Theses and Dissertations

ORCID

https://orcid.org/0009-0005-0131-9560

Issuing Body

Mississippi State University

Advisor

Thomasson, J. Alex

Committee Member

Reddy, Kambham R.

Committee Member

Wijewardane, Nuwan K.

Committee Member

Rangappa, Raju B.

Committee Member

Samiappan, Sathishkumar

Date of Degree

12-13-2024

Original embargo terms

Visible MSU only 1 year

Document Type

Dissertation - Campus Access Only

Major

Engineering (Biosystems Engineering)

Degree Name

Doctor of Philosophy (Ph.D.)

College

James Worth Bagley College of Engineering

Department

Department of Agricultural and Biological Engineering

Abstract

Accurate crop yield estimation early in season can help in identifying specific areas within a field that may require different management practices, thus enhancing overall efficiency and productivity. However, early estimation has been a challenging task mainly due to integrated complex association between genetic × environment × management factors. Over the decades, machine learning techniques, especially artificial neural network, has been widely adopted in conjunction with uncrewed aerial system (UAS) for high-throughput plant phenotyping and crop yield estimation. This study introduces the plus neural network (pNN) to incorporate temporal dynamics across different crop growth stages. Cotton field data from 2020-2023 were collected using uncrewed aerial vehicle (UAV) mounted multispectral sensors at Texas A&M University (site A) and Mississippi State University (site B) research farms. Data from 2020-2021 at site A were used to develop the model, tested with 2022 data from the same site, and further tested using site B data for geographic validation. Two feature selection methods were employed: a data-driven approach utilizing all available features, and a hybrid method based on Shannon information criteria for correlated features. Comparative analysis was conducted against LSTM and Transformer models. All the models developed performed well during testing. During cross-year testing, the Transformer model outperformed others with an R2 of 0.72, mean average percent error (MAPE) of 6.72%, and IA of 0.92. The pNN and LSTM models had lower performance (pNN: R2 0.49, MAPE 10.01%, IA 0.85; LSTM: R2 0.71, MAPE 7.33%, IA 0.91). Although MAPE differences were not significant between model, the pNN model had a MAPE greater than 10% but less than 15%, whereas Transformer and LSTM had MAPE less than 10%. Using a hybrid feature selection method, the Transformer model showed the highest performance (R2 0.65, MAPE 8.28%, IA 0.91), followed by pNN (R2 0.61, MAPE 8.28%, IA 0.87), while LSTM underperformed. The MAPE of pNN and Transformer model developed with hybrid method was not significantly different. Moreover, geographic testing for all models developed with data driven and hybrid method failed to estimate yield accurately, indicating the need for further refinement and data collection.

Download

Share

COinS