Non-Destructive Testing Based on Hyperspectral Imaging for Determination of Available Silicon and Moisture Contents in Ginseng Soils of Different Origins

Abstract

ABSTRACT

Soil-available silicon (SAS) and soil moisture (SM) contents are critical parameters for crop growth; however, traditional detection methods are time-consuming and inefficient. This study aimed to develop a non-destructive testing method using hyperspectral imaging (HSI) technology for the rapid and real-time detection of SAS and SM in ginseng soils of various origins. Twenty-two batches of soil samples and 51 batches of ginseng samples were collected, and spectral data in the visible near-infrared (VNIR) and shortwave infrared (SWIR) ranges were acquired simultaneously using an HSI system. To reduce data redundancy, principal component analysis for variable dimensionality reduction and a genetic algorithm (GA) involving iterative and voting methods were employed to process spectral data. The results showed that for SAS, the raw ELM performed best (SWIR R_v² = 0.88, RMSE = 28.19), while BP-GA3 peaked after GA (SWIR Rv² = 0.93, RMSE = 15.47). For SM, the raw BP (SWIR R_v² = 0.89, RMSE = 3.16), BP-GA3 achieved the highest GA result (SWIR Rv² = 0.94, RMSE = 1.80). PCA consistently underperforms (lowest SAS PCA-ELM SWIR R_v² = 0.41). Combined PCA and SAM analysis revealed distinct ginseng classification by origin, with RF achieving 77.78% (test) and 100% (train) accuracy for soil in SWIR, while BP model yielded 73.33% (test) and 80.56% (train) accuracy for ginseng in VNIR, demonstrating effective differentiation. This study provides theoretical support and a practical basis for the non-destructive testing of ginseng soil from the three provinces of Northeast China based on hyperspectral imaging; however, further expansion of the studied research samples is required to verify the generalization ability of the developed model.