Deriving Forest Growth Variables with Lidar
Forest growth variables are important for forest and ecosystem monitoring. LiDAR provides fast and accurate access to deriving forest growth variables. The aim of the work is to explore the ability of LiDAR data to get the forest growth variables and further monitor the forest conditions in Estonia.
Part A: Forest Height as Direct Information
The part A used raw LiDAR data together with the Digital Terrain Model (DTM) and Digital Surface Model (DSM) derived from LiDAR from Estonian Land Board . The data cover the whole Estonia and were collected from 2017-2020. The ground truth data were selected within 2017-2020 among the pure stands.
DTM and DSM were used to calculate the canopy height model (CHM). CHM shows the distance from the ground to the top of the object for each pixel. Within a continuous area of forest, CHM shows how the top of the forest is above the ground, from which forest height can be estimated.
Then validation followed, as well as the results of the different tree species. All processes were conducted in Matlab.
The result denotes that the forest heights of different tree species have obvious differences. While the height of all species is more than 10m, the pine has the highest value of around 17m, followed by birch with around 15m. Fir, aspen and black alder are around 12m-14m. And the gray alder is the shortest species with a little above 10m.
Additionally, the estimated forest height is commonly lower than the true one. Aspen and gray alder have the least difference between the estimated and true data while pine, birch and fir have the most.
1) Black alder (highest R2 and ICC, and low MAE and RMSE): most consistent with true data
2) Pine (lower R2 but the lowest MAE and RMSE, and high ICC): consistency as second
3) Birch and fir (moderate R2, MAE and RMSE value and ICC>0.75 but <0.8): average consistency
4) Gray alder and aspen (low R2 <0.6, high MAE and RMSE, and low ICC <0.75): comparatively low consistency with true data
Part B: Basal Area, Wood Volume and Forest Height as Indirect Information
LiDAR data of 2021 from the Estonian Land Board for the pure stand of Aegna island north of Tallinn, Estonia were selected. Together related ground true data were obtained from Metsaportal.
The pre-processing was done by LiDAR 360 by extracting the first return, removing the noise points, separating the ground points and normalizing of the points based on the ground point.
From LiDAR point clouds, 33 forest metrics were extracted. I utilized the random forest to build the model for deriving forest variables. As for the validation, I calculated R2 and MAE between true value and estimated value.
The R2 for forest height, wood volume and basal area is all over 0.9 and MAE of it relatively low. This suggests the training process of random forest is successful. The model can explain the variables to some degree. For the test dataset, all variables have lower R2 than the train dataset but over 0.6, which means the models for all 3 variables are acceptable. Forest height has the highest R2 as 0.81, with MAE a bit higher than train dataset, which shows forest height estimation from LiDAR is relatively accurate. Wood volume one is second to the forest height one while basal area is least accurate.
Additionally, we could assess the importance of independent variables by calculating %IncMSE and IncNodePurity.
(1)Forest height: it is the high percentile of height (90th,95th,99th) that has the most influence. This proves that the high percentile of height can estimate forest height well.
(2) Wood volume: The mean height and middle percentile (30th, 40th) are most important. This shows that the middle part of the LiDAR points have more information about wood volume.
(3) Basal area: the low percentile (1st, 5th) of height is the most influential.