Virpi Junttila

Algorithm for Extracting Digital Terrain Models under Forest Canopy from Airborne LiDAR Data

Extracting digital elevationmodels (DTMs) from LiDAR data under forest canopy is a challenging task. This is because the forest canopy tends to block a portion of the LiDAR pulses from reaching the ground, hence introducing gaps in the data. This paper presents an algorithm for DTM extraction from LiDAR data under forest canopy. The algorithm copes with the challenge of low data density by generating a series of coarse DTMs by using the few ground points available and using trend surfaces to interpolate missing elevation values in the vicinity of the available points. This process generates a cloud of ground points from which the final DTM is generated. The algorithm has been compared to two other algorithms proposed in the literature in three different test sites with varying degrees of difficulty. Results show that the algorithm presented in this paper is more tolerant to low data density compared to the other two algorithms. The results further show that with decreasing point density, the differences between the three algorithms dramatically increased from about 0.5m to over 10m.

Linear Models for Airborne-Laser-Scanning-Based Operational Forest Inventory With Small Field Sample Size and Highly Correlated LiDAR Data

Modern operational forest inventory often uses remotely sensed data that cover the whole inventory area to produce spatially explicit estimates of forest properties through statistical models. The data obtained by airborne light detection and ranging (LiDAR) correlate well with many forest inventory variables, such as the tree height, the timber volume, and the biomass. To construct an accurate model over thousands of hectares, LiDAR data must be supplemented with several hundred field sample measurements of forest inventory variables. This can be costly and time consuming. Different LiDAR-data-based and spatial-data-based sampling designs can reduce the number of field sample plots needed. However, problems arising from the features of the LiDAR data, such as a large number of predictors compared with the sample size (overfitting) or a strong correlation among predictors (multicollinearity), may decrease the accuracy and precision of the estimates and predictions. To overcome these problems, a Bayesian linear model with the singular value decomposition of predictors, combined with regularization, is proposed. The model performance in predicting different forest inventory variables is verified in ten inventory areas from two continents, where the number of field sample plots is reduced using different sampling designs. The results show that, with an appropriate field plot selection strategy and the proposed linear model, the total relative error of the predicted forest inventory variables is only 5%-15% larger using 50 field sample plots than the error of a linear model estimated with several hundred field sample plots when we sum up the error due to both the model noise variance and the models lack of fit.

Moving Voxel Method for Estimating Canopy Base Height from Airborne Laser Scanner Data

Canopy base height (CBH) is a key parameter used in forest-fire modeling, particularly crown fires. However, estimating CBH is a challenging task, because normally, it is difficult to measure it in the field. This has led to the use of simple estimators (e.g., the average of individual trees in a plot) for modeling CBH. In this paper, we propose a method for estimating CBH from airborne light detection and ranging (LiDAR) data. We also compare the performance of several estimators (Lorey’s mean, the arithmetic mean and the 40th and 50th percentiles) used to estimate CBH at the plot level. The method we propose uses a moving voxel to estimate the height of the gaps (in the LiDAR point cloud) below tree crowns and uses this information for modeling CBH. The advantage of this approach is that it is more tolerant to variations in LiDAR data (e.g., due to season) and tree species, because it works directly with the height information in the data. Our approach gave better results when compared to standard percentile-based LiDAR metrics commonly used in modeling CBH. Using Lorey’s mean, the arithmetic mean and the 40th and 50th percentiles as CBH estimators at the plot level, the highest and lowest values for root mean square error (RMSE) and root mean square error for cross-validation (RMSEcv) and R2 for our method were 1.74/2.40, 2.69/3.90 and 0.46/0.71, respectively, while with traditional LiDAR-based metrics, the results were 1.92/2.48, 3.34/5.51 and 0.44/0.65. Moreover, the use of Lorey’s mean as a CBH estimator at the plot level resulted in models with better predictive value based on the leave-one-out cross-validation (LOOCV) results used to compute the RMSEcv values.

LiDAR-Assisted Multi-Source Program (LAMP) for Measuring Above Ground Biomass and Forest Carbon

Forest measurement for purposes like harvesting planning, biomass estimation and mitigating climate change through carbon capture by forests call for increasingly frequent forest measurement campaigns that need to balance cost with accuracy and precision. Often this implies the use of remote sensing based measurement methods. For any remote-sensing based methods to be accurate, they must be validated against field data. We present a method that combines field measurements with two layers of remote sensing data: sampling of forests by airborne laser scanning (LiDAR) and Landsat imagery. The Bayesian model-based framework presented here is called Lidar-Assisted Multi-source Programme—or LAMP—for Above Ground Biomass estimation. The method has two variants: LAMP2 which splits the biomass estimation task into two separate stages: forest type stratification from Landsat imagery and mean biomass density estimation of each forest type by LiDAR models calibrated on field plots. LAMP3, on the other hand, estimates first the biomass on a LiDAR sample using models calibrated with field plots and then uses these LiDAR-based models to generate biomass density estimates on thousands of surrogate plots, with which a satellite image based model is calibrated and subsequently used to estimate biomass density on the entire forest area. Both LAMP methods have been applied to a 2 million hectare area in Southern Nepal, the Terai Arc Landscape or TAL to calculate the emission Reference Levels (RLs) that are required for the UN REDD+ program that was accepted as part of the Paris Climate Agreement. The uncertainty of these estimates is studied with error variance estimation, cross-validation and Monte Carlo simulation. The relative accuracy of activity data at pixel level was found to be 14 per cent at 95 per cent confidence level and the root mean squared error of biomass estimates to be between 35 and 39 per cent at 1 ha resolution.

Robustness of model-based high-resolution prediction of forest biomass against different field plot designs

BackgroundParticipatory forest monitoring has been promoted as a means to engage local forest-dependent communities in concrete climate mitigation activities as it brings a sense of ownership to the communities and hence increases the likelihood of success of forest preservation measures. However, sceptics of this approach argue that local community forest members will not easily attain the level of technical proficiency that accurate monitoring needs. Thus it is interesting to establish if local communities can attain such a level of technical proficiency. This paper addresses this issue by assessing the robustness of biomass estimation models based on air-borne laser data using models calibrated with two different field sample designs namely, field data gathered by professional forester teams and field data collected by local communities trained by professional foresters in two study sites in Nepal. The aim is to find if the two field sample data sets can give similar results (LiDAR models) and whether the data can be combined and used together in estimating biomass.Results Results show that even though the sampling designs and principles of both field campaigns were different, they produced equivalent regression models based on LiDAR data. This was successful in one of the sites (Gorkha). At the other site (Chitwan), however, major discrepancies remained in model-based estimates that used different field sample data sets. This discrepancy can be attributed to the complex terrain and dense forest in the site which makes it difficult to obtain an accurate digital elevation model (DTM) from LiDAR data, and neither set of data produced satisfactory results.Conclusions Field sample data produced by professional foresters and field sample data produced by professionally trained communities can be used together without affecting prediction performance provided that the correlation between LiDAR predictors and biomass estimates is good enough.

Bayesian principal component regression model with spatial effects for forest inventory variables under small field sample size

Abstract Remote sensing observations are extensively used for analysis of environmental variables. These variables often exhibit spatial correlation, which has to be accounted for in the calibration models used in predictions, either by direct modelling of the dependencies or by allowing for spatially correlated stochastic effects. Another feature in many remote sensing instruments is that the derived predictor variables are highly correlated, which can lead to unnecessary model over-training and at worst, singularities in the estimates. Both of these affect the prediction accuracy, especially when the training set for model calibration is small. To overcome these modelling challenges, we present a general model calibration procedure for remotely sensed data and apply it to airborne laser scanning data for forest inventory. We use a linear regression model that accounts for multicollinearity in the predictors by principal components and Bayesian regularization. It has a spatial random effect component for the spatial correlations that are not explained by a simple linear model. An efficient Markov chain Monte Carlo sampling scheme is used to account for the uncertainty in all the model parameters. We tested the proposed model against several alternatives and it outperformed the other linear calibration models, especially when there were spatial effects, multicollinearity and the training set size was small.