*Presenting author (Email: girish.stat@gmail.com; girish.jha@icar.gov.in)

Crop improvement programs aim to develop high yielding varieties coupled with resistance to biotic and abiotic stresses and nutritional superiority. Grain yield is a complex trait governed by genotype, environment, and their interaction. Growing a large number of genotypes under multiple environments and measurement for grain yield and its components are tedious and resource consuming tasks.

Therefore, there is a great need for novel, cost-effective techniques to evaluate the performance of crops at field scale through indirect selection of easily scorable traits using artificial intelligence based algorithms on comprehensive data. Hand-held or vehicle-mounted active proximal sensing offers a non-destructive method for monitoring crops within fields by providing real-time spectral vegetation indices.

The study aims to incorporate such proximal sensing data into a deep learning architecture for field scale wheat yield prediction. The novelty of this study is to predict the wheat yield based on normalized difference vegetation indices (NDVIs), canopy temperatures (CTs) and plant height (PH) by employing deep neural network (DNN) optimized through genetic algorithm (GA). NDVIs and CTs data were collected during the growing season at different growth stages from tillering through senescence (ground cover, heading, anthesis, grain filling and maturity).

Models are developed on crop yield data using 70% of a total 3350 germplasm of wheat planted in two growing environments (irrigated and rainfed) and two different locations during the winter season 2020-21.

Yield data on 1005 germplasm of wheat are used as validation dataset to assess the performance of developed model on unseen data. The developed GA optimized DNN model is compared and found superior to different machine learning models such as random forest regression (RFR), least absolute shrinkage and selection operator (LASSO), and support vector regression (SVR) for wheat yield prediction.

Notably, when analyzing individual feature groups, NDVIs at five different growth stages of wheat provide adequate predictions (R2 values ≥ 50%) of the yield under both environments. Additionally, the random forest method was employed to identify significant features among each feature group. This study introduces the application of a GA-optimized deep neural network, leveraging handheld or vehicle-mounted proximal sensing data for predicting wheat yield at field scale.

This innovative approach promises to significantly benefit the breeding community, providing a powerful and efficient tool for the pre-harvest evaluation of a large volume of germplasm, marking a major advancement in agricultural practices.