Developing a robust yield prediction model for potatoes (Solanum tuberosum L.) using multi-faceted and multi-year data

Abstract

Robust yield prediction models can help farmers with fertilization decisions to maintain yield while reducing impact of crop production on the environment. For potatoes, the No No 1 consumed underground crop that need high nitrogen (N) fertilizer input, accurate yield prediction during the growing season will help reduce over-use of N fertilizer and mitigate groundwater contamination issues. This multi-year study collected hyperspectral imagery (400 – 2500 nm) across different potato (Solanum tuberosum L.) growth stages and seasons under varied nitrogen (N) treatments. It developed random forest (RF) models to predict final tuber yield using different model inputs, including genotype (G) (cultivar), management (M) practices (N treatment), environmental (E) factors (soil temperature and precipitation), and hyperspectral data that was obtained by smart agriculture technologies (T). Our findings revealed that: 1) the top 45 (approximately ⁓ 10 % of total bands: coefficient of determination: R² = 0.575) bands generated from feature selection of RF could result in similar yield prediction accuracy as when using all 474 narrow-bands from the hyperspectral images, and these selected spectral signatures were distributed across the visible (VIS), near-infrared (NIR), and short-wave infrared (SWIR) regions; 2) for plants under the higher N rate, lower reflectance in the VIS region (due to more chlorophyll accumulation), higher reflectance in the NIR region (due to more dense plant canopy), and lower reflectance in the SWIR (due to higher water content) were observed; 3) The model that included multi-faceted inputs, suggesting the G × E × M × T interactions, produced highly satisfying performance (R² = 0.716) in comparison to the models that incorporated less inputs (using single input). Specifically, models using single input showed a performance range of R² = 0.009–0.728 for individual years and R² = 0.481–0.616 for combined years. Models incorporating two inputs demonstrated improved performance with R² = 0.011–0.751 for individual years and R² = 0.308–0.766 for combined years. Notably, models with multiple inputs achieved the highest performance, with R² = 0.154–0.834 for individual years and R² = 0.647–0.716 for combined years. 4) All years of data are needed to develop a robust and generalized potato yield prediction model. In conclusion, this study highlights the need to use a holistic approach that considers multiple facets (varietal choice, environmental conditions, management practice such as N supply, and smart agriculture technologies) of the crop production systems across different growing seasons to develop robust yield prediction models for potatoes.