Recently, convolutional neural networks (CNNs) showed excellent performance in many tasks, such as computer vision and remote sensing semantic segmentation. Especially, the ability to learn high-representation features of CNN draws much attention. And random forest (RF) algorithm, on the other hand, is widely applied for variables selection, classification, and regression. Based on the previous fusion models that fused CNN with the other models, such as conditional random fields (CRFs), support vector machine (SVM), and RF, this article tested a method based on the fusion of an RF classifier and the CNN for a very high resolution remote sensing (VHRRS) based forests mapping. The study area is located in the south of China and the main purpose was to precisely distinguish Lei bamboo forests from the other subtropical forests. The main novelties of this article are as follows. First, a test was conducted to confirm if a fusion of CNN and RF make an improvement in the VHRRS information extraction. Second, based on RF, variables with high importance were selected. Then, a test was again conducted to confirm if the learning from the selected variables will further give better results.

China is one of the countries with the most abundant bamboo forest resources in the world, and Zhejiang province is among the top-3 Chinese provinces with richest bamboo forests. For rational bamboo forests management, it is of great significance to study the spatiotemporal dynamic changes of Aboveground Carbon (AGC) stocks of bamboo forest in Zhejiang. In this study, remote sensing variables, such as spectral, vegetation indices and texture features of bamboo forest in Zhejiang, were extracted from 32 Landsat TM and OLI images got from four different years (2000, 2004, 2008 and 2014). These variables were subsequently selected with stepwise regression method to build an estimation model of AGC of the bamboo forests. The results showed that (1) the accuracy of bamboo forest remote sensing information extracted from the four different years was high with a classification accuracy of >76.26% and an accuracy of users of >91.62%. The classification area of bamboo forest was highly consistent with the area from forest resource inventory, and the area accuracy was over 96.50%; (2) the estimation model performed well in predicting the AGC in Zhejiang for different years. The correlation coefficient for estimated and measured AGC was between 63% and 72% with low root mean square error; (3) the derived AGC of the bamboo forests in Zhejiang province increased gradually from 2000 to 2014, with the AGC density of 6.75 Mg·ha⁻¹, 10.95 Mg·ha⁻¹, 15.25 Mg·ha⁻¹ and 19.07 Mg·ha⁻¹ respectively, and the average annual growth of 0.88 Mg·ha⁻¹. The spatiotemporal evolution of bamboo forest AGC in Zhejiang province had a close relationship with the gradual expansion of bamboo forest in the province and the differentiation of management levels in different regions.

Future land use and cover change (LUCC) simulations play an important role in providing fundamental data to reveal the carbon cycle response of forest ecosystems to LUCC. Subtropical forests have great potential for carbon sequestration, yet their future dynamics under natural and human influences are unclear. Zhejiang Province in China is an important distribution area for subtropical forests. For forest management, it is of great significance to explore the future dynamic changes of subtropical forests in Zhejiang. As a popular LUCC spatial simulation model, the cellular automata (CA) model coupled with machine learning and LUCC quantitative demand models such as system dynamics (SD) can achieve effective LUCC simulation. Therefore, we first integrated a back propagation neural network (BPNN), a CA, and a SD model as a BPNN_CA_SD (BCS) coupled model for future LUCC simulation and then designed a slow development scenario (SD_Scenario), a harmonious development scenario (HD_Scenario), a baseline development scenario (BD_Scenario), and a fast development scenario (FD_Scenario), combining climate change and human disturbance. Thirdly, we obtained future land-use patterns in Zhejiang Province from 2014 to 2084 under multiple scenarios, and finally, we analyzed the temporal and spatial changes of land use and discussed the subtropical forest dynamics of the future. The results showed the following: (1) The overall accuracy was approximately 0.8, the kappa coefficient was 0.75, and the figure of merit (FOM) value was over 28% when using the BCS model to predict LUCC, indicating that the model could predict the consistent change of LUCC accurately. (2) The future evolution of the LUCC under different scenarios varied, with the growth of bamboo forests and the decline of coniferous forests in the FD_Scenario being prominent among the forest dynamics changes. Compared with 2014, the bamboo forest in 2084 will increase by 37%, while the coniferous forest will decrease by 25%. (3) Comparing the area and spatial change of the subtropical forests, the SD_Scenario was found to be beneficial for the forest ecology. These results can provide an important decision-making reference for land-use planning and sustainable forest development in Zhejiang Province.

Bamboo forests, especially the Moso bamboo forest (MBF) and the Lei bamboo forest (LBF), have a strong carbon sequestration capability and play an important role in the global forest carbon cycle. The leaf area index (LAI) is an important structural parameter for simulating the spatiotemporal pattern of the carbon cycle in bamboo forests. However, current LAI products suffer from substantial noise and errors, and data assimilation methods are the most appropriate way to improve the accuracy of LAI data. In this study, two data assimilation methods (the Dual Ensemble Kalman filter (DEnKF) and Particle filter (PF) methods) were applied to improve the quality of MODIS LAI time-series data, which removed noises and smoothed the results using a locally adjusted cubic-spline capping method for the MBF and LBF during 2014–2015. The method with the highest correlation coefficient (r) and lowest root-mean-square error (RMSE) was used to generate highly accurate LAI products of bamboo forests in Zhejiang Province. The results show that the LAI assimilated using two methods saw greatly reduced fluctuations in the MODIS LAI product for both the MBF and the LBF. The LAI assimilated using DEnKF significantly correlated with the observed LAI, with an r value of 0.90 and 0.95, and an RMSE value of 0.42 and 0.42, for the MBF and the LBF, respectively. The PF algorithm achieved a better accuracy than the DEnKF algorithm, with an average increase in r of 8.78% and an average decrease in the RMSE of 33.33%. Therefore, the PF method was applied for LAI assimilation in Zhejiang Province, and the assimilated LAI of bamboo forests achieved a reasonable spatiotemporal pattern in Zhejiang Province. The PF algorithm greatly improves the accuracy of MODIS LAI products and provides a reliable structural parameter for the large-scale simulation of the carbon cycle in bamboo forest ecosystems.

The highly accurate multiresolution leaf area index (LAI) is an important parameter for carbon cycle simulation for bamboo forests at different scales. However, current LAI products have discontinuous resolution with 1 km mostly, that makes it difficult to accurately quantify the spatiotemporal evolution of carbon cycle at different resolutions. Thus, this study used MODIS LAI product (MOD15A2) and MODIS reflectance data (MOD09Q1) of Moso bamboo forest (MBF) from 2015, and it adopted a hierarchical Bayesian network (HBN) algorithm coupled with a dynamic LAI model and the PROSAIL model to obtain high-precision LAI data at multiresolution (i.e., 1000, 500, and 250 m). The results showed the LAIs assimilated using the HBN at the three resolutions corresponded with the actual growth trend of the MBF and correlated significantly with the observed LAI with a determination coefficient (R²) value of >0.80. The highest-precision assimilated LAI was obtained at 1000-m resolution with R² values of 0.91. The LAI assimilated using the HBN algorithm achieved better accuracy than the MODIS LAI with increases in the R² value of 2.7 times and decreases in the root mean square error of 87.8%. Therefore, the HBN algorithm applied in this study can effectively obtain highly accurate multiresolution LAI time series data for bamboo forest.