Carbon Property of Soil Prediction By VIS/NIR Spectroscopy Using DrSeqANN

Abstract

Carbon(C) levels have a direct impact on plant health and productivity. 200 soil samples from the Indian state of Uttar Pradesh were utilized in this study as a database to assess the efficacy of employing visible/near-infrared (VIS/NIR) spectroscopy data. The samples wavelengths ranged from 350 to 2,500 nm. The spectral features used to predict C were chosen using Ensemble Lasso Ridge Regression (ELRR), Random Forest (RF), and the more complicated Artificial Neural Network. The preprocessing employed the log derivative, Log10x derivative, and inverse derivative to replicate the wavelength of the spectrum. The essential feature wavelengths for C were discovered to be between 350 and 450 nm, according to the results. The recommended Dropout Sequential Artificial Neural Network (DrSeqANN) technique combined with the Log10x pre-processed data produced the most accurate results.