Search

European forest monitoring is a central topic nowadays due to the critical role that forests can play in combatting climate change. Crucial information on forests is the number of tree species and the area covered by each of them, as they vary concerning growth rates, wood value, value for biodiversity conservation, and susceptibility to disturbances and global warming. The primary source of forest information is national forest inventories (NFIs). However, they are updated too infrequently to accommodate climate change-related analyses, and their estimates are not based on wall-to-wall information. Remotely sensed data offer new opportunities for up-to-date and large-scale forest monitoring and for enhancing NFI estimates. However, despite the huge scientific efforts, it is still challenging to accurately map forest species through satellite imagery analysis.This study introduces a method for large-scale forest species mapping in the Netherlands using Sentinel-2 (S2) harmonic predictors and demonstrates a scientific procedure for reliably estimating area proportions from remote sensing-based species maps and comparing these estimates with NFI-based estimates.Compared to more standard predictors, harmonic predictors increased the model performance by 8% in terms of overall accuracy and the kappa coefficient by 9% while reducing omission and commission errors by as much as 18% and 13%, respectively. We estimated the area proportion of forest species for each 10-km cell covering the Netherlands first using NFI data and then using the predicted maps. Although the resulting estimates differ by source data and methods, we found an average deviation between NFI and remote sensing-based area proportion estimates of 9%, with deviations approaching 0% when increasing the number of NFI plots per cell.The outcomes of this research play an important role in understanding the relative strengths and limitations of remote sensing-based products and NFI data, as well as be a solid basis for forest species area proportion estimation when (i) no field data are available, (ii) more frequently updated estimates are required, or (iii) wall-to-wall and fine resolution spatially explicit estimates are needed.

Abstract Large‐scale ecological sampling networks, such as national forest inventories (NFIs), collect in situ data to support biodiversity monitoring, forest management and planning, and greenhouse gas reporting. Data harmonization aims to link auxiliary remotely sensed data to field‐collected data to expand beyond field sampling plots, but outliers that arise in data harmonization—questionable observations because their values differ substantially from the rest—are rarely addressed. In this paper, we review the sources of commonly occurring outliers, including random chance (statistical outliers), definitions and protocols set by sampling networks, and temporal and spatial mismatch between field‐collected and remotely sensed data. We illustrate different types of outliers and the effects they have on estimates of above‐ground biomass population parameters using a case study of 292 NFI plots paired with airborne laser scanning (ALS) and Sentinel‐2 data from Sawyer County, Wisconsin, United States. Depending on the criteria used to identify outliers (sampling year, plot location error, nonresponse, presence of zeros and model residuals), as many as 53 of the 292 Forest Inventory and Analysis plot observations (18%) were identified as potential outliers using a single criterion and 111 plot observations (38%) if all criteria were used. Inclusion or removal of potential outliers led to substantial differences in estimates of mean and standard error of the estimate of biomass per unit area. The simple expansion estimator, which does not rely on ALS or other auxiliary data, was more sensitive to outliers than model‐assisted approaches that incorporated ALS and Sentinel‐2 data. Including Sentinel‐2 predictors showed minimal increases to the precision of our estimates relative to models with ALS predictors alone. Outliers arise from many causes and can be pervasive in data harmonization workflows. Our review and case study serve as a note of caution to researchers and practitioners that the inclusion or removal of potential outliers can have unintended consequences on population parameter estimates. When used to inform large‐scale biomass mapping, carbon markets, greenhouse gas reporting and environmental policy, it is necessary to ensure the proper use of NFI and remotely sensed data in geospatial data harmonization.

Remote sensing (RS) facilitates forest inventory across a wide range of variables required by the UNFCCC as well as by other agreements and processes. The Conventional model-based (CMB) estimator supports wall-to-wall RS data, while Hybrid estimators support surveys where RS data are available as a sample. However, the connection between these two types of monitoring procedures has been unclear, hindering the reconciliation of wall-to-wall and non-wall-to-wall use of RS data in practical applications and thus potentially impeding cost-efficient deployment of high-end sensing instruments for large area monitoring. Consequently, our objectives are to (1) shed further light on the connections between different types of Hybrid estimators, and between CMB and Hybrid estimators, through mathematical analyses and Monte Carlo simulations; and (2) compare the effects and explore the tradeoffs related to the RS sampling design, coverage rate, and cluster size on estimation precision. Primary findings are threefold: (1) the CMB estimator represents a special case of Hybrid estimators, signifying that wall-to-wall RS data is a particular instance of sample-based RS data; (2) the precision of estimators in forest inventory can be greater for stratified non-wall-to-wall RS data compared to wall-to-wall RS data; (3) otherwise cost-prohibitive sensing, such as LiDAR and UAV, can support large scale monitoring through collecting RS data as a sample. These conclusions may reconcile different perspectives regarding choice of RS instruments, data acquisition, and cost for continuous observations, particularly in the context of surveys aiming at providing data for mitigating climate change.

Remotely sensed data are frequently used for predicting and mapping ecosystem characteristics, and spatially explicit wall-to-wall information is sometimes proposed as the best possible source of information for decision-making. However, wall-to-wall information typically relies on model-based prediction, and several features of model-based prediction should be understood before extensively relying on this type of information. One such feature is that model-based predictors can be considered both unbiased and biased at the same time, which has important implications in several areas of application. In this discussion paper, we first describe the conventional model-unbiasedness paradigm that underpins most prediction techniques using remotely sensed (or other) auxiliary data. From this point of view, model-based predictors are typically unbiased. Secondly, we show that for specific domains, identified based on their true values, the same model-based predictors can be considered biased, and sometimes severely so.We suggest distinguishing between conventional model-bias, defined in the statistical literature as the difference between the expected value of a predictor and the expected value of the quantity being predicted, and design-bias of model-based estimators, defined as the difference between the expected value of a model-based estimator and the true value of the quantity being predicted. We show that model-based estimators (or predictors) are typically design-biased, and that there is a trend in the design-bias from overestimating small true values to underestimating large true values. Further, we give examples of applications where this is important to acknowledge and to potentially make adjustments to correct for the design-bias trend. We argue that relying entirely on conventional model-unbiasedness may lead to mistakes in several areas of application that use predictions from remotely sensed data.

Forest parameter estimation is required to support the sustainable management of forest ecosystems. Currently, forest resource assessment is increasingly linked to auxiliary information obtained from remote sensing (RS) technologies. In forest parameter estimation, airborne laser scanning (ALS) data have been demonstrated to be an invaluable source of information. However, ALS data are often not available for the entire forest area, whereas images from multiple satellite systems offer new opportunities for large-scale forest surveys. This study aims to assess and estimate the contribution of field plot data and ALS data along with Landsat data to the precision of growing stock volume (GSV) estimates. We compared different approaches for model-assisted estimation of mean forest GSV per unit area using different proportions of field sample data, ALS cover data, and wall-to-wall Landsat data. Model-assisted estimators were used with NFI sample data in an Italian study area using 10 RS predictors, specifically the seven Landsat 7 ETM+ bands and three fine-resolution metrics based on ALS-derived canopy height. We found that relative to the standard simple expansion estimator, the model-assisted estimators produced relative efficiency of 1.16 when using only Landsat data and relative efficiencies as great as 1.33 when using increasing levels of ALS coverage.

Remote sensing (RS) has enhanced forest inventory with model-based inference, that is, a family of statistical procedures rigorously estimates the parameter of a variable of interest (VOI) for a spatial population, e.g., the mean or total of forest carbon for a study area. Upscaling in earth observation, alias to this estimation, aggregates VOI from a finer spatial resolution to a coarser one with reduced uncertainty, serving decision making for natural resource management at larger scales. However, conventional model-based estimation (CMB) confronts a major challenge: it only supports RS wall-to-wall data, meaning that remotely sensed data must be available in panorama and non-wall-to-wall but quality data such as lidar or even cloud-masked satellite imagery are not supported due to incomplete coverage, impeding precise upscaling with cutting-edge instruments or for large scale applications. Consequently, this study aims to develop and demonstrate the use and usefulness of RS non-wall-to-wall data for upscaling with Hierarchical model-based estimation (HMB) which incorporates a two-stage model for bridging RS non– and wall-to-wall data; and for optimizing cost-efficiency, to evaluate the effects of non-wall-to-wall sample size on upscaling precision. Three main conclusions are relevant: (1) the HMB is a variant of the CMB estimator through trading in the uncertainty of the second-stage model to enable estimation using RS non-wall-to-wall data; (2) a quality first-stage model is key to exerting the advantage of HMB relative to the CMB estimator; (3) the variance of the HMB estimator is dominated by the first-stage model variance component, indicating that increasing the sample size in the first-stage is effective for increasing the overall precision. Overall, the HMB estimator balances tradeoffs between cost, efficiency and flexibility when devising a model-based upscaling in earth observation.

Deep learning (DL) models are gaining popularity in forest variable prediction using Earth observation (EO) images. However, in practical forest inventories, reference datasets are often represented by plot- or stand-level measurements, while high-quality representative wall-to-wall reference data for end-to-end training of DL models are rarely available. Transfer learning facilitates expansion of the use of deep learning models into areas with sub-optimal training data by allowing pretraining of the model in areas where high-quality teaching data are available. In this study, we perform a “model transfer” (or domain adaptation) of a pretrained DL model into a target area using plot-level measurements and compare performance versus other machine learning models. We use an earlier developed UNet based model (SeUNet) to demonstrate the approach on two distinct taiga sites with varying forest structure and composition. The examined SeUNet model uses multi-source EO data to predict forest height. Here, EO data are represented by a combination of Copernicus Sentinel-1 C-band SAR and Sentinel-2 multispectral images, ALOS-2 PALSAR-2 SAR mosaics and TanDEM-X bistatic interferometric radar data. The training study site is located in Finnish Lapland, while the target site is located in Southern Finland. By leveraging transfer learning, the SeUNet prediction achieved root mean squared error (RMSE) of 2.70 m and R2 of 0.882, considerably more accurate than traditional benchmark methods. We expect such forest-specific DL model transfer can be suitable also for other forest variables and other EO data sources that are sensitive to forest structure.

National forest inventories (NFI) provide essential forest-related biomass and carbon information for country greenhouse gas (GHG) accounting systems. Several tropical countries struggle to execute their NFIs while the extent to which space-based global information on aboveground biomass (AGB) can support national GHG accounting is under investigation. We assess whether the use of a global AGB map as auxiliary information produces a gain in precision of subnational AGB estimates for the Peruvian Amazonia. We used model-assisted estimators with data from the country’s NFI and explored hybrid inferential techniques to account for the sources of uncertainty associated with the integration of remote sensing-based products and NFI plot data.Our results show that the selected global biomass map tends to overestimate AGB values across the Peruvian Amazonia. For most strata, directly using the map in its published form did not reduce the precision of AGB estimates. However, after calibrating the map using the NFI data, the precision of our map-assisted AGB estimates increased by up to 50% at stratum level and 20% at Amazonia level. We further demonstrate how different sources of uncertainties can be incorporated in the map-NFI integrated estimates. With the hybrid inferential analysis, we found that the small spatial support of the NFI plots compared to the remote sensing-based sample units of aggregated pixels (within block variability) contributed the most to the total uncertainty associated with the AGB estimates from our map-NFI integration. Uncertainties caused by measurement variability and allometric model prediction uncertainty were the second largest contributors. When these uncertainties were incorporated, the increase in precision of our calibrated map-assisted AGB estimates was negligible, probably hindered by the great contribution of the within block variability to our map-plot assessment. We developed a reproducible method that countries can build upon and further improve while the global biomass products continue to evolve and better characterize the AGB distribution under large biomass conditions. We encourage further cross-country case studies that reflect a wider range of AGB distributions, especially within humid tropical forests, to further assess the contribution of global biomass maps to (sub)national AGB estimates and finally GHG monitoring and reporting.

Copernicus Sentinel-1 images are widely used for forest mapping and predicting forest growing stock volume (GSV) due to their accessibility. However, certain important aspects related to the use of Sentinel-1 time series have not been thoroughly explored in the literature. These include the impact of image time series length on prediction accuracy, the optimal feature selection approaches, and the best prediction methods. In this study, we conduct an in-depth exploration of the potential of long time series of Sentinel-1 SAR data to predict forest GSV and evaluate the temporal dynamics of the predictions using extensive reference data. Our boreal coniferous forests study site is located near the Hyytiälä forest station in central Finland and covers an area of 2500 km2 with nearly 17,000 stands. We considered several prediction approaches and fine-tuned them to predict GSV in various evaluation scenarios. Our analyses used 96 Sentinel-1 images acquired over three years. Different approaches for aggregating SAR images and choosing feature (predictor) variables were evaluated. Our results demonstrate a considerable decrease in the root mean squared errors (RMSEs) of GSV predictions as the number of images increases. While prediction accuracy using individual Sentinel-1 images varied from 85 to 91 m3/ha RMSE, prediction accuracy with combined images decreased to 75.6 m3/ha. Feature extraction and dimension reduction techniques facilitated the achievement of near-optimal prediction accuracy using only 8–10 images. Examined methods included radiometric contrast, mutual information, improved k-Nearest Neighbors, random forests selection, Lasso, and Wrapper approaches. Lasso was the most optimal, with RMSE reaching 77.1 m3/ha. Finally, we found that using assemblages of eight consecutive images resulted in the greatest accuracy in predicting GSV when initial acquisitions started between September and January.

Emerging satellite radar and lidar platforms are being developed to produce gridded aboveground biomass (AGB) predictions that are poised to expand our understanding of global carbon stocks and changes. However, the spatial resolution of AGB map products from these platforms is often larger than the available field plot data underpinning model calibration and validation efforts. Intermediate-resolution/extent remotely sensed data, like airborne lidar, can serve as a bridge between small plots and map resolution, but methods are needed to estimate and propagate uncertainties with multiple layers of data. Here, we introduce a workflow to estimate the pixel-level mean and variance in AGB maps by propagating uncertainty from a lidar-based model using small plots, taking into account prediction uncertainty, residual uncertainty, and residual spatial autocorrelation. We apply this workflow to estimate AGB uncertainty at a 100 m map resolution (1 ha pixels) using 0.04 ha field plots from 11 sites across four ecoregions. We compare uncertainty estimates using site-specific models, ecoregion-specific models, and a general model using all sites. The estimated AGB uncertainty for 1 ha pixels increased with mean AGB, reaching 7.8–33.3 Mg ha−1 for site-specific models (one standard deviation), 11.1–28.2 Mg ha−1 for ecoregion-specific models, and 21.1–22.1 Mg ha−1 for the general model for pixels in the AGB range of 80–100 Mg ha−1. Only 3 of 11 site-specific models had a total uncertainty of −1 in this biomass range, suitable for the calibration or validation of AGB map products. Using two additional sites with larger field plots, we show that lidar-based models calibrated with larger field plots can substantially reduce 1 ha pixel AGB uncertainty for the same range from 18.2 Mg ha−1 using 0.04 ha plots to 10.9 Mg ha−1 using 0.25 ha plots and 10.1 Mg ha−1 using 1 ha plots. We conclude that the estimated AGB uncertainty from models estimated from small field plots may be unacceptably large, and we recommend coordinated efforts to measure larger field plots as reference data for the calibration or validation of satellite-based map products at landscape scales (≥0.25 ha).

Abstract In 2019, 100 years had elapsed since the first National Forest Inventory (NFI) was established in Norway. Motivated by a fear of over-exploitation of timber resources, NFIs today enable informed policy making by providing data vital to decision support at international, national, regional, and local scales. This Collection of articles celebrates the 100th anniversary of NFIs with a description of past, present, and future research aiming at improving the monitoring of forest and other terrestrial ecosystems.

Abstract Background Large area forest inventories often use regular grids (with a single random start) of sample locations to ensure a uniform sampling intensity across the space of the surveyed populations. A design-unbiased estimator of variance does not exist for this design. Oftentimes, a quasi-default estimator applicable to simple random sampling (SRS) is used, even if it carries with it the likely risk of overestimating the variance by a practically important margin. To better exploit the precision of systematic sampling we assess the performance of five estimators of variance, including the quasi default. In this study, simulated systematic sampling was applied to artificial populations with contrasting covariance structures and with or without linear trends. We compared the results obtained with the SRS, Matérn’s, successive difference replication, Ripley’s, and D’Orazio’s variance estimators. Results The variances obtained with the four alternatives to the SRS estimator of variance were strongly correlated, and in all study settings consistently closer to the target design variance than the estimator for SRS. The latter always produced the greatest overestimation. In populations with a near zero spatial autocorrelation, all estimators, performed equally, and delivered estimates close to the actual design variance. Conclusion Without a linear trend, the SDR and DOR estimators were best with variance estimates more narrowly distributed around the benchmark; yet in terms of the least average absolute deviation, Matérn’s estimator held a narrow lead. With a strong or moderate linear trend, Matérn’s estimator is choice. In large populations, and a low sampling intensity, the performance of the investigated estimators becomes more similar.

To combat global deforestation, monitoring forest disturbances at sub-annual scales is a key challenge. For this purpose, the new Planetscope nano-satellite constellation is a game changer, with a revisit time of 1 day and a pixel size of 3-m. We present a near-real time forest disturbance alert system based on PlanetScope imagery: the Thresholding Rewards and Penances algorithm (TRP). It produces a new forest change map as soon as a new PlanetScope image is acquired. To calibrate and validate TRP, a reference set was constructed as a complete census of five randomly selected study areas in Tuscany, Italy. We processed 572 PlanetScope images acquired between 1 May 2018 and 5 July 2019. TRP was used to construct forest change maps during the study period for which the final user’s accuracy was 86% and the final producer’s accuracy was 92%. In addition, we estimated the forest change area using an unbiased stratified estimator that can be used with a small sample of reference data. The 95% confidence interval for the sample-based estimate of 56.89 ha included the census-based area estimate of 56.19 ha.

Afforestation is one of the most effective processes for removing carbon dioxide from the atmosphere and combating global warming. Landsat data and machine learning approaches can be used to map afforestation (i) indirectly, by constructing two maps of the same area over different periods and then predicting changes, or (ii) directly, by constructing a single map and analyzing observations of change in both the response and remotely sensed variables. Of crucial importance, no comprehensive comparisons of direct and indirect approaches for afforestation monitoring are known to have been conducted to date. Afforestation maps estimated through the analysis of remotely sensed data may serve as intermediate products for guiding the selection of samples and the production of statistics. In this and similar studies, a huge effort is dedicated to collecting validation data. In turn, those validation datasets have varying sampling intensities in different areas, which complicates their use for assessing the accuracies of new maps. As a result, the work done to collect data is often not sufficiently exploited, with some validation datasets being used just once. In this study, we addressed two main aims. First, we implemented a methodology to reuse validation data acquired via stratified sampling with strata constructed from remote sensing maps. Second, we used this method for acquiring data for comparing map accuracy estimates and the precision of estimates for direct and indirect approaches for country-wide mapping of afforestation that occurred in Italy between 1985 and 2019. To facilitate these comparisons, we used Landsat imagery, random forest classification, and Google Earth Engine. The herein-presented method produced different accuracy estimates with 95% confidence interval and for different map classes. Afforestation accuracies ranged between 53 ± 5.9% for the indirect map class inside the buffer—defined as a stratum within 120 m of the forest/non-forest mask boundaries—and 26 ± 3.4% for the direct map outside the buffer. The accuracy in non-afforestation map classes was much greater, ranging from 87 ± 1.9% for the indirect map inside the buffer to 99 ± 1.3% for the direct map outside the buffer. Additionally, overall accuracies (with 95% CI) were estimated with large precision for both direct and indirect maps (87 ± 1.3% and 89 ± 1.6%, respectively), confirming (i) the effectiveness of the method we introduced for reusing samples and (ii) the relevance of remotely sensed data and machine learning for monitoring afforestation.

Afforestation processes, natural and anthropogenic, involve the conversion of other land uses to forest, and they represent one of the most important land use transformations, influencing numerous ecosystem services. Although remotely sensed data are commonly used to monitor forest disturbance, only a few reported studies have used these data to monitor afforestation. The objectives of this study were two fold: (1) to develop and illustrate a method that exploits the 1985–2019 Landsat time series for predicting afforestation areas at 30 m resolution at the national scale, and (2) to estimate afforestation areas statistically rigorously within Italian administrative regions and land elevation classes. We used a Landsat best-available-pixel time series (1985–2019) to calculate a set of temporal predictors that, together with the random forests prediction technique, facilitated construction of a map of afforested areas in Italy. Then, the map was used to guide selection of an estimation sample dataset which, after a complex photointerpretation phase, was used to estimate afforestation areas and associated confidence intervals. The classification approach achieved an accuracy of 87%. At the national level, the afforestation area between 1985 and 2019 covered 2.8 ± 0.2 million ha, corresponding to a potential C-sequestration of 200 million t. The administrative region with the largest afforested area was Sardinia, with 260,670 ± 58,522 ha, while the smallest area of 28,644 ± 12,114 ha was in Valle d’Aosta. Considering elevation classes of 200 m, the greatest afforestation area was between 400 and 600 m above sea level, where it was 549,497 ± 84,979 ha. Our results help to understand the afforestation process in Italy between 1985 and 2019 in relation to geographical location and altitude, and they could be the basis of further studies on the species composition of afforestation areas and land management conditions.

Abstract The State Forest Inventory (SFI) in the Russian Federation is a relatively new project that is little known in the English-language scientific literature. Following the stipulations of the Forest Act of 2006, the first SFI sample plots in this vast territory were established in 2007. The 34 Russian forest regions were the basic geographical units for all statistical estimates and served as a first-level stratification, while a second level was based on old inventory data and remotely sensed data. The sampling design was to consist of a simple random sample of 84,700 circular 500 m2 sample plots over forest land. Each sample plot consists of three nested concentric circular subplots with radii of 12.62, 5.64 and 2.82 m and additional subplots for assessing and describing undergrowth, regeneration and ground vegetation. In total, 117 variables were to be measured or assessed on each plot. Although field work has begun, the methodology has elicited some criticism. The simple random sampling design is less efficient than a systematic design featuring sample plot clusters and a mix of temporary and permanent plots. The second-level stratification is mostly ineffective for increasing precision. Qualitative variables, which are not always essential, are dominant, while important quantitative variables are under-represented. Because of very slow progress, in 2018 the original plan was adjusted by reducing the number of permanent sample plots from 84,700 to 68,287 so that the first SFI cycle could be completed by 2020.

Field surveys are often a primary source of aboveground biomass (AGB) data, but plot-based estimates of parameters related to AGB are often not sufficiently precise, particularly not in tropical countries. Remotely sensed data may complement field data and thus help to increase the precision of estimates and circumvent some of the problems with missing sample observations in inaccessible areas. Here, we report the results of a study conducted in a 15,867 km² area in the dry miombo woodlands of Tanzania, to quantify the contribution of existing canopy height and biomass maps to improving the precision of canopy height and AGB estimates locally. A local and a global height map and three global biomass maps, and a probability sample of 513 inventory plots were subject to analysis. Model-assisted sampling estimators were used to estimate mean height and AGB across the study area using the original maps and then with the maps calibrated with local inventory plots. Large systematic map errors – positive or negative – were found for all the maps, with systematic errors as great as 60–70 %. After being calibrated locally, the maps contributed substantially to increasing the precision of both mean height and mean AGB estimates, with relative efficiencies (variance of the field-based estimates relative to the variance of the map-assisted estimates) of 1.3–2.7 for the overall estimates. The study, although focused on a relatively small area of dry tropical forests, illustrates the potential strengths and weaknesses of existing global forest height and biomass maps based on remotely sensed data and universal prediction models. Our results suggest that the use of regional or local inventory data for calibration can substantially increase the precision of map-based estimates and their applications in assessing forest carbon stocks for emission reduction programs and policy and financial decisions.

Spatial predictions of forest variables are required for supporting modern national and sub-national forest planning strategies, especially in the framework of a climate change scenario. Nowadays methods for constructing wall-to-wall maps and calculating small-area estimates of forest parameters are becoming essential components of most advanced National Forest Inventory (NFI) programs. Such methods are based on the assumption of a relationship between the forest variables and predictor variables that are available for the entire forest area. Many commonly used predictors are based on data obtained from active or passive remote sensing technologies. Italy has almost 40% of its land area covered by forests. Because of the great diversity of Italian forests with respect to composition, structure and management and underlying climatic, morphological and soil conditions, a relevant question is whether methods successfully used in less complex temperate and boreal forests may be applied successfully at country level in Italy.For a study area of more than 48,657 km2 in central Italy of which 43% is covered by forest, the study presents the results of a test regarding wall-to-wall, spatially explicit estimation of forest growing stock volume (GSV) based on field measurement of 1350 plots during the last Italian NFI. For the same area, we used potential predictor variables that are available across the whole of Italy: cloud-free mosaics of multispectral optical satellite imagery (Landsat 5 TM), microwave sensor data (JAXA PALSAR), a canopy height model (CHM) from satellite LiDAR, and auxiliary variables from climate, temperature and precipitation maps, soil maps, and a digital terrain model.Two non-parametric (random forests and k-NN) and two parametric (multiple linear regression and geographically weighted regression) prediction methods were tested to produce wall-to-wall map of growing stock volume at 23-m resolution. Pixel level predictions were used to produce small-area, province-level model-assisted estimates. The performances of all the methods were compared in terms of percent root mean-square error using a leave-one-out procedure and an independent dataset was used for validation. Results were comparable to those available for other ecological regions using similar predictors, but random forests produced the most accurate results with a pixel level R2 = 0.69 and RMSE% = 37.2% against the independent validation dataset. Model-assisted estimates were more precise than the original design-based estimates provided by the NFI.

Abstract Background Carbon accounting in forests remains a large area of uncertainty in the global carbon cycle. Forest aboveground biomass is therefore an attribute of great interest for the forest management community, but the accuracy of aboveground biomass maps depends on the accuracy of the underlying field estimates used to calibrate models. These field estimates depend on the application of allometric models, which often have unknown and unreported uncertainties outside of the size class or environment in which they were developed. Results Here, we test three popular allometric approaches to field biomass estimation, and explore the implications of allometric model selection for county-level biomass mapping in Sonoma County, California. We test three allometric models: Jenkins et al. (For Sci 49(1): 12–35, 2003), Chojnacky et al. (Forestry 87(1): 129–151, 2014) and the US Forest Service’s Component Ratio Method (CRM). We found that Jenkins and Chojnacky models perform comparably, but that at both a field plot level and a total county level there was a ~ 20% difference between these estimates and the CRM estimates. Further, we show that discrepancies are greater in high biomass areas with high canopy covers and relatively moderate heights (25–45 m). The CRM models, although on average ~ 20% lower than Jenkins and Chojnacky, produce higher estimates in the tallest forests samples (> 60 m), while Jenkins generally produces higher estimates of biomass in forests

Information about forest cover and its characteristics are essential in national and international forest inventories, monitoring programs, and reporting activities. Two of the most common forest variables needed to support sustainable forest management practices are forest cover area and growing stock volume (GSV m3 ha−1). Nowadays, national forest inventories (NFI) are complemented by wall-to-wall maps of forest variables which rely on models and auxiliary data. The spatially explicit prediction of GSV is useful for small-scale estimation by aggregating individual pixel predictions in a model-assisted framework. Spatial knowledge of the area of forest land is an essential prerequisite. This information is contained in a forest mask (FM). The number of FMs is increasing exponentially thanks to the wide availability of free auxiliary data, creating doubts about which is best-suited for specific purposes such as forest area and GSV estimation. We compared five FMs available for the entire area of Italy to examine their effects on the estimation of GSV and to clarify which product is best-suited for this purpose. The FMs considered were a mosaic of local forest maps produced by the Italian regional forest authorities; the FM produced from the Copernicus Land Monitoring System; the JAXA global FM; the hybrid global FM produced by Schepaschencko et al., and the FM estimated from the Corine Land Cover 2006. We used the five FMs to mask out non-forest pixels from a national wall-to-wall GSV map constructed using inventory and remotely sensed data. The accuracies of the FMs were first evaluated against an independent dataset of 1,202,818 NFI plots using four accuracy metrics. For each of the five masked GSV maps, the pixel-level predictions for the masked GSV map were used to calculate national and regional-level model-assisted estimates. The masked GSV maps were compared with respect to the coefficient of correlation (ρ) between the estimates of GSV they produced (both in terms of mean and total of GSV predictions within the national and regional boundaries) and the official NFI estimates. At the national and regional levels, the model-assisted GSV estimates based on the GSV map masked by the FM constructed as a mosaic of local forest maps were closest to the official NFI estimates with ρ = 0.986 and ρ = 0.972, for total and mean GSV, respectively. We found a negative correlation between the accuracies of the FMs and the differences between the model-assisted GSV estimates and the NFI estimate, demonstrating that the choice of the FM plays an important role in GSV estimation when using the model-assisted estimator.

The need for timely, spatially, and thematically accurate information regarding forests is increasing because of the key role of forests in the global carbon balance and sustainable social, economic, ecological, and cultural development [...]

A Landsat time series has been recognized as a viable source of information for monitoring and assessing forest disturbances and for continuous reporting on forest dynamics. This study focused on developing automated procedures for detecting disturbances in Mediterranean coppice forests which are characterized by rapid regrowth after a cut. Specifically, new methods specific to Mediterranean coppice forests are needed for mapping clearcut disturbances over time and for estimating related indicators in the context of Sustainable Forest Management and Biodiversity International monitoring frameworks. The aim of this work was to develop a new change detection algorithm for mapping clearcut disturbances in Mediterranean coppice forests with Landsat time series (LTS) using a short time window. Accuracy for the new algorithm, characterized as the Two Thresholds Method (TTM), was evaluated using an independent clearcut reference dataset over a temporal period of the 13 years between 2001 and 2013. TTM was also evaluated against two benchmark approaches: (i) LandTrendr, and (ii) the forest loss category of the Global Forest Change Map. Overall Accuracy for LandTrendr and TTM were greater than 0.94. Meanwhile, smaller accuracies were always obtained for the GFC. In particular, Producer’s Accuracy ranged between 0.45 and 0.84 for TTM and between 0.49 and 0.83 for LT, while for the GFC, PA ranged between 0 and 0.38. User’s Accuracy ranged between 0.86 and 0.96 for TTM and between 0.73 and 0.91 for LT, while for the GFC UA ranged between 0.19 and 1.00. Moreover, to illustrate the utility of TTM for mapping clearcut disturbances in Mediterranean coppice forests, we applied TTM to a Landsat scene that covered almost the entirety of the Tuscany region in Italy.

Forest/non-forest and forest species maps are often used by forest inventory programs in the forest estimation process. For example, some inventory programs establish field plots only on lands corresponding to the forest portion of a forest/non-forest map and use species-specific area estimates obtained from those maps to support the estimation of species-specific volume (V) totals. Despite the general use of these maps, the effects of their uncertainties are commonly ignored with the result that estimates might be unreliable. The goal of this study is to estimate the effects of the uncertainty of forest species maps used in the sampling and estimation processes. Random forest (RF) per-pixel predictions were used with model-based inference to estimate V per unit area for the six main forest species of La Rioja, Spain. RF models for predicting V were constructed using field plot information from the Spanish National Forest Inventory and airborne laser scanning data. To limit the prediction of V to pixels classified as one of the main forest species assessed, a forest species map was constructed using Landsat and auxiliary information. Bootstrapping techniques were implemented to estimate the total uncertainty of the V estimates and accommodated both the effects of uncertainty in the Landsat forest species map and the effects of plot-to-plot sampling variability on training data used to construct the RF V models. Standard errors of species-specific total V estimates increased from 2–9% to 3–22% when the effects of map uncertainty were incorporated into the uncertainty assessment. The workflow achieved satisfactory results and revealed that the effects of map uncertainty are not negligible, especially for open-grown and less frequently occurring forest species for which greater variability was evident in the mapping and estimation process. The effects of forest map uncertainty are greater for species-specific area estimation than for the selection of field plots used to calibrate the RF model. Additional research to generalize the conclusions beyond Mediterranean to other forest environments is recommended.

For tropical countries that do not have extensive ground sampling programs such as national forest inventories, the gain-loss approach for greenhouse gas inventories is often used. With the gain-loss approach, emissions and removals are estimated as the product of activity data defined as the areas of human-caused emissions and removals and emissions factors defined as the per unit area responses of carbon stocks for those activities. Remotely sensed imagery and remote sensing-based land use and land use change maps have emerged as crucial information sources for facilitating the statistically rigorous estimation of activity data. Similarly, remote sensing-based biomass maps have been used as sources of auxiliary data for enhancing estimates of emissions and removals factors and as sources of biomass data for remote and inaccessible regions. The current status of statistically rigorous methods for combining ground and remotely sensed data that comply with the good practice guidelines for greenhouse gas inventories of the Intergovernmental Panel on Climate Change is reviewed.

Information on land use and land cover (LULC) including forest cover is important for the development of strategies for land planning and management. Satellite remotely sensed data of varying resolutions have been an unmatched source of such information that can be used to produce estimates with a greater degree of confidence than traditional inventory estimates. However, use of these data has always been a challenge in tropical regions owing to the complexity of the biophysical environment, clouds, and haze, and atmospheric moisture content, all of which impede accurate LULC classification. We tested a parametric classifier (logistic regression) and three non-parametric machine learning classifiers (improved k-nearest neighbors, random forests, and support vector machine) for classification of multi-temporal Sentinel 2 satellite imagery into LULC categories in Dak Nong province, Vietnam. A total of 446 images, 235 from the year 2017 and 211 from the year 2018, were pre-processed to gain high quality images for mapping LULC in the 6516 km2 study area. The Sentinel 2 images were tested and classified separately for four temporal periods: (i) dry season, (ii) rainy season, (iii) the entirety of the year 2017, and (iv) the combination of dry and rainy seasons. Eleven different LULC classes were discriminated of which five were forest classes. For each combination of temporal image set and classifier, a confusion matrix was constructed using independent reference data and pixel classifications, and the area on the ground of each class was estimated. For overall temporal periods and classifiers, overall accuracy ranged from 63.9% to 80.3%, and the Kappa coefficient ranged from 0.611 to 0.813. Area estimates for individual classes ranged from 70 km2 (1% of the study area) to 2200 km2 (34% of the study area) with greater uncertainties for smaller classes.

Despite the popularity of random forests (RF) as a prediction algorithm, methods for constructing confidence intervals for population means using this technique are still only sparsely reported. For two regional study areas (Spain and Norway) RF was used to predict forest volume or aboveground biomass using remotely sensed auxiliary data obtained from multiple sensors. Additionally, the changes per unit area of these forest attributes were estimated using indirect and direct methods. Multiple inferential frameworks have attracted increased recent attention for estimating the variances required for confidence intervals. For this study, three different statistical frameworks, design-based expansion, model-assisted and model-based estimators, were used for estimating population parameters and their variances. Pairs and wild bootstrapping approaches at different levels were compared for estimating the variances of the model-based estimates of the population means, as well as for mapping the uncertainty of the change predictions. The RF models accurately represented the relationship between the response and remotely sensed predictor variables, resulting in increased precision for estimates of the population means relative to design-based expansion estimates. Standard errors based on pairs bootstrapping within or internal to RF were considerably larger than standard errors based on both pairs and wild external bootstrapping of the entire RF algorithm. Pairs and wild external bootstrapping produced similar standard errors, but wild bootstrapping better mimicked the original structure of the sample data and better preserved the ranges of the predictor variables.

The boreal tree line is in many places expected to advance upwards into the mountains due to climate change. This study aimed to develop a general method for estimation of vegetation height change in general, and change in tree height more specifically, for small geographical domains utilizing bi-temporal airborne laser scanner (ALS) data. The domains subject to estimation may subsequently be used to monitor vegetation and tree height change with detailed temporal and geographical resolutions. A method was developed with particular focus on statistically rigorous estimators of uncertainty for change estimates. The method employed model-dependent statistical inference. The method was demonstrated in a 12 ha study site in a boreal−alpine tree line in southeastern Norway, in which 316 trees were measured on the ground in 2006 and 2012 and ALS data were acquired in two temporally coincident campaigns. The trees ranged from 0.11 m to 5.20 m in height. Average growth in height was 0.19 m. Regression models were used to predict and estimate change. By following the area-based approach, predictions were produced for every individual 2 m2 population element that tessellated the study area. Two demonstrations of the method are provided in which separate height change estimates were calculated for domains of size 1.5 ha or greater. Differences in height change estimates among such small domains illustrate how change patterns may vary over the landscape. Model-dependent mean square error estimates for the height change estimators that accounted for (1) model parameter uncertainty, (2) residual variance, and (3) residual covariance are provided. Findings suggested that the two latter sources of uncertainty could be ignored in the uncertainty analysis. The proposed estimators are likely to work well for estimation of differences in height change along a gradient of small monitoring units, like the 1.5 ha cells used for demonstration purposes, and thus may potentially be used to monitor tree line migration over time.

Multisource forest inventory methods were developed to improve the precision of national forest inventory estimates. These methods rely on the combination of inventory data and auxiliary information correlated with forest attributes of interest. As these methods have been predominantly tested over coniferous forests, the present study used this approach for heterogeneous and complex deciduous forests in the center of France. The auxiliary data considered included a forest type map, Landsat 8 spectral bands and derived vegetation indexes, and 3D variables derived from photogrammetric canopy height models. On a subset area, changes in canopy height estimated from two successive photogrammetric models were also used. A model-assisted inference framework, using a k nearest-neighbors approach, was used to predict 11 field inventory variables simultaneously. The results showed that among the auxiliary variables tested, 3D metrics improved the precision of dendrometric estimates more than other auxiliary variables. Relative efficiencies (RE) varying from 2.15 for volume to 1.04 for stand density were obtained using all auxiliary variables. Canopy height changes also increased RE from 3% to 26%. Our results confirmed the importance of 3D metrics as auxiliary variables and demonstrated the value of canopy change variables for increasing the precision of estimates of forest structural attributes such as density and quadratic mean diameter.

This paper focuses on the use of models for increasing the precision of estimators in large-area forest surveys. It is motivated by the increasing availability of remotely sensed data, which facilitates the development of models predicting the variables of interest in forest surveys. We present, review and compare three different estimation frameworks where models play a core role: model-assisted, model-based, and hybrid estimation. The first two are well known, whereas the third has only recently been introduced in forest surveys. Hybrid inference mixes designbased and model-based inference, since it relies on a probability sample of auxiliary data and a model predicting the target variable from the auxiliary data..We review studies on large-area forest surveys based on model-assisted, modelbased, and hybrid estimation, and discuss advantages and disadvantages of the approaches. We conclude that no general recommendations can be made about whether model-assisted, model-based, or hybrid estimation should be preferred. The choice depends on the objective of the survey and the possibilities to acquire appropriate field and remotely sensed data. We also conclude that modelling approaches can only be successfully applied for estimating target variables such as growing stock volume or biomass, which are adequately related to commonly available remotely sensed data, and thus purely field based surveys remain important for several important forest parameters. Keywords: Design-based inference, Model-assisted estimation, Model-based inference, Hybrid inference, National forest inventory, Remote sensing, Sampling