Effect of the number of calibration samples on the prediction of several soil properties at the farm-scale

Abstract Precision agriculture (PA) is a management method that measures and manages within-field variability. Previously, PA has required expansive and time consuming measurement of soil physical and chemical properties. In this paper we use a new and more rapid method of data collection based on Visible and Near-Infrared Spectroscopy (VIS–NIRS) in the 400–2200 nm spectral range to predict soil organic carbon (SOC), plant available [Mg, P, K], pH and texture at the farm scale. The experimental work was done at the experimental Station at Baborówko (52.583778°N, 16.647353°E) in Poland. The focus of the paper was to look at the effect of the number of samples on the calibration. Different calibration schemes using PLS regression with calibration datasets of different sizes were applied. The best predictions were obtained using K-means clustering for calibration sample selection. Using this scheme and 79 calibration samples, satisfactory results were obtained predicting SOC (r2 = 0.63; RMSEP = 0.13%) and soil texture (e.g. clay, r2 = 0.71; RMSEP = 0.36%). The use of the entire dataset did not improve significantly the prediction ability (r2 = 0.72; RMSEP = 0.12% for SOC and r2 = 0.73; RMSEP = 0.32% for clay). Reasonable results were obtained for available Mg content (r2 = 0.53; RMSEP = 1.54 mg.100 g− 1) and pH (r2 = 0.52; RMSEP = 0.34 pH unit). Available [P, K] gave unsatisfactory results (r2 < 0.5 for both; RMSEP 6.27 and 3.31 mg.100 g− 1 respectively). The maps (SOC and pH) generated with the K-means clustering scheme were compared with those obtained with reference data. The results show that the VIS–NIRS method is suitable to adequately predict SOC and texture using 1.5 samples per ha (79 samples). The method can also be useful as a rough screening for pH and available Mg thereby significantly reducing the cost of mapping.