Abstract Bayesian inference offers a flexible framework for parameter estimation and uncertainty quantification in eco‐hydrological models. However, simultaneously achieving robust posterior exploration and high computational efficiency for multimodal, high‐dimensional, and computationally intensive targets remains challenging for the widely used Markov chain Monte Carlo (MCMC) and sequential Monte Carlo (SMC) methods. In this study, we developed the Parallel Adaptive Transition Particle Evolution Metropolis Sequential Monte Carlo (PATPEMS) algorithm, which is an adaptive and parallel SMC sampler for posterior distributions of model parameters in offline calibration. PATPEMS employs an adaptive sequence of intermediate distributions to control weight degeneracy and automatically select stages, a flexible scheduling of MCMC proposal kernels used to rejuvenate particles, together with reflection boundary handling to maintain particle diversity, and a particle‐level parallelization scheme to exploit multicore architectures and reduce wall‐clock time for computationally intensive models. Performance is assessed on four case studies: two synthetic targets probing multimodality and high‐dimensional dependence, and a land surface model (LSM) with six parameters constrained by synthetic and real observations. Across all cases, PATPEMS provides close approximations to the target posteriors, judged against the analytic ground truth or reference solutions. For the LSM, parallelization yields substantial wall‐clock speedups over the original non‐parallel implementation. Compared with the original particle evolution Metropolis sequential Monte Carlo (PEM‐SMC) algorithm, these results indicate that PATPEMS provides a more adaptive and parallel framework for robust Bayesian calibration of multimodal, correlated, and computationally demanding land surface and environmental models.

Grassland aboveground biomass (AGB) is a key indicator of grassland ecosystem structure and function, and its accurate monitoring is of great importance for assessing grassland ecological conditions and supporting sustainable grassland management. Traditional biomass estimation methods based on vegetation indices (VIs) often suffer from saturation due to canopy shading. However, comparative studies on VI saturation and the saturation height of AGB detectable by different indices remain limited. In this study, we evaluated 12 commonly used VIs based on field-measured AGB and hyperspectral data in the Hulunbuir meadow steppe. Relationships between vertically accumulated biomass and VIs were analyzed to identify optimal AGB fitting models and to determine the saturation height of each index. Results showed that vertical distribution of AGB followed a unimodal pattern, with biomass peaking at approximately 36 cm in this region. This study employed four models (namely the Linear model, the Logarithmic model, the Power Function model and the Gompertz model) to fit the relationship between the vegetation index and AGB. Among them, Gompertz models consistently outperformed other models, indicating saturation across all indices. Based on saturation height, the 12 VIs were classified into two groups: ARVI, GNDVI, NDRE, OSAVI, and SAVI saturated at 40 cm, whereas DVI, EVI, MSAVI, NDPI, NDVI, RVI, and VARI maintained sensitivity up to 50 cm, demonstrating a stronger anti-saturation capacity. NDVI and NDPI exhibited the highest fitting accuracy and resistance to saturation. These findings validate the saturation limitations of VIs and provide guidance for selecting appropriate indices to improve the accuracy of grassland biomass retrieval.

Spatially explicit patterns of forest species diversity are essential for effective conservation planning. Satellite remote sensing enables large-scale monitoring but relies heavily on sufficient field observations, which are difficult to obtain in complex, less accessible tropical forests. Unmanned aerial vehicles (UAVs) offer fine-resolution, flexible observations that can bridge this ground-satellite gap, yet their potential for cross-scale species diversity upscaling remains underexplored. In this study, we integrated UAV and satellite data from a typical tropical forest to investigate whether UAV-based estimates improve satellite-based species diversity upscaling. We first evaluated multiple UAV-based diversity estimation approaches, including regression methods based on spectral diversity indices, height heterogeneity indices, and their combination, as well as clustering methods applied at both pixel and individual tree crown (ITC) scales. Building upon the UAV-derived diversity estimates, we further developed a new two-step upscaling framework (Ground-UAV-Satellite) for species diversity extrapolation based on PlanetScope and Sentinel-2 imagery. The results demonstrated that variable configuration significantly influenced estimation performance. At the UAV scale, regression models incorporating features related to multidimensional traits yielded the optimal performance (r2 = 0.43 for Richness, and r2 = 0.55 for the Shannon-Wiener index). The findings confirmed the feasibility of unsupervised clustering algorithm in tropical forests, although pixel-based clustering exhibited saturation effects. Critically, at the satellite scale, our two-step upscaling optimal models outperform the traditional Ground-Satellite modeling (r2 = 0.41 > 0.31 for Richness, and r2 = 0.46 > 0.39 for the Shannon-Wiener index). Our study demonstrated the potential of UAVs as effective intermediate observational platforms for cross-scale species diversity monitoring and provided a practical framework for integrating limited field data with wall-to-wall satellite data for large-scale biodiversity mapping.

Terrestrial evapotranspiration (ET) and its components (Transpiration, T; Soil evaporation, Es; Canopy interception, Ec) play a crucial role in the land-climate interaction and climate system. There have been multiple state-of-the-art climate models that were included in Phase 6 of the Coupled Model Inter-comparison Project (CMIP6), but the performance in estimating the ET and its components of the models is still unclear, particularly in China. Here, we evaluate the performance of ET and how the partitioning of ET into three components; and investigate the impacts of vegetation dynamics and its sensitivity on ET partitioning. Our results show that spatial patterns of ET in most CMIP6 models are basically consistent with that in the mainstream ET products, but most CMIP6 models clearly overestimate ET compared with national-scale estimates and site-observations. Overall, the EC-Earth3-AerChem, KACE-1-0-G, and FIO-ESM-2-0 models perform well in simulating total ET, and the CMIP6 models ensemble mean showed generally better agreement. Moreover, a large discrepancy exists in the historical ET partitioning across CMIP6 models; even transpiration and soil evaporation account for 18 - 66% and 25 - 68% of national-mean ET across models, respectively. CMIP6 models underestimate the T/ET compared to the mainstream ET products (63.0 f 6.1%) and site observations. Vegetation (that is leaf area index, LAI) plays a critical role in ET partitioning, and has strong correlations with all ET components in models. However, the mean annual LAI simulated by the CMIP6 model (1.79 f 0.69 m2/m2) is about twice that observed by satellite (0.98 f 0.06 m2/m2), revealing a significant overestimation and indicating that models fail to capture how strongly T/ET responds to changes in LAI (i.e., d(T/ET)/dLAI). The d(T/ET)/dLAI from CMIP6 models ranged from 0.01 - 0.18 per m2/m2, while the result calculated by the satellite observations was 0.19 f 0.04 per m2/m2. The bias of d(T/ET)/dLAI in models rises in tandem with the increase in LAI bias. To better represent the coupling between vegetation dynamics and the hydrological cycle, we suggest that the modeling community should promptly correct these biases, especially the LAI.

Plant diversity underpins wetland ecosystem stability and functional sustainability, and its reliable assessment is vital for effective conservation and management. Remote sensing provides an efficient mean of plant diversity monitoring, however the potential of remotely sensed functional traits (RS-traits) and spectral metrics for plant diversity estimation in wetland ecosystems has not been fully investigated. In this study, we integrated UAV hyperspectral and LiDAR data to extract spectral band metrics, Rao’s quadratic entropy and principal components, texture features, and retrieve physiological and morphological RS-traits. We then used these features to predict multi-dimensional (species, functional, and phylogenetic) plant diversity using multiple stepwise regression (MSR), generalized additive models (GAMs), and random forest regression (RF). The results demonstrated that the retrieved RS-traits were generally consistent with field measurements (R2 = 0.36–0.78 for physiological, R2 = 0.47–0.87 for morphological traits). Among the different models, MSR performed best for species diversity, GAMs for functional diversity, and RF for phylogenetic diversity. The highest predictive performance was achieved for species diversity (R2adj = 0.62–0.73), followed by functional (R2adj = 0.41–0.83) and phylogenetic diversity (R2adj = 0.55–0.64). Models based on RS-traits consistently performed better than those based on spectral metrics, while combining spectral metrics and RS-traits did not lead to statistically significant improvements. While our results provide preliminary evidence towards a unified RS-trait framework for the multi-dimensional monitoring of wetland plant diversity, further work is needed to generalise these findings to other sites and wetland types.

Accurate large-scale forecasting of tea anthracnose (TA) (Colletotrichum camelliae) is crucial for tea cultivation disease management. TA infection exhibits a strong synchrony with phenological development regulated by accumulated growing degree days (AGDD). However, the complex microclimatic conditions of mountainous regions lead to highly heterogeneous spatiotemporal phenological response patterns in tea plants, posing significant challenges to the effective extraction and utilization of habitat factors in disease forecasting models. To address these issues, this study proposes a novel forecasting approach for TA by integrating multi-source habitat information with spatiotemporal phenological corrections. Using survey data on TA collected in Zhejiang Province from 2016 to 2020, a spatiotemporal phenological correction strategy based on AGDD was developed. This strategy facilitated the alignment of AGDD-calibrated multi-source habitat data to construct a comprehensive feature dataset for disease forecasting. A forecasting model was subsequently developed through the application of the Relief-F algorithm for feature selection, combined with representative machine learning techniques, including Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbor (KNN), and Naive Bayes (NB). The results demonstrated that the strategy combining RF with AGDD-aligned multi-source features significantly enhanced forecasting performance, achieving an average overall accuracy (OA) of 77 % and an average kappa coefficient of 0.65. This approach outperformed conventional methods using either calendaraligned features or single meteorological factors combined with machine learning algorithms. From a spatiotemporal heterogeneity perspective, this study elucidated the response characteristics of multi-source habitat factors under varying geographical and terrain conditions using the AGDD-aligned phenological correction strategy. Furthermore, the integration of remote sensing data, which reflects the physiological state of tea plants, with meteorological and geographical factors substantially improved the comprehensiveness and precision of TA forecasting. These findings highlight the importance of incorporating spatiotemporal phenological corrections and multi-source habitat data for advancing disease forecasting methodologies in complex agricultural landscapes.

Accurate mapping of rice cultivation is vital for ensuring food security, reducing greenhouse gas emissions, and achieving sustainable development goals. However, large-scale deep learning–based crop mapping remains limited due to the demand for vast, uniformly distributed, high-quality samples. To address this challenge, we propose a Progressive Deep Learning Crop Mapping (PDLCM) framework for national-scale, high-resolution rice mapping. Beginning with a small set of localized rice and non-rice samples, PDLCM progressively refines model performance through iterative enhancement of positive and negative samples, effectively mitigating sample scarcity and spatial heterogeneity. By combining time-series Sentinel-2 optical data with Sentinel-1 synthetic aperture radar imagery, the framework captures distinctive phenological characteristics of rice while resolving spatiotemporal inconsistencies in large datasets. Applying PDLCM, we produced 10 m rice maps from 2022 to 2024 across the middle and lower Yangtze River Basin, covering more than one million square kilometers. The results achieved an overall accuracy of 96.8% and an F1 score of 0.88, demonstrating strong spatial and temporal generalization. All datasets and source codes are publicly accessible, supporting SDG 2 and providing a transferable paradigm for operational large-scale crop mapping.

Foliar insect herbivory is a growing global threat to the health and productivity of forests. Timely and spatially explicit monitoring is essential for effective silvicultural interventions. Remote sensing (RS) technologies are powerful tools for detecting, mapping, and monitoring insect herbivory, offering scalable alternatives to traditional ground-based methods. This systematic review synthesises findings from 60 studies published between 2010 and February 2026, categorising them by insect feeding guilds and operational scales to identify key advancements, research gaps, and future opportunities. Research has predominantly focused on a limited number of host-pest systems and geographic regions. Results reveal a strong emphasis on landscape-scale assessments of leaf-chewing guilds, while tree-level studies remain underrepresented. Post-2020 adoption of Sentinel-2 has demonstrated strong potential for herbivory characterisation across feeding guilds. Leaf-chewing studies used spectral, structural, textural, and polarimetric features achieving high accuracy (R2 = 0.34-0.9, overall accuracy = 73-97.7

The accurate point cloud completion of individual tree crowns is critical for quantifying crown complexity and advancing precision forestry, yet it remains challenging in dense plantations due to canopy occlusion and LiDAR limitations. In this study, we extended the scope of conventional point cloud completion techniques to artificial planted forests by introducing a novel approach called Multi−feature Fusion Completion of Populus (MFCPopulus). Specifically designed for Populus Tomentosa plantations with uniform spacing, this method utilized a dataset of 1050 manually segmented trees with expert−validated trunk−canopy separation. Key innovations include the following: (1) a hierarchical adversarial framework that integrates multi−scale feature extraction (via Farthest Point Sampling at varying rates) and biologically informed normalization to address trunk−canopy density disparities; (2) a structural characteristics split−collocation (SCS−SCC) strategy that prioritizes crown reconstruction through adaptive sampling ratios, achieving a 94.5% canopy coverage in outputs; (3) a cross−layer feature integration enabling the simultaneous recovery of global contours and a fine−grained branch topology. Compared to state−of−the−art methods, MFCPopulus reduced the Chamfer distance variance by 23% and structural complexity discrepancies (ΔDb) by 33% (mean, 0.12), while preserving species−specific morphological patterns. Octree analysis demonstrated an 89−94% spatial alignment with ground truth across height ratios (HR = 1.25−5.0). Although initially developed for artificial planted forests, the framework generalizes well to diverse species, accurately reconstructing 3D crown structures for both broadleaf (Fagus sylvatica, Acer campestre) and coniferous species (Pinus sylvestris) across public datasets, providing a precise and generalizable solution for cross−species trees’ phenotypic studies.

Accurate estimation of plantation aboveground biomass (AGB) is critical for quantifying carbon cycles and informing sustainable forest resource management, but enhancing estimation accuracy remains a key challenge. Although tree height and stand age are recognized as critical predictors for enhancing AGB models in addition to spectral vegetation indices, their individual and combined contributions in regional plantation forests remain insufficiently quantified, especially concerning the potential for leveraging the distinct characteristics of fast-growing plantations to facilitate AGB estimation. This study developed multi-source remote sensing-based Eucalyptus AGB estimation models for Nanning, Guangxi, integrating stand age and tree height to assess their impacts. Stand age was mapped from Landsat time-series imagery, and tree height was derived from UAV-LiDAR data. Plot-level reference AGB was obtained using fused UAV and terrestrial LiDAR point clouds. A random forest model, incorporating these variables with Sentinel-2 spectral information and topography, then achieved regional AGB estimation. The findings demonstrate that (1) tree height serves as the most influential predictor for AGB estimation at the regional scale, yielding a robust model performance (R2 = 0.84). (2) Tree height captures the majority of the explanatory power associated with stand age. Once tree height was included as a predictor, the subsequent addition of stand age offered no significant improvement in model accuracy (R2 = 0.85). (3) Given the challenges in obtaining precise tree height data and the robust correlation between stand age and tree height in fast-growing plantations, the integration of stand age substantially improved the accuracy of AGB estimations (from the spectral model of R2 = 0.54 to R2 = 0.74), with performance approaching that of tree height-based models (ΔR2 = 0.10). Consequently, in fast-growing plantations, which are often characterized by high stand homogeneity, a hybrid model incorporating stand age can offer a reliable and cost-effective solution for AGB estimation.

Influenced by climate change and human activities, shrub encroachment in global arid and semi-arid grasslands profoundly affects ecosystem functions and livestock farming. Remote sensing technology is crucial for evaluating shrub encroachment across spatial and temporal scales, providing a broader perspective for understanding its driving mechanisms. The shrub encroachment dominated by Caragana vegetation in the Inner Mongolia grasslands is typical both in China and globally. However, the lack of remote sensing studies hinders a deeper understanding of the current status and causes of shrub encroachment in this region. This study focuses on the semi-arid grasslands in central Inner Mongolia, estimating shrub coverage based on vegetation indices, SAR backscatter coefficients, phenological metrics, and the extreme gradient boosting algorithm. A new Remote Sensing-based Shrub Encroachment Index (RSSEI) was established to assess the stage of shrub encroachment. The complex driving mechanisms of shrub encroachment were innovatively elucidated through the integrated application of Geographical Detector (GD) and Geographical Convergent Cross Mapping (GCCM). The results show that the shrub coverage estimation model achieved high accuracy (R2 = 0.70, RMSE = 4.0%, MAE = 2.7%), with the SHapley Additive exPlanations (SHAP) value indicating the significant role of phenological metrics in estimating shrub coverage. The RSSEI was found to grade shrub encroachment stages more accurately, with overall accuracy reaching 74.6%. Moderate to severe encroached grasslands dominate the study area, exhibiting an east-high and west-low spatial distribution. GD and GCCM indicated that moisture-related factors, including precipitation, soil water content, and vapor pressure deficit, were the main driving forces of shrub encroachment. And shrub encroachment has a significant causal influence on 2 m-height air temperature. This study provides an improving technical reference for remote sensing observation of shrub encroachment in semi-arid grasslands and is also valuable for understanding the formation causes and feedback effects of shrub encroachment in these areas.

Fast-growing eucalyptus species, used as vineyard posts in New Zealand's Marlborough region, offer both durability and potential carbon sequestration benefits. However, the scale of carbon sequestration by these species remains unexplored. This study aimed to estimate individual tree dimensions (diameter at breast height, DBH) and above-ground biomass (AGB) for Eucalyptus globoidea and E. bosistoana using light detection and ranging (LiDAR) data acquired by an unpiloted aerial vehicle (UAV). LiDAR data were captured before destructive sampling, and 96 individual tree LiDAR metrics were extracted. Three machine learning (ML) models, including Partial Least Squares Regression (PLSR), Random Forest, and Extreme Gradient Boosting (XGBoost), were trained. Model performance was evaluated using the root mean square error and coefficient of determination (R2). SHapley Additive exPlanations (SHAP) analysis was employed to explain model predictions and evaluate input variables. Results showed that among the ML models, XGBoost and PLSR demonstrated superior performance, with the former yielding the highest R2 values for AGB (0.903) and the latter getting the highest R2 values for DBH (0.829). SHAP analysis highlighted that LiDAR height and voxel metrics were the most important factors influencing AGB and DBH predictions. These findings demonstrate that UAV LiDAR can provide efficient and accurate AGB estimates in eucalyptus plantations, supporting the wine industry's carbon neutrality efforts.

Spatial and temporal clear-cutting distribution is an important data source for investigating forest dynamics and formulating policies related to timber production, land use, and sustainable development. Subtropical and tropical regions of China due to their favorable weather conditions for tree growth are important timber providers, but the spatial distribution of annual clear-cutting areas was unavailable. This research leveraged dense time-series Landsat images (1986-2022) and a novel approach combining Continuous Change Detection and Classification (CCDC) with Random Forest to produce the first annual 30 m resolution clear-cutting map for the China's subtropical and tropical regions. The per-polygon approach based on the sub-compartment and field survey data was used to evaluate the developed clear-cutting product. The spatiotemporal distribution of clear-cutting areas was further analyzed along longitude and latitude as well as in different provinces. The results indicated that the combination of CCDC and Random Forest successfully detected spatial and temporal distribution of clear-cutting areas with an overall accuracy of 75.6% based on sub-compartments and 93.3% based on field polygons. The cumulative clear-cutting area between 1987 and 2021 was approximately 21.26 x 106 ha, with a mean annual clear-cutting area of 0.61 x 106 ha in China's subtropics and tropics. About 10.24% of subtropical/tropical vegetation experienced clear-cutting and most areas (85.94%) were clear-cut once. Clear-cutting activities gradually clustered at low latitudes and moderate longitudes, represented by Guangxi Province. More frequent clear-cutting was observed during the later period (2001-2021). This research contributes to an improved understanding of forest dynamics over time at large scale, and provides scientific data to make proper decisions for forest sustainability.

Background: Geospatial technologies have emerged as powerful tools for optimising forest management, improving operational precision, and supporting data-driven decision-making. This study aims to understand the technologies adopted by the New Zealand plantation forest industry and identify any barriers to the uptake of geospatial tools. This is the third such study, following comparable surveys in 2013 and 2018. Methods: An online survey was sent to 29 organisations in New Zealand's forestry sector. Topics included organisation demographics, data acquisition, positioning technology, remote sensing technologies, software, and Artificial Intelligence (AI). Specifically, the survey focused on five remote sensing technologies: aerial photography, aerial videography, multispectral imagery, hyperspectral imagery, and LiDAR. Each section contained questions relating to the acquisition and application of the remote sensing technology and the software used for data processing. Questions were included to ascertain barriers to adoption. To identify changes in technology usage and uptake, results were compared to the 2013 and 2018 studies. Results: Twenty-seven of the 29 queried organisations responded, resulting in a 93% response rate. Responding organisations managed 1,283,000 hectares (74% of New Zealand's plantation forest estate), with estate sizes ranging from about 7,000 to 200,000 hectares. Data acquisition from online portals included aerial imagery (100%), property ownership data (96%), and elevation data (89%), primarily from the Land Information New Zealand (LINZ) Data Service. Global Navigation Satellite Systems (GNSS) technology was universally employed. All respondents acquired aerial photography. In addition, 67% acquired multispectral imagery, 4% acquired hyperspectral imagery, and 93% acquired LiDAR data. The AI topic was surveyed for the first time and the technology was used by 30% of respondents when working with geospatial data. The main barrier to using remotely sensed data was the lack of perceived benefits, while the primary barrier to AI adoption was a lack of staff knowledge and training. Except for hyperspectral imagery, all remote sensing technologies saw increased uptake since 2013. LiDAR experienced the largest growth, with uptake increasing from 17% in 2013 to 93% in 2023. ArcGIS remains the primary tool for geospatial analysis, used by 96% of respondents. Notably, the use of open-source software such as QGIS increased by 31% over the past decade. Conclusions: This study demonstrated an overall increase in the usage of geospatial technology in the forestry sector. To promote further uptake, it is important not only to increase exposure to available tools and provide training, particularly on emerging technologies such as AI, but also to demonstrate the practical and economic value these technologies can offer.

Satellite remote sensing data is essential for large-scale, timely, and repeatable monitoring of forest species diversity. While various methods have been applied to satellite-based diversity estimation at regional scales, selecting suitable sensor and monitoring period remains challenging, especially in tropical forests. This study aims to identify the optimal time window, spatial resolution, and metrics for species diversity estimation in the Jianfengling tropical forest in southern China. We constructed stepwise linear regression models for estimating Richness, Simpson, and Shannon-Wiener indices using in-situ species diversity and heterogeneity metrics of spectra and structure. For analyzing phenology influence, we utilized six Sentinel-2 images acquired bimonthly from January to November. For evaluating scale dependency, we resampled the GF2 image to five spatial resolutions ranging from 0.8 to 10 m. The results indicated that the suitable phenological periods for species diversity estimation were at the beginning and end of the growing season, especially September performing the best for all diversity indices. Among four types of heterogeneity metrics, spectral information consistently explained most variance in species diversity indices across all periods. The optimal spatial resolution for estimating Richness and Shannon-Wiener index was 4-5 m, which corresponded to the average tree crown size. The texture features made a significant contribution compared to other metrics. Our study highlights that species diversity monitoring is highly dependent on the spatiotemporal scales of remote sensing data. It may offer practical guidance for selecting appropriate data and methods for species diversity monitoring in tropical forests.

Bayesian inference is crucial for optimizing parameters in complex models, but often requires sampling due to high-dimensional, intractable posteriors. Beyond Markov-Chain Monte Carlo (MCMC) methods, Sequential Monte Carlo (SMC) algorithms offer an alternative. This paper introduces a Matlab toolbox for the Particle Evolution Metropolis Sequential Monte Carlo (PEM-SMC) algorithm, which combines the strengths of population-based MCMC and SMC. Two case studies—a complex multi-modal probability and a land surface model—demonstrate the toolbox’s capabilities. This tool is valuable for Bayesian inference across fields like statistics, ecology, hydrology, and land surface processes.

Forest aboveground biomass (AGB) is a key indicator for evaluating carbon sequestration capacity and forest productivity. Accurate regional-scale AGB estimation is crucial for advancing research on global climate change, ecosystem carbon cycles, and ecological conservation. Traditional methods, whether based on LiDAR or optical remote sensing, estimate AGB using planar density (t/ha) multiplied by pixel area, which fails to account for vertical forest structure variability. This study proposes a novel “stereoscopic (stereo) density × volume” approach, upgrading planar density to stereo density (t/ha/m) by integrating canopy height information, thereby improving estimation accuracy and exploring the feasibility of this new method. In the Daxing’anling region, plot-scale AGB estimation models were developed using stepwise linear regression (SLR) for both “planar density × area” and “stereo density × volume” methods. Results indicated that the stereo model using arithmetic mean height (HAM) achieved comparable accuracy (R2 = 0.83, RMSE = 2.77 t) with the planar model (R2 = 0.83, RMSE = 2.52 t). At the regional scale, high-precision AGB estimates derived from airborne LiDAR were combined with vegetation indices from the Landsat Thematic Mapper (TM), and topographic factors from DEM to develop regional-scale AGB estimation models, using SLR and random forest (RF) algorithms. The results of 10-fold cross-validation demonstrated the superiority of the stereo method over the planar method, with RF outperforming SLR. The optimal RF-based stereo model of HAM (R2 = 0.65, rRMSE = 26.05%) significantly improved AGB estimation compared to the planar model (R2 = 0.59, rRMSE = 30.41%). Independent accuracy validation using 75 field plots demonstrated that the stereo model achieved a higher validation R2 of 0.45 compared to the planar model’s R2 of 0.35. These findings suggest that the stereo approach mitigates the underestimation of AGB caused by forest height variability in planar methods, with no significant differences observed across forest types. In conclusion, the use of the stereo method to estimate forest AGB is superior to the planar method in optical remote sensing. This approach offers a scalable solution for forest AGB estimation and carbon stock assessment.

Accurate and timely detection of forest disturbance types is crucial for evaluating ecosystem health and global climate stability. Time series remote-sensing data offers valuable spatiotemporal information. However, frequent cloud cover in subtropical regions disrupts the temporal consistency of optical satellite data. In addition, the impact of different data sources on modeling accuracy and the challenge of acquiring large and labeled datasets for deep learning are considerable obstacles. In this study, a novel positional encoding module was designed to handle the irregular Sentinel-2 time series. The Transformer model based on this positional encoding module effectively fused Sentinel-1 and Sentinel-2 data. Meanwhile, self-supervised learning was used to address the issue of insufficient samples. The improved Transformer model was successfully applied to detect clear-cutting and disease disturbances with a F1 score of 0.95. Our results highlighted that appropriate encoding techniques increased the model’s performance by between 5% and 14%. This research also found that while Sentinel-2 data alone yields good accuracy, combining Sentinel-1 and Sentinel-2 data improved model accuracy. The self-supervised learning model achieved higher accuracy (0.95) than the supervised model (0.86) on a limited data (20% of the training samples) and exhibited stable accuracy across different training data proportions.