Comparison of the Performance of Artificial Neural Networks with Regression Analysis in Estimating the Density of Forest Stands in Saravan, Guilan Province

Abstract

Tree density as one of the most importantfeatures of forest structuralis necessary for management, conservation and reforestation of northern Iran forest. In this research, tree densities were estimated using physiographic, soil and human factors using artificial neural network supervised self-organized, multi-layer perceptron and multiple linear regression model and compared according to their performance evaluation criteria. For this purpose, homogeneous units in GIS environment were prepared. Sampling was performed by random-systematic method with 150 200 m network dimensions and a total of 779 0.1 ha circular shape plots were implemented. By measuring the diameter at breast height of all trees above 7.5 cm, tree density was calculated for each sample plot and homogeneous units. The results showed that SSOM 5 neural network (R2 = 0.9117, R2adj = 0.9909, RMSE% = 9.16, Bias% = 4.26) compared to MLP 4 neural network (R2 = 0.8321, R2adj = 0.8760, RMSE% = 15.14, Bias% = 10.96) and multiple linear regression model (R2 = 0.6812, R2adj = 0.6910, RMSE% = 28.71, Bias% = 24.26) had more accuracy and less error. To select the best model, T- test was performed and the results showed that the neural network, of the competitive and supervisory type, had values similar to the actual values. This is due to Gaussian functions, which are not seen in MLP neural networks with sigmoid functions. Therefore, SSOM neural network can be a suitable alternative to multilayer perceptron neural network in estimating the density of trees in the northern forests of Iran.

Keywords

• multi-layer perceptron
• number of trees per hectare
• self - organizing map
• supervised learning
• winning neuron.