Application of artificial neural networks for predicting tree survival and mortality in the Hyrcanian forest of Iran

• Published in: Computers and electronics in agriculture 2019-09, Vol.164, p.104929, Article 104929
• Main Authors: Bayat, Mahmoud, Ghorbanpour, Mansour, Zare, Rozita, Jaafari, Abolfazl, Thai Pham, Binh
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Beech; Biotic and abiotic variables; Ecosystems; Forest management; Learning theory; Machine learning; Mortality; Neural networks; Oriental beech; Predictions; Relative humidity; Solar radiation; Statistical analysis; Survival; Survival probability; Trees; Wind speed
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2019.104929

•Comparing the predictive abilities of ANN and MLR for tree survival probability.•Incorporating a wide range of biotic and abiotic variables into the models.•Developing several ANN topologies with different combination of input variables.•Identifying the most and least effective biotic and abiotic variables.•Proving 92% accuracy for predicting survival probability using the ANN model. Tree survival and mortality play a critical role in forest ecosystems. Being able to predict the survival probability provides significant insights into forest management decision support. In this study, we employed an artificial neural network (ANN) to predict tree survival and mortality in an Oriental beech (Fagus orientalis Lipsky) forest in northern Iran, using long-term data and wide range of biotic (average tree diameter increment, and basal area in larger trees (BAL)) and abiotic variables (slope, average solar radiation during the growing season, wind velocity and direction, height above the nearest drainage point, topographic wetness index, temperature, and relative humidity). The ANN model revealed that the greatest survival rates belong to the trees with diameter >20–100 cm. An increase in competition leads to reduced survival, and the Carpinus betulus species showed a slightly lower survival rate than the other species. Further, our findings demonstrated that BAL and diameter increment have a strong correlation with the probability of tree survival in the study area. The ANN performance was evaluated using the R2 and RMSE metrics and compared to a multiple logistic regression that demonstrated the capability of the ANN model for the prediction of survival probability (R2 = 0.92; RMSE = 0.985). From these results, we recommend the employment of ANN, and perhaps other machine learning methods, for the prediction of survival probability in many different ecosystems around the world.