Algoritmos evolucionários na predição de estoque de carbono acima do solo em florestas de Mopane - Moçambique

Abstract

Tropical forests play an important role in the global climate regulation and the carbon cycle. Mopane is a tropical dry forest, occurring in southern Africa, with socioeconomic importance at the local level. Mopane harvesting for charcoal production in Mozambique is a main driver for forest degradation and carbon stocks reduction. Estimating carbon stocks in this forests can help monitoring carbon emissions, in this type of forest can help to assess and monitoring CO2 emissions, in the context of climate change, including Reduction of Emissions from Deforestation and Forest Degradation (REDD+). In this study, we tested Machine Learning methods by applying evolutionary algorithms and remote sensing, forest cover data, biophysical and bioclimatic data to predtic Aboveground Carbon (AGC) in the Mopane forest, in the districts of Mabalane and Chicualacuala, Gaza province, Mozambique. The sample was composed of 114 clusters and we used satellites images from Sentinel-2, Sentinel-1, MODIS and World.Clim dataset to extract the predictor variables. A set of 139 variables of different nature has been tested to predict the AGC, using (i) the hybrid method between Genetic Algorithm-AG for variable selection and Random Forest - RF for prediction (GARF) and (ii) Genetic Programming (PG) via symbolic regression. Both methods were able to reduce the database size by 95.6%. The GARF adhered more to bioclimatic variables and optical sensors, while the PG combined variables regardless of their nature and can generate mixed and segmented models. The AGC values (in MgC.ha-1) from field survey ranged from 1.313 to 28.476, mean = 10.988. The AGC estimated by GARF ranged from 2.910 to 19.459, mean = 10.235, normalized root mean square error – nRMSE = 0.427 and mean bias error - BEM = 0.08. For PG it ranged from 1.721 to 23.503, nRMSE = 0.428 and BEM = 2.731×10-17. Both methods showed efficiency for variables selection and potential for predicting AGC in tropical dry forests. The PG algorithm is more practical than GARF, as it provides a model with a visible and easily replicable structure.