United States Department of Agriculture Forest Service, Southern Research Station Diameter Distributions for Young Longleaf Pine Plantations: Initial Conditions for a Growth and Yield Model D.J. Leduc, Information Technology Specialist, & J.C.G. Goelz, Principal Forest Biometrician
United States Department of Agriculture Forest Service, Southern Research Station Why Longleaf?
United States Department of Agriculture Forest Service, Southern Research Station
United States Department of Agriculture Forest Service, Southern Research Station Why is this a problem?
United States Department of Agriculture Forest Service, Southern Research Station Mature longleaf pine stands can be unimodal, ….
United States Department of Agriculture Forest Service, Southern Research Station … but they can also be bi- or tri- modal.
United States Department of Agriculture Forest Service, Southern Research Station There are two known causes of this.
United States Department of Agriculture Forest Service, Southern Research Station Suppressed longleaf trees do not die easily.
United States Department of Agriculture Forest Service, Southern Research Station Not all trees exit the grass stage at the same time
United States Department of Agriculture Forest Service, Southern Research Station To produce the irregular diameter distributions observed in older stands, it is essential that initial diameter distributions be irregular.
United States Department of Agriculture Forest Service, Southern Research Station Techniques Weibull distribution by parameter recovery Artificial neural networks Model cohorts of trees beginning height growth
United States Department of Agriculture Forest Service, Southern Research Station What we have to work with Age Basal area Site index Container stock or not Trees planted per acre Number of trees in 0-inch diameter class?
United States Department of Agriculture Forest Service, Southern Research Station Weibull distribution Included as baseline parametric technique P i =0.97 and P j =0.17 as suggested by Zanakis (1979) Only used for trees with dbh > 0
United States Department of Agriculture Forest Service, Southern Research Station Artificial Neural Network
United States Department of Agriculture Forest Service, Southern Research Station Artificial Neural Network Number of grass stage trees is known Predict proportion of dbh 0 trees Do not predict proportion of dbh 0 trees Number of grass stage trees is unknown Predict proportion of dbh 0 trees Do not predict proportion of dbh 0 trees
United States Department of Agriculture Forest Service, Southern Research Station Predicting number of dbh 0 trees Necessary for Weibull distribution that we used and two neural network models. Used standard logistic model and an evolutionary algorithm
United States Department of Agriculture Forest Service, Southern Research Station Logistic model logit = age * ( ) + baa * ( ) + si * container * ( ) + tpa * age*baa * age*si * age*container * baa*si * ( ) + baa*container * baa*tpa * si*container * ( ) + si*tpa * container*tpa * predp =exp(logit)/(1+exp(logit)) pdc00 =predp*tsa
United States Department of Agriculture Forest Service, Southern Research Station Evolutionary algorithm tmp= (container*11.24+baa)/9.84 tmp2 =5.08*age+ ((tmp+tsa)/tmp) container pdc00=((( tmp2)*tmp2)/( *si)*baa+tmp2)*age/( )+tmp2
United States Department of Agriculture Forest Service, Southern Research Station Crossover (sexual recombination) X reproduction Inversion Mutation Hill climbing Migration and intermarriage Evolutionary algorithm
United States Department of Agriculture Forest Service, Southern Research Station Explicitly Modeling Cohorts Seedlings exit the grass stage over several years. This is one of the main factors causing diameter distributions to be irregular. Model diameter distribution as a mixture of distributions for each cohort. As there are potentially several cohorts, it seems wise to use a very simple distribution.
United States Department of Agriculture Forest Service, Southern Research Station Epanechnikov Kernal K i (u) = 0.75 (1-u 2 ) For (X min -.05)<X<(X max +.05) Complete distribution is: Where p i is proportion of stand in cohort i.
United States Department of Agriculture Forest Service, Southern Research Station Using mixture of Epanechnikov-kernals in prediction Predict the proportion of trees in each cohort. User-supplied input regarding length of time in grass stage (average length, or years for 75% to leave grass stage…). Select “Guiding” D max (or D min ). “oldest” cohort or most populous. Develop equations to predict other D max and D min ’s from guiding value, and stand variables (age,site index, etc). Recover guiding D max from predicted basal area and trees/acre.
United States Department of Agriculture Forest Service, Southern Research Station CohortPiPi D max D min 1p1p1 GF(G,SI,TPA…) 2p2p2 3p3p3 4p4p4
United States Department of Agriculture Forest Service, Southern Research Station
United States Department of Agriculture Forest Service, Southern Research Station
United States Department of Agriculture Forest Service, Southern Research Station Preliminary Results
United States Department of Agriculture Forest Service, Southern Research Station Predicting the number of trees in the grass stage CriterionLogistic Function Evolutionary Algorithm Bias Largest Deviation Mean Absolute Deviation Root Mean Squared Error
United States Department of Agriculture Forest Service, Southern Research Station Criterion Neural Networks Weibull Predict 0Don’t predict 0 Known 0Unknown 0 Known 0Unknown 0 MSE FI 2 lowest KS lowest Mean closest
United States Department of Agriculture Forest Service, Southern Research Station All methods work
United States Department of Agriculture Forest Service, Southern Research Station Weibull works best
United States Department of Agriculture Forest Service, Southern Research Station Neural net works best
United States Department of Agriculture Forest Service, Southern Research Station Conclusions Evolutionary algorithm better than logistic function for predicting trees in grass stage. Neural networks show promise for modeling young stand diameter distributions. Modeling cohorts looks promising, but remains untested The biggest problem is finding enough easily measured variables to base predictions on