Vertical structural and spectral heterogeneity are two key remote sensing characteristics of complex forests. To enable effective forest health management and provide early warnings of abnormal disturbance, monitoring forest biochemical content with a vertically layered spectral perspective is critically needed. However, commonly used remote sensing technique still have limited capacity to study the biochemical status of the middle and lower canopy layers. This study provides the first insight into the potential of the full-waveform large-footprint hyperspectral LiDAR (LFHSL) system for retrieving the vertical heterogeneity of forest chlorophyll using 3D radiative transfer modeling. In our newly constructed LFHSL model, virtual three-dimensional (3D) complex forest scenes, comprising trees, bushes, and grass, were defined with varying positions and biochemical content inputs. Hyperspectral waveforms within the large-footprint were then simulated for each combination of vegetation position and biochemical level. The concept of spectral index time profiles (SITP), referred to as spectral index variation along the laser path, were introduced and used to assess the vertical distribution of chlorophyll for the first time in forest scenes. The main findings of this study are as follows: (1) Full-waveform LFHSL owns great potential for retrieving vertical chlorophyll content across trees, bushes, and grass layers in complex forest ecosystems. (2) SITP is a novel and essential reference indicator that fully registers chlorophyll variations along the laser path. (3) Simulations with random positions and chlorophyll contents indicate that the peak points of SITP yield higher chlorophyll prediction accuracy in trees layer and grass layer (R2: 0.996 vs. 0.971, RMSE: 1.39 vs. 3.81 μgcm−2) than that of bushes layer (R2 of 0.801, RMSE of 9.97 μgcm−2). (4) Compared to position patterns, LFHSL system is more sensitive to chlorophyll content sets. This study demonstrates that full-waveform LFHSL is a promising and surely reliable tool for acquiring and monitoring vertical vegetation health in complex forests. It not only provides significant guiding for the development of laser radar models but also holds promise for adoption in design of large-footprint multi-spectral or hyperspectral LiDAR.

The edge effect, caused by partial illumination of leaf boundaries, leads to significant hyperspectral LiDAR (HSL) echo-intensity loss and severely restricts the precise expression of the spectral characteristics and retrieval accuracy of biochemical components in the leaf-edge area. Despite being one of the major radiometric effects in LiDAR sensing, the edge effect has not been systematically investigated for hyperspectral intensity data. Along these lines, this study presents the first comprehensive exploration, detection, and filtering of edge effects in HSL-collected data. We analyze the physical mechanisms responsible for multilayer edge-effect formation and propose a new integrated detection and filtering framework, incorporating a histogram of intensity distribution, edge detection, and spherical-spatial filtering (HIDEDaSF). Results obtained from the HSL system on broadleaf plants demonstrate that our framework robustly detects edge-affected spectral points across 32 wavelengths. After filtering and correction, the standard deviation and coefficient of variation (CV) of edge-region intensities are reduced by 22.68% and 28.30%, respectively, and the mean ratio of CV is 0.7288, less than 1, confirming the stability and effectiveness of the proposed algorithm framework again. The HIDEDaSF framework preserves valid spectral information and significantly improves the consistency and quantitative reliability of hyperspectral intensity data at leaf-edge zones. Although tailored for the HSL system, this framework is also adaptable to other multi and HSL platforms. This study enriches the fundamental knowledge of edge-effect radiometric behaviors and expands the methodology foundation for fine-scale HSL vegetation remote sensing. We are willing to make our codes freely available via https://github.com/Jie-Bai/HIDEDaSF

Temporal analysis from time-series (TS) observations has emerged as an effective approach for improving forest aboveground biomass carbon (AGC) estimation. However, most existing methods generally rely on primary global temporal features of reflectance TS at the annual scale (e.g., annual fluctuation amplitude), while neglecting finer local temporal features within or between seasons, such as forest phenology, which are directly linked to forest carbon sink processes. This underutilization of temporal information constrains AGC estimation accuracy and remains susceptible to saturation effects. To address this gap, we propose a methodology that leverages fine-scale local temporal features (i.e., phenological features) to achieve accurate and reliable largescale AGC mapping. First, we employed the eXtreme Gradient Boosting (XGBoost) algorithm to verify the feasibility of the phenological features in AGC estimation. Second, we proposed a pixel-specific calibration method to optimally fuse Landsat and Sentinel-2 TS data, enabling more accurate extraction of phenological features. Third, the relationships between phenological features and AGC were characterized using SHapley Additive exPlanations (SHAP) analysis. Validation revealed that the initial model using primary global temporal features produced R2 of 0.53/0.56 (by Landsat/Sentinel-2, respectively) and RMSE of 26.56/26.28 Mg C/ha. Incorporating phenology-related features significantly improved accuracy, attaining R2 of 0.60/0.66 and RMSE of 24.86/22.50 Mg C/ha. Further enhancement was achieved by fusing Landsat and Sentinel-2 data using the proposed pixel-specific calibration method, which yielded an R2 of 0.75 and an RMSE of 19.40 Mg C/ha. This model also partially mitigated saturation effects within the investigated AGC range, reducing RMSE by 14.52 Mg C/ha in high-AGC forests (95.6-170.0 Mg C/ha) compared with traditional models using only global features derived from single-sensor data. Feature importance ranking and SHAP analysis further confirm the critical role of local features, with the five most influential predictors all being phenology-related and exhibiting biologically meaningful relationships with AGC variation. Specifically, higher AGC is associated with delayed phenological timing and reduced short-term spectral fluctuations in reflectance TS data.

Accurate large-scale estimation of forest surface Dead Fuel Moisture Content (DFMC) is critical for wildfire risk warning and scientific decision-making. While existing process-based and empirical models leveraging satellite data show utility at local scales, they exhibit inherent limitations: process-based models suffer from physical simplifications in numerical simulations, while empirical approaches lack mechanistic integration due to shallow learning architectures. This persistent gap necessitates-yet lacks-integrative frameworks that synergize physical realism with deep learning flexibility. To address this challenge, we propose a physics-guided deep learning framework that synergistically integrates geostationary meteorological satellite data and reanalysis data for regional-scale forest surface DFMC estimation. Our methodology fuses Long Short-Term Memory (LSTM) neural network features with physical features derived from the process-based Fuel Stick Moisture Model (FSMM). Critically, the physical feature fuel surface relative humidity (RHsurf) is incorporated into the loss function to constrain model weights, yielding our final Physics-guided LSTM (PyLSTM) model. Validation using single-site 3079 h DFMC data from Chengdu, Sichuan, China, demonstrated PyLSTM's superior temporal performance (R2 = 0.70, RMSE = 10.60%). Spatial validation across 241 sites in Xizang, Yunnan, Guizhou, and Sichuan provinces confirmed its robust spatial accuracy (R2 = 0.71, RMSE = 16.96%), outperforming both standalone FSMM and LSTM models. PyLSTM successfully captured the declining DFMC trend preceding the Yajiang fire event, with significantly lower estimated DFMC in the burned area compared to surrounding pixels in Yajiang County. These results demonstrate PyLSTM's capability to enhance wildfire risk early warning and identify high-risk areas. Therefore, this study serves as a foundational step toward estimating hourly regionalscale DFMC dynamics-a vital factor in assessing fire danger and behavior.

Live Fuel Moisture Content (LFMC) is a critical determinant of wildfire ignition and spread. Accurate forecasting of LFMC dynamics, particularly at a two-week timescale, is essential for early wildfire danger assessment. While satellite remote sensing provides valuable current and historical observations, it lacks the ability to predict future LFMC dynamics. Meanwhile, although weather forecasts are relatively reliable over short timescales (up to two weeks), LFMC models based solely on meteorological inputs often fall short, particularly when predicting conditions at the species level. To address these limitations, this study introduces a species-specific approach that integrates MODIS-derived LFMC data into the biophysical process-based MEDFATE model to optimize LFMC simulations and enable short-term forecasting based on daily weather projections. A global sensitivity analysis was conducted to identify key input parameters for different tree species within MEDFATE. These parameters guided the development of a cost function that quantifies discrepancies between model-simulated and fieldmeasured LFMC, enabling species-specific model calibration. To enhance model optimization, the global optimal DEoptim algorithm was combined with four-dimensional variational data assimilation (4D-Var) to integrate MODIS-derived LFMC estimates into MEDFATE. Using weather projections, the optimized MEDFATE model produced LFMC forecasts at about a two-week timescale. Time-series measurements of LFMC dynamics for Quercus faginea, Quercus ilex, and Pinus halepensis in Spain, Pinus ponderosa in the USA, and Eucalyptus species in Australia demonstrated that model calibration improved daily LFMC estimates (R2 increased from 0.22 to 0.31; RMSE reduced from 18.71% to 16.04%). Further incorporation of MODIS-derived LFMC data significantly enhanced accuracy (R2 = 0.56; RMSE = 9.75%). Validation across seven wildfire events in Spain, Australia, and the USA confirmed the effectiveness and operational relevance of the approach for early fire warning. These findings underscore the potential of integrating satellite remote sensing and meteorological data into biophysical process-based models to improve tree species-specific LFMC prediction and support proactive fire management.

Forest fires pose significant threats to human lives and the environment globally. As the material basis and energy source of forest fires, forest fuel load is a key parameter influencing fire behavior, intensity, and spread. Forest fuels can be managed to some degree and play a vital role in fire risk assessment and the formulation of fire prevention strategies. The large-area monitoring of forest fuel loads remains a key research priority, given their temporal dynamics as well as the horizontal and vertical spatial variability in fuel properties that exert distinct influences on various types of forest fires. This review synthesizes literature on forest fuel load inventories from the 19th century to the present, tracing the evolution from traditional approaches to multi-source remote sensing-based methods. Early fuel load estimations relied on static fuel models derived from field observations. With advances in the dynamic estimation of forest structural parameters, fuel modeling has transitioned toward dynamic frameworks. Initial remote sensing-based methods mainly used passive optical data to estimate canopy fuel loads, while recent developments include applications of active techniques such as Synthetic Aperture Radar (SAR) and Light Detection and Ranging (LiDAR), as well as the integration of multi-source remote sensing for estimating fuel loads across vertical strata. We discuss the current challenges and future perspectives for estimating forest fuel loads, highlighting the need for improved estimation of canopy vertical structure, identification of ladder fuels connecting surface and canopy layers, and accurate surface fuel load mapping for surface fires and prescribed burn management. This review aims at providing theoretical insights and technical guidance for forest fuel load estimation, thereby supporting the development of forest fire monitoring and early warning systems.

High-intensity forest fires have significant destructive impacts on ecosystems and society, and are an increasing concern worldwide. Accurate probabilistic risk assessment of these fires can effectively enhance the ability to guide wildfire management, particularly for large and extreme fires. However, forecasting large-scale fire behavior characteristics remains challenging, limiting the effectiveness of spatial estimations of high-intensity forest fire potential (HIFFP). This study aims to integrate fire spread simulations and machine learning (ML) algorithms to enhance HIFFP estimations through multi-step time-series forecasting on fire rate of spread and fireline intensity at regional scales. We first established a high-intensity forest fire dataset based on remote sensing-informed fire spread simulations from the Weather Research and Forecasting coupled fire-spread model (WRF-SFIRE), incorporating explanatory variables on fuel, weather, climate, and topography. Then, the knowledge-guided framework (multi-step time series-based ML, MTS-ML) was designed to estimate HIFFP within different hours after fires occur, integrating with Bayesian Network (BN), Random Forest (RF), and copula models. Results indicate that MTS-ML improved HIFFP modeling compared with ML-based methods, achieving AUC (the area under the receiver operating characteristic curve) > 0.95 (with similar to 0.04 increments), F1 score > 0.85 (with similar to 0.08 increments), and MAE < 0.15. Topographic index, foliage fuel load, and wind speed are identified as primary contributors to HIFFP. Probabilistic mapping of HIFFP represents wildfire danger, which is closely linked to burn severity and fire-induced carbon emissions. This study presents a novel framework for enhancing regional risk assessment of high-intensity forest fires, providing valuable guidance in wildfire control and management.

Extreme climatic events, particularly prolonged droughts and high-intensity wildfires, are becoming increasingly frequent, drawing attention to the compounding impacts of these events on forest ecosystems. Following a severe early-season drought, southeastern Australia experienced sudden and extensive forest canopy dieback between September and October 2023. In this study, we addressed two main objectives: (1) to develop a high-resolution canopy dieback prediction model for classifying the severity of the 2023 canopy dieback event, and (2) to analyze the key drivers of its spatial pattern. For objective 1, we selected and tested a suite of canopy moistureand greenness-related variables from Sentinel-2A/2B data (10 m). A random forest model was trained using these variables, with a classification accuracy of 0.93. Live fuel moisture content (LFMC) emerged as the most significant predictor for predicting dieback, yielding a classification accuracy of 0.91 when using it alone. For objective 2, a SHAP analysis of the spatial drivers of dieback revealed that north-facing slopes and ridges experienced the highest canopy loss, while areas affected by the 2019-2020 megafires were significantly associated with more severe canopy dieback, suggesting a potential legacy effect that warrants further investigation. A time-series analysis from 2019 to 2023 identified multiple canopy dieback events coinciding with periods of drought and wildfire, highlighting a compounding stress cycle of drought-wildfire-drought. These findings underscore the robustness of the remotely sensed LFMC in predicting canopy dieback and highlight the importance of integrating climate, topographic, and disturbance factors into forest management to mitigate canopy dieback and enhance ecosystem resilience.

Landslides, as a common geological disaster, pose a significant threat to the safety of people’s lives and property, as well as the stability of infrastructure in mountainous areas. Based on a detailed investigation of the Shibanping landslide, this paper first analyzes the causes and mechanisms of the landslide. Secondly, the plane sliding method and numerical simulation method are adopted to explore the changes in the stability of the landslide under natural, rainfall, and seismic conditions. The research results show that the instability of the Shibanping landslide is closely related to the complex geological structure, topography, geomorphology, lithology, excavation activities, hydrological changes caused by rainfall, and seismic effects in the area. The deformation of the landslide starts with tensile cracking and downward displacement at the top. Then, it gradually expands to the deep layer and the front edge under the influence of rainfall and excavation; its mechanism is progressive-bedding slip. Rainfall increases the self-weight, saturation, and pore water pressure of the rock and soil mass, leading to a continuous decrease in effective stress, thereby exacerbating the instability of the landslide. Earthquakes reduce the stability of the slope through shear stress caused by seismic waves. The stability calculation results indicate that the landslide has a certain degree of stability under natural conditions, but its stability decreases significantly when rainfall or earthquakes occur. This study provides theoretical support for the analysis and prevention of landslide stability, and is of great significance especially for the prediction, prevention, and control of landslides in areas with long-term rainfall and high earthquake frequency.

Urban traffic accident will cause national economic and property losses, people’s lives are threatened, and traffic accident is a hot issue concerned by all countries. Therefore, it is important for traffic management departments to study the main factors affecting urban traffic accidents and to accurately predict the severity of traffic accidents. Using the data set of Seattle City from 2004 to 2020 released by SDOT, In this paper, the traditional machine learning model, artificial neural network MLP model, Random Forest (RF) model and Gradient Boosting model were compared to predict the severity of traffic accidents. Among them, RF and Gradient Boosting models were better, with an accuracy of $75.9 \%$. By output important factors affecting the severity of traffic accidents and partial dependence plots, the model was more explanatory. Based on location, time, number of people, number of vehicles, collision type, weather, road conditions and other factors, traffic management departments can accurately predict the severity of traffic accidents and reduce property losses and casualties.

Climate change and extreme weather have increased fire frequency and intensity in Pakistan, but the national wildfire risk assessment remains underdeveloped, with existing studies focusing only on limited regions and few exploring comprehensive regional wildfire risk assessment. Furthermore, there is a lack of exploration into the temporal characteristics of contributing factors. This study constructs a comprehensive database using fuel, meteorological, topographic, and humanistic data to analyze historical wildfire risk factors. A wildfire risk assessment model is developed by integrating meteorological and fuel-related time series data, employing the eXtreme Gradient Boosting algorithm. Results show that the model incorporating temporal data outperforms the one without, with AUC values of 0.929 and 0.906, respectively. These findings highlight the importance of temporal features in improving wildfire risk prediction accuracy, contributing to more effective early assessment and risk management strategies in Pakistan.

Wildfires occur more frequently with warmer climate worldwide. Explainable wildfire susceptibility assessments can improve our understanding of wildfire regimes and provide efficiently decision-making support to fire management. Researchers have attempted to integrate the temporal characteristics of fire-related variables into wildfire susceptibility modeling, but the accuracy still with limited improvement. The Long Short-Term Memory network (LSTM) has shown significant advantages in extracting features from time-series data. This study aimed to integrate LSTM networks and machine learning models (i.e., random forest (RF), extreme gradient boosting (XGBoost)) to enhance the accuracy of wildfire susceptibility assessment, and to employ the SHapley Additive exPlanations (SHAP) method for the explanation of variable importance. Results indicate that the ensemble model combining LSTM and XGBoost showed superior performance across all metrics. Compared to model without LSTM integration, the accuracy increased by 3.3%, and a similar improvement of 5.4% was observed with RF-based model. The SHAP method identified the most critical factors for each category as: total evaporation (meteorological), elevation (topographic), distance from residential areas (anthropogenic), and NDVI (vegetation). In wildfire susceptibility mapping based on the XGBoost model integrated with LSTM, there is a strong association between wildfire susceptibility levels and the occurrence of wildfires.

The thawing of ice-rich permafrost leads to the formation of thermokarst landforms.Precise mapping of retrogressive thaw slumps(RTSs)is imperative for assessing the deg-radation and carbon exchange of permafrost at both local and regional scales on the Tibetan Plateau(TP).However,previous methods for RTSs mapping rely on a large number of sam-ples and complex classifiers with low automation level or unnecessary complexity.We pro-pose an automatic mapping network(AmRTSNet)for producing decimeter-level RTSs maps from GaoFen-7 images based on deep learning.Both the quantitative metrics and qualitative evaluations show that AmRTSNet trained in the Beiluhe offers significant advantages over previous methods.Without further fine-tuning,we conducted RTSs automatic mapping based on AmRTSNet in the Wulanwula,Chumarhe,and Gaolinggo.Over 141,312 ha on the TP have been automatically mapped,comprising 926 RTS regions with a total RTS area of 2318.72 ha.The average statistics of the mapped RTSs show low roundness(0.38),moder-ate rectangularity(0.61),and high convexity(0.79).About 90％of the RTSs are smaller than 6 ha.The average aspect ratio is 2.18.RTSs are unevenly distributed in belt-like aggregations with dominant density peaks.RTSs often concentrate in hillslopes and along lateral streams,with more dense areas more likely to have larger RTSs.

Volcanic eruptions are highly destructive natural geological hazards, often accompanied by lava flows and gas emissions, which lead to surface deformation. The Lewotobi volcano erupted again on December 23, 2023. To investigate the impact of volcanic eruptions on surface deformation, this study employs multi-temporal Sentinel-1A data and the Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) technique to monitor temporal changes in surface deformation in the Lewotobi volcano region one year before and after its first eruption. The results indicate that continuous volcanic activity leads to surface subsidence or collapse, while large explosive eruptions are associated with widespread surface uplift prior to the event.

The integration of data from multi-source Geostationary Earth Orbit (GEO) satellites can significantly enhance disaster monitoring capabilities, such as wildfires, volcano eruptions and disastrous weather events, due to its advantages in real-time and high frequency observation. In this study, the radiance consistency and deviation characteristics are analyzed for Advanced Himawari Imager (AHI) on board Himawari-9 and Advanced Meteorological Imager (AMI) of GK2A infrared (IR) channels, including. The results show that the brightness temperature (BT) generally consistent for IR channels with r > 0.95 in the overlapping regions of GK2A and Himawari-9. The BT bias is -1.08K to 0.07K with no significant temporal variations, and channel B16 (13.3 mu m) has maximum absolute bias. In addition, the results indicate that BT bias of GK2A and Himawari-9 IR channels present different characteristics with changes in satellite zenith angle, and surface types. The radiance consistency and deviation characteristics need to be further analyzed in detail, such as, in different surface types. Additionally, a fusion method should be designed to reduce the differences between the two sensors, enabling more effective integration of multi-source GEO satellites with high-frequency.

In recent years, the continuous development of deep learning has significantly advanced its application in the field of remote sensing. However, the semantic segmentation of high-resolution remote sensing images remains challenging due to the presence of multi-scale objects and intricate spatial details, often leading to the loss of critical information during segmentation. To address this issue and enable fast and accurate segmentation of remote sensing images, we made improvements based on SegNet and named the enhanced model CSNet. CSNet is built upon the SegNet architecture and incorporates a coordinate attention (CA) mechanism, which enables the network to focus on salient features and capture global spatial information, thereby improving segmentation accuracy and facilitating the recovery of spatial structures. Furthermore, skip connections are introduced between the encoder and decoder to directly transfer low-level features to the decoder. This promotes the fusion of semantic information at different levels, enhances the recovery of fine-grained details, and optimizes the gradient flow during training, effectively mitigating the vanishing gradient problem and improving training efficiency. Additionally, a hybrid loss function combining weighted cross-entropy and Dice loss is employed. To address the issue of class imbalance, several categories within the dataset are merged, and samples with an excessively high proportion of background pixels are removed. These strategies significantly enhance the segmentation performance, particularly for small-sample classes. Experimental results from the Five-Billion-Pixels dataset demonstrate that, while introducing only a modest increase in parameters compared to SegNet, CSNet achieves superior segmentation performance in terms of overall classification accuracy, boundary delineation, and detail preservation, outperforming established methods such as U-Net, FCN, DeepLabv3+, SegNet, ViT, HRNe and BiFormert.

Achieving harmony between the ecological environment and urbanization is essential for sustainable development. In China, the rapid pace of urbanization and intensified human activities have heightened the conflict between human progress and environmental sustainability. Using MODIS and socioeconomic data, this study analyzed the spatiotemporal evolution and interaction between the Modified Remote Sensing Ecological Index (MRSEI) and Urbanization Level Index (ULI) in Jiangsu Province from 2002 to 2022. The findings are as follows: (1) Northern Jiangsu has experienced consistent improvement in ecological environment, while southern Jiangsu saw significant degradation between 2002 and 2010, followed by recovery and stabilization from 2010 to 2022. (2) Southern Jiangsu maintains a higher urbanization level with rapid growth, whereas urbanization in northern Jiangsu accelerated after 2014 but still lags behind the south. (3) Both the urbanization lag in northern Jiangsu and the ecological environment lag in southern Jiangsu have improved over time, contributing to enhanced coordination between ecological environment and urbanization across the province.

The Canopy Live Fuel Moisture Content (LFMC) is a pivotal factor in wildfire risk assessment within the fire triangle model, representing the ratio of canopy moisture content to its dry weight. Against the backdrop of degraded Moderate Resolution Imaging Spectroradiometer (MODIS) performance and the underutilization of Visible Infrared Imaging Radiometer Suite (VIIRS) in LFMC inversion, this study harnessed the coupled radiative transfer models (RTMs) to probe the spectral sensitivity of the VIIRS to LFMC and pinpoint the optimal band combination for LFMC inversion. To tackle the challenge of ill-posed inversion, we leveraged the correlation coefficient matrix to mitigate erroneous combinations of free parameters in the construction of the lookup table. Results showcase that VIIRS-based LFMC inversion yields marginally superior accuracy (R2= 0.57, R2= 0.32) for both grassland and forest types, with VIIRS-based inversion demonstrating a lower relative root mean square error (rRMSE = 5.84%), compared to results from the MODIS. By scrutinizing LFMC trends alongside precipitation (PP) data in four forest fires spanning from 2019 to 2022 in southwest China, varied degrees of LFMC decrease preceding fire outbreaks. Those results substantiated the validity of the proposed method for wildfire warning. Consequently, our study asserts the reliability of VIIRS in LFMC inversion, positioning it as a viable substitute and extension of MODIS. VIIRS offers continuous and effective product support for wildfire warning assessment, enhancing our ability to monitor and mitigate wildfire risks.

Accurate evaluation of forest carbon sequestration is crucial for understanding its role in mitigating climate change. Previous methods to estimate forest biomass carbon storage were dominantly based on single-phase data, while there remained large uncertainties in the forest carbon flux simulation. In this study, we employed Landsat 8 (L8) and Sentinel-1 (S1) time-series data to explore the potential of Optical and SAR spatiotemporal features for estimating forest aboveground carbon (AGC) density and developed a forest AGC density model. The results showed that the employment of temporal features significantly improved the performance of AGC density estimation compared with only spatial features, with R-2 increasing 0.25 similar to 0.36 and rRMSE decreasing 8%similar to 11%. The combination of the two sensors' data has achieved the best performance with R-2 of 0.72 and rRMSE of 25%. This study demonstrated that the proposed method based on multi-sensor spatiotemporal features was promising for forest aboveground biomass carbon estimation.

Temporal information derived from remote sensing time-series (TS) data, such as growth trends and periodic change characteristics, has been proven to be valuable in improving forest aboveground biomass carbon (AGC) estimation. Accurate time-series reconstruction is vital for comprehensive analysis and maximizing the potential of temporal information in understanding dynamic vegetation changes. However, significant variations exist in the reconstructed results of various existing models, impeding accurate analysis of temporal dynamics. This study proposed an ensemble reconstruction (EM) model inspired by ensemble methods to enhance robustness. Stepwise comparisons revealed that previous models performed optimally under specific conditions. Notably, it remains challenging for any single model to achieve both high accuracy and robustness simultaneously. The proposed EM model demonstrated the best performance, with an average R2 value of 0.930, while also enhancing robustness. Additionally, the impact of TS reconstruction on forest AGC estimation was significant. When temporal information derived from the reconstruction models was applied to AGC modeling, the EM model produced the best AGC estimation (R2 = 0.749, rRMSE = 30.316%). In conclusion, this study proposed the EM model that enhanced both the accuracy and robustness of TS reconstruction, thereby improving AGC estimation.