Epsilon Archive for Student Projects

Abstract

Forests are estimated to sequester and emit respectively 15% and 20% of the CO2
emissions. REDD+ aims at establishing a financial framework to compensate developing
countries for reducing Green House Gasses emissions due to decreased deforestation and
land degradation. An accurate Monitoring, Reporting and Verification (MRV) of the forest
carbon pools is needed. The adoption of State-Of-The-Art remote sensing technologies,
such as Lidar in combination with participatory approaches can potentially produce an
accurate assessment of the forest resources, ensuring the sustainability of the process.
The study aims at defining the feasibility of Lidar assisted Above Ground Biomass (AGB)
assessment with a participatory approach. The study compares AGB regression models
built with wall-to-wall, low density (0.8 points m-2) laser scanning data and two field
datasets collected by professionals and Community Forest User Groups (CFUGs) teams.
The models were built using ArboLiDAR©, a tool-box developed in ESRI environment by
Arbonaut Oy, that uses a Sparse Bayesian approach to define a set of weights for each
independent variable based on the variance of the field measured AGB and the Lidar
metrics. Finally the models were validated with Leave-One-Out Cross Validation
(LOOCV). The adjusted R2, relative RMSE and BIAS as well as the analyses of the
residuals were used to compare the models. In addition the study also analyzed the
reliability of the models across different forest structures. The professional model
described a greater part of the variability of the AGB (adj.R2=0.75) compared the CFUG
model (adj. R2=0.55), moreover the first was slightly more accurate (professional: rel.
RMSE= 45.6 %; CFUG: rel. RMSE= 47.2 %). Although both of the models proved to have
the mean of the error term not equal to zero and did not follow a normal distribution, the
CUFG model showed heteroschedastic residuals. The accuracy improved when applying
the models to forests characterized by a more uniform height distribution (rel. RMSE=
32.1 – 45.2 %), whereas it drastically decreased for sparse forests (rel. RMSE= 91.4 -130.5
%). The study concludes that with the limitation of having different sampling designs and
measuring techniques the CFUGs models were slightly worst than the professional ones.
However, it is likely that with a more accurate retrieval of the GPS plot center and increase
of plot size the results can be as good as the ones obtained with professionally collected
data.