Academic literature on the topic 'Multiplicative scatter correction'

Soil texture is a crucial property that can provide insight into its productive capacity. However, determining soil texture can be a complex and time-consuming process. To address this challenge, we aimed to develop machine learning (ML) models that could predict the composition of sand, clay, and silt in soil based on hyperspectral imaging (HSI) data. We collected and analyzed 500 soil samples and processed the HSI data by masking samples with reflectance and transforming the texture variables with Box-Cox. We also employed various techniques, such as moving average, Savitzky–Golay filtering (SG), first and second derivatives (FD and SD), gap-segment (GS) with FD and SD, standard normal variate (SNV), SNV with detrending (DT), SNV-SG, multiplicative scatter correction (MSC), and GS-DT for the covariates. We applied cubist models (CUB), principal component regression (PCR), partial least squares regression (PLSR), and artificial neural networks (ANN). A total of 1240 models were obtained. For clay, the models with the best performance in the TRAIN and TEST sets were a CUB-tuned model, where R2-TEST = 0.93 and the test root-mean-square error (RMSE-TEST) = 3.63%. For silt and sand, CUB models without refinement were selected, with R2-TEST values of 0.63 and 0.61 and RMSE-TEST of 5.55 and 8.65%, respectively. In conclusion, clay prediction with HSI is feasible if purified spectra are used, with outlier detection techniques, evaluation in the overlap zone, and transformations such as GS-FD. The models for silt and sand prediction are not recommended due to their low performance.