Spectral discrimination of vegetable crops using in situ hyperspectral data and reference to organic vegetables

Remote sensing has been evolving as a general method for multi-scale crop information extraction. Large scale multi-crop discrimination using airborne, or satellite remote sensing is required for farm level intervention and for decision making by various stake holders such as agriculture insurance companies, risk assessment agencies and local governments. Multispectral and hyperspectral remote sensing data from various platforms have been used for discriminating and mapping various crops. However, the number of crops, and the scale at which the discrimination is often limited to a few crops and is often approached as discrimination against other land cover classes. Further, multi-temporal datasets and spectral indices form the bulk of the datasets for crops discrimination. For potential operational application at field level, the ability to discriminate numerous crops-at least ten different crops at the same timeframe are vital. Furthermore, discrimination of crops grown under organic practices has promising application in the certification and quality assurance of agricultural produced sold as organic product. Theoretically, high resolution hyperspectral data has the capability to difference few tens of classes unambiguously. However, given the context of systematic spectral similarity in vegetation, especially crops, the potential of discrimination several crops are unclear. We, therefore, has assessed the spectral discrimination of as many as 23 different vegetable crops and attempted discriminating a few vegetable crops grown under organic and inorganic crop growing practices. For this, we have applied 12 different statistical and machine learning algorithms establishing the spectral discrimination and assessing its relative stand across the range of crops considered. The results indicate complex patterns of spectral discrimination wherein a few crops exhibit spectral similarity with several other crops at any scale of spectral characterization. The discrimination analysis of vegetable crops grown under organic and chemical input-based practices indicate a good discrimination. However, the quality of discrimination is substantially affected by the type of machine learning model used. We recommend coordinated multi-site and multi-phenology-based crop discrimination for establishing the stability of the discrimination observed across space and time.