Dynamic ensemble approach for estimating organic carbon using computational intelligence

Organic Carbon (OC) is a soil property that describes the amount of carbon broken down from plants or other life forms stored in the soil below. Traditionally to quantify organic carbon content, a soil sample is collected and analyzed in a laboratory. This is a time-consuming and expensive process. Recently there has been interest in finding cost-efficient solutions to predicting organic carbon using remotely sensed data. Australia-wide readings of organic carbon and associated input factors were used in the study. The input factors included climate, landform, Landsat Multi-Spectral Scanner bands, lithology and land use. Experiments have been undertaken using Artificial Neural Networks (ANNs), Multiple Linear Regression (MLR) and ensembles of ANNs. A dynamic ensemble combination model is presented. The ensemble assesses the performance of the ensemble members on individual cases and weights each ensemble member relative to their performance on similar cases. The weighting scheme is dynamic, changing for each case in the testing set. The ensemble output is derived from the combination of weighted ensemble members. The results encourage further research into estimating organic carbon using ensembles. 505-052 in display pdf