Modelos preditivos de biomassa em Floresta Amazônica a partir de dados LiDAR

Abstract

LiDAR (Light Detection and Ranging) remote sensing data combined with machine learning techniques has presented great potential for modeling large-scale forest atributes. In this context, the work aims to evaluate the application of machine learning techniques in the construction of models that relate LiDAR metrics and forest inventory data in the prediction of biomass in tropical forest. Initially, above-ground biomass was computed by an adjusted allometric equation, using biometric variables inventoried in 85 sample units at Fazenda Cauaxi, in the municipality of Paragominas / PA. The biomass of the plots (variable of interest) was related to 87 LiDAR metrics (explanatory variables), obtained by processing the LiDAR points clouds. This database was randomly divided into 70% for model adjustment and 30% for validation. The predictive performance of three different machine learning techniques (Random Forest - RF, Support Vector Machine - SVM and Artificial Neural Networks - ANN) was compared to a Generalized Linear Model (GLM) technique, traditionally used in nonparametric estimations. The results indicated that the information derived from the airborne LiDAR survey proved to be efficient and perfectly applicable to the modeling process of biomass in a tropical environment. With the exception of the RF model, with R² of 0.60, the machine learning models obtained better performance in the training stage. The value of 0.99 for the R² and the superior performance in the other adjustment quality indicators (RMSE, Syx, BIAS and DM), gave the ANN model the condition of better adaptation to the training data. In the validation stage, the GLM and RF models that presented the worst indicators in relation to the adjustment, showed superior performance, while the ANN estimates showed the greatest distortion. In general, the Spearman correlation between the estimated and observed values presented a behavior inversely proportional to the degree of adjustment of the models in the training stage, varying from 0.57 to 0.87 for the ANN and GLM models respectively. In spite of the lower adjustment of the RF model and the lower generalization capacity of the ANN and SVM models, the Wilcoxon Rank Sum Test did not detect a significant difference between the biomass values observed and predicted by the different models. In this way, it was possible to observe that the machine learning algorithms were able to detect and reproduce well the nonparametric data structure and to cope with generalized regression, without the need for data dimensionality reduction techniques, which gave more agility to the modeling process.