Cultivated soils have the potential to store large amount of organic carbon and thus soils have the potential to mitigate increasing atmospheric CO2 concentration. To date, the knowledge on carbon stocks in the Brazilian soils and their dynamics and stability is limited. In this context, a soil carbon inventory is in development based on the near infrared (NIR) spectroscopy technique. Therefore, we used for the first time NIR spectroscopy and chemometrics methods to determine the total carbon content in soils samples from the Amazonian region and to constitute models of soil management characterization utilizing artificial neural networks. Two soil sample sets were collected in the topsoil (0-20 cm) under native forest and pasture and cultivated fields with tillage or no-tillage practices in Rondônia State. A total of 307 soil samples were collected in União Farm and 240 in Nova Vida Ranch. The study areas are representative of the land use and land cover changes in the Amazon. Before the NIR-analysis, samples were dried and ground at 100 mesh. Carbon contents were determined by the dry combustion method (CNS-2000, LECO Equipment, St Joseph, MI) utilized as reference. Soil C contents were in the range of 0.51 to 4.60 %. All samples were analyzed in the NIR by diffuse reflectance using a Foss NIRSystems 5000 spectrophotometer. Spectral data were collected from 1100 to 2498 nm at 2 nm resolution. Principal component analysis (PCA) was performed on all the samples. Scores were obtained for evaluation of population distribution and spectral outliers identification. The spectra with a Mahalanobis distance (GH) higher than 3 were eliminated. Then, each sample set was divided randomly into two groups, two-thirds of the samples for the calibration set and one-third for the validation set. The most appropriate spectrum pre-treatment was standard normal variate and detrend (SNVD) scatter correction using first derivative. Calibrations for total C obtained by using modified partial least squares regression, gave a low standard error of cross-validation (SECV) with a high R² and a RPD (SD/SECV) higher than 5.0 (Table 1). Validation process was done by a statistical comparison between laboratory reference values and the NIR predicted values, resulting a low standard error of prediction (SEP) with a high R², as well as a low bias (Table 1). The results demonstrate that the NIR spectroscopy methodology is able to determine total C in the Amazonian soils. Counter-propagation artificial neural networks provided the best results for the different soil management.