Real-Time Monitoring and Assessment of Natural Resources in Tourist Attractions Using Computer Vision

9660722f-0be3-4a1c-bb72-47cccb1ae866.png

Figure 1. Core steps of the proposed method

fd17c76d-e2d0-43a0-9ef9-0398ac37dd27.png

Figure 2. Flowchart of the core technical chain of the proposed method

The superpixel segmentation method for real-time monitoring of natural resources in tourist attractions proposed in this paper follows the technical chain of "preprocessing-feature optimization-parameter adaptation-precise segmentation." Figure 1 shows the schematic diagram of the core steps of this method. In the preprocessing stage, the structured edge detection algorithm and the watershed transform algorithm based on multi-scale morphological gradient reconstruction are used in coordination to enhance image edge information: the structured edge detection algorithm enhances the edge response of different resource types, highlighting key boundaries; the watershed transform algorithm based on multi-scale morphological gradient reconstruction suppresses environmental noise while retaining resource texture details, generating preprocessed images with clear edge contours and strong anti-interference capability. In the feature optimization stage, multi-dimensional feature vectors are constructed: by integrating the local detail description ability of LBP texture features with the multi-scale, multi-directional representation advantages of Gabor texture features, a mixed texture feature set is formed, while color features are also integrated to characterize different resource types. In the parameter adaptation stage, the density peak algorithm is introduced to automatically identify the density peak points in the feature space to determine the optimal number of clusters, adapting to the resource distribution characteristics of different monitoring areas. In the precise segmentation stage, based on the optimized multi-dimensional feature vectors and the adaptive clustering parameters, the fuzzy C-means clustering algorithm is used to perform the superpixel segmentation operation, using its soft segmentation feature to handle the pixel assignment problem in resource transition areas, and obtaining precise segmentation results with accurate boundaries through a single segmentation. Figure 2 shows the flowchart of the core technical chain of this method.

2.2 Superpixel segmentation

The proposed superpixel segmentation method for real-time monitoring of natural resources in tourist attractions focuses on improving edge extraction accuracy. By integrating structured edge detection with multi-scale morphological gradient reconstruction-based watershed transform, it addresses the problem of blurred edges and poor noise resistance in traditional segmentation methods for complex tourist scenes, laying the foundation for subsequent precise segmentation of natural resource types. This stage directly serves the research goal of real-time monitoring and assessment of natural resources in tourist attractions, namely, to accurately capture the boundary information of resources such as vegetation, water bodies, and rocks, supporting the quantitative analysis of key metrics such as resource distribution and dynamic changes.

The core improvement of the superpixel segmentation stage lies in the optimization of the edge extraction method. The traditional multi-scale morphological gradient reconstruction-based watershed transform algorithm relies on the Sobel operator to generate edge images. However, the Sobel operator only identifies edges by calculating color gradient magnitude, which has significant limitations in complex tourist scenes: for example, visual prominent edges such as texture boundaries formed by overlapping leaves in dense vegetation, light and dark alternating areas caused by light reflection on water bodies, and fine contours formed by rock weathering, often have weak correlations with color gradients, causing edges extracted by the Sobel operator to be broken, false, or mislocated. These problems directly affect the segmentation results of the watershed transform based on multi-scale morphological gradient reconstruction, potentially misclassifying adjacent vegetation and bare land areas as the same category or incorrectly segmenting reflective areas of water bodies as independent regions, ultimately interfering with the accuracy of natural resource status assessment. Therefore, the introduction of structured edge detection algorithms to replace the Sobel operator becomes the key improvement direction for enhancing edge extraction accuracy.

The reason why structured edge detection algorithms are suitable for edge extraction of natural resources in tourist attractions lies in their deep ability to mine "structured features." This algorithm computes based on the intrinsic structural information of local image blocks, using a structured learning framework combined with a random decision forest, specifically designed to solve the problem of predicting local edge masks. In tourist scenes, the edges of natural resources often have significant local correlations. For example, the edges between water bodies and shorelines maintain consistent directions in continuous image blocks, and the boundaries of vegetation communities show continuous transitional trends in adjacent pixels. By learning these associations, structured edge detection can break through the limitations of a single-color gradient, capturing "real edges" that align more closely with human visual perception, especially suitable for handling the edge extraction challenges caused by complex textures, lighting changes, and noise interference in tourist attractions.

The computational mechanism of structured edge detection achieves precise encoding and selection of edge features through two key steps: "intermediate mapping" and "information gain selection," which are highly adaptable to the edge characteristics of natural resources in tourist attractions. In the intermediate mapping stage, the algorithm maps the edge labels of the output space Y to a high-dimensional space Z and approximates the dissimilarity of the labels by calculating the Euclidean distance in Z. For example, for a 16×16 image block, binary vector encoding is used to represent whether pixel pairs belong to the same resource region, such as "vegetation-vegetation" or "vegetation-bare land," thus precisely capturing edge details in small areas, such as the boundary between coniferous and broad-leaved forests. The specific mapping formula is:

$\Pi: Y \rightarrow Z$    (1)

Considering the computational cost of high-dimensional space, the algorithm simplifies the mapping and uses PCA dimensionality reduction to retain distance approximation while removing noise, addressing interference issues caused by camera noise and atmospheric scattering in tourism monitoring images.

In the information gain selection stage, structured edge detection maps the high-dimensional space labels to a discrete set, ensuring global consistency of edge features. The algorithm uses K-means clustering or PCA quantization for mapping, allowing similar edge features to be assigned the same discrete label, and then selects the most discriminative features using Shannon entropy or Gini impurity. This process is independently performed for each decision tree node and can adapt to the resource distribution characteristics of different regions. For example, in the water area, the algorithm prioritizes learning "smooth continuous edge" features to distinguish shorelines; in the vegetation area, it focuses on "texture gradient edge" features to identify community boundaries. This adaptive capability allows the algorithm to simultaneously handle diverse edge types in tourist attractions, with the output edge images retaining fine textures while maintaining global continuity.

Finally, the high-precision edge images generated by structured edge detection are input into the watershed transform algorithm based on multi-scale morphological gradient reconstruction, forming an improved superpixel segmentation process. Compared with traditional methods, the superpixel boundaries output by this method are more aligned with the actual distribution of natural resources. For example, in lake monitoring, it accurately segments the fuzzy transition zone between water bodies and shoreline vegetation; in forest monitoring, it clearly distinguishes the boundaries of different tree species communities. This precise superpixel segmentation provides reliable spatial units for subsequent feature extraction and clustering analysis, directly supporting the real-time monitoring and quantitative assessment of natural resources in tourist attractions and achieving a deep coupling of technological improvements and research goals.

2.3 Color and texture feature extraction and fusion

In the color and texture feature extraction and fusion stage, this paper adopts a strategy combining Gabor texture features and LBP texture features to fully capture texture information at different scales and details, laying the foundation for accurate differentiation of resource types. Natural resources in tourist attractions, such as coniferous forests, broad-leaved forests, calm water bodies, turbulent water, and weathered rocks, have significant texture differences. The global texture of vegetation communities exhibits regular distribution, while local details such as overlapping leaves and rock cracks form fine textures. Additionally, lighting changes, seasonal transitions, and other factors easily lead to dynamic fluctuations in texture features. Gabor filters, with their excellent ability to select directions and scales, can effectively capture global texture information at different directions and scales and have strong robustness to lighting changes and slight image rotation deformations, making them highly adaptable to dynamic texture changes such as water reflections and vegetation swaying in the wind in tourism scenes. Specifically, assume that the input image is represented by d(a, b), where L is the L channel in the Lab color space. The scale and direction of the Gabor filter are represented by μ and ϕ, and the Gabor filter is denoted by h. The expression formula for Gabor texture features is:

$D(a, b, \phi, \mu)=d(a, b) * h(a, b, \phi, \mu)$    (2)

Assume the normalization factor is represented by X, and the Gabor filter definition formula is:

$\left\{\begin{array}{l}h(a, b, \phi, \mu)=X \exp \left(-a^2 / \delta_a^2-b^{\prime 2} / \delta_b^2\right) \cos \left(\mu a^{\prime}\right) \\ a^{\prime}=a \operatorname{COS} \phi+b \operatorname{SIN} \phi \\ b^{\prime}=a \operatorname{SIN} \phi-b \operatorname{COS} \phi\end{array}\right.$    (3)

To further optimize the quality of Gabor features, the algorithm smooths its amplitude information using a Gaussian low-pass filter, reducing interference caused by local texture fluctuations, and uses the PCA algorithm to reduce the 16-dimensional feature map to a 1-dimensional vector to improve computational efficiency, ensuring the performance requirements of real-time monitoring.

LBP features focus on the extraction of local texture details. Their high sensitivity to subtle texture changes and scale invariance perfectly compensates for Gabor features' limitations in characterizing local details. For example, when distinguishing between different tree species in vegetation areas, LBP features can precisely capture the local differences in leaf arrangement, while Gabor features characterize the global texture distribution pattern of vegetation communities. The complementary combination of both effectively reduces the misclassification rate of single features in complex tourism scenes, providing a comprehensive texture representation foundation for subsequent accurate clustering of resource types.

To overcome the interference of lighting, seasonal, and other factors on color features in tourism scenes, the algorithm selects three color spaces—RGB, HSV, and Lab—and extracts the first-order and second-order moments of each color component as color features to provide comprehensive and robust representation of natural resource color information. Different color spaces have varying sensitivities to environmental changes: RGB space reflects raw color information, HSV space better aligns with human color perception, and Lab space has good color separation characteristics, which can effectively resist color drift caused by lighting changes. For example, in vegetation monitoring during winter and summer, color features in the Lab space can more stably reflect changes in vegetation growth; in water body monitoring, the saturation component in the HSV space can accurately distinguish between clean and polluted water. At the same time, the algorithm incorporates spatial location features, fully considering the spatial distribution patterns of natural resources in tourist attractions, such as water bodies being concentrated in low-lying areas and vegetation being distributed on hillsides or plains. By using spatial location information to assist in distinguishing resources of similar color and texture features but different types, the discriminative ability of features is further enhanced.

To achieve effective integration of multi-dimensional features, the algorithm uses feature concatenation to combine color features, texture features, and spatial location features into high-dimensional feature vectors. These vectors are then normalized using Z-score normalization to eliminate dimensional differences between features, avoiding feature weight imbalance caused by differences in value ranges. Considering the computational efficiency requirements of real-time monitoring and the potential redundancy in high-dimensional features, the algorithm uses PCA to reduce the dimensionality of the fused feature vector. This reduces data dimensions and the computational complexity of subsequent clustering tasks while retaining key discriminative information, ensuring that the algorithm can quickly process massive monitoring images of large-scale tourist attractions.

The feature vectors optimized through the above process comprehensively integrate the color, texture, and spatial attributes of the superpixel blocks, enabling precise characterization of different types of natural resources and providing high-quality input for the automatic determination of the number of clusters and fuzzy C-means clustering segmentation.

2.4 Determination of the number of clusters

In the clustering number determination stage, this paper introduces the density peak algorithm. The core objective is to address the dynamic heterogeneity of natural resource distribution in tourist sites, achieving precise matching between the number of clusters and actual resource types, avoiding subjective biases and efficiency bottlenecks caused by manual setting of clustering parameters, and providing a scientific parameter basis for subsequent fuzzy C-means clustering segmentation. The natural resource types in tourist attractions are complex and subject to dynamic changes in distribution: on the one hand, there are significant differences in resource combinations across different regions, such as mountainous areas dominated by vegetation and rocks, and waterfront areas dominated by water bodies and wetlands; on the other hand, seasonal changes and environmental factors cause dynamic adjustments to resource type boundaries and distribution ranges, such as the expansion of vegetation in summer and the change in the area of frozen water bodies in winter. Manually setting the number of clusters is difficult to adapt to these dynamic changes. If the number of clusters is too few, different resource types will be misclassified as the same category, affecting the accuracy of evaluation indicators. If the number of clusters is too many, excessive segmentation will occur, increasing computational costs and interfering with the effective identification of resource types, which cannot meet the real-time monitoring requirements. The Density Peak Algorithm, driven by data, automatically uncovers the distribution pattern of superpixel features, which is well-suited to the core demands of the tourism scene.

The Density Peak Algorithm determines the number of clusters automatically by calculating two key indicators: local density ϑ_u and minimum distance σ_u. For each superpixel block, the local density ϑ_u is calculated based on the Euclidean distance f_uk between superpixels and truncation distance f_z, reflecting the density of similar superpixels around that superpixel. Specifically, assuming the number of the k-th superpixel area is represented by T_k, and the total number of superpixels by V, the calculation formula is:

$\vartheta_u=\sum_{k=1, k \neq u}^V T_k \exp \left(-f_{u k}^2 / f_z\right)$     (4)

In the tourism scene, superpixels of the same resource type typically have similar features and small Euclidean distances, resulting in higher local densities, while superpixels of different resource types show significant feature differences and larger Euclidean distances, resulting in lower local densities.

The minimum distance σ_u is defined as the minimum Euclidean distance between the superpixel and all higher-density superpixels, used to distinguish the core areas of different resource types. The specific calculation formula is:

$\sigma_u=\underset{k: \vartheta_k>\vartheta_u}{\operatorname{MIN}}\left(f_{u k}\right)$    (5)

In practice, the core superpixels in water body regions, due to their strong feature specificity, will have a significantly larger distance from other high-density superpixels than other regions. For the superpixels with the highest density, the algorithm defines their minimum distance σ_u separately to ensure the accurate identification of core resource types. For the superpixel region with the highest density, f_uk is defined as:

$f_{u k}=\left\|\frac{1}{T_u} \sum_{o \in E_w} a_o-\frac{1}{T_k} \sum_{o \in E_k} a_w\right\|$     (6)

The decision graph of superpixels is derived from ϑ_u and σ_u:

$\theta_u=\vartheta_u \sigma_u$    (7)

After constructing the decision graph based on ϑ_u and σ_u, the algorithm analyzes the distribution characteristics of points in the graph to identify the density peak points. Each peak point corresponds to a core cluster of a natural resource type, automatically determining the optimal number of clusters. For example, in a tourist scene containing water bodies, coniferous forests, broad-leaved forests, and bare land, the decision graph will form four distinct density peak points corresponding to four core resource types. This data-driven automatic decision-making mechanism not only avoids the cumbersome process of manual parameter adjustment but also adapts to the resource distribution characteristics of different tourism scenes, ensuring that the number of clusters aligns closely with the actual resource types. Ultimately, this stage provides precise parameter input for the subsequent fuzzy C-means clustering, ensuring the accuracy of segmentation results.

2.5 Superpixel-based fuzzy C-means clustering

The core objective of this stage is to use the soft clustering properties of the fuzzy C-means algorithm to precisely handle the boundary transition problems of natural resource types in tourist attractions and achieve efficient clustering and segmentation at the superpixel level, providing reliable category division results for subsequent real-time monitoring and quantitative evaluation of resource conditions. The natural resource distribution in tourist attractions generally exhibits significant transition characteristics: for example, the gradient zone between vegetation and bare land, the boundary between shallow water areas and tidal flats, and the mixed areas of different vegetation communities. The pixel features of these regions are fuzzy, and traditional hard clustering algorithms force each pixel to be assigned to a single category, which often leads to distorted boundary segmentation and affects the accuracy of resource area calculation, type identification, and other evaluation indicators. The fuzzy C-means clustering algorithm, by introducing the concept of membership degree, allows each superpixel to belong to multiple clustering categories simultaneously and quantifies its degree of membership. This is well-suited for the feature fuzziness of natural resource transition areas, allowing for a more realistic reflection of the spatial distribution patterns of resource types. Additionally, its stronger hierarchy and stability compared to hard clustering further ensure the reliability of segmentation results in complex tourism scenes.

The core logic of this stage revolves around the optimization of the objective function of the fuzzy C-means clustering algorithm. The objective function minimizes the weighted sum of squared distances between superpixel features and clustering centers, achieving optimal category division, with parameter designs highly compatible with the characteristics of natural resources in tourist attractions. The specific objective function expression is:

$K_l=\sum_{k=1}^V \sum_{s=1}^j T_k i_{k s}^l\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|^2$    (8)

In the objective function, the membership matrix i_ks quantifies the degree of membership of the k-th superpixel to the s-th resource category. For example, superpixels in the transition zone between vegetation and bare land will have membership degrees to both the "vegetation" class and the "bare land" class between 0 and 1, with their sum satisfying the fuzzy set constraint. The clustering center z_s represents the core feature vector of a certain natural resource category, integrating the color, texture, and spatial location features of that resource. The fuzziness factor l is usually set to 2, balancing the fuzziness and distinguishability of the clustering, avoiding category confusion due to excessive fuzziness, and reserving reasonable space for the feature expression of transition areas. In addition, the objective function introduces the number of pixels T_k in the superpixel area as a weight, so that larger resource areas occupy more important weight in the clustering optimization, aligning with the dominant position of dominant resource types in tourist attractions, and ensuring that the clustering results accurately reflect the spatial proportion relationships of resources.

Superpixel-based fuzzy C-means clustering achieves convergence of clustering results through a systematic iterative optimization process. At the same time, relying on the preprocessing advantages of superpixels, it meets the efficiency requirements of real-time monitoring in tourism scenes. Its iterative process follows the logic of "initialization-update-convergence judgment": first, the membership matrix is randomly initialized, combining the number of clusters k, which is automatically determined by the density peak algorithm in the previous section (i.e., the actual number of core resource types in the tourist attraction), to provide initial parameters for the iteration; then, using the analytical solution derived from the Lagrange multiplier method, the clustering center and membership matrix are alternately updated. The optimization problem is transformed using the Lagrange multiplier η, as follows:

$\begin{aligned} & K_l=\sum_{k=1}^V \sum_{s=1}^j T_k i_{k s}^l\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|^2 -\eta\left(\sum_{k=1}^j i_{k s}-1\right)\end{aligned}$     (9)

Furthermore, partial differential equations of K_l with respect to z_s and i_ks are calculated as follows:

$\begin{aligned} & \frac{\partial K_l}{\partial i_{k s}}=\sum_{k=1}^V \sum_{s=1}^j \frac{\partial T_k i_{k s}^l\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z s\right\|^2}{\partial i_{k s}}-\eta =\sum_{k=1}^V \sum_{s=1}^j l T_k i_{k s}^{l-1}\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|^2-\eta^{\prime} =0\end{aligned}$    (10)

$\begin{aligned} & \frac{\partial K_l}{\partial z_s}=\sum_{k=1}^V \sum_{s=1}^j \frac{\partial T_k i_{k s}^l\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|^2}{\partial z_s} =\sum_{k=1}^V \sum_{s=1}^k T_k i_{k s}^l \frac{\partial\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|^2}{\partial z_s} =\sum_{k=1}^V T_k i_{k s}^l \frac{\partial\left\|\left(\frac{1}{T_k} \sum_{o \in \varepsilon_s} a_o\right)-z_s\right\|^2}{\partial z_s} =-2 \sum_{k=1}^V T_k i_{k s}^l\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|=0\end{aligned}$    (11)

By integrating the above equations, the corresponding solutions for z_s and i_ks are:

$z_s=\frac{\sum_{k=1}^V i_{k s}^l \sum_{o \in E_k} a_o}{\sum_{k=1}^V T_k i_{k s}^l}$    (12)

$i_{k s}=\frac{\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|^{-2 /(l-1)}}{\sum_{s=1}^j\left\|\left(\frac{1}{T_k} \sum_{o \in E_k} a_o\right)-z_s\right\|^{-2 /(l-1)}}$    (13)

The updating of clustering centers makes the core features of each category more aligned with the actual resource types, while the updating of the membership matrix dynamically adjusts the degree of membership of superpixels to categories. This is particularly important for addressing feature fluctuations caused by lighting changes and seasonal transitions, and the iterative process can adaptively adjust parameters to maintain segmentation accuracy. Finally, the convergence condition is used to determine whether the iteration should stop, ensuring the stability of segmentation results and avoiding classification bias caused by insufficient iteration. The convergence condition expression is:

$\operatorname{MAX}\left\{I^{(y)}-I^{(y+1)}\right\}<\lambda$    (14)