Machine learning modeling in temporal variability of soil respiration in planted forest areas

• Autor(a) principal: Vicentini, Maria Elisa [UNESP]
• Data de Publicação: 2021
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://hdl.handle.net/11449/215295
• topic: Carbon dynamics; Soil-atmosphere; Land use; Greenhouse gases; Mathematical models; Dinâmica do carbono; Atmosfera do solo; Uso do solo; Gases do efeito estufa; Modelos matemáticos

Understanding the temporal dynamics of land respiration in tropical ecosystems is challenging, especially when it is associated with Land Use, Land-Use Change and Forestry (LULUCF). Many studies have modeled the dynamics of CO2 emission from soil (FCO2), but few studies have modeled the temporal dynamics of soil O2 influx (FO2). Therefore, the objective of this study was to evaluate the predictive performance of artificial neural networks (ANN), support vector regression (SVR), adaptive neuro-fuzzy inference system (ANFIS), and random forest (RF) machine learning (ML) techniques in modeling the temporal variability of FCO2 and FO2 in forests planted in three ecosystems of planted forests: Pinus (Pinus spp), Eucalyptus (Eucalyptus spp), and native species, converted more than 30 years ago in the Cerrado biome, Brazil. We used a database composed of agro-meteorological data, improved vegetation index (EVI), and soil chemical and physical attributes as predictor variables and principal component analysis as the main data mining technique. For each monoculture and native species forest the numbers of FCO2 and FO2 recordings were (n = 500) and (n = 175), respectively. For pine forest, ANNs showed better predictive performance than SVR. The multilayer perceptron (MLPNN) with 12 input variables explained R2 = 42% of the temporal variability in FCO2. The general regression neural network (GRNN) with 10 input variables explained temporal variability in FO2 with an R2 of 56%. For eucalyptus, in the estimation of FCO2, the best predictive performance was obtained with MLP with validation (R² = 0.59; RMSE = 1.034 µmol m-2s-1). FO2 estimation: validation (R² = 0.36; RMSE = 0.076 mg m-2s-1). SVR with radial basis function kernel (SVR-RBF) was superior to the sigmoid (SVR-SIG), and polynomial kernels (SVR-PL), with the following values for FCO2; validation (R² = 0.53; RMSE = 0.990 µmol m-2s-1).In Native Species areas, the best results were: FCO2 with Radial Basis Function Neural Network (RBFNN) (R2 = 0.54, RMSE = 1.015 µmol m-2s-1) and FO2 with RBFNN (R2 = 0.74, 0.079 mg m-2s-1). Estimates of FCO2 showed better predictive performance than FO2. RBFNN was best estimate for FCO2. MLPNN is the best architecture for FO2 (R2 = 0.45, RMSE = 0.94 mg m-2s-1). In relation to ANFIS, FO2, did not show good generalizability and presented the worst performance, showing the highest mean absolute percentage error and lowest accuracy (R2 = 0.12, MAPE 51.27% and R2 = 0.28, MAPE 47.48%) calibration and validation respectively). Analyzing the performance of the two estimates, SVR-RBF for FCO2 performed better than SVR-RBF for FO2. The RF for FCO2 in the calibration and validation phases presented values of the (R2 = 0.80 and R2 = 0.60 respectively). The type of forest influenced temporal variability in soil respiration. We found that soil temperature (Ts) EVI, global solar radiation (GSR), macroporosity (macro), and organic matter (SOM) were the variables that most influenced the two estimates.