Predição do estoque e dinâmica da biomassa acima do solo na floresta Amazônica utilizando inteligência artificial e dados de sensores remotos

Abstract

The Amazon forest is characterized by expressive biomass and, therefore, stores high amounts of carbon, which is an important variable for climate monitoring. So, it is crucial to develop an efficient method to estimate biomass accurately, especially in tropical forests, where dense vegetation makes modeling difficult. Thus, the objective of the present study was to estimate the aboveground biomass, at plot and landscape level, in areas of the Amazon forest with selective logging, using machine learning algorithms and data from the LiDAR and OLI/Landsat 8 sensors, and map the biomass for the years 2014 and 2017 allowing to analyze its dynamics during the analysis period. For that, 79 plots of 50x50 m were used, at Fazenda Cauaxi, in the municipality of Paragominas, Pará. The prediction of biomass was performed using the machine learning algorithms Random Forest (RF), Support Vector Machine (SVM) and Artificial Neural Network (ANN) using LiDAR data and these combined with variables (spectral and texture) from the OLI/Landsat 8. To verify the best performance algorithm, the RMSE (Root Mean Square Error) was used, for this, 30 repetitions were performed with data separation in 80% for training and 20% for validation. The RF and SVM algorithms obtained the lowest average RMSE values in all data sets, with emphasis on the models using LiDAR and the combination of this with the spectral variables. The association of these variables allowed to increase the performance of the RF decreasing the RMSE from 48.01 Mg/ha to 45.24 Mg/ha. However, SVM using only LiDAR had the lowest average RMSE (44.99 Mg/ha). Thus, it was selected to map the biomass of 2014 and 2017 for analysis of the temporal dynamics. The older exploration units (2006, 2007 and 2008) had lower biomass stocks, in the years under analysis, however, the largest biomass losses in 2017 were obtained in the most recent exploration units (2012 and 2013). Thus, with the method employed in the present study, it was possible to infer that the machine learning algorithms proved to be efficient for estimating biomass, with emphasis on RF and SVM. Thus, the study may serve as a basis for improvements in biomass predictions in areas of Amazon forest with selective extraction, where the amplitudes of biomass and dense vegetation make it difficult to model attributes.