Development of Soil Fertility Index Using Machine Learning and Visible-Near-Infrared Spectroscopy

Abstract

An accurate assessment of soil fertility is crucial for monitoring environmental dynamics, improving agricultural productivity, and achieving sustainable land management and utilization. The inherent complexity and spatiotemporal heterogeneity of soils result in significant challenges in soil fertility assessment. Therefore, this study focused on developing a rapid, economical, and precise approach to evaluate soil fertility through the application of visible-near-infrared spectroscopy (VNIR). To achieve this, we utilized the Land Use and Cover Area Frame Survey (LUCAS) dataset and employed a variety of prediction models, including partial least squares regression, support vector machines (SVMs), random forest, and convolutional neural networks, to estimate various soil properties and overall soil fertility. The results showed that the SVM model had the highest prediction accuracy, particularly for clay content (coefficient of determination (R2) = 0.79, ratio of performance to interquartile range (RPIQ) = 3.04), pH (R2 = 0.84, RPIQ = 4.54), total nitrogen (N) (R2 = 0.80, RPIQ = 2.40), and cation exchange capacity (CEC) (R2 = 0.83, RPIQ = 3.16). A soil fertility index (SFI) was developed based on factor analysis, integrating nine essential soil properties: clay content, silt content, sand content, pH, carbonate content, N, soluble phosphorus, soluble potassium, and CEC. We compared direct and indirect prediction models for estimating SFI and found that both models showed high accuracy (mean value of R2 = 0.80, mean value of RPIQ = 2.21). Additionally, SFI was classified into five classes to provide insights for precision agriculture. The kappa coefficient was 0.63, which indicated that the SFI evaluation results between VNIR and chemical analysis were relatively consistent. This study provides a theoretical foundation of real-time soil fertility monitoring for the optimization of agricultural practices.