TABLE 5.

Model performance for comparable studies using UAV imagery for yield prediction.

References	Crop	Model	Performance	Description
This study	Rice	2D-CNN	7.9% (rRMSE)	Yield predicted from thermal and six VIs using data at late booting stage.
			5.8% (MAPE)	Performance based on 4-fold cross validation from the same field and season.
			0.22 (R2)
Duan et al. (2021)	Rice	Neural network	5.3–7.1% (rRMSE)	Yield predicted on two individual VIs from 6 or more imaging days.
			0.48–0.62 (R2)	Performance based on leave-one-out cross validation from the same field and season.
Wan et al. (2020)	Rice	random forest	2.75% (rRMSE)	Yield predicted from four RGB- and multispectral-derived features. Data set included substantial yield variation due to experimental nitrogen treatment.
			0.83 (R2)	Performance based on random held-out set from the same field and season.
Yang et al. (2019)	Rice	2D-CNN	26.6% (MAPE)	Yield predicted from raw RGB and multispectral imagery at ripening stage.
			0.49 (R2)	Performance based on held-out set of independently managed plots from the same season.
Maimaitijiang et al. (2020)	Soybean	2D-CNN	15.9% (rRMSE)	Yield predicted from 72 features derived from multispectral, thermal, and RGB sensors on a single day. Data set included substantial yield variation due to cultivar-specific differences.
			0.72 (R2)	Performance based on held-out set from the same field and season.
Nevavuori et al. (2019)	Wheat/barley	2D-CNN	8.8–12.6% (MAPE)	Yield predicted from RGB or a single VI measured on a single day. UAV data were combined for two crops, nine fields, and multiple imaging dates. Images for “early” or “late” season models were sub-sampled, shuffled, and split into test and train sets.

rRMSE, relative root mean squared error; MAPE, mean absolute percentage error.