Predicting wheat yield and grain quality with UAV multispectral imagery and deep learning.

Abstract

Bread wheat ( Triticum aestivum L.) is a major staple crop, and timely, in-season prediction of grain yield (GY) and grain quality traits, grain protein content (GP), and grain test weight (TW), is critical for informed management and field-based high-throughput phenotyping (HTP). Unmanned Aerial Vehicle (UAV) remote sensing, coupled with artificial intelligence and deep learning (DL), offers a practical pathway for rapid, plot-scale trait estimation. Here, we investigate the value of multitemporal, multispectral UAV imagery for predicting winter wheat GY, GP, and TW, and we systematically compare two modeling paradigms: (1) handcrafted feature-based workflows that use plot-aggregated spectral and texture descriptors derived from UAV imagery, and (2) image-based, end-to-end workflows that learn directly from plot-level reflectance image chips. During the 2022 growing season, multispectral UAV data were collected repeatedly over seven experimental wheat sites in South Dakota, USA. For handcrafted feature-based modeling, we evaluated Support Vector Regression (SVR) and Random Forest Regression (RFR), along with DL models including a feedforward Deep Neural Network (DNN) and a one-dimensional Convolutional Neural Network (1D-CNN). For end-to-end image-based modeling, we implemented 2D-CNN, 3D-CNN, and a hybrid 2D-CNN–LSTM architecture to leverage both spatial information and multi-date dependencies. Our results show that: 1) the image-based modeling workflow yielded comparable to slightly better performance than the handcrafted feature-based modeling workflow across wheat GY, GP, and TW predictions; 2) 3D-CNN outperformed all other methods with R 2 of 0.65, 0.61 and 0.69 for GY, GP and TW estimations, respectively; 3) multitemporal UAV data outperformed the data collected from a single growth stage; and UAV data from wheat Feekes 10 (booting) stage yielded slightly better estimation results compared to the data collected from other growing stages, with R 2 of 0.62, 0.55, and 0.62 for GY, GP, and TW estimations, respectively. The results indicate that DL applied to high-resolution multitemporal and multispectral UAV imagery holds strong promise for predicting winter wheat yield and grain quality during the growing season, while also informing HTP efforts and site-specific management.

Plant phenotyping relevance

UAVマルチスペクトル画像から収量・品質形質を推定するワークフローを開発・比較・評価しており、植物表現型取得が研究の中心である。

Code and data availability

The supplied blocks describe UAV multispectral imagery collection, ground-truth GY/GP/TW measurements, and deep learning modeling, but contain no data availability statement, public dataset deposit, or author code/model release with a URL. No qualifying paper-specific public assets are present.

No evidence-backed public reproduction asset is currently recorded.