. For example, under poor lighting conditions, craters may be either invisible or only partially visible. Since DEMs provide height information, craters can be located using threshold filtering methods. However, DEMs have low resolution and may not be able to display small craters. In summary, although existing technologies have disclosed methods for detecting craters of different scales, the following problems still exist for small-scale craters with diameters in the range of meters or hundreds of meters:

[0005] 1. Unlike the mid- and low-latitude regions, the lunar south pole has a highly undulating terrain, a large and dynamically changing shadow area, and extremely uneven lighting conditions. High-resolution images are more suitable than low-resolution DEMs for detecting small impact craters down to the meter level. However, current research has only focused on experiments in mid- and low-latitude regions, making it challenging to extract small impact craters from the lunar polar regions.

[0006] 2. Supervised methods learn features from a large set of labeled training data that requires a large number of samples. To obtain good detection results, well-labeled training data is needed to adapt to various conditions, such as terrain, lighting conditions, and degradation.

[0007] 3. Deep learning methods have achieved breakthroughs in the accuracy and reliability of detecting multi-scale impact craters, especially large-scale craters even globally. However, they have not performed well in detecting small-scale impact craters. This is because the receptive fields of the top-level neurons are large, and information about small-scale targets is not fully preserved. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a lunar south pole impact crater detection method based on an improved TransUnet network. This network, built upon TransUnet, incorporates a Convolutional Block Attention (CBAM) module, Depthwise Separable Convolution (DSC), and an enhancement module. This model can simultaneously perform channel and spatial attention, thus better exploring the inter-channel and spatial relationships of features. Furthermore, it more effectively captures the contour features of impact craters and provides better identification of smaller and overlapping craters at different illumination azimuths on the lunar south pole.

[0009] The technical solution to achieve the purpose of this invention is as follows:

[0010] A method for detecting lunar south pole impact craters based on an improved TransUnet network, characterized by the following steps:

[0011] Step 1: Obtain LRONAC images of the lunar south pole;

[0012] Step 2: Preprocess the input LRO NAC image;

[0013] Step 3: Based on the preprocessed LRO NAC images, use the CraterTools plugin to manually identify real meteorite craters in ArcGIS software as label data, and divide the obtained LRO NAC image sample data into training set, validation set and test set;

[0014] Step 4: Construct the AM-TransUnet+ network model and train it on the training set;

[0015] Step 5: Update the weight matrix through backpropagation, and repeat Step 4 until the AM-TransUnet+ network accuracy reaches the preset accuracy on the validation set or the loss function converges.

[0016] Step 6: Test the trained AM-TransUnet+ network model on the test set;

[0017] Step 7: Input the image of the lunar south pole to be detected into the trained AM-TransUnet+ network, and finally output the detection results.

[0018] Furthermore, the specific operational steps of step 2 include:

[0019] Step 21: Perform random sliding window sampling on the LRO NAC image data, and simultaneously randomly scale the size of the sliding window;

[0020] Step 22: Resample the randomly sampled NAC window data to a fixed size;

[0021] Step 23: Flip and mirror the resampled LRO NAC image.

[0022] Furthermore, the AM-TransUnet+ network model constructed in step 4 includes an encoder, a decoder, and an enhancement module. The encoder is used to encode the features of the input image vector; the decoder is used to decode the features of the image vector; and the enhancement module is used to enhance the features of the encoded image vector.

[0023] Furthermore, the encoder includes a convolutional block attention module, a CNN convolutional module, and a Transformer layer. It first generates image block weights from column vectors to construct a score matrix. In the MSA of the Transformer layer, all weight matrices obtained from the score matrix are summed to obtain the weight matrix of the Transformer layer. Then, each encoder feature is enhanced by the weight matrix in the enhancement module, and the skipped features are sent to the decoder. Finally, in the decoder, the weight matrix is ​​upsampled to a size corresponding to the skipped features by a cascaded upsampler.

[0024] Furthermore, the specific steps for detection by the AM-TransUnet+ network described in step 7 include:

[0025] Step 71: Input the image of the lunar south pole to be detected;

[0026] Step 72: Perform three downsampling operations in the encoder to obtain the corresponding feature matrices. Each downsampling process includes a convolution, regularization, ReLU activation, and max pooling layer.

[0027] Step 73: After obtaining the feature maps extracted by the CNN in the encoder, the spatial information of the image patch is obtained by adding position embeddings in the image patch embedding. Then the output of the l-th layer encoder is represented as: ;

[0028] Where MSA(·) is the multi-self-attention head operation, MLP(·) is the multilayer perceptron operation, LS(·) represents the layer normalization operation, and r l The feature representation of the reconstructed l-th layer transformer;

[0029] Step 74: Perform multi-stage upsampling and decoding in the decoder through cascaded upsamplers, and output the segmentation result using features;

[0030] Step 75: Generate the final segmentation mask using the features transmitted by the fusion encoder, and identify small impact craters in the lunar NAC image based on the segmentation mask.

[0031] Furthermore, in step 4, when training the AM-TransUnet+ network model, the initial learning rate is set to 0.001, the number of model training iterations is set to 300, the number of filters is set to 112, the filter length is set to 3, and the drop value is set to 0.15.

[0032] Furthermore, the AM-TransUnet+ network model uses the binary cross-entropy (BCE) loss function during training:

[0033] ;

[0034] Among them, y i It is the label of pixel i in the AM-TransUNet+ prediction result, t i This is the label for this pixel in the ground truth.

[0035] Beneficial effects:

[0036] First, this invention proposes a backbone network adapted to small impact craters on the lunar south pole. This network is mainly used to detect impact craters by extracting rich contextual information while maintaining detailed information, which is very effective in improving the detection results.

[0037] Second, this invention enhances the skip connection functionality through a redesigned skip connection and combines it with fractional matrix column vectors, enabling the extraction of small impact crater contour information in the lunar south pole region of interest. Furthermore, combining the CBAM module with the DSC module from the classic TransUNet improves performance while significantly reducing the number of parameters in the model encoder.

[0038] Third, the transferability of the model of this invention is very significant, and it has achieved good results in the identification of impact craters on Mars and Mercury. Attached Figure Description

[0039] Figure 1 A timeline of the development of impact crater identification both domestically and internationally;

[0040] Figure 2 This is a flowchart of the present invention;

[0041] Figure 3 The model architecture is AM-TransUNet+.

[0042] Figure 4 The results of impact crater extraction in the region of interest (ROI) of the lunar south pole using AM-TransUNet are shown below. (a) and (b) are the input image and detection results for the Haworth region, respectively; (c) and (d) are the input image and detection results for the Amundsen region, respectively; (e) and (f) are the input image and detection results for the Shackton region, respectively; and (g) and (h) are the input image and detection results for the Faustini region, respectively.

[0043] Figure 5 The Epoch-Loss curve for the AM-TransUNet+ network;

[0044] Figure 6 The results of AM-TransUNet+ detection on different datasets are shown; where (a) and (b) are the input image and detection result for Mars, respectively; and (c) and (d) are the input image and detection result for Mercury, respectively.

[0045] Figure 7 A schematic diagram of the scoring matrix and the redesigned jump connections. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0047] This invention proposes an end-to-end lunar south pole crater extraction algorithm (AM-TransUNet+), whose main objective is to detect multi-scale craters, especially small-scale craters, at the lunar south pole. It includes the following steps:

[0048] 1. Image preprocessing

[0049] Imagery was acquired from the Narrow Angle Camera (NAC) of the Lunar Reconnaissance Orbiter (LRO) at a resolution of 0.5 m / pixel. Verification data were obtained using images from the High-Resolution Stereo Camera (HRSC) of the Mars Express spacecraft and images from the MESSENGER spacecraft (Mercury_MESSENGER_mosaic_global_250m).

[0050] Since the surface NAC (Neural Image Conversion) data of the Moon has a fixed size, while deep learning in image segmentation applications typically uses square images as input, LRO (Lunar Resonance Occurrence) NAC data was selected for random sliding window sampling to accommodate deep learning models. Furthermore, the sliding window size was randomly scaled to accommodate craters of varying diameters. The randomly sampled NAC window data was then resampled to a fixed size, such as 114×114, to meet the input requirements of the deep learning model. To further increase training data and improve network recognition performance, data augmentation operations, such as flipping and mirroring, were used for preprocessing.

[0051] 2. Create data labels and groupings

[0052] Data labels were created based on preprocessed LRO NAC images, using the CraterTools plugin in ArcGIS software to manually identify real meteorite craters as label data. The validation and test datasets were generated in the same manner, with the regions used to generate the test images differing from those in the training images. This ensured that 80% of the dataset was used as the training set, 10% as the validation set, and 10% as the test set in both training and validation datasets.

[0053] 3. Constructing an AM-TransUnet+ lunar south pole impact crater detection model

[0054] AM-TransUNet+ consists of an encoder, decoder, and enhancement modules. AM-TransUNet+ replaces the encoder layer of the original TransUNet with an attention module convolution, and generates a weight matrix using the column vectors of the score matrix to enhance features and increase attention to critical patches. Then, it performs upsampling to achieve a segmentation result of the same size as the original image. Figure 7 As shown, the score matrix and the redesigned jump connection are:

[0055] First, image patch weights are generated from the column vectors and constructed into matrices. In the MSA of the Transformer encoder, all weight matrices obtained from the fractional matrix are summed to obtain the weight matrix for that Transformer layer. Since each Transformer layer has an independent weight matrix, there are N weight matrices. These weight matrices are summed to obtain the final weight matrix.

[0056] ;

[0057] Among them, w patch M is the weight of each image patch. i It is the i-th th The score matrix in the Transformer layer, f(·), involves operations on column vectors.

[0058] Next, the weight matrix is ​​upsampled to a size corresponding to the skipped features. Then, each encoder feature is augmented using the weight matrix, and the skipped features are passed to the decoder.

[0059] ;

[0060] Where F in and F out These represent the features before and after the enhancement module, respectively. Wpatch is the weight matrix; ups(·) is the upsampling operation.

[0061] The specific process of the AM-TransUnet+ network includes:

[0062] (1) The input image is a 144×144 lunar south pole image. Through three downsampling operations, the corresponding feature matrices F1, F2, and F3 are obtained respectively. Their sizes are (x / 2, y / 2, c), (x / 4, y / 4, c), and (x / 8, y / 8, c), where c is the number of channels. Each downsampling process includes one convolution, regularization, ReLU activation, and max pooling layer.

[0063] (2) In the encoder path, CBAM is merged into the CNN part of the CNN-Transformer hybrid layer. After obtaining the feature maps extracted by the CNN, the features are mapped to a new embedding space through a trainable linear projection. By adding positional embeddings in the image patch embedding, the image patch spatial information y can be obtained. input .

[0064] ;

[0065] Among them, y input It is the input to the Transformer layer, x i It is a feature image patch extracted by CNN, where E is the linear projection, E loc This represents positional embedding, where n is the number of slices, and [;…;] represents concatenation operations.

[0066] In Figure 3 (left), each encoder Transformer module contains a layer specification, a multi-head self-attention module (MSA, which contains multiple self-attention modules), a multilayer perceptron (MLP), and residual connections. In each Transformer layer, a vector preceding the self-attention mechanism or the feedforward neural network is introduced through residual connections to enhance the output vector of the self-attention mechanism or the feedforward network. The output of the l-th encoder layer can be expressed as:

[0067] ;

[0068] Where MSA(·) is the multi-self-attention head operation, MLP(·) is the multilayer perceptron operation, LS(·) represents the layer normalization operation, and r l This represents the feature representation of the reconstructed l-th layer transformer.

[0069] (3) In the decoder, a cascaded upsampler is used for multi-stage upsampling and decoding, and the segmentation result is output using features. The CBAM module is placed after all convolutions, as shown in Figure 3 (right). Each decoder block in each layer contains a 2x upsampling, a feature concatenation, and a convolution operator. After the decoder block, the length and width of the features are doubled, and the number of channels is halved. When passing through all three decoder blocks, the length and width are half of the original image. Finally, the result is obtained through a 1×1 convolutional layer.

[0070] In AM-TransUNet+, the convolutional block attention module is a lightweight, general-purpose attention module within a feedforward convolutional neural network. Given an intermediate feature map, this module sequentially infers a two-dimensional attention map, which can be used for feature extraction. An enhancement module uses column vectors of the fractional matrix to augment skip connections, thus redesigning the skip connections. The final segmentation mask is generated by fusing features transmitted from the encoder, enabling the identification of small impact craters in lunar NAC imagery.

[0071] 4. Model Training

[0072] The training set of lunar south pole images was fed into the AM-TransUNet+ network. The initial learning rate was set to 0.001, the number of model training iterations was set to 300, the number of filters was set to 112, the filter length was set to 3, and the drop value was set to 0.15.

[0073] 5. Calculate the loss function

[0074] The essence of AM-TransUNet+ network prediction is determining whether each pixel is on the edge of an impact crater. It is essentially a binary classification problem. The loss function used during AM-TransUNet+ training is the binary cross-entropy (BCE) loss.

[0075] ;

[0076] Where y i It is the label of pixel i in the AM-TransUNet+ prediction result, t i This is the label of this pixel in the ground truth. The image loss is the sum of the losses of all pixels. If the difference between the predicted image and the labeled image is large, the loss function value will be larger.

[0077] 6. Backpropagation

[0078] Calculate the gradient of the loss function with respect to each parameter, and then adjust and update the parameters based on the gradient.

[0079] 7. Update the weight matrix

[0080] The weight matrix of the backpropagation is updated based on the gradient of the parameters obtained from backpropagation, thereby reducing the loss function.

[0081] 8. Impact crater detection and result output

[0082] The image of the lunar south pole to be detected is loaded into the trained model to obtain the final target detection result image.

[0083] 9. Applications of transfer learning on Mars and Mercury

[0084] The images of Mars and Mercury to be detected are loaded into the trained model to obtain the images of the impact craters of Mars and Mercury.

[0085] 10. Model Evaluation

[0086] To evaluate the performance of the impact crater recognition algorithm, a comprehensive test was conducted using precision-recall (PR) curves and average precision (AP) values.

[0087] ;

[0088] In the formula, N tp It is to identify the correct number of impact craters, N fp It represents the number of incorrectly identified impact craters.

[0089] In the PR curve, Recall represents the algorithm's false negative rate:

[0090] ;

[0091] In the formula, Nfn represents the missed impact crater targets. The PR curve is fitted by changing the threshold conditions, with Precision as the vertical axis and Recall as the horizontal axis. In addition, in order to reflect the accuracy of impact crater identification, the IOU between the predicted position and the actual position of the target needs to be considered when calculating the PR curve. In this invention, the IOU is set to 0.5 when calculating the PR curve.

[0092] The F1 score is a statistical metric used to measure the precision of a binary classification model. It measures both the model's accuracy and recall. The F1 score can be viewed as a weighted average of the model's precision and recall, and its formula is:

[0093] ;

[0094] In the formula: P is precision; R is recall.

[0095] Example

[0096] To verify the overall performance of the proposed algorithm, the following aspects are analyzed.

[0097] 1. Impact crater extraction results

[0098] Figure 4 Results of impact crater extraction from the AM-TransUNet+ lunar south pole region of interest, from Figure 4As can be seen in (b), (d), (f), and (h), the AE-TransUNet+ network enhances the skip connection by redesigning the skip connection and uses a fractional matrix column vector combined with the skip connection to extract the contour information of small impact craters in the lunar south pole region of interest.

[0099] 2. Performance Indicator Comparison

[0100] Table 1 compares the performance metrics of UNet+, TransUNet+, and the AM-TransUNet+ algorithm proposed in this invention. As shown in Table 1, the AM-TransUNet+ network achieves a recall of 0.822 and an accuracy of 0.890. Furthermore, this network achieves good results in approximately 150 batches, and its total parameter count is 0.98 M, significantly less than TransUNet+, consistent with recent research.

[0101] Table 1 Comparison of performance metrics of different algorithms

[0102]

[0103] To understand the training process of the AM-TransUNet+ network, epoch-loss curves were plotted for analysis, as shown in Figure 5. Figure 5 shows that the AM-TransUNet+ network converges in approximately 250 epochs, indicating a relatively fast convergence speed. This demonstrates that the network trains quickly, significantly reducing training time. Furthermore, the introduced Transformer model did not lead to gradient vanishing or exploding, indicating high applicability of the model.

[0104] 3. Model transferability

[0105] To verify the model's transferability, tests were performed on different data sources on Mars and Mercury, and the results are as follows: Figure 6 As shown. From Figure 6 As can be seen, for heterogeneous data, the model can detect impact craters of different scales on the Martian surface, and also has a certain detection rate for overlapping impact craters. For impact crater identification on the surface of Mercury, despite the differences in Mercury's topographic features, the model can also detect a certain number of impact craters, laying the foundation for further research on the geological structure of Mercury's surface.

[0106] Contents not described in detail in this specification are existing technologies known to those skilled in the art. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting lunar south pole impact craters based on an improved TransUnet network, characterized in that, Includes the following steps: Step 1: Obtain LRO NAC images of the lunar south pole; Step 2: Preprocess the input LRO NAC image; Step 3: Based on the preprocessed LRO NAC images, use the CraterTools plugin to manually identify real meteorite craters in ArcGIS software as label data, and divide the obtained LRO NAC image sample data into training set, validation set and test set; Step 4: Construct the AM-TransUnet+ network model and train it on the training set; Step 5: Update the weight matrix through backpropagation, and repeat Step 4 until the AM-TransUnet+ network accuracy reaches the preset accuracy on the validation set or the loss function converges. Step 6: Test the trained AM-TransUnet+ network model on the test set; Step 7: Input the image of the lunar south pole to be detected into the trained AM-TransUnet+ network, and finally output the detection results; Step 2 includes the following specific steps: Step 21: Perform random sliding window sampling on the LRO NAC image data, and simultaneously randomly scale the size of the sliding window; Step 22: Resample the randomly sampled NAC window data to a fixed size; Step 23: Flip and mirror the resampled LRO NAC image; The AM-TransUnet+ network model constructed in step 4 includes an encoder, a decoder, and an enhancement module. The encoder is used to encode the features of the input image vector; the decoder is used to decode the features of the image vector; and the enhancement module is used to enhance the features of the encoded image vector. The encoder includes a convolutional block attention module, a CNN convolutional module, and a Transformer layer. It first generates image block weights from column vectors to construct a score matrix. In the MSA of the Transformer layer, all weight matrices obtained from the score matrix are summed to obtain the weight matrix of the Transformer layer. Then, each encoder feature is enhanced using the weight matrix in the enhancement module, and the skipped features are sent to the decoder. Finally, in the decoder, a cascaded upsampler upsamples the weight matrix to a size corresponding to the skipped features. The specific steps for detection by the AM-TransUnet+ network described in step 7 include: Step 71: Input the image of the lunar south pole to be detected; Step 72: Perform three downsampling operations in the encoder to obtain the corresponding feature matrices. Each downsampling process includes a convolution, regularization, ReLU activation, and max pooling layer. Step 73: After obtaining the feature maps extracted by the CNN in the encoder, the spatial information of the image patch is obtained by adding position embeddings in the image patch embedding. Then the output of the l-th layer encoder is represented as: ； Where MSA(·) is the multi-self-attention head operation, MLP(·) is the multilayer perceptron operation, LS(·) represents the layer normalization operation, and r l The feature representation of the reconstructed l-th layer transformer; Step 74: Perform multi-stage upsampling and decoding in the decoder through cascaded upsamplers, and output the segmentation result using features; Step 75: Generate the final segmentation mask using the features transmitted by the fusion encoder, and identify small impact craters in the lunar NAC image based on the segmentation mask.

2. The lunar south pole impact crater detection method based on an improved TransUnet network as described in claim 1, characterized in that, In step 4, when training the AM-TransUnet+ network model, the initial learning rate is set to 0.001, the number of model training iterations is set to 300, the number of filters is set to 112, the filter length is set to 3, and the drop value is set to 0.

15.

3. The lunar south pole impact crater detection method based on an improved TransUnet network as described in claim 1, characterized in that, The AM-TransUnet+ network model is trained using the binary cross-entropy (BCE) loss function. ； Among them, y i It is the label of pixel i in the AM-TransUNet+ prediction result, t i This is the label for this pixel in the ground truth.

Images