Assessing soil carbon using laboratory and field sensors in a Typic Hapludox of Colombia

The soil is a large and dynamic reservoir of carbon. Changes in the soil carbon (SC) concentration influence the quality of the soil and the increase of greenhouse gases in the atmosphere. The intensification of agricultural production and its expansion to new areas, for example, in the Eastern Plains of Colombia, followed by the growing environmental concern require monitoring of natural resources and implementation of new techniques in agriculture to better use resources and optimize production processes, such as precision farming. The traditional methods for determining SC are expensive and generate residues with high environmental impact. Using diffuse reflectance spectroscopy in NIR region is an accurate, economical, and environmentally friendly nondestructive technique for SC estimation. The objective of this work was to calibrate the model for the prediction of SC by NIR spectroscopy comparing the spectral responses obtained with laboratory spectrometer versus field spectrometer.
The study area corresponds to Carimagua Experimental Station - Colombian Corporation for Agricultural Research (CORPOICA) in the municipality of Puerto Gaitán, Meta, Colombia (04° 37 'N and 71 ° 19' W). Based on the Koppen climate classification study area it corresponds to a rainy tropical savanna climate Awf, with average temperature of 27,8ºC and annual rainfall of 2240 mm. The predominant soils are Typic Hapludox strongly acidic (pH <5), slightly undulating relief with slopes between 2 and 5%. A sampling of rigid network covering an area of 5100 ha, taking soil samples of surface and subsurface horizon was raised, taking 850 soil samples. The spectral curves were acquired through two NIRS sernsors: One of laboratory and another to field work, which spectral resolution of 2 nm, from 1000 and 2500 nm region, and 64 scans per wavelength. For the development and calibration of the models 13 subgroups were made, where the first group was formed of 50 samples, 50 samples increased for each subgroup, for a total of 650 samples for the last group. A group of 150 remaining samples was selected to validate the models. The SC was determined by elemental analyzer, spectral responses were analyzed with the ParleS software. Prediction models were calibrated by the method of PLSR, statistical indicators used to determine the prediction model were R2, RMSE and the residual prediction deviation (RPD). The prediction model generated from the laboratory measurements presented better fit, reaching values of R2 (0.949), RMSE (0.123) and RPD (4.41), compared with the measurements made with the field sensor, who presented an R2 of 0.869, RPD of 0.189 and RMSE 2.54, when the model with all the samples was calibrated. While a better model with spectral responses obtained in the laboratory, the model with the spectral responses of field sensor obtained also provided good results. This do make necessary evaluate the costs and efficiency of the acquisition of soil curve spectral in field or laboratory. It is noteworthy that with a low samples number (<50) can also get good models with RPD values higher than 2.0 and R2 above 0.75, regardless of the sensor that is used.