In the study area, there are 30 5A-level scenic spots, accounting for 3.31%, 293 4A-level scenic spots, accounting for 32.34%, and 583 3A-level scenic spots, accounting for 64.35%. Because the influence of 5A-level scenic spots is much greater than that of 3A-level scenic spots and the number of 5A-level scenic spots is relatively small. Therefore, in order to accurately express the actual situation of scenic spots (Huang and Bian, 2009; Cheng et al., 2013), this study uses the influence degree F of scenic spots to represent the spatial distribution of scenic spots in the study area, that is, change the number of scenic spots in the original statistical unit window to the scenic spot influence degree F of the statistical unit window. The influence degree F of the scenic spot is calculated as follows: F = ( M 1 × W 1 + M 2 × W 2 + M 3 × W 3 ) / S where F is the influence degree of scenic spots in Northwest China; M1, M2, and M3, respectively, represent the number of 5A scenic spots, 4A scenic spots, and 3A scenic spots in a region; W1, W2, and W3, respectively, represent the weight of 5A-level scenic spot, 4A-level scenic spot, and 3A-level scenic spot in the study area; and S represents the area of scenic spots in a region. This study uses AHP, and after a series of decision-making and consistency tests, the weights are as follows: Table 1.

Geographic detector is an important method to detect the cause and mechanism of spatial pattern of geographical elements (Wang and Xu, 2017; Huo and Sun, 2021). This study uses the factor detector in the geographic detector to analyze the contribution rate of the factors affecting the spatial distribution of tourist attractions in Northwest China measured by the value of q-statistic, that is, the larger the value of q-statistic, the stronger the influence of factor X on tourist attractions, the smaller the value of q-statistic, the weaker the influence of factor X on tourist attractions, and the value of q-statistic means that X explains 100 × q% of the Y.

Deep learning is a new field in machine learning research and is currently widely applied in facial recognition, voice recognition, medical treatment, driverless vehicles, and geosciences (Cepeda-Pacheco and Domingo, 2022). Convolutional neural network (CNN) is an important algorithm used in deep learning (Liu and Tao, 2020). CNN consists of an input layer, a hidden layer with three sublayers (the convolutional, pooling, and fully connected sublayers), and an output layer and is characterized by a simple structure, few training parameters, and strong adaptability.

SqueezeNet is a classic variation of the CNN model, which can achieve the approximate effect of AlexNet on the ImageNet dataset (Liu and Tao, 2020) but with 50 times fewer parameters than AlexNet. Its core is the fire module, which consists of two layers, the SqueezeNet layer and the +expandNet layer. Among them, the SqueezeNet layer is a 1 × 1 convolutional layer, mainly used to reduce the number of input channels input to this module; the +expandNet layer contains 1 × 1 and 3 × 3 convolutional layers, mainly used for refusion and re-expression of data features so that more information can be retained; hence, it has the advantages of simple structure, high computational efficiency, and high accuracy (He et al., 2019). Therefore, the present study selected this network for simulation and prediction. First, the input layer and the first convolutional layer of SqueezeNet were modified to ensure that the input size and number of channels of the input layer could be customized according to the characteristics of the input image. This allowed for the number of channels of the first convolutional layer to equal the number of image bands. The fully connected Softmax and the classification layers were modified to obtain an output size of the fully connected layer of one. The Softmax and the classification layers were replaced by the regression layer. Please refer to the attachment for the specific code. Attachment 1 is the SqueezeNet modification code, Attachment 2 is the SqueezeNet training code, and Attachment 3 is the SqueezeNet prediction code.

The present study applied the geographic detector model under the administrative division-scale in Northwest China. The sampling tool of ArcGIS was used in which the influence F on scenic spot was taken as the Y factor, and the seven main influencing factors (transportation, rivers, cities, population, GDP, DEM, and NDVI) were taken as the X factors. Under this approach, 71 groups of sampling point data were obtained, which were input into the factor detector model of the geographic detector to identify the influence of each factor on the spatial distribution of tourist attractions (Table 2). The obtained value of q was proportional to the strength of the influence of factor X on tourist attractions. As shown in Table 2, the dominant factors affecting the spatial distribution of tourist attractions in Northwestern China were NDVI, GDP, and cities. The order of the factors with the greatest influence on the spatial distribution of tourist attractions was NDVI (0.4475) > GDP (0.4053) > cities (0.3298). Population, transportation, rivers, and the DEM had relatively minor influences on the distribution of tourist attractions.

The approach taken to determine the contributions of influencing factors on the distribution of scenic spots at a county level in Northwest China was similar to that under the municipal level. Here, 287 groups of sampling point data were obtained and input into the factor detector model to identify the impact of each factor on the spatial distribution of scenic spots (Table 3). As shown in Table 3, the dominant factors affecting the spatial distribution of tourist attractions in Northwest China were cities and NDVI, in the order of city (0.4890) > NDVI (0.2931). In contrast, rivers, DEM, GDP, population, and transportation had relatively minor influences on the distribution of scenic spots. These results under a county scale differed from those under a municipal scale (Table 2), with the GDP factor having a strong impact under the latter scale but not under the former.

In theory, the authenticity of data and the experimental effect should increase with increasing sampling points. The present study investigated the contributions of various factors to the spatial distribution of tourist spots in Northwest China under a scale of 10 km × 10 km. Here, the 30,170 groups of sampling point data obtained were input into the geographic detector model to identify the influences of various factors on the spatial distribution of tourist attractions in Northwest China (Table 4). As shown in Table 4, the dominant factors affecting the spatial distribution of tourist attractions in the northwestern region were NDVI, cities, and transportation, with an order of NDVI (0.4657) > cities (0.4236) > transportation (0.3180). In contrast, population, river, GDP, and DEM had relatively minor influences on the spatial distribution of tourist spots. The results under this spatial scale were consistent with those under the municipal- and district-level spatial scales in that NDVI and cities were factors that affected the spatial distribution of scenic spots under all three spatial scales. However, the differences between results under the three scales related to the different degrees of influence of the remaining factors and their order of influence.

In summary, the results showed that the contributions of the different factors to the spatial distribution of scenic spots changed with changing spatial scale. The results indicated that optimal results can be obtained under a particular spatial scale.

The selection of the optimal spatial scale followed the procedure outlined below. First, the scenic spot influence F (see Section 2.1.1) was calculated under different spatial scales (nonoverlapping windows) within the study area. The mean value sequence of each factor in the window was then obtained, and the correlation analysis was performed on the sequence (Figure 3). The results showed that the variation in correlation coefficients of the F values of rivers, transportation, GDP, and population with the distribution of scenic spots increased with the increasing spatial scale, and all of the peaks appeared at a scale of 150 km.

Population has the strongest correlation with the distribution of scenic spots at various scales, followed by GDP, transportation, and rivers. The changes in the correlation coefficients between the spatial distribution of scenic spots and DEM, cities, and NDVI were both consistent and different. The correlation coefficients between the spatial distribution of scenic spots and NDVI and DEM were highest under a scale of 200 km, whereas those between the scenic spots and cities were highest under a spatial scale of 400 km. The rank of the three factors according to their correlations with the spatial distribution of scenic spots was NDVI > City > DEM. The mean value of the correlation coefficients between each factor and the spatial distribution of scenic spots under different spatial scales showed that a peak correlation coefficient occurred under a spatial scale of 150 km.

After setting the spatial scale to 150 km × 150 km, the 134 groups of sampling point data were obtained and input into the geodetector model. The results showed that the rank of factors according to their influence on the spatial distribution of scenic spots was NDVI (0.5958) > cities (0.4925) > population (0.3589) > transportation (0.3075) > rivers (0.1440)> DEM (0.1048) > GDP (0.0006) (Table 5). This result showed that NDVI, cities, population, and transportation were the dominant factors affecting the spatial distribution of scenic spots in Northwest China, whereas the remaining factors had relatively little influence.

The dominant factors affecting the spatial distribution of tourist attractions in the actual situation are basically consistent with the conclusions based on the optimal scale.

The development of tourism resources in Northwest China is more limited than that in other regions. Northwest China is characterized by high terrain and harsh climate conditions, with the Tianshan mountains in the north and the Himalayas in the south. Therefore, development of tourism resources in Northwest China faces more challenges compared to that in the southern region. Northwest China also has a vast area, a relatively small population, and relatively discrete distribution of urban centers. Scenic spots are typically distributed in areas not far from populated centers and cities. Therefore, these two factors play an extremely important role in the spatial distribution of tourist attractions in Northwest China. In addition, the unique geographical characteristics of Northwest China have resulted in accessibility to transport becoming an important factor affecting the spatial distribution and the formation of tourist attractions in the region, and this factor also plays an important role in promoting the development and scale of tourist attractions.

Green vegetation in the natural environment is an important factor promoting the quality of scenic spots. The quality of the ecological environment of a scenic spot is directly related to the viewing experience of tourists. The results of the present study showed that rivers and terrain were important landscape elements influencing the distribution of scenic spots. The geomorphology of the northwestern region is complex, and this region also experiences extreme climatic conditions. These climatic conditions and the steep inclines and high altitudes of mountain regions can pose a risk to the safety of tourists, thereby restricting the development of scenic spots to some extent. In addition, Northwest China has a lower distribution of rivers, resulting in a relatively lower performance of tourist attractions. Therefore, the results of the present study showed that rivers and topography had little influence on the spatial distribution of tourist attractions in Northwest China.

Deep learning method shows two characteristics not reflected in traditional learning methods. First, the deep learning method contains a higher number of hidden layers (6–10 layers). Second, feature learning is applied in the deep learning during which the original data features are transformed from one feature space to another through feature transformation at each layer. This approach simplifies the prediction and the classification process. In addition, the process of deep learning uses big datasets as samples, and not fictitious data. Therefore, deep learning is able to produce more realistic simulations. The convolutional neural network (CNN) in deep learning is to build a simple convolutional network for image recognition; this method reduced the complexity of the network and improved its efficiency. At the same time, this approach reduced model overfitting and improved simulation accuracy. Others used the CNN network to extract image features and have shown this method to have high accuracy (Cepeda-Pacheco and Domingo, 2022). The high accuracy of the method is expressed as sample information closer to real values (Liu and Tao, 2020). Under the CNN algorithm, many samples are trained through a supervised learning mode, whereas the model is fine-tuned against other samples to reduce model overfitting during training (He et al., 2019).

In recent years, the number of layers of CNN has been increasing, from 7-layer AlexNet to 16-layer VGGNet, then to 22-layer GoogleNet, 152-layer ResNet, and more complex networks with thousands of layers. Although the network performance has been improved, with it comes the problem of efficiency. The problem of efficiency mainly refers to the problem of model storage and model prediction calculation. Only by solving the problem of efficiency can the CNN go out of the laboratory and be more widely used in daily life. In this regard, the usual method is to compress on the already trained model so that the network carries fewer parameters, thus solving the memory and speed problems. In contrast, the main idea of the lightweight model is to design a more efficient “network computing method” so that the network parameters are reduced without excessive loss of network operations. This article mainly studies the application of convolutional network SqueezeNet in scenic spot adaptability assessment. Due to the advantages of high accuracy, less time-consumption, and limited personal computer configuration, the network can quickly iterate and try different settings during data training. Zhao et al. (2019) applied SqueezeNet to conduct forest fire detection research and also came to the same conclusion as aforementioned. Future research on scenic spots should apply more cutting-edge artificial intelligence science and technology, such as other neural network models of deep learning (Random Forest and Long Short-Term Memory Network). This would allow for the sustainable development of scenic spots.

The method proposed in the present study is an improvement to the traditional research methods used for studying tourist attractions. Previous studies of the spatial distribution of tourism attractions mostly applied mathematical statistics, GIS, and the analysis of influencing factors through geographic detectors. These studies suffer from a certain level of uncertainty due to the application of key technologies (Zhu and Chen, 2008; Wu et al., 2013; Wu et al., 2017; Li et al., 2018; Zhu et al., 2019; Xie et al., 2021). For example, the use of a geographic detector model to analyze the factors influencing the spatial distribution of scenic spots will provide different results under different spatial scales. Previous studies have generally examined these relationships under a county-level or municipal administrative regions spatial scale. Moreover, the geographic detector model requires an independent variable to be a categorical variable. Difference in the data discretization method within the quantification of different indicators to a unified scale will also lead to differences in results. When applying GIS sampling, the number of samples will increase as the spatial resolution increases. The geographic detector tends to give errors under a large quantity of sample data (Cheng and Hu, 2014; Fengjiao and Wang, 2014; Tang et al., 2018; Jia et al., 2019; Tang et al., 2019). The present study used Python scripting to conduct multiscale sampling in the study area. This approach allowed for the scale-dependence of the results to be avoided by selecting the most suitable spatial scale for sampling in the study area through a correlation analysis. Future research should consider a wider range of influencing factors in the model, such as policy factors, cultural factors, preferences of tourists, and other emotional factors, thereby increasing the accuracy of the model results and providing a solid scientific reference for the development of regional tourism according to local conditions. The use of deep learning methods is suitable for identifying new areas with high suitability for development into high-level scenic spots and is relatively novel within the field of tourism science (Liu et al., 2017; Romão et al., 2018; Wu et al., 2018). In addition, the present study identified new areas with a high potential for tourism development based on existing resources with the aim of sustainable tourism development (Blancas et al., 2011). Innovative green tourism should be promoted in these areas, such as ecotourism, agricultural tourism, forest tourism, grassland tourism, marine tourism, and desert tourism.