Estimating potato aboveground biomass using unmanned aerial vehicle RGB imagery and analyzing its relationship with tuber biomass

• Published in: Field crops research 2024-12, Vol.319, p.109657, Article 109657
• Main Authors: Ye, Yanran, Jin, Liping, Bian, Chunsong, Xian, Guolan, Lin, Yongxin, Liu, Jiangang, Guo, Huachun
• Format: Article
• Language: English
• Subjects: aboveground biomass; algorithms; ammonium nitrogen; Bayesian information criterion; Bayesian theory; Biomass estimation; canopy; canopy height; Canopy structure; color; cost effectiveness; crop management; data collection; dry matter partitioning; field experimentation; genotype; nitrate nitrogen; nitrogen; potatoes; Recursive feature elimination; regression analysis; texture; UAV RGB imagery; unmanned aerial vehicles; variance; Variance inflation factor
• ISSN: 0378-4290
• DOI: 10.1016/j.fcr.2024.109657

Monitoring the aboveground biomass (AGB) is critical for assessing crop growth status, predicting yield, and making informed crop management decisions. This study aimed to develop an efficient and robust model for predicting potato AGB using data derived from unmanned aerial vehicle (UAV) RGB imagery, and to clarify the relationship between AGB and tuber biomass (TB). Remote sensing images of the potato canopy at multiple growth stages were acquired over two consecutive years (2022–2023), together with synchronous ground-based AGB and TB measurements. Sixty-four candidate variables encompassing spectral, color, structure, and texture features were extracted from the 2022 RGB images. We identified five single variables most sensitive to AGB through correlation analysis, which were then subjected to linear, polynomial, logarithmic, exponential, and power regressions. Recursive feature elimination (RFE) and variance inflation factor (VIF) analyses were used to select multivariate combinations as input parameters for Partial Least Squares (PLS) and Random Forest (RF) models. The optimal single-variable and multivariate regression models were selected based on the Bayesian information criterion (BIC), and subsequently applied to predict AGB in field trial plots for 2023. Additionally, we analyzed the dynamic relationship between AGB and TB, as well as the effects of genotype and nitrogen management on the accuracy of AGB predictions and its relationship with TB. The results showed that: (1) Structural indicators had the highest correlation with AGB among the four features. The linear regression using canopy volume (CVol) as an input parameter (Model 1) exhibited superior performance among the single-variable regression models (R2 = 0.75, RMSE = 0.42 kg m−2, BIC = −272.92). Meanwhile, the RF regression model with canopy cover (CC), maximum canopy height (CHmax), and average canopy height (CHmean) as input parameters (Model 2) had the lowest BIC value of −314.15 (R2 = 0.82, RMSE = 0.36 kg m−2), and its predicted values for the new dataset were significantly correlated with the measured AGB values (correlation coefficient of 0.84). Furthermore, Model 2 showed a stronger predictive power for AGB in plots with the high-erectability genotype ('Zhongshu18', R2 = 0.78, RMSE = 1.02 kg m−2) or those treated with ammonium nitrogen (NH₄⁺-N) (R2 = 0.75, RMSE = 1.24 kg m−2). (2) A significant positive correlation was observed between TB and cumulative AGB, with R² values of