Forest above-ground biomass estimation based on strongly collinear variables derived from airborne laser scanning data

Estimating forest aboveground biomass (AGB) and its components (wood, branch, bark, foliage) is critical for forest inventories and provides important information for timber harvesting and carbon accounting. Current approaches for modelling forest AGB at the stand scale often employ airborne laser scanning (ALS) data which provide robust AGB estimates. However, in structurally complex forest ecosystems, ALS-based models may not estimate forest biomass and its components with sufficient accuracy. One method to improve ALS-based model performance is through data fusion. Deep neural networks (DNNs) are effective for data fusion because they can combine different data modalities without the need to modify the original data resolution. This study evaluated the effectiveness of a data fusion DNN that combines ALS, multispectral, and topographic data for forest biomass estimation (total and component). We implemented a DNN architecture consisting of three convolutional neural network (CNN) modules: Octree-CNN for ALS data; 1-D CNN for Landsat-8 multispectral data; and 2-D CNN for topographic data. Variants of the DNN architecture combining different input data modalities were trained and tested using sample plots from New Brunswick, Canada (n = 2,336). The model, including all three data modalities, performed best overall for total AGB estimation (R2 = 0.77; RMSE = 28.38 Mg/ha) and explained an additional 2−5% variation in wood, bark, and foliage biomass compared to the ALS-only model. This study demonstrates the effectiveness of a novel data fusion DNN architecture that extracts information directly from input data modalities for improving forest biomass estimates. However, relatively small performance gains should be weighed against computational resources and domain knowledge required to implement and interpret DNNs.

The accurate estimation of forest aboveground biomass (AGB) is essential for effective forest resource management and carbon stock assessment. However, the estimation accuracy of forest AGB is often constrained by scarce in situ measurements and the limitations of using a single data source or retrieval model. This study proposes a multi-source data integration framework using Sentinel-1 (S-1) and Sentinel-2 (S-2) data along with eight predictive models (i.e., multiple linear regression—MLR; Elastic-Net; support vector regression (with a linear kernel and polynomial kernel); k-nearest neighbor; back-propagation neural network—BPNN; random forest—RF; and gradient-boosting tree—GBT). With airborne light detection and ranging (LiDAR)-derived AGB as a reference, a three-stage optimization strategy was developed, including stepwise feature selection (SFS), hyperparameter optimization, and multi-decision vector fusion (MDVF) model construction. Initially, the optimal feature subsets for each model were identified using SFS, followed by hyperparameter optimization through a grid search strategy. Finally, eight models were evaluated, and MDVF was implemented to integrate outputs from the top-performing models. The results revealed that LiDAR-derived AGB demonstrated a strong performance (R2 = 0.89, RMSE = 20.27 Mg/ha, RMSEr = 15.90%), validating its effectiveness as a supplement to field measurements, particularly in subtropical forests where traditional inventories are challenging. SFS could adaptively select optimal variable subsets for different models, effectively alleviating multicollinearity. Satellite-based AGB estimation using the MDVF model yielded robust results (R2 = 0.652, RMSE = 31.063 Mg/ha, RMSEr = 20.4%) through the synergy of S-1 and S-2, with R2 increasing by 4.18–7.41% and the RMSE decreasing by 3.55–5.89% compared to the four top-performing models (BPNN, GBT, RF, MLR) in the second optimization stage. This study aims to provide a cost-effective and precise strategy for large-scale and spatially continuous forest AGB mapping, demonstrating the potential of integrating active and passive satellite imagery with airborne LiDAR to enhance AGB mapping accuracy and support further ecological monitoring and forest carbon accounting.

Allometric regression models are widely used to estimate tropical forest biomass, but balancing model accuracy with efficiency of implementation remains a major challenge. In addition, while numerous models exist for aboveground mass, very few exist for roots. We developed allometric equations for aboveground biomass (AGB) and root biomass (RB) based on 300 (of 45 species) and 40 (of 25 species) sample trees respectively, in an evergreen forest in Vietnam. The biomass estimations from these local models were compared to regional and pan-tropical models. For AGB we also compared local models that distinguish functional types to an aggregated model, to assess the degree of specificity needed in local models. Besides diameter at breast height (DBH) and tree height (H), wood density (WD) was found to be an important parameter in AGB models. Existing pan-tropical models resulted in up to 27% higher estimates of AGB, and overestimated RB by nearly 150%, indicating the greater accuracy of local models at the plot level. Our functional group aggregated local model which combined data for all species, was as accurate in estimating AGB as functional type specific models, indicating that a local aggregated model is the best choice for predicting plot level AGB in tropical forests. Finally our study presents the first allometric biomass models for aboveground and root biomass in forests in Vietnam.

Medium spatial resolution biomass is a crucial link from the plot to regional and global scales. Although remote-sensing data-based methods have become a primary approach in estimating forest above ground biomass (AGB), many difficulties remain in data resources and prediction approaches. Each kind of sensor type and prediction method has its own merits and limitations. To select the proper estimation algorithm and remote-sensing data source, several forest AGB models were developed using different remote-sensing data sources (Geoscience Laser Altimeter System (GLAS) data and Thematic Mapper (TM) data) and 108 field measurements. Three modeling methods (stepwise regression (SR), support vector regression (SVR) and random forest (RF)) were used to estimate forest AGB over the Daxing’anling Mountains in northeastern China. The results of models using different datasets and three approaches were compared. The random forest AGB model using Landsat5/TM as input data was shown the acceptable modeling accuracy (R2 = 0.95 RMSE = 17.73 Mg/ha) and it was also shown to estimate AGB reliably by cross validation (R2 = 0.71 RMSE = 39.60 Mg/ha). The results also indicated that adding GLAS data significantly improved AGB predictions for the SVR and SR AGB models. In the case of the RF AGB models, including GLAS data no longer led to significant improvement. Finally, a forest biomass map with spatial resolution of 30 m over the Daxing'anling Mountains was generated using the obtained optimal model.

Accurate estimation of forest aboveground biomass is essential for the assessment of regional carbon cycle and the climate change in the terrestrial ecosystem. Currently, ensemble learning algorithms and cross-validation methods have been widely applied to estimate regional forest Above Ground Biomass (AGB). However, the effects of ensemble learning algorithms, validation methods, and their interactions on forest AGB estimation were rarely investigated. Based on Landsat 8 Operational Land Imager (OLI) imagery, Airborne Laser Scanning (ALS) data and China’s National Forest Continuous Inventory data, this study explored the effects of five ensemble learning algorithms, including Simple Averaging (SA), Weighted Averaging (WA), Stacked Generalization (SG), Random Forest (RF) and Extreme Gradient Boosting (XGBoost), and two validation approaches (i.e. 10-fold and leave-one-out cross-validation) on the AGB estimation of the Natural Secondary Forests (NSFs) in northeast China. The results revealed that the ensemble learning algorithms that combine heterogenous-based models (i.e. SA, WA, SG) generally produced higher accuracy than the base models (i.e. Decision Tree (DT), K-Nearest Neighbor (KNN), Support Vector Regression (SVR), Convolutional Neural Network (CNN)). Among all ensemble learning algorithms, the SG algorithm has the highest accuracy whereas the XGBoost algorithm has the lowest accuracy. Although prediction models considerably impact the accuracy of AGB estimation, the validation approach also plays a non-negligible role in AGB estimation. The leave-one-out cross-validation produced much higher accuracy than the 10-fold cross-validation using the same prediction model and tends to generate over-optimistic AGB estimates compared to 10-fold cross-validation, especially for the averaging and stacking ensemble learning algorithms (i.e. SA, WA, SG). This study highlights the potential challenges of applying a leave-one-out cross-validation approach and provides a scientific foundation for the feasibility of different ensemble learning algorithms and cross-validation approaches for accurate AGB estimation.

Forest aboveground biomass (AGB) estimation is crucial for carbon cycle studies and climate change mitigation actions. However, because of limitations in timely and reliable forestry surveys and high-resolution remote sensing data, producing a fine resolution and spatial continuous forest AGB map of China is challenging. Here, we combined 4789 ground-truth AGB measurements and multisource remote sensing data such as a recently released forest canopy-height product, optical spectral indexes, topographic data, climatological data, and soil properties to train a random forest regression model for forest AGB estimation of China at 30-m resolution. The accuracy of the estimated AGB can yield R2 = 0.67 and RMSE = 70.71 Mg/ha. The nationwide estimates show that the average forest AGB and total forest carbon storage were 97.57 ± 23.85 Mg/ha and 11.06 Pg C for the year 2019, respectively. The value of AGB uncertainty ranges from 0.68 Mg/ha to 37.80 Mg/ha, and the average AGB uncertainty was 4.32 ± 1.75 Mg/ha. The forest AGB estimates of China in this study correspond reasonably well with the AGB estimates derived from the forestry and grassland statistical yearbook at the provincial level (R2 = 0.61, RMSE = 30.15 Mg/ha). In addition, we found that previous AGB products generally underestimate the forest AGB compared with our estimated AGB at the pixel-level and ground-truth AGB measurements. The high-resolution forest AGB map provides an important alternative data source for forest carbon cycle studies and can be used as a baseline map for forest management and conservation practices.

This research examines the role of canopy cover in influencing above ground biomass (AGB) dynamics of an open canopied forest and evaluates the efficacy of individual-based and plot-scale height metrics in predicting AGB variation in the tropical forests of Angkor Thom, Cambodia. The AGB was modeled by including canopy cover from aerial imagery alongside with the two different canopy vertical height metrics derived from LiDAR; the plot average of maximum tree height (Max_CH) of individual trees, and the top of the canopy height (TCH). Two different statistical approaches, log-log ordinary least squares (OLS) and support vector regression (SVR), were used to model AGB variation in the study area. Ten different AGB models were developed using different combinations of airborne predictor variables. It was discovered that the inclusion of canopy cover estimates considerably improved the performance of AGB models for our study area. The most robust model was log-log OLS model comprising of canopy cover only (r = 0.87; RMSE = 42.8 Mg/ha). Other models that approximated field AGB closely included both Max_CH and canopy cover (r = 0.86, RMSE = 44.2 Mg/ha for SVR; and, r = 0.84, RMSE = 47.7 Mg/ha for log-log OLS). Hence, canopy cover should be included when modeling the AGB of open-canopied tropical forests.

A critical component of the forest ecosystem, the understorey supports the vast majority of wildlife habitat and total ecosystem floristic diversity. Remote sensing data have been developed to provide information at different scales for surveys of forest ecosystems, but obtaining information about the understorey remains a challenge. As rapid and efficient tools for forest structure attribute estimation, Airborne Laser Scanning (ALS) and Terrestrial Laser Scanning (TLS) have attracted much attention. We examine the relationship between ALS and TLS data and detect changes in the forest understorey caused by forest-tending events in the study area. We conducted trials in five plots within a young Khasi pine (Pinus kesiya Royle ex Gord.) plantation in Yunnan province, China, before and after forest tending. We collected bi-temporal ALS data in this area and TLS data from 10 scanning stations. Canopy height profiles were retrieved from ALS and TLS data, and understorey material volume retrieved from filled TLS voxels volume agreed well with the understorey point clouds percentile distribution (PD) obtained from ALS data. The PD value for the understorey from ALS multiplied by the percentage of ALS return points in the overstorey had a stronger correlation (R2 = 0.90) with the TLS-derived understorey material volume than did the ALS PD value for the understorey only (R2 = 0.71). Taking the effect of the overstorey into consideration will improve evaluations of forest understorey parameters from ALS data. This study confirmed the potential of TLS as a validation tool to assess the accuracy of forest understorey material volume estimation at large scales using ALS data.

This dissertation addresses the problem of automated vector extraction from Airborne Laser Scanner (ALS) data in urban areas. The recent popularity of Geographic Information Systems (GIS) has stimulated research on automated object extraction in order to simplify the data acquisition and update process. By automatically generating GIS inputs and updates from a single data source, the cost of data acquisition and processing will be kept to a minimum. Compared with other remote sensing techniques, extraction of objects from ALS data is in its infancy. ALS sensor technology has evolved rapidly and now allows the acquisition of very dense point clouds in a short period of time. ALS data is unique in that it explicitly contains 3D information and is acquired from an active sensor. As such, there are several benefits that can be immediately realised by using an ALS only data approach; data acquisition will not be limited to daylight hours as with other sensors and accurate height information is contained in the data. This means that registration of different data sources is not required and as only one data source is used, acquisition costs are minimised. Apart from these facts, ALS data has some other unique properties that have not been utilised to their full potential. An ALS sensor can record both height and intensity information from multiple returns in a single swath. To date, all the available information has rarely been used. For example, the intensity of a laser return has regularly been dismissed as it was considered under-sampled and noisy. This data still contains usable information and until this information is used, the optimum object recognition results will not be achieved; thus the use of as much of this information as possible is a major focus of this thesis. As ALS is an explicit 3D data source, the early stages of development were primarily focused on topographic mapping of terrain in forested areas in order to generate Digital Terrain Models (DTMs). As sensor technology has improved, so has the achievable resolution of point clouds from ALS data, and methods to extract objects from stand-alone ALS data have emerged. Attempts have been made to create city models and maps from ALS data that included buildings, roads, trees and powerlines. Each of these spatial object types has unique attributes, which means that automatic map creation is not an easy task. For example, buildings in general are easily detected in ALS data but the building outline is not easily delineated. Another difficulty with building detection is the separation of buildings and bridges as they have many of the same properties as observed by an ALS system. Bridges can usually be found in a road network but road extraction techniques typically produce poor detection rates and often require existing data and / or user interaction in semi-automatic techniques. These simple examples highlight the complexity of automatically generating vectorised maps from ALS data or in fact any data source. In this thesis, new methods are presented for the automatic creation of vectorised maps from ALS data. A two-step processing paradigm is adopted for this purpose, namely the classification of the ALS data and the vectorisation of the classification results. A classification strategy is introduced that creates a hierarchy for object detection with respect to the ALS data itself. This approach develops an ontology between the spatial object classes and the ALS data. The hierarchical framework highlights the fact that one object might not be discernable within the data without considering another. New classification algorithms are then presented within this framework. Each algorithm attempts to exploit the attributes of the data that are consistent within the spatial object class being considered as described by the classification framework. New algorithms for the classification of roads, trees and powerlines are all introduced whilst an extension to existing building classification methods is presented. Once classification is complete, a vectorisation process specific to the task at hand can be employed to yield vectorised results. These developed vectorisation processes are new and include an algorithm that has been generalised to allow the vectorisation of thick lines in images by detecting the centreline, direction and width. The primary goal of this thesis is to present a framework of new algorithms that will allow automatic spatial object detection and vectorisation whilst providing results of an acceptable quality. The algorithms presented rely solely on ALS data and require minimal operator knowledge. Each algorithm has been designed to exploit the way in which the object exhibits itself in the data. The new algorithms are integrated into a software package called JTD (Join The Dots) that will facilitate the effective automatic processing of ALS data. The results of the new algorithms have been evaluated over four 2 x 2 km areas that have been sampled with medium resolution ALS data. The results for each area are displayed and analysed to show the applicability of the whole process in an exemplary way.

The demand for cost-efficient forest aboveground biomass (AGB) prediction methods is growing worldwide. The National Land Survey of Finland (NLS) began collecting airborne laser scanning (ALS) data throughout Finland in 2008 to provide a new high-detailed terrain elevation model. Similar data sets are being collected in an increasing number of countries worldwide. These data sets offer great potential in forest mapping related applications. The objectives of our study were (i) to evaluate the AGB component prediction accuracy at a resolution of 300 m2 using sparse density, leaf-off ALS data (collected by NLS) derived metrics as predictor variables; (ii) to compare prediction accuracies with existing large-scale forest mapping techniques (Multi-source National Forest Inventory, MS-NFI) based on Landsat TM satellite imagery; and (iii) to evaluate the accuracy and effect of canopy height model (CHM) derived metrics on AGB component prediction when ALS data were acquired with multiple sensors and varying scanning parameters. Results showed that ALS point metrics can be used to predict component AGBs with an accuracy of 29.7%–48.3%. AGB prediction accuracy was slightly improved using CHM-derived metrics but CHM metrics had a more clear effect on the estimated bias. Compared to the MS-NFI, the prediction accuracy was considerably higher, which was caused by differences in the remote sensing data utilized.

Accurate estimation of forest aboveground biomass (AGB) is crucial for monitoring ecosystem responses to environmental change. Optical remote sensing is the most widely used method for obtaining AGB information. However, there is a need for improving the accuracy of AGB estimates obtained in this way. In this paper, a synergistic estimation model through the integration of spectral and textural features from GF-1 satellite image for AGB estimation in the arid region of Ejin, China was put forward. The synergistic model was verified by comparing with the ground measurements and the results of the best spectrum-alone and texture-alone models. The results indicated that the proposed synergistic estimation model was more effective than the spectrum-alone or texture-alone model, and showed considerable potential in forest aboveground biomass estimation by combining spectral and textural information.

Accurate regional-scale estimation of forest aboveground biomass (AGB) is critical for effective forest management and terrestrial carbon cycle research. However, applications integrating multiple machine learning models (MLs) for forest AGB estimation in mountainous forests remain limited. In this study, we introduced a practical method to estimate diameter at breast height (DBH < 5 cm) for under-threshold trees using National Forest Inventory (NFI) data. By combining Sentinel-2 remote sensing imagery and DEM data, we employed individual MLs (RF, XgBost, CatBoost and SVM) and a stacking approach to estimate forest AGB in Chongqing under two scenarios: with and without under-threshold trees. The DBH estimation method achieved high accuracy (R²=0.93, RMSE=1.46 cm). Feature importance analysis showed spectral bands dominated predictors, while vegetation and topographic indices varied across models. CatBoost outperformed RF and XgBoost in both scenarios. The stacked ensemble model demonstrated best performances in including under-threshold trees in cross-validation (CV) and external verification (EV) (R²=0.65, RMSE=24.34 Mg·ha -¹; R²=0.68, RMSE=25.45 Mg·ha -¹), generating 10m-resolution AGB maps with consistent spatial patterns suitable for mountainous urban terrain. This work advances AGB estimation in southwestern China’s mountains regions and provides insights for forest ecology and management.

Improving species-specific plot volume estimates based on airborne laser scanning and image data using alpha shape metrics and balanced field data

&amp;lt;p&amp;gt;The large uncertainty characterizing the terrestrial carbon (C) cycle is a consequence of the sparse and irregular observations on the ground. In terms of observations, spaceborne remote sensing has been achieving global, repeated coverages of the Earth since the late 1970s, with a continuous increase in terms of density of observations in time and spatial resolution, thus potentially qualifying as data source to fill such gap in knowledge. Above-ground biomass is a baseline for quantifying the terrestrial C pool; however, remote sensing observations do not measure the organic mass of vegetation. Above-ground biomass (AGB) of forests can only be inferred by inverting numerical models relating and combining multiple remote sensing observations. One of the longest time record of observations from space is represented by the backscattered intensity from the European Remote Sensing Wind Scatterometer (ERS WindScat) and the MetOp Advanced Scatterometer (ASCAT), both operating at C-band (wavelength of 6 cm). An almost unbroken time series of backscatter observations at 0.25&amp;amp;#176; spatial resolution exists since 1991 and data continuity is guaranteed in the next decades. In spite of the weak sensitivity of C-band backscatter to AGB, wall-to-wall estimates of AGB have been derived from high-resolution SAR observations by exploiting multiple observations acquired in a relatively short time period&amp;amp;#160; (Santoro et al., Rem. Sens. Env., 2011; Santoro et al., Rem. Sens. Env., 2015). We have now applied this approach to generate a global time series of AGB estimates for each year between 1992 and 2018 from the C-band scatterometer data at 0.25&amp;amp;#176; spatial resolution. The spatial patterns of AGB match known patterns from in situ records and other remote sensing datasets. The uncertainty of our AGB estimates is between 30% and 40% of the estimated value at the pixel level, providing strong confidence in multi-decadal AGB trends. We identify a constant increase of biomass across most boreal and temperate forests of the northern hemisphere. In contrast, we detect severe loss of biomass throughout the wet tropics during the 1990s and the beginning of the 2000 decade in consequence of massive deforestation. This loss in biomass is followed by a steady increase during the 2000s and the beginning of the most recent decade, coming more recently into saturation. Overall, we find that the global AGB density at 0.25&amp;amp;#176; steadily increased by 9% from 71.8 Mg ha&amp;lt;sup&amp;gt;-1&amp;lt;/sup&amp;gt; Pg in the 1990s to 78.1 Mg ha&amp;lt;sup&amp;gt;-1&amp;lt;/sup&amp;gt; in the 2010s. Combining our AGB density estimates with the annual maps of the Climate Change Initiative (CCI) Land Cover dataset, we show that total AGB in forests decreased slightly from 566 Pg in the 1990s to 560 Pg in the 2000s, then increased to 593 Pg in the 2010s, resulting in an almost 5% net increase during the last three decades.&amp;lt;/p&amp;gt;

The United States Forest Service Forest Inventory and Analysis (FIA) Program provides a diverse selection of data used to assess the status of the nation’s forests using sample locations dispersed throughout the country. Airborne laser scanning (ALS) systems are capable of producing accurate measurements of individual tree dimensions and also possess the ability to characterize forest structure in three dimensions. This study investigates the potential of discrete return ALS data for modeling forest aboveground biomass (AGBM) and gross volume (gV) at FIA plot locations in the Malheur National Forest, eastern Oregon utilizing three analysis levels: (1) individual subplot (r = 7.32 m); (2) plot, comprising four clustered subplots; and (3) hectare plot (r = 56.42 m). A methodology for the creation of three point cloud-based airborne LiDAR metric sets is presented. Models for estimating AGBM and gV based on LiDAR-derived height metrics were built and validated utilizing FIA estimates of AGBM and gV derived using regional allometric equations. Simple linear regression models based on the plot-level analysis out performed subplot-level and hectare-level models, producing R2 values of 0.83 and 0.81 for AGBM and gV, utilizing mean height and the 90th height percentile as predictors, respectively. Similar results were found for multiple regression models, where plot-level analysis produced models with R2 values of 0.87 and 0.88 for AGBM and gV, utilizing multiple height percentile metrics as predictor variables. Results suggest that the current FIA plot design can be used with dense airborne LiDAR data to produce area-based estimates of AGBM and gV, and that the increased spatial scale of hectare plots may be inappropriate for modeling AGBM of gV unless exhaustive tree tallies are available. Overall, this study demonstrates that ALS data can be used to create models that describe the AGBM and gV of Pacific Northwest FIA plots and highlights the potential of estimates derived from ALS data to augment current FIA data collection procedures by providing a temporary intermediate estimation of AGBM and gV for plots with outdated field measurements.