Karst mountain bare rock information extraction method of unmanned aerial vehicle high-resolution image

Abstract

The invention relates to the technical field of geological exploration, in particular to a karst mountain bare rock information extraction method based on an unmanned aerial vehicle high-resolution image, which comprises the following steps: firstly, making a bare rock data set based on an unmanned aerial vehicle original image, then on the basis of a DeepLabV3 + model, taking MobileNetV2 as a backbone network, carrying out bare rock information extraction and training on the bare rock data set, and obtaining a karst mountain bare rock information extraction model; and finally, supervising and classifying the orthoimage of the research area by using a support vector machine, and comparing the bare rock extraction rate of the improved model through an experimental result. Experimental results show that the lightweight bare rock extraction model constructed by the method has relatively high accuracy and robustness in the aspects of detection and extraction of bare rock information, and the improved model greatly reduces calculation parameters and complexity, effectively reduces dependence on hardware resources, and can stably operate on a general calculation platform.

Description

Technical Field

[0001] The present invention relates to the technical field of geological exploration, and particularly relates to a method for extracting bare rock information from high-resolution images of karst mountains by using an unmanned aerial vehicle (UAV). Background Art

[0002] Dangerous rocks refer to geological bodies that are cut and separated by multiple sets of structural planes, have poor stability, and may collapse in forms such as toppling, falling, and sliding. They are a kind of geological disaster that occurs globally. In China, the terrain is undulating and the geological conditions are complex. Most of the limestone mountainous areas are located in Guangxi, where dangerous rock blocks are widely distributed and the rock strata are relatively unstable. Traditional geological exploration and on-site monitoring are common methods for obtaining bare rock information, which can provide detailed descriptions of bare rocks, but they require a large amount of time and material resources, are extremely vulnerable to the terrain environment, and it is difficult to quickly and accurately obtain relevant data information. In contrast, people can use remote sensing satellites to obtain bare rock images of the target area over a large area and interpret the bare rock area using its spectral characteristics, reducing the intensity of field work. However, the resolution of remote sensing satellite data is insufficient, and it usually cannot capture small-scale bare rock features. The low resolution may cause the bare rocks to be mixed with surrounding ground objects, making it difficult to clearly extract the boundaries and textures of the bare rocks. In addition, satellite images usually require complex preprocessing processes, have a long processing cycle, and consume a large amount of time and computing resources.

[0003] With the continuous improvement of UAV remote sensing technology, UAV images, with their advantages of high resolution, flexibility, and low cost, provide a new technical means for geological disaster monitoring and quantitative assessment. More and more studies have begun to use UAV images for precise extraction and monitoring research of bare rock information. By using the high-resolution images obtained by UAVs, the bare rocks in the images are vectorized, the exposed rock areas are outlined, and the spatial distribution characteristics of the exposed rock strata on the ground surface are obtained. Then, combined with various methods such as geographic information system (GIS) and model simulation, the bare rock rate of the area is calculated. Based on the extracted bare rock information, the vulnerable parts and potential instability areas of the surface bare rocks are accurately identified and quantitatively evaluated, providing an important scientific basis for the risk prevention of geological disasters. The development of UAV remote sensing images has better filled the deficiencies of satellite images in aspects such as fine identification of bare rocks and delineation of ranges. However, its extraction method still mainly relies on manual interpretation, and the identification method is time-consuming and laborious, lacking intelligence. Deep learning models can provide a new solution idea for the automated processing of UAV images.

[0004] However, the current research on using deep learning to extract bare rock information in karst rocky desertification areas is still relatively limited. On the one hand, deep learning models have high requirements for sample datasets and often require the participation of numerous real samples, which requires on-site collection of sample data. However, in karst landforms, the mountain bodies have large undulations and the vegetation is patchily distributed, which not only increases the difficulty of UAV sampling but also may lead to the imbalance of sample data, further increasing the workload and difficulty of sample annotation. On the other hand, since bare rocks, vegetation, and shadows often exhibit similar or overlapping spectral and texture features in multi-band images, it is easy to cause category confusion, thereby reducing the recognition accuracy and robustness of the model. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for extracting bare rock information of karst mountain bodies from UAV high-resolution images. By introducing a lightweight network on the basis of the existing deep learning model, the calculation parameters and complexity are reduced, the dependence on hardware resources is effectively reduced, and it can also run stably on general computing platforms.

[0006] To achieve the above purpose, the present invention provides a method for extracting bare rock information of karst mountain bodies from UAV high-resolution images, including the following steps:

[0007] Step 1: Use a UAV to obtain high-resolution images of the mountain bodies in the research area;

[0008] Step 2: Construct a bare rock dataset;

[0009] Step 3: Select the DeepLabV3+ model as the main framework of the lightweight bare rock extraction model;

[0010] Step 4: Use the MobileNetV2 network to replace the backbone network to train the bare rock dataset for bare rock information extraction and conduct accuracy evaluation;

[0011] Step 5: Apply the trained lightweight bare rock extraction model to the orthophoto of the research area for bare rock extraction, and then compare it with the bare rock information obtained by using SVM for supervised classification of the orthophoto, and improve the bare rock extraction rate of the model through experimental results comparison.

[0012] Optionally, the high-resolution image in Step 1 is an orthophoto of the mountain body, which is obtained after the image taken by the UAV aerial photography is subjected to initialization processing, point cloud encryption, orthorectification, and mosaicking.

[0013] Optionally, in the process of constructing the bare rock dataset, first screen out suitable images and manually label the bare rock areas in the images using LabelMe, and then divide the objects included in the images into two categories: bare rock and background, and classify all the ground objects except the bare rock as the background category.

[0014] Optionally, the main framework of the lightweight bare rock extraction model adopts an encoder-decoder structure. The encoder part includes a deep convolutional neural network and a spatial pyramid pooling part. In the decoder part, a multi-scale feature fusion mechanism is introduced. After the deep features obtained by the feature extraction network are upsampled by a factor of 4, they are fused with the shallow features, and then through 3×3 convolution fusion and upsampling, the spatial information is gradually restored.

[0015] Optionally, in step 4, an inverted residual structure is introduced to adjust the order of dimensionality reduction and dimensionality increase. At the same time, a linear bottleneck layer is adopted, and the ReLU activation function is replaced by a linear activation function.

[0016] Optionally, during the accuracy evaluation process, pixel accuracy, mean intersection over union, precision, recall, and F1 score are used for index evaluation.

[0017] Optionally, the process of supervised classification of orthoimages based on SVM includes the following steps:

[0018] Step 5.1: Load the UAV orthoimage in envi;

[0019] Step 5.2: Make samples, that is, region of interest (ROI) files, and outline the ground object information to be classified on the image;

[0020] Step 5.3: Calculate the separability of the samples, and judge whether the samples are qualified according to the calculated values;

[0021] Step 5.4: Select the support vector machine as the classifier for this training;

[0022] Step 5.5: Post-classification processing, including changing the category color, post-classification statistics, small patch processing, and raster-vector conversion;

[0023] Step 5.6: Evaluate the classification results to determine the accuracy and reliability of the classification.

[0024] Optionally, during the process of supervised classification of orthoimages based on SVM, the overall classification accuracy and Kappa coefficient are used for evaluation.

[0025] The present invention provides a method for extracting karst mountain bare rock information from UAV high-resolution images. First, a bare rock data set is made based on the original UAV images. Then, on the basis of the DeepLabV3+ model, with MobileNetV2 as the backbone network, the bare rock information of the bare rock data set is extracted and trained, and then accuracy evaluation is carried out. Finally, a support vector machine is used to perform supervised classification on the orthoimage of the study area, and the bare rock extraction rate of the improved model is compared through experimental results. The experimental results show that the lightweight bare rock extraction model constructed by the present invention has high accuracy and robustness in the detection and extraction of bare rock information, and the improved model greatly reduces the calculation parameters and complexity, effectively reduces the dependence on hardware resources, and enables it to run stably on general computing platforms. Description of the Drawings

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0027] Figure 1 It is a schematic diagram of the study area and object of a specific embodiment of the method for extracting karst mountain bare rock information from UAV high-resolution images of the present invention.

[0028] Figure 2 It is a schematic diagram of the orthoimage of the mountain and the UAV situation in a specific embodiment of the present invention.

[0029] Figure 3 It is a schematic diagram of the comparison between the data set image and the label image in a specific embodiment of the present invention.

[0030] Figure 4 It is a schematic diagram of the DeepLabv3+ model framework in a specific embodiment of the present invention.

[0031] Figure 5 It is a schematic diagram of the bottleneck structure of the MobileNetV2 network in a specific embodiment of the present invention.

[0032] Figure 6 It is a schematic diagram of the comparison between the supervised classification result and the supervised classification accuracy in a specific embodiment of the present invention.

[0033] Figure 7 It is a schematic diagram of the comparison between the supervised classification and the deep learning bare rock rate results in a specific embodiment of the present invention.

[0034] Figure 8Schematic diagram of the training results of the DeepLabV3+ model based on the MobileNetV2 network in a specific embodiment of the present invention.

[0035] Figure 9 Schematic diagram of the comparison of bare rock recognition results using different backbone networks in a specific embodiment of the present invention.

[0036] Figure 10 Schematic diagram of the recognition results of different backbone networks in a specific embodiment of the present invention. Detailed implementation manners

[0037] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present invention, and should not be construed as limiting the present invention.

[0038] The present invention provides a method for extracting bare rock information of karst mountains from high-resolution UAV images, including the following steps:

[0039] Step 1: Use a UAV to obtain high-resolution images of the mountains in the research area;

[0040] Step 2: Construct a bare rock data set;

[0041] Step 3: Select the DeepLabV3+ model as the main framework of the lightweight bare rock extraction model;

[0042] Step 4: Use the MobileNetV2 network to replace the backbone network to perform bare rock information extraction training on the bare rock data set and conduct accuracy evaluation;

[0043] Step 5: Apply the trained lightweight bare rock extraction model to the orthophoto of the research area for bare rock extraction, and then compare it with the bare rock information obtained by supervised classification of the orthophoto using SVM, and improve the bare rock extraction rate of the model through experimental result comparison.

[0044] The following is further illustrated by specific embodiments in combination with the implementation steps:

[0045] In this embodiment, a UAV image of three mountains is obtained using the DJI Phantom 4 RTK, the camera model is FC6310_8.8_5472x3648, and the oblique photography parameters are shown in the following table. The regional information of the three mountains is as Figure 1As shown, (a) is Lingui Jinshan, (b) is Dongbian Mountain, and (c) is Yangjiao Mountain. Among them, Yangjiao Mountain is located on Liuhe Road, Qixing District, Guilin City, with a peak height of about 50 meters, higher in the east and lower in the west. Lingui Jinshan is located on Jinshui Road, Lingui Town, Lingui District, Guilin City, with an altitude of about 160 meters and a peak height of about 70 meters. The slope of the mountain body on the northwest side is relatively large, and there are a large number of buildings below. Dongbian Mountain is located on the north side of Cuizhu Road, Lingui District, Guilin City, with an altitude of about 160 meters and multiple peaks, and the peak heights are about 80 meters, 130 meters, 150 meters, and 210 meters respectively. The slope of the mountain body is relatively large near the road on the north side and near the residential community on the east side.

[0046] Table 1 UAV Oblique Photogrammetry Parameters

[0047]

[0048] In step 1, the source and preprocessing of UAV image data are carried out.

[0049] For different mountain bodies, the flight heights of the UAVs are different. The relative flight height between Yangjiao Mountain and Guijin Mountain is 120 meters. Due to the complex situation of the surrounding mountain bodies of Dongbian Mountain, the relative flight height is 200 meters for safety considerations. The flight heading overlap rate is 80%, and the side overlap rate is 75%. Five-way flights are carried out for the three mountain bodies respectively, and a total of 2792 images are taken for the three mountain bodies. Qualified images are selected to ensure that most of the mountain bodies can be presented in the images. The reference plane of the image geographical location is World Geodetic System 1984. The aerial images are imported into the Pix4D Mapper software and sequentially subjected to initialization processing, point cloud encryption, orthorectification, and mosaicking to generate orthophotos of the mountain bodies as Figure 2 shown, where (a) is Lingui Jinshan, (b) is Dongbian Mountain, (c) is Yangjiao Mountain, and (d) is the Phantom 4 RTK UAV used in this embodiment. Combining the orthophotos with on-site exploration, it is found that there are a large number of residential areas around Lingui Jinshan and Yangjiao Mountain, and Dongbian Mountain is located at the intersection of three national roads, namely Lushan Road, Jinshui Road, and Cuizhu Road.

[0050] Step 2 is the production process of the dataset.

[0051] To ensure a relatively good accuracy of the deep learning model, it is necessary to ensure the quantity and quality of the dataset. Making a dataset of drone bare rocks is a complex process that requires considering multiple steps such as data collection, annotation, and preprocessing. During data collection, ensure that the image collection has sufficient coverage and includes samples with different angles and lighting conditions. The size of the drone images obtained is 5472 pixels × 3648 pixels. Select suitable images and manually annotate the bare rock areas in the images using LabelMe. Classify the objects included in the images into two categories: bare rocks and background. Classify all features other than bare rocks as the background category. Convert the generated json file into image and label files, as Figure 3 shown, for subsequent model training.

[0052] To reduce computational complexity and memory consumption, control the complexity of the model, and accelerate the training and inference speed, it is necessary to adjust the size of the photos. Resample to adjust the size of the image and label files to 128 pixels × 128 pixels. At the same time, to improve the model performance and solve problems such as overfitting due to insufficient dataset, it is necessary to perform image enhancement operations on the training data, such as adjusting brightness, contrast, color balance, etc. Expand the dataset to 500 images to increase the diversity and robustness of the dataset. Divide the enhanced dataset into a training set, a validation set, and a test set according to 7:2:1 to ensure that the model has good generalization ability.

[0053] The main types of bare rocks in this dataset are limestone and dolomite. The acquisition time of the dataset is mainly from 9 am to 11 am, but the lighting conditions are not exactly the same. Among them, the datasets of Lingui Jinshan and Dongbian Mountain were acquired on cloudy days. The lighting on cloudy days is relatively uniform, reducing strong shadow and highlight effects. Yangjiao Mountain was photographed in sufficient light, with higher image brightness and more obvious light and shadow contrast, which helps to clearly present details such as bare rock textures and fissures. The diversity of lighting conditions enables the model to adapt to different lighting environments during the training process and improves its generalization ability in practical applications. In addition, some of the rocks in the dataset are located in complex backgrounds such as buildings, which improves the model's ability to handle complex scenes in practical applications. By covering a variety of rock types, rich lighting conditions, and diverse environmental factors, this dataset ensures the diversity and representativeness of the dataset and provides a solid data foundation for model training and evaluation.

[0054] The process of constructing the main framework of the lightweight bare rock extraction model in step 3 is as follows (illustrated in combination with relevant deep learning models):

[0055] DeepLabV3+ is a deep learning model for semantic segmentation tasks. The model structure is as Figure 4As shown, it is improved and extended based on the DeepLab series of models. The DeepLabV3+ model adopts an encoder-decoder structure. The encoder part is mainly composed of two parts: a deep convolutional neural network (DCNN) and atrous spatial pyramid pooling (ASPP). DeepLabV3+ uses atrous convolution to expand the receptive field so that the model can better understand the context information in the image. Traditional convolutional operations use a fixed-size convolutional kernel at each position, while atrous convolution introduces a variable sampling rate or "dilation rate", which can increase the receptive field without increasing the number of parameters. This enables the model to capture a larger range of context information, thereby improving the accuracy of the segmentation results. The atrous spatial pyramid pooling (ASPP) module is used to further enhance the model's perception ability of context information at different scales. By using different atrous convolution rates and pooling operations at different scales, multi-scale features are obtained, so as to better adapt to target objects of different sizes. To further improve the performance of the model, DeepLabV3+ introduces a multi-scale feature fusion mechanism in the decoder. After the deep features obtained by the feature extraction network are upsampled by a factor of 4, they are fused with the shallow features, and then through 3×3 convolutional fusion and upsampling (Upsample), the spatial information is gradually restored to obtain a clearer object boundary to obtain a more accurate semantic segmentation result.

[0056] In step 4, the MobileNetV2 network is used to replace the backbone network to perform the training of bare rock information extraction on the bare rock dataset and conduct accuracy evaluation.

[0057] 4.1. Replace the backbone network

[0058] Since the DeepLabv3+ model was proposed by Google researchers in 2018, its original text uses Xception as the backbone network. However, compared with some lighter network architectures, it still has a relatively large number of parameters and a high computational cost, and it will face huge challenges when running on resource-constrained devices. Therefore, in this invention, the MobileNetV2 network is used as the backbone network of the DeepLabv3+ model to reduce the number of model parameters and improve the recognition efficiency of bare rock. MobileNetV2 is a lightweight neural network structure mainly used for image recognition tasks on mobile and embedded devices. It is an improved version of MobileNetV1, using ReLU6 as the activation function to improve numerical stability by restricting the output range, aiming to improve the efficiency and performance of the model.

[0059] The innovation of MobileNetV2 is the introduction of the inverted residual structure (Inverted residual block) and the linear bottleneck layer. Based on the use of depthwise separable convolution, it greatly reduces the number of model parameters and computational complexity, and speeds up the network convergence rate. The main feature of the inverted residual structure is that it first performs a 1×1 convolution for dimensionality increase to increase the number of channels of the feature map, then performs feature extraction through a 3×3 depthwise separable convolution layer, and finally performs dimensionality reduction through a 1×1 convolution to reduce the number of channels of the feature map. This structure is called "inverted residual" because it adjusts the order of dimensionality reduction and increase compared with the traditional residual network. The linear bottleneck uses a linear activation function instead of the ReLU activation function commonly used in layers with smaller output dimensions in the network, which can minimize information loss as much as possible when compressing features, simplify the network structure, and improve computational efficiency. The main network bottleneck structure of MobileNetV2 is as Figure 5 , and there is a shortcut connection only when stride = 1 and the shapes of the input feature matrix and the output feature matrix are the same. The parameters of the MobileNetV2 network are shown in Table 2, where t represents the expansion factor, c represents the number of output channels, n represents the number of repetitions, and s represents the stride.

[0060] Compared with other lightweight architectures such as SqueezeNet, MobileNetV2 finds a better balance between accuracy and computational efficiency and performs excellently in multiple benchmark tests. By introducing the inverted residual module and linear bottleneck design, MobileNetV2 can efficiently capture rich features in images, thereby reducing the computational and memory burden while maintaining high accuracy. This makes MobileNetV2 an ideal choice for processing high-resolution drone images, especially in the bare rock information extraction environment that requires near-real-time applications. SqueezeNet significantly reduces the number of parameters and model size by reducing the number of channels of the convolutional kernels, but its feature extraction ability is relatively weak and it is difficult to handle tasks that require fine feature discrimination. In contrast, MobileNetV2 has more advantages for tasks with high accuracy requirements such as bare rock extraction.

[0061] Table 2 Parameters of the MobileNetV2 Network

[0062]

[0063] 4.2. Accuracy Evaluation

[0064] To evaluate the accuracy of the DeepLabV3+ model for bare rock recognition, this paper uses the following semantic segmentation evaluation metrics to effectively judge the effect of bare rock segmentation.

[0065] The Pixel Accuracy represents the proportion of the number of pixels with correct predicted classes to the total number of pixels. The calculation formula is as follows:

[0066]

[0067] The mean Intersection over Union (MIoU) first calculates the ratio of the intersection to the union of the ground truth and the prediction for each class to obtain the Intersection over Union (IoU), and then sums up the IoUs of all classes and takes the average to get the MIoU. The calculation formula is as follows:

[0068]

[0069] Precision refers to the proportion of the number of correctly predicted samples among the samples predicted as positive examples. The calculation formula is as follows:

[0070]

[0071] Recall refers to the proportion of the number of correctly predicted samples among the actual positive example samples. The calculation formula is as follows:

[0072]

[0073] The F1 score is the harmonic mean of precision and recall, which combines the accuracy and recall ability of the semantic segmentation model. The value of the F1 score ranges from 0 to 1, and the closer the value is to 1, the better the performance of the model.

[0074]

[0075] Among the above-mentioned evaluation metrics, k represents the total number of segmented classes (excluding the background value), p ii represents the number of pixels correctly identified as class i, p ji represents the predicted value of class i, p ij represents the actual value of class i.

[0076] Extraction of bare rock information based on SVM supervised classification in Step 5

[0077] Support Vector Machine (SVM) is a supervised classification algorithm based on the VC dimension theory of statistical learning theory and the principle of structural risk minimization, which is used to solve the problem of small sample nonlinear high-dimensional pattern learning and recognition. Its core idea is to construct an optimal hyperplane by mapping data to a high-dimensional space, thereby achieving efficient classification of data. The advantage of SVM is that it can effectively solve the classification problem of high-dimensional data and has good generalization ability for small sample data. In the present invention, SVM is applied to the orthophoto classification of the study area. By manually annotating and extracting features from the images, the SVM model performs nonlinear classification, maps the data to a high-dimensional space, finds the corresponding optimal classification surface in the high-dimensional space, effectively separates exposed rocks from other landform categories, and obtains the optimal bare rock information extraction result.

[0078] Specifically, after completing the classification and extraction of bare rock information, in order to objectively test the quality of the classification results, the accuracy of the image classification results must be evaluated, and the overall classification accuracy (OverallAccuracy) and Kappa coefficient are used to evaluate the accuracy of this bare rock classification.

[0079] Figure 6 (a)-6(c) are the result images obtained by the support vector machine classification. The classification results are divided into bare rock, vegetation, and other landforms. The boundaries between bare rock and vegetation areas in the figure are relatively clear. The selected SVM can classify the mountains in the study area relatively well. 6(d) is the accuracy of this supervised classification. In this classification process, the overall classification accuracy of the three mountains reached above 0.9, among which Dongbian Mountain and Yangjiao Mountain reached 0.96 and 0.97 respectively. In terms of the Kappa index, the closer the value is to 1, the higher the classification accuracy. The Kappa values ​​of Lingui Jinshan, Dongbian Mountain, and Yangjiao Mountain are 0.88, 0.86, and 0.92 respectively, and the values ​​are all above 0.85, indicating that there is a high consistency between the classification results and the real data. The overall accuracy and Kappa coefficient both show high values, indicating that the supervised classification of bare rock information extraction in the study area is reliable.

[0080] The bare rock information obtained by supervised classification is used as a reliable value to evaluate the performance of this deep learning model in orthophoto bare rock recognition, which is convenient for the optimization of the network structure in the later stage. The orthophoto area identified by the supervised classification and deep learning models is the same, and the bare rock rate is selected as the accuracy indicator for experimental comparison. In a specific area, the bare rock area identified is divided by the total area in the area to obtain the bare rock rate. Figure 7is the specific value of the bare rock rate obtained in this experiment. The absolute errors of Lingui Jinshan and Dongbian Mountain are -3.09% and 1.43% respectively, and the error ranges are both within 5%. Among them, the absolute error of Dongbian Mountain is only 1.43%. The absolute error of Yangjiao Mountain is slightly higher than that of the other two mountains, reaching -5.78%. Generally speaking, the absolute errors of the three mountains are within the acceptable range, fully proving that the DeepLabV3+ model based on the MobileNetV2 network can effectively extract bare rock information.

[0081] Further, the present invention also conducts a comparison by evaluating the training results, replacing different backbone networks, and using different models for bare rock extraction results:

[0082] The hardware environment of this experiment: a 64-bit operating system of Windows, the CPU configuration is AMD Ryzen 9 5900X 12-Core Processor with a main frequency of 3.70 GHz and a computing memory of 64 GB; the GPU configuration is NVIDIA GeForce RTX 3060. The software environment of the experiment is built by Cuda 11.6 + conda 4.5.4 + Python 3.6.5. Their compatibility and interoperability are good, and they can seamlessly use various deep learning frameworks and libraries. BCEWithLogitsLoss is used as the loss function for this model training. This is a commonly used loss function for binary classification problems. It combines the Sigmoid activation function and binary cross-entropy loss, which can provide better numerical stability and avoid gradient explosion or disappearance caused by numerical problems during loss calculation.

[0083] Since MobileNetV2 has fewer parameters and a shallower hierarchical structure, and the dynamic range of parameter updates is larger, the Adam optimizer is selected in this embodiment to automatically adjust the learning rate to ensure the stability and efficiency of the training process. When creating the optimizer, the initial value of the model's initial learning rate should usually be set smaller to avoid skipping the optimal point of the loss function during the parameter update process. Therefore, the initial learning rate is set to 0.0005 to prevent the model from not converging due to too fast parameter updates; as for the weight decay setting, it is set to 0.001 to prevent overfitting of the model. When creating the learning rate decay strategy, the exponential decay is used to adjust the learning rate, and the decay coefficient is 0.988. The exponential decay enables the model to quickly approach the optimal solution in the initial stage of training and make fine adjustments with a smaller learning step in the later stage of training to avoid oscillations caused by too large a learning rate. Finally, the number of epochs of this deep model is 500 to ensure that each parameter can be effectively updated to reach the global or a relatively optimal local optimal solution.

[0084] The training results of the DeepLabV3+ model based on the MobileNetV2 network in this time are asFigure 8 As shown. It can be seen from the figure that the training loss gradually stabilizes after 300 training times, indicating that the model gradually improves its fitting degree on the training data during the continuous learning process; the validation loss remains relatively stable after 200 training times, meaning that the model has a certain generalization ability and there is no obvious overfitting phenomenon during the training process. After 500 iterations, the training accuracy of the model reaches more than 97%. There are slight oscillations during this process, indicating that the performance of the model on the training set is gradually improving and the accuracy has reached a relatively high standard. In addition, the relatively high validation accuracy indicates that the model also performs well on unseen data. Both convey that the model has good convergence, which will help improve the recognition accuracy of bare rock.

[0085] In terms of comparing the replaced backbone networks, in this embodiment, the backbone network is replaced with Vgg16, Xception, and the lightweight network SqueezeNet, and the extraction results of each backbone network are compared. Vgg16 has a network structure with 16 layers deep, including multiple convolutional layers and fully connected layers. This depth helps the model learn complex features from shallow textures to deep semantics, which can help the model better capture the subtle features of bare rock and improve the accuracy of classification and recognition. Xception is an improvement based on Inception-V3 and combines depthwise separable convolutions to enhance the model's ability to capture features at different scales and improve the overall performance of the model, which may be more powerful for identifying images with complex textures and structures such as bare rock.

[0086] Figure 9 Shows the accuracy results of using MobileNetV2 as the backbone network based on the DeepLabV3+ model and other classic semantic segmentation networks in bare rock recognition. By Figure 9It can be seen that the Accuracy of the 4 networks for bare rock recognition all reaches over 96.95%, indicating that the 4 networks all have good performance in bare rock information extraction, and it is also feasible to apply deep learning technology to mountain bare rock information extraction. In addition, the Accuracy, F1 score, and MIoU of MobileNetV2 reach 97.39%, 78.91%, and 82.11% respectively, which are 0.16%, 1.6%, and 2.05% higher than those of Vgg16, 0.21%, 0.11%, and 1.13% higher than those of Xception, and 0.41%, 0.33%, and 1.35% higher than those of SqueezeNet. In terms of Precision, MobileNetV2 reaches 85.58%, which is 1.34% higher than that of Xception and 4.74% higher than that of SqueezeNet, and is only 0.01% lower than that of Vgg16. In terms of Recall, MobileNetV2 reaches 73.21%, slightly lower than that of Xception and SqueezeNet, and 2.72% higher than that of Vgg16. Table 3 shows the model parameters and computational complexity of different backbone network models in this experiment. When the backbone network is Vgg16, the number of model parameters is 21.55M, and the computational complexity is as high as 45.68G, and the bare rock recognition efficiency is not excellent. When Xception is replaced with the lightweight MobileNetV2, the number of model parameters decreases from 54.51M to 5.81M, a reduction of about 9.4 times, and the computational complexity also decreases from 20.66G to 6.61G, about 3 times, greatly reducing the consumption of computer memory and improving the bare rock recognition efficiency. When the backbone network becomes the lightweight SqueezeNet, its computational complexity is 0.21G lower than that of MobileNetV2, but the number of parameters and the recognition accuracy are inferior to those of MobileNetV2.

[0087] Table 3 Comparison of Model Parameters and Computational Complexity of Different Backbone Networks

[0088]

[0089] Figure 10It is the result map of the extraction of bare rock information by four different backbone networks of the DeepLabV3+ model. The first column is the original image after resampling, the second column is the manually annotated bare rock area, and the 3rd to 6th columns correspond to the result maps identified by Vgg16, Xception, SqueezeNet, and MobileNetV2 respectively. From these result maps, it can be found that all four networks can identify the main bare rock areas in the original image. However, compared with the reference labels, there are certain degrees of missed classification and misclassification in the bare rock recognition for all networks. During the recognition process, the recognition results of the other three backbone networks show fragmented or scattered blocks in some mountain areas. The recognition result of MobileNetV2 is more spatially coherent and closer to the reference label, and closer to the real bare rock area.

[0090] In addition, in order to verify the advantages of the present invention over other deep learning methods in bare rock extraction, two representative deep learning models, SegNet and U-Net, are selected for comparative analysis with the method of the present invention. SegNet and U-Net are both the most commonly used networks in remote sensing image segmentation, and the parameters are kept consistent with those of DeepLabV3+_MobileNetV2 during training. Table 4 shows the performance comparison of these three methods for bare rock information extraction. It can be seen from the table that DeepLabV3+_MobileNetV2 has the best comprehensive performance, with core indicators such as Accuracy, Precision, and MIoU slightly higher than those of U-Net and SegNet, and the number of parameters and computational complexity are significantly lower than those of the other two networks. The number of parameters and computational complexity are only 5.81M and 6.61G, which indicates that the method of the present invention has more advantages in model lightweight and bare rock recognition efficiency while ensuring accuracy.

[0091] Table 4 Performance comparison of different models for extraction

[0092]

[0093]

[0094] The comparison results of the implementation examples show that:

[0095] (1) When the backbone network is MobileNetV2, its Accuracy, F1 score, and MIoU are 97.39%, 78.91%, and 82.11% respectively, all higher than those of the Vgg16, Xception, and SqueezeNet networks, and the recognition accuracy of the improved model is also slightly higher than that of SegNet and U-Net. The overall ability to capture bare rock features and the ability to adapt to complex backgrounds are stronger.

[0096] (2) The lightweight MobileNetV2 provides an efficient feature extractor for the DeepLabV3+ model, reducing the number of model parameters and computational complexity to 5.81M and 6.61G respectively, greatly improving the model's recognition efficiency, and enabling the model to efficiently and accurately identify bare rocks in resource-constrained environments.

[0097] (3) Applying the trained deep learning model to the bare rock recognition of the orthophoto of the study area, the bare rock rates of each mountain body in two dimensions are obtained. The bare rock rates of Lingui Jinshan, Dongbian Mountain, and Yangjiao Mountain are 10.77%, 8.84%, and 12.15% respectively. The error of the bare rock rate can be controlled within about 5%, which is within the confidence interval. The extracted bare rock information can provide reliable data support for fields such as geological research, environmental monitoring, and resource management.

[0098] The above-disclosed are only one or more preferred embodiments of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. A method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles, characterized in that: The following steps are involved: Step 1: Use drones to obtain high-resolution images of the mountains in the study area; Step 2: Construct a bare rock dataset; Step 3: Select the DeepLabV3+ model as the main framework of the lightweight bare rock extraction model; Step 4: Use the MobileNetV2 network to replace the backbone network to perform bare rock information extraction training on the bare rock dataset and perform accuracy evaluation; Step 5: Apply the trained lightweight bare rock extraction model to the orthophotos of the study area for bare rock extraction, and then compare it with the bare rock information obtained by supervised classification of the orthophotos using SVM. The bare rock extraction rate of the improved model is compared through experimental results.

2. The method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles as claimed in claim 1, characterized in that: The high-resolution image in step 1 is an orthophoto of the mountain, which is obtained by the image acquired by drone aerial photography after initialization processing, point cloud encryption, orthophoto correction and mosaic stitching.

3. The method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles as claimed in claim 2, characterized in that: In the process of constructing the bare rock dataset, suitable images are first selected and the bare rock areas in the images are manually annotated using LabelMe. Then the objects contained in the images are divided into two categories: bare rock and background. All objects except bare rock are classified as background.

4. The method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles as claimed in claim 3, characterized in that: The main framework of the lightweight bare rock extraction model adopts an encoder-decoder structure. The encoder part includes a deep convolutional neural network and a spatial pyramid pooling part. A multi-scale feature fusion mechanism is introduced in the decoder part. The deep features obtained by the feature extraction network are upsampled by 4 times and then fused with shallow features. Then, the spatial information is gradually restored through 3×3 convolution fusion and upsampling.

5. The method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles as claimed in claim 4, characterized in that: In step 4, the inverted residual structure is introduced to adjust the order of dimensionality reduction and dimensionality increase. At the same time, a linear bottleneck layer is used, and the linear activation function is used instead of the ReLU activation function.

6. The method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles as claimed in claim 5, characterized in that: In the process of accuracy evaluation, pixel accuracy, average intersection-over-union ratio, precision, recall rate and F1 score are used for indicator evaluation.

7. The method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles as claimed in claim 6, characterized in that: The supervised classification process of orthophotos based on SVM includes the following steps: Step 5.1: Load the drone orthophoto in ENVI; Step 5.2: Make a sample, i.e., a region of interest (ROI) file, and outline the features that need to be classified on the image; Step 5.3: Calculate the separability of the sample and determine whether the sample is qualified based on the calculated value; Step 5.4: Select support vector machine as the classifier for this training; Step 5.5: Post-classification processing, including changing category color, post-classification statistics, small patch processing, and grid-vector conversion; Step 5.6: Evaluate the classification results to determine the accuracy and reliability of the classification.

8. The method for extracting bare rock information of karst mountains from high-resolution images of unmanned aerial vehicles as claimed in claim 7, characterized in that: In the process of supervised classification of orthophotos based on SVM, the overall classification accuracy and Kappa coefficient were used for evaluation.

Images