Predicting the Carbon Property of Soil Using DrSeqANN and VIS/NIR Spectroscopy

Abstract

Plant productivity and health are directly impacted by carbon(C) levels. This study evalu-ated the potential of visible/near-infrared (V/NIR) spectroscopy (350-2,500 nm) for soil characterization, utilizing a dataset of 200 soil samples from Uttar Pradesh, India. The pre-dictive performance of spectral data was compared across three modeling approaches: an Ensemble of Lasso and Ridge Regression models (ELRR), Random Forest (RF), and a more complex Artificial Neural Network (ANN) were employed to choose the spectral character-istics that were utilized in the C prediction.. To reproduce the spectrum's wavelength The log derivative, log to base 10 derivative log10x and inverse derivative were employed in the preprocessing.. The results showed that the availability of C was found to be between 350 and 450 nm. Using the Log10x pre-processed data and the suggested DrSeqANN-Dropout Sequential Artificial Neural Network technique, the most accurate results were obtained by accessing parameters with the aid of RMSE = 0.08, R2 = 0.82, and RPIQ = 4.32 for our sug-gested DrSeqANN model. Compared to the other two approaches