A multi-year study to use hyperspectral remote sensing and machine learning for predicting potato nutrient status and yield

Hyperspectral imaging system provides accurate high-resolution maps that can be used to monitor nutrient status of vegetation and estimate crop yield. Nutrient (such as nitrogen (N)) status monitoring and final yield estimation are critical for crop management in precision agriculture. Hyperspectral remote sensing together with machine learning have been effective in facilitating precision farming decision-making that can increase resource use efficiency and reduce impact of agricultural production on the environment. In this three-year study, thirty-four vegetation indices calculated from the full spectrum (400-2500 nm) were used to predict crop N status (indicated by petiole nitrate and whole leaf total N) and final tuber yield of four cultivars of potatoes (Solanum tuberosum L), over three growing seasons (2018 to 2020) with seven different machine learning models. Besides vegetation indices, agronomical factors (such as potato growth stages, cultivar and N rates), and environmental factors (such as growing degree days and precipitation) have also served as the inputs to the machine learning models. Our results showed that for chipping potatoes, petiole nitrate can be predicted with the highest R2 ranging between 0.630 and 0.940, followed by whole leaf total N (R2 = 0.561-0.881), and final yield (R2 = 0.323-0.671). For fresh market potatoes, whole leaf total N prediction had the best results with R2 between 0.704 and 0.913, followed by petiole nitrate (R2 = 0.638-0.905) and final tuber yield (R2 = 0.241-0.799). The best vegetation indices to predict each potato trait varied depending on the cultivar and growing season. Overall random forest has the strongest power to predict potato traits studied in this research. Our results have demonstrated the great potential of using vegetation indices based on hyperspectral imagery and machine learning for precision agriculture management of potato production.