Spatio-temporal generation of morphological Plant features for yield
prediction before harvest from Visual Image input using Progressively
Growing GANs
Abstract
Recent Innovations in Precision Agriculture (PA) are driven by Computer
Vision and Data Processing systems to quantify plant parameters.
Quantitative analysis of Plant Phenotyping in PA and monitoring
morphological traits is a protracting process, precluding the objective
and phenotyping pipeline. Greenhouses growing Genetically Modified (GM)
crops need to be maintained at constant environmental and simulated
conditions. Multiple parameters have to be controlled and regulated
inside a greenhouse for effective growth of crops and yield
maximisation. Not at all times are these factors derived and so, yield
maximisation in greenhouse is an experimental approach to new varieties.
For deduced environmental parameters and conditions for certain crops,
few other biotic and abiotic factors can hinder or affect growth in
certain ways that are not always factored in during calculating
parameters conducive for plant growth. Such factors may not always be
affecting parametral calculations, but transpose visual cues on plant
growth environment such as spectral change in soil values, or minute
changes like leaf reflectance or visible changes in plant stimuli to
biotic factors. Plant growth is inclusive of multiple environmental
variables, and yield maximisation approaches are experimental to finding
the optimum derived value for these variables. Computer Vision provides
a catalytic approach to predicting optimum parameters for yield
maximization in phenomics. Computer Vision and Generative Adversarial
Networks (GAN)’s offer a catalytic approach to the time-consuming
process, providing a solution to the phenotyping bottleneck. This
research proposes a concept of curating data of plant growth over time
to predict conditional growth and responsive stimuli of the plant under
different situations and how this can affect crop yield. The method
proposed here is a non-invasive approach to the existing destructive
biomass estimation methods and Frameworks. This methodology of the
research focuses on utilizing image parameters modelled using a time
series Progressively Growing Generative Adversarial Networks PGGAN to
map plant growth patterns and progressive variance in biomass of plant
in the Spatio-Temporal Domain. These Generative networks evaluate and
predict based on merely raw pixel input excluding dependence on further
constraints, feature vectors or parameters influencing data.