Hyperspectral discrimination of vegetable crops grown under organic and conventional cultivation practices: a machine learning approach

Abstract

A verifiable and regional level method for mapping crops cultivated under organic practices holds significant promise for certifying and ensuring the quality of farm products marketed as organic. The prevailing method for the identification of organic crops involves labor-intensive manual inspections, detailed record-keeping of crop stages, and certification. Hyperspectral remote sensing is an evolving general sensing technique for extracting crop information across various scales. High-resolution hyperspectral data theoretically can distinguish numerous crops unambiguously at various levels of detail. The aim of this study is to investigate the possibility of spectral discrimination of a few vegetable crops (brinjal ( Solanum melongena) and red spinach ( Amaranthus dubius)) grown under organic and conventional cultivation practices and assess the inclusion of numerous landscape-level co-occurring crop species in the discrimination analysis. We acquired high-resolution in situ hyperspectral measurements on the research farms of the College of Agriculture, Kerala Agricultural University, Thiruvananthapuram, India, in the 2022 crop-growing season. Methodologically, quantifying the spectral discrimination as the multi-crop classification problem, we applied 12 different machine learning algorithms to assess the spectral discrimination and evaluated their relative performance across the diverse range of the crops considered. The results reveal intricate patterns of spectral discrimination. Vegetable crops grown under both organic and conventional chemical inputs-based practices indicate a high level (accuracy: 85–95%) of spectral discrimination. The effectiveness of the discrimination observed is significantly influenced, with a reduction in the accuracy of discrimination by 10%, by choice of the machine learning model and the presence of several co-occurring crop species. We advocate for coordinated, multi-site, and multi-phenology-based crop discrimination studies to ensure the stability of observed discrimination across different spatial and temporal contexts. The findings indicate that, due to physiological and biochemical differences, organically cultivated crops exhibit distinct spectral features than conventionally cultivated crops, and with a suitable ML method, it is plausible to map crops over geographically extant areas using hyperspectral remote sensing.