Estimating Crop Grain Yield and Seed Composition Using Deep Learning from UAV Multispectral Data

Maitiniyazi Maimaitijiang, Bruce Millett, Sidike Paheding, Shahid Nawaz Khan, Kamila Dilmurat, Abel A. Reyes, Péter Kovács

Abstract: The objective of this work is to predict corn grain yield and seed compositions (i.e., protein, oil and starch concentration) using Unmanned Aerial Vehicles (UAV)-based multispectral imagery and deep learning models. UAV multispectral imagery was acquired throughout the growing season of 2022 over a cornfield near Brookings, South Dakota, USA. Deep learning methods such as Convolutional Neural Network (CNN), Long Short-Term Memory networks (LSTM), and Transformer were used to predict corn grain yield and seed protein, oil and starch compositions based on UAV multispectral imagery. The results show that: (1) both original and attention-based CNN and LSTM, as well as Transformer models are able to successfully predict corn yield and seed compositions. Transformer yielded superior performance over other methods. (2) attention CNN and LSTM consistently outperformed original CNN and LSTM. (3) multitemporal data outperformed each single-day imagery-based estimations.

External IDs:dblp:conf/igarss/MaimaitijiangMPKDRK23