Forest statistics presents forest information on the quantitative and qualitative conditions of forest structural features, which underpin forest planning and policy (Fatehi et al., 2017). Traditional canopy estimation methods involve field surveys and measurements, which can be labor-intensive and time-consuming (Jing et al., 2023). In this research, therefore, a using a two-stage approach was used to analyze the relationships between variables and estimate the dependent variable. This approach includes selecting the most related variables to the SSOM neural network and then estimating the dependent variable (canopy) with this network.
Our results of the observed patterns reveal that changes in elevation, slope, and aspect are linked to changes in the diameter at breast. Similarly, Jin et al. (2008), Daghestani et al. (2017), Goodarzi et al. (2012), and Zhang et al. (2016) concluded that different environmental factors, including the elevation, slope, and aspect could play a critical role in the development and formation of plant communities and vegetation. These factors indirectly affect soil agents and cause major changes in the quantitative and qualitative features of trees, such as the diameter at breast and the canopy percentage. Wanga et al. (2015) found a linear relationship between topography and plant distribution and introduced the slope direction as one of the major factors affecting vegetation distribution. Hosseini et al. (2008) presented evidence that the tree canopy percentage increased initially and then decreased with increasing elevation. They observed the highest canopy in middle-height classes because of harvesting trees at low-height classes. According to Farhadur Rahman et al. (2022), topography variables (elevation, slope, and aspect) affect canopy height by indirect effects on solar radiation, disturbance, wind direction, and soil erosion.
Based on the results of previous studies, a nonlinear relationship apparently exists between independent variables and tree canopies. Using ANNs is one of the most fundamental methods to understand complex relationships. Since forest ecosystems are highly complex with interrelated variables, neural networks can better manage complex nonlinear relationships and learn large datasets than traditional statistical methods (Kucuk and Sevinc, 2023; Zhu et al., 2024). In recent years, neural networks have been increasingly used to estimate the canopy because of their ability to process complex and extract meaningful data. Their ability to generalize training data to unobserved samples has made them a useful tool for forest management (Bayat et al., 2020; Ghasemi et al., 2022; Tian et al., 2022). The literature indicates that ANNs as a novel modeling method are of interest for major reasons, including pattern recognition ability, good input-output relationship, less sensitivity to errors in input data, fully parallel processing, the need for less input data, ability to discover and predict between-variable relationships, simulation ability despite incomplete input data, and generalization and learning ability (Ercanlı et al., 2018; Ferraz Filho et al., 2018Reis et al., 2018; Bayat et al., 2020). Among the main advantages of the SOM neural network is its unique ability to recognize patterns and nonlinear relationships in data that groups similar data points with each other in an unsupervised style, which is particularly very useful for exploratory data. Their topology maintenance nature ensures that the spatial relationships of input data are maintained in the output map and present a meaningful demonstration of the data structure analysis (Kohonen, 1990; Kohonen, 2001). However, the SOM neural network is sensitive to the selection of parameters such as map size, learning rate, and neighborhood function. These parameters can largely affect the learning process and final output (Das et al., 2016). The SOM neural network is extensively used for representing and analyzing vast and multi-dimensional data (Kurasova and Molyt 2011). This has found applications in ecological sciences (Chon, 2011; Kim and Kwak, 2022), in clustering complex relationships among the variables of ecological and environmental datasets (Chon, 2011; Hong et al., 2020), in clustering forest land areas (Fujino and Yoshida, 2006), and in processing time-series data and image compression (Ley et al., 2011; Gao et al., 2014; Liu et al., 2018; Fratarcangeli et al., 2019; Lakshminarayanan, 2020; Liu et al., 2021).
In this study, the SSOM ANN performance was evaluated with topography and the diameter at breast variables to estimate tree canopies. This neural network could estimate more than 80–90% of tree canopies, with more considerable accuracy in estimating the Fagus Orientalis canopy than the other tree species (Carpinus betulus, Diospyros lotus., Alnus subcordata, and Parrotia persica trees). This result suggests the interrelation of the selected features, and SSOM can examine the relationships between input variables and the canopy. In previous studies, Eskandari (2020) concluded that the support vector machine (SVM) algorithm presented the highest accuracy for the regional canopy. Dabija et al. (2021) reported that the SVM algorithm performed better than the random forest in mapping tree canopies. Nazariani et al. (2022) presented evidence that the MLP neural network and the radial basis function showed more accuracy in tree canopy estimation than the k-nearest neighbors (KNN), SVM, and random forest algorithms. Liu et al. (2022) used multiple regression, ANN, KNN, and random forest based on remote sensing data to precisely estimate forest tree canopy height. They claimed that the ANN presented the best performance in upper-ground biomass estimation, with a root mean square error of 19.9%.