To reveal the differences in growth rhythms and ecological strategies among three dominant Castanopsis species in the subtropical broadleaved secondary forests of Chongyi County, Jiangxi Province, and to provide a theoretical basis for classified management, structural regulation, and sustainable management of this region and similar subtropical secondary forests, as well as methodological references for multidimensional tree growth simulation under small-sample conditions.

Three dominant Castanopsis species-C. fargesii, C. faberi, and C. carlesii-in the subtropical secondary forests of Chongyi County were selected for analysis. Sample trees were subjected to stem analysis to obtain precise time-series data of annual radial growth. Five classic growth functions (Logistic, Gompertz, Hossfeld IV, Korf, and Richards) were applied to fit and compare the growth processes of DBH, height, and volume for each species. The Richards model, identified as the best-fitting function in most cases, was used to construct a system of nonlinear compatible growth equations. These equations were then estimated using the seemingly unrelated regression (SUR) method. Additionally, logarithmic transformation was introduced to correct heteroscedasticity in allometric growth models, thereby improving model stability. Based on the modeling results, growth rate and maturation patterns of the three species were further compared and analyzed.

Among the five tested models, the Richards function provided the best fit across multiple growth dimensions. The SUR-based compatible model revealed significant residual correlations between DBH and volume (r=0.875 1), and between DBH and height (r=0.682 0), validating the necessity of joint modeling. After logarithmic transformation, the allometric models exhibited a substantial reduction in mean systematic error (18.7% to 56.7%), while maintaining high R² values and stable fitting performance. Growth pattern analysis showed that C. faberi and C. carlesii had stronger early growth potential and faster maturation, making them suitable for timber-oriented management. In contrast, C. fargesii exhibited a slower growth rate but demonstrated strong ecological adaptability, making it ideal for deployment in ecological conservation forests.

Under limited sample conditions, this study successfully developed nonlinear compatible models for DBH, height, and volume of three dominant Castanopsis species, and effectively identified interrelationships among growth dimensions using the SUR method. The introduction of logarithmic transformation significantly improved the performance of the allometric models by reducing heteroscedasticity. Furthermore, through a comprehensive analysis of current annual increment and mean annual increment, the study further reveals significant differences among the three Castanopsis species in growth rates of diameter at breast height, tree height, and stem volume, as well as in maturation patterns. C. faberi and C. carlesii exhibit a fast-growing strategy, whereas C. fargesii shows a stable and conservative strategy, reflecting interspecific differences in resource utilization and environmental adaptation.

As an important indicator of the biomass storage capacity of forest ecosystems, forest stock is a key parameter for measuring forest health, ecological function and carbon stock. Accurate inversion of forest stock can provide scientific data support for forest resource management, climate change, ecological protection and sustainable management, which is one of the core tasks of forest resource investigation and monitoring.

In this study, Liuyang City, Hunan Province was taken as the study area, and the remote sensing inversion of forest stock was carried out by combining the National Forest Inventory (NFI) data and Sentinel-2 remote sensing images. Firstly, key feature variables closely related to forest stock were extracted by Pearson correlation analysis and importance screening of feature variables. Subsequently, support vector machine (SVM), random forest (RF) and fully convolutional neural network (FCN) models were used for forest stock inversion. The model accuracy was calculated using the cross-validation method, and several assessment indices were applied to compare the model performance, and the model with the best fitting effect was finally selected for the accurate inversion of forest stock in the study area.

The results showed that: 1) each feature variable showed significant correlation with the forest stock; 2) the introduction of the canopy height model (CHM) could significantly improve the accuracy of the forest stock inversion, in which the coefficient of determination (R²) of the support vector machine (SVM) model was increased from 0.34 to 0.52, representing an improvement of 52.95%. The root mean square error (R_MSE) decreased from 30.52 m³·hm^-2 to 25.59 m³·hm^-2, representing a reduction of 16.15%. And the random forest (RF) model was increased from 0.39 to 0.55, representing an improvement of 41.03%. The R_MSE decreased from 29.46 m³·hm^-2 to 25.07 m³·hm^-2, representing a reduction of 14.90%, which indicated that CHM could significantly improve the accuracy of the inversion of forest stock. This indicates that CHM has an important contribution to improving the accuracy of forest stock inversion; 3) Compared with other models, the full convolutional neural network (FCN) model has the best performance in forest stock inversion, with an R² value of 0.66 and the R_MSE is 21.20 m³·hm^-2. In addition, the CHM data based on the year 2022 significantly improved the prediction performance of the other models, which effectively overcomes the accuracy bottleneck of the traditional inversion methods due to the lack of measured tree height data and provides a good opportunity for the large-scale inversion. This effectively overcomes the bottleneck of accuracy caused by the lack of measured tree height data in traditional inversion methods and provides new ideas and methods for large-scale forest stock monitoring.

The FCN model combining measured data and Sentinel-2 imagery can realize accurate inversion and monitoring of forest stock volume, and the introduction of CHM data significantly improves the prediction accuracy of the inversion model, which provides an important reference for accurate monitoring and scientific management of forest resources in the future.