The content of this article is from the Journal of Surveying and Mapping, Issue 1, 2024 (drawing review number: GS Jing (2024) No. 0107)
Knowledge-guided intelligent recognition of fragmented raster topographic map scales
Ren Jiaxin 1, 2, 3
, Liu Wanzeng2,3,4
, Chen Jun2,3, Zhang Lan2,3,5, Tao Yuan2,3,5, Zhu Xiuli2,3,4, Zhao Tingting2,3,4, Li Ran2,3,4, Zhai Xi2,3,4, Wang Haiqing2,3,4, Zhou Xiaoguang1, Hou Dongyang1, Wang Yong6 1. School of Earth Sciences and Information Physics, Central South University, Changsha 410083, Hunan, China;
2. National Basic Geographic Information Center, Beijing 100830, China;
3. Key Laboratory of Spatio-temporal Information and Intelligent Services, Ministry of Natural Resources, Beijing 100830, China;
4. Hubei Luojia Laboratory, Wuhan 430079, China;
5. School of Environment and Geomatics, China University of Mining and Technology, Xuzhou 221116, Jiangsu, China;
6. Chinese Academy of Surveying and Mapping, Beijing 100830, China: National Natural Science Foundation of China (42394062); National Key R&D Program of China(2022YFB3904205); Abstract of the open fund project (220100037) of Hubei Luojia Laboratory: The scale bar is an important basis for determining the secret level of topographic maps. In order to solve the problem of scale judgment of fragmented raster topographic map, this paper condenses the prior knowledge of map-scale features to guide the construction of Expert Knowledge Image Pyramid Dataset (EKIPD), and then uses the deep convolutional neural network algorithm to model to construct a knowledge-guided, data-driven, algorithm-centered hybrid intelligence model of knowledge, data and deep convolutional neural network coupling. The optimal recognition size (ORS) was obtained by counting the sample distribution of fragmented topographic maps of different sizes in EKIPD, and then the topographic maps to be recognized were segmented with ORS as the step size. The model was used to predict each subgraph, and the scale of the fragmented raster topographic map was obtained by integrating the prediction results of the subgraph. After experimental verification, the recognition accuracy of the proposed method is about 97%, which proves the effectiveness of the proposed method. Keywords: Intelligent Surveying and Mapping Expert Knowledge Hybrid Intelligence Raster Topographic Map Scale Recognition Deep Convolutional Neural Network
Citation format:Ren Jiaxin, Liu Wanzeng, Chen Jun, et al. Knowledge-guided Intelligent Scale Recognition of Fragmented Raster Topographic Maps[J]. Journal of Surveying and Mapping,2024,53(1):146-157. DOI: 10.11947/j.AGCS.2024.20230005
REN Jiaxin, LIU Wanzeng, CHEN Jun, et al. Knowledge-guided intelligent recognition of the scale for fragmented raster topographic maps[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(1): 146-157. DOI: 10.11947/j.AGCS.2024.20230005 阅读全文：http://xb.chinasmp.com/article/2024/1001-1595/20240113.htm
Introduction Topographic map is to abstract, summarize and express the natural and human geographical elements such as residential areas, roads, water systems, borders, topography, and vegetation on the earth's surface in accordance with the cartographic specifications, and are widely used in economic construction, national defense construction, and scientific research, and are important basic and strategic information resources and production factors of the country. At the same time, topographic maps are the basis for battlefield situational awareness and visualization, an important strategic resource for winning the initiative in modern warfare, and an important "important weapon of the country, which cannot be compared with others". According to Articles 6, 7 and 18 of the notice of the Ministry of Natural Resources and the State Secrets Administration on the issuance of the "Provisions on the Scope of State Secrets in the Management of Surveying, Mapping and Geographic Information" (Natural Resources Development [2020] No. 95), the mainland topographic map is a state secret, and its production, management, circulation, use and destruction must comply with the relevant national confidentiality regulations, and the whole life cycle can be traced and managed. Confidentiality appraisal is an important technical link in the traceability management of topographic maps, which is to evaluate the confidentiality level of suspected confidential topographic map data in accordance with the relevant regulations on the management of national confidential surveying and mapping results, and give an appraisal opinion on whether it is confidential and the corresponding confidentiality level. Fragmented raster topographic maps mainly refer to the deliberate deletion of the map number, coordinates, scale and other information on the topographic map that can be used to explicitly judge the scale of the topographic map in order to avoid inspection, or to cut out part of the incomplete topographic map from the complete topographic map (Fig. 1(b)). It is necessary for experienced cartographers to conduct a comprehensive evaluation according to the characteristics expressed by different geographical elements on different scale topographic maps, combined with their own cartographic experience, to infer the scale of topographic maps, and then determine their confidentiality level.
| |
| 图 1 不同类型的碎片化栅格地形图Fig. 1 Different types of fragmented raster topographic maps |
| Diagram options |
The shape, color and size of map symbols are often closely related to the scale of topographic maps, so they become an important basis for map scale identification. However, on topographic maps of different scales, the same feature has different shapes and sizes, and conversely, different figures may also have the same shape and size, resulting in the phenomenon of "homogeneous objects" and "homogeneous objects" (Fig. 1(c)). Furthermore, topographic maps with similar scales (such as 1:25 000, 1:50 000, 1:100 000) use the same map symbology and show very little difference on the map, and map experts need to spend more energy to determine the scale of the topographic map, resulting in slow manual detection, greatly reducing the efficiency of topographic map verification, and is not conducive to the protection of classified topographic maps. In recent years, artificial intelligence technology has been widely used in the field of surveying and mapping because it can replace experts to automatically design features [1-4]. Ref. [5] uses deep convolutional neural network (DCNN) to realize the intelligent detection of significant error areas in the "problem map" at multiple scales, with an accuracy rate of more than 80%. Ref. [6] combines vector data with raster images, and uses neural networks to learn high-level fuzzy features to distinguish overpass types, so as to classify complex overpass structures in OSM (open street map). In Ref. [7], a hyperspectral image classification method based on graph convolutional network is proposed, in which a topological map is constructed based on the spatial spectral information of the image, and the feature information of neighboring nodes is aggregated by the graph convolutional network, which can achieve high classification accuracy with fewer training samples. In Ref. [8], a non-subsampled shear wave transform fusion method based on the combination of parametric adaptive pulse-coupled neural network model and maintaining energy attribute fusion strategy was proposed, which can fuse complex remote sensing images with greatly enhanced adaptability. Since the scale of fragmented topographic maps is a kind of implicit information, the classification of fragmented topographic maps requires multi-node coupling and multi-factor comprehensive evaluation of map symbols, annotations, information loads, map semantics, cartographic synthesis and other knowledge of different scales with artificial intelligence algorithms, so as to obtain relatively reasonable results, and there is no good solution at present. Therefore, this paper proposes a knowledge-data-deep convolutional neural network coupled hybrid intelligence model (KDD-HIM) based on the knowledge-data-deep convolutional neural network for fragmented topographic map scale discrimination. Experts condense the knowledge of topographic maps, combine the characteristics of DCNN to gradually abstract the image and automatically design features, guide the construction of expert knowledge image pyramid dataset (EKIPD), and then use the popular DCNN for multi-scale training, and count the optimal recognition size (ORS) The sub-map is segmented, and the prediction results of the sub-map are integrated to obtain the final topographic map scale. 1 The general idea of scale intelligent recognition of fragmented raster topographic mapThe biggest difference between raster topographic map and vector topographic map is that raster topographic map is composed of regular arrays, and the same spatial target will be decomposed into multiple elements (pixels) in the array that are independent of each other and do not affect each other; Therefore, the recognition of the scale of fragmented raster topographic map is not a simple spatial calculation, but a kind of map cognition problem, the core of which lies in how to mine the mapping relationship between the characteristics of topographic map elements and the corresponding scale based on the existing cartographic knowledge. For ease of explanation, the topographic maps in this paper are limited to fragmented raster topographic maps. As shown in Figure 2, this paper draws on the ideas of intelligent surveying and mapping [9-10] and knowledge service [11-12] to excavate, extract, and extract the knowledge of topographic maps and deep convolutional neural networks. Describe and express, and then summarize and condense the relevant knowledge of topographic map scale; use prior knowledge to construct an expert knowledge image pyramid dataset for topographic map scale recognition, and use the dataset to proxy expert prior knowledge to realize the deep coupling of prior knowledge and convolutional neural network, which improves the algorithm's attention to scale; a set of end-to-end hybrid intelligent computing ideas is designed to mine the mathematical relationship between topographic map elements and scale, and realize the logical self-consistency of high-level semantic concepts between topographic map elements and scale, which can be fast, accurately identify the scale of the topographic map.
| |
| 图 2 知识引导的碎片化地形图比例尺智能识别总体思路Fig. 2 Overall idea of knowledge-guided intelligent recognition of the scale for fragmented topographic maps |
| Diagram options |
2 Knowledge-guided Intelligent Recognition Model and Algorithm of Fragmented Raster Topographic Map ScaleIn order to realize the knowledge-guided intelligent recognition of fragmented raster topographic map scale, this paper constructs an expert knowledge image pyramid dataset guided by knowledge, and trains the dataset through deep learning algorithms to form knowledge, The hybrid intelligence model coupled with data and deep convolutional neural network is mainly divided into three parts: (1) the construction of expert knowledge image pyramid dataset, which uses the prior knowledge of topographic map to preprocess the original topographic map and construct a knowledge-guided multi-scale pyramid dataset to enhance the sensitivity of the algorithm to size and improve the accuracy of scale recognition, and (2) the real-time data augmentation method that takes into account expert knowledge, taking into account the prior knowledge of topographic map experts, and combines mixup [13] and real-time data augmentation method in small sample scenarios [5] (3) Integrated inference under the optimal recognition size, the optimal recognition size is explored through statistics and testing, and the recognition results are optimized by the ensemble algorithm. The model architecture is shown in Figure 3.
| |
| 图 3 KDD-HIM混合智能计算模型架构Fig. 3 Hybrid intelligent computing architecture of KDD-HIM |
| Diagram options |
2.1 The construction of topographic maps from expert knowledge image pyramid datasets is drawn in accordance with strict cartographic specifications and a unified symbol and color system, which has strong artificial prior knowledge constraints. The size of the DCNN feature map decreases with the deepening of the network depth, and the same object has different expressions in different network layers, which coincides with the characteristics of the same feature in the "surface-line-point" gradually reduction under different scale topographic maps. Through the condensation of topographic map knowledge by experts, combined with the above characteristics of DCNN, the construction of expert knowledge-guided dataset can amplify the sensitivity of DCNN to size and improve the accuracy of scale recognition. From the expert knowledge condensed in Table 1, it can be concluded that size is the key to identify the scale of fragmented topographic maps, and it is necessary to cut and scale the topographic maps to the same size in different ways while ensuring the information of the topographic map, so that DCNN can recognize the topographic map scale according to the differential expression of different scales of the same type of features under the same size. Accordingly, a fixed size, aspect ratio and deformation cutting method with fixed sub-graph number (Fig. 4) were proposed to construct an expert knowledge image pyramid dataset for topographic map scale recognition.
表 1 用于比例尺识别的DCNN和地形图专家知识Tab. 1 Expert knowledge of DCNN and topographic maps for scale recognition
Table options
| |
| 图 4 专家知识图像金字塔数据集构建流程Fig. 4 Flowchart of expert knowledge image pyramid dataset |
| Diagram options |
2.1.1 The proportional cutting method of fixed size and aspect ratio is shown in Figure 4(a), the cutting size is fixed, the original topographic map is divided, the sub-map obtained by segmentation is saved and scaled to the unified input size (UIS) to the expert knowledge image pyramid dataset. Because the cut size and the input size are square, the width and height ratio of the scaled sub-topographic map remains unchanged and will not be deformed, so it is called the equal scale cutting method
(1) In the formula, BS represents the base size, and the UIS of the current popular DCNN classification algorithm is generally 299 pixels, if the cut is made at this size, it cannot carry enough ground object information and will bring a huge amount of computation, so the BS is set to 2 × UIS, that is, 598 pixels; factor represents the cutting coefficient, in order to ensure that the subgraph obtained by cutting has a large change, it is generally taken as a multiple of 2;split size indicates the cutting size, which is determined by the cutting coefficient based on BS before each round of cutting. By changing the cut size several times and then cycling through the above steps, the equiscale data set D1 was obtained.
2.1.2 The fixed number of sub-maps is shown in Figure 4(b), refer to the proportional cutting method in section 2.1.1, and the number of sub-maps is fixed before each cut, and the original topographic map is divided. Because a topographic map with a complete format is generally rectangular, and the sub-map obtained by segmentation is often rectangular, and the UIS will be deformed when scaled into a square, it is called the deformation cutting method
(2) In the formula, the side represents the number of rows and columns of the number of subgraphs, and the number of rows and columns is generally equal, in order to ensure that the cut subgraphs have great differences and take into account the computational efficiency, the side is limited to the prime number within 10, and the subimage is the number of subgraphs, that is, the square of the side, which is determined by the side before each round of cutting. By changing the number of subgraphs several times and then cycling the above steps, the deformation dataset D2 is obtained. Through the above two methods, EKIPD is constructed to simulate fragmented topographic maps of different sizes in the real world, and the attention mechanism of the model to the scale of the topographic map is generated, so as to accurately identify the scale of the fragmented topographic map.
2.2 Real-time data augmentation method taking into account expert knowledgeData augmentation uses various transformation methods such as rotation, clipping, and noise enhancement to artificially expand the training dataset on the basis of existing data, which can improve the accuracy and generalization ability of DCNN, and is widely used in DCNN training. However, there are significant differences between topographic maps and natural images, and it is necessary to optimize the data augmentation method based on topographic map knowledge to expand EKIPD. According to the expert knowledge condensed in Section 2.1, size is the key to the scale of DCNN topographic map recognition, and changing the image size will affect the scale recognition accuracy. Unlike traditional data augmentation methods, EKIPD cannot be processed using size-dependent data augmentation methods (e.g., random cropping, horizontal stretching, vertical stretching, etc.). As shown in Figure 5, considering the expert knowledge condensed in Section 2.1, the knowledge-based topographic map augmentation (KTMA) method was selected to exclude the data augmentation method that changed the size
(3)
| |
| 图 5 基于知识优选的地形图增强方法集合Fig. 5 Knowledge-based topographic map augmentation collection |
| Diagram options |
As shown in Fig. 3(c), combined with the real-time data augmentation method proposed in the small-sample scenario proposed in the mixup and literature [5], the mixup algorithm is improved on the basis of KTMA, and a real-time data augmentation method considering expert knowledge is developed
(4) where k indicates that there are k data augmentation methods in this round of training;
Indicates rounding down;|KTMA|denotes the number of augmentation methods in the topographic map augmentation method set;n represents the current training round;E represents the total number of training rounds;aug represents the set of randomly selected data augmentation methods in the current round of training;x′ and y′ represent the synthesized image data and labels;(xi, yi) and (xj, yj) are the two randomly selected samples in the training data, xi and xj represent the original image data, yi and yj represent the original labels;Beta represents the beta distribution;α∈(0, ∞); λ∈[0, 1] indicates the mixing ratio, and in this paper, if all images contribute the same contribution, then λ=0.5 is taken, and Figure 6 shows the samples generated by different λs.
| |
| 图 6 不同混合比例生成的训练样本Fig. 6 Training samples generated by different mixing ratios |
| Diagram options |
In the training process, the training samples are generated in real time and paired in pairs to obtain the final mixed samples and labels, which can increase the number of samples without bringing additional time and space overhead, and can increase the accuracy and robustness of the model.
2.3 Topographic map scale recognition model based on deep convolutional neural networkThe selection of the model needs to be combined with the characteristics of the specific problem and the sample data [6, 17]. The performance of DCNN is closely related to the resolution of the sample data, the depth of the model, and the width of the model, while the model design of DCNN often relies on engineering experience and there is no clear mathematical formula to refer to. Researchers often improve the performance of the model within a certain range by modifying a certain parameter (e.g., increasing the depth of the model), but they will soon encounter performance bottlenecks or even loss of accuracy [18]. In order to solve the above problems, EfficientNet [15] uses a composite scaling method to explore the interaction between sample resolution, model depth, and model width at the same time, so as to achieve the optimal trade-off between model accuracy and speed, and achieve state-of-the-art (SOTA) performance in image recognition related tasks
(5) where d, w, and r are the coefficients used to scale the width, depth, and resolution of the model, respectively, and ⊙ represent continuous operations.
X represents the input tensor;
Represents the height, width, and number of channels of X, target_memory is the limited memory size, and target_flops is the limited amount of computation. Considering that the scale of the topographic map is related to the seriousness of confidential identification, the accuracy of the model is required, and the size of the topographic map to be identified is generally large, which needs to be split into sub-topographic maps for scale recognition, and the model needs to have a fast inference speed. In order to balance accuracy and speed, this paper uses the lightweight and high-performance EfficientNet for feature extraction and model training.
2.4 The performance of the ensemble inference DCNN model at the optimal recognition size is related to the resolution of the image, and the larger the image resolution, the stronger the performance of the model [15, 19]. However, in the real world, even the fragmented topographic map size far exceeds the unified input size of the DCNN algorithm, so it is necessary to study the optimal recognition size integration strategy to avoid the recognition size being too small and misleading the model to produce false predictions. As shown in Fig. 3(d), in order to maintain the consistency of image size in the training and prediction stages, and to solve the problem of information loss caused by scaling of large-size topographic maps, the image of the topographic map to be recognized is segmented according to the optimal recognition size ORS, and then the recognition results of the sub-map are ensembled to learn to obtain the final prediction results.
2.4.1 Topographic map cutting under the optimal recognition size should maintain the same resolution in the model training and prediction stages in order to give full play to the inference effect of the model [20-22]. In order to overcome the problem of large differences in recognition results under different sizes, the sample size of topographic maps of different sizes in EKIPD was counted, and the top 5 distribution of sample size shown in Table 2 was obtained. In order to maintain the consistency of the training and prediction process and take into account the prediction efficiency, the ORS was set to 598 pixels with the largest number of samples, and the dynamic cutting or scaling was carried out according to the size of the topographic map to be detected
(6)
表 2 EKIPD中样本数量top 5分布Tab. 2 Distribution of top 5 samples in EKIPD
Note: † indicates that due to the slight difference in the resolution of different original images, there is also a slight difference in the resolution of the sub-image obtained by the fixed number of sub-image deformation cutting method, so the cutting parameters are used instead of the resolution of this part of the image, and the same is true below. Table options
where size(·) is the shortest side of the topographic map to be tested;I is the topographic map to be tested;Iwidth and Iheight are the width and height of the topographic map to be tested, respectively;resize(·) When the shortest side of the topographic map to be detected is less than or equal to ORS, the topographic map is directly scaled to the UIS for prediction. When the shortest side of the topographic map to be detected is larger than the ORS, the topographic map is cut window by window according to the ORS, and then the sub-image is scaled to the UIS for prediction.
2.4.2 Probability-weighted ensemble inference ensemble learning [23] allows researchers to design combinatorial solutions for specific machine learning problems by combining multiple simple models to obtain a more powerful combinatorial model [24-27]. Ref. [28] mathematically illustrates three basic reasons why ensemble learning is better than a single classifier: statistics, computation, and representativeness. Drawing on the idea of ensemble learning, after obtaining the inference results of each segmented subgraph, the inference vectors obtained by integrating each subgraph are integrated to obtain the final topographic map scale, which makes the ensemble results have lower variance and better generalization ability
(7) where pi is the inference vector of the ith segmented subgraph, T represents the total number of topographic maps to be detected is cut into T subgraphs, and argmax represents the index when the maximum probability value is obtained, i.e., the scale. 3 Test and Analysis3.1 The hardware environment for the test equipment and data test is 64 GB memory, Intel(R) Xeon(R) Gold 5222 [email protected] GHz processor, and Nvidia Quadro P4000 8 GB graphics card. All DCNNs in this paper are implemented based on the Python language and the deep learning library Tensorflow. According to the notice of the Ministry of Natural Resources and the State Administration of Secrets on the issuance of the "Provisions on the Scope of State Secrets in the Management of Surveying, Mapping and Geographic Information" (Natural Resources Development [2020] No. 95), the topographic map with a standard scale of 1:5000 and larger cannot meet the area requirement of 25 km2, and the fragmented topographic map derived from it cannot meet the area requirement of 25 km2, so it is not within the scope of this paper. Accordingly, 100 topographic maps of 1:10 000, 1:25 000, 1:50 000, 1:100 000, 1:250 000 and 1:500 000 scale were selected for the experimental dataset, with a resolution of 11 000×7000 to 13 000×8000, and a total of 557 original topographic maps were eliminated after excluding the damaged topographic maps. The EKIPD constructed by using the above original topographic maps includes 141 798 fragmented topographic maps with a resolution of 598×598 to 13 000×8000 different sizes, which are divided into training sets and test sets at a ratio of 4:1.
3.2 In order to fully verify the effectiveness of the method in this paper, the method is designed to be precision, recall, F1 score, accuracy, macro-average precision, macro-average recall, macro-average F1 score, Based on the weighted-average precision, weighted-average recall and weighted-average F1 score of the sample size, the comprehensive performance of the model was evaluated in an all-round way. In the experiment, EfficientNet, which was pre-trained on ImageNet, was used to perform transfer learning on EKIPD. Among them, the EfficientNet of KDD-HIM for transfer learning proposed in this paper is not used as the benchmark algorithm, and is denoted as Baseline-EfficientNet. The training iteration round (epoch) of the algorithm was uniformly set to 30, and the stochastic gradient descent (SGD) method was used to optimize the model, and the initial learning rate was 0.003, which was attenuated according to Eq. (8).
(8) In the formula, lr represents the learning rate adopted in each round, lr0 represents the initial learning rate, decay represents the decay coefficient, which is taken as 0.95, epoch represents the current training round, factor represents the decay frequency, and 2 represents the learning rate that is attenuated every two rounds.
3.3 Experimental results and analysisBy comparing with external algorithms, Table 3 shows the prediction results of KDD-HIM and Baseline-EfficientNet on EKIPD, and it can be seen that KDD-HIM is significantly better than Baseline-EfficientNet in all 7 indicators, which effectively proves the effectiveness of the proposed algorithm. Figure 7 shows the confusion matrix between KDD-HIM and Baseline-EfficientNet on EKIPD, which can accurately identify the topographic maps at a scale of 1:10 000, but there is a serious misclassification of other scales, especially the 615 topographic maps with a scale of 1:500 000. KDD-HIM achieved the best classification effect in all 6 scales, and significantly reduced the misclassification of 1:25 000~1:500 000 scales.
表 3 Baseline-EfficientNet与KDD-HIM对比Tab. 3 Comparison between Baseline-EfficientNet and KDD-HIM
Table options
| |
| 图 7 Baseline-EfficientNet与KDD-HIM在EKIPD测试集上的混淆矩阵Fig. 7 The confusion matrix of Baseline-EfficientNet and KDD-HIM with EKIPD |
| Diagram options |
Table 4 shows the recognition performance of KDD-HIM at all 6 scales, all of which are above 90%, especially the 1:10 000 and 1:25 000 scales, and a small number of recognition errors on other scales. The analysis shows that there are two main cases in the above-mentioned misidentification cases: (1) the clutter element is too single (such as containing a large number of water bodies), which is an extreme case and rarely occurs in the training set, and the model needs to give priority to the distribution of clutter elements in most of the samples, so the misclassification of this part of the sample is generated; (2) the smaller scale (relative to 1∶ 10 000) carries too rough clutter symbols and lacks the detailed information of cleats, resulting in the wrong recognition of the model.
表 4 KDD-HIM在各比例尺下的性能对比Tab. 4 Performance comparison of KDD-HIM at various scales
Table options
In order to further study the influence of topographic map size on recognition accuracy, the topographic map obtained by each cutting method in EKIPD was independently trained, and the model test results under a single size were obtained (Fig. 8). Among them, the black line represents the model prediction results obtained by the proportional cutting method of fixed size and aspect ratio. As can be seen in Figure 8, as the size of the image increases, the performance of the model increases (black, solid markings), probably because more information is included in the sample. When the size of the topographic map is increased to 897×897, the model performance reaches its highest, and further increasing the size of the topographic map will reduce the performance of the model (black, hollow markings). This may be due to the fact that the size of the topographic map is too large, although it contains a large number of features, but a large amount of detail information is lost due to the zoom for UIS, resulting in recognition errors. The red line represents the deformation and cutting method with a fixed number of subplots, and as the number of subplots increases, the image size becomes smaller and the performance improves. In particular, the performance of the 299×299 and 2 rows and 2 columns models is significantly lower than that of other sizes, which proves that too small or too large training size will weaken the performance of the model, respectively.
| |
| 图 8 单一尺寸地形图对模型性能的影响Fig. 8 Influence of single-size topographic map on model performance |
| Diagram options |
Fig. 9 shows the specific confusion matrix of the single-size data model on the EKIPD test set, and it can be seen that the single-size topographic map cannot completely take into account the six scales of the whole world, and can only achieve excellent performance in a few scales, and even the 897×897 model with the best performance still has the situation that 116 1:250 000 topographic maps are misdivided into 1:500 000 topographic maps. By fully blending the single-size training data to form a multi-scale dataset, there are only 36 topographic maps of the same type that are misdivided into 1:500 000, and the error ratio is reduced by nearly 70%, which significantly improves the recognition performance of the proposed method on the global scale.
| |
| 图 9 单一尺寸数据模型在EKIPD测试集上的混淆矩阵Fig. 9 The confusion matrix of a single-scale model with EKIPD |
| Diagram options |
4 Conclusion: Geographic information confidentiality appraisal is an urgent need for the implementation of national geographic information security supervision. In order to solve the problem of intelligent determination of the confidentiality level of fragmented raster topographic maps, this paper proposes a hybrid intelligent model coupled with knowledge, data and deep convolutional neural network by condensing the knowledge of DCNN and topographic maps, and successfully excavates the mapping relationship between the features of topographic map elements and the corresponding scales, and the recognition accuracy of the proposed method is 0.974 91, which is nearly twice the performance compared with the original DCNN network, and can effectively identify the scale of fragmented topographic maps. The classification identification system constructed based on this method has been applied in the National Basic Geographic Information Center, which replaces the traditional time-consuming and laborious manual verification work, and provides an effective means for the security supervision of classified geographic information. Although the method in this paper has been able to effectively identify the scale of fragmented topographic maps, there are still a small number of topographic maps to be detected that will be misclassified, and the method to further improve the accuracy will be studied in the future. About author:REN Jiaxin (1993—), male, Ph.D. candidate, research direction is intelligent surveying and mapping. E-mail:[email protected] Corresponding author: Liu Wanzeng E-mail:[email protected]
First trial: Zhang Yanling review: Song Qifan
Final Judge: Jin Jun
information