Detecting forest tree species distribution using satellite imagery with deep-learning models is essential for effective forest management. While sufficient training samples are crucial for developing deep-learning-based tree species classification models, creating these training samples requires significant resources. Therefore, understanding the optimal balance between model accuracy and training sample size is essential for efficient resource allocation. Here, we determined the optimal training sample size for forest tree species classification using Sentinel-2 imagery and the U-Net model. The study area comprised the Seoul–Gyeonggi region of South Korea, where the nine dominant tree species were selected for classification. We utilized multi-temporal Sentinel-2 imagery, incorporating spectral, vegetation, and textural features. Optimal points were identified using Locally Estimated Scatterplot Smoothing (LOESS) regression. The maximum overall accuracy reached 61%, with 90% and 95% of the maximum accuracy with training sample sizes of 2.37%–2.67% and 4.42%–5.89%, respectively. The congeneric Pinus and Quercus groups had major confusion, with species-specific F1-scores ranging from 0.40 (Robinia pseudoacacia) to 0.75 (Pinus koraiensis). These results provide practical guidelines for efficient resource allocation in tree species classification. Rather than pursuing excessive data collection beyond the optimal point, integrating multiple sensor types can overcome existing limitations and enhance classification accuracy.

A methodological framework is provided for characterizing large-scale forest resource distribution in South Korea, along with a baseline for sustainable forest management practices. This study aimed to establish a baseline framework that integrates satellite and ground-based data for nationwide growing stock volume (GSV) estimation. Several machine learning models were applied and compared for estimating GSV across South Korea using Sentinel-2 imagery, national forest inventory data, and topographic information. Four algorithms, namely, k-nearest neighbors (kNN), random forest (RF), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost), were evaluated. The ensemble methods outperformed kNN, with RF demonstrating the highest accuracy (coefficient of determination and root mean squared error of 0.56 and 66.9 m³/ha, respectively). Accuracy assessment shows that kNN performed relatively well near the mean GSV (≒200 m³/ha), but its accuracy decreased sharply toward the extremes, failing to represent plots above 400 m³/ha. Estimation accuracy also varied substantially with stand height, which was identified as the primary predictor, and kNN was the most affected. These findings suggest that the structural complexity and mountainous terrain of South Korean forests may amplify the limitations of distance-based methods, reinforcing the need for improved 3D structural predictors such as satellite-derived stand height.