In-season nitrogen status and pre-harvest potato yield assessment using air-spaceborne imagery with AI techniques

Abstract

Context

In modern agriculture, timely and precise nitrogen monitoring and pre-harvest tuber yield assessment for potato (Solanum tuberosum L.) crop is essential to optimize resource management, in-time decision-making and trade benefits. Traditional nitrogen assessment methods are labor-intensive and time consuming. Utilizing multispectral data from unmanned aerial vehicles (UAV) and satellite imagery integrated with artificial intelligence (AI) modeling offers high resolution solution for efficient field monitoring throughout the growing season.

Objective

The aim of this study was to use AI algorithms trained and validated solely on high-resolution multispectral UAV and Sentinel-2 satellite data to evaluate their potential in estimating in-season plant nitrogen status and pre-harvest tuber yield assessment in commercial potato production fields.

Methods

A study was conducted on four commercial potato fields located in Prince Edward Island, Canada. UAV and Sentinel-2 images were used at the early flowering (S1) and pre-harvest (H) stages of potato to extract multispectral bands and vegetative indices. Extracted multispectral bands and vegetative indices were used to assess ground truth data (petiole nitrate concentration as crop N status and tuber yield) at S1 and H phenological stages using five different machine learning algorithms.

Results

Results indicated that the bagged tree machine learning algorithm trained on UAV S1 images revealed a relative root mean square error (RRMSE) of 12.7 % and a relative mean absolute error (RMAE) of 9.6 %. In contrast, the random forest model trained on Sentinel-2 S1 data showed an RRMSE of 15.5 % and an RMAE of 12.6 %. The random forest models trained on UAV H and Sentinel-2H data demonstrated an RRMSE of 16.1 % and 13.8 %, and an RMAE of 12.5 % and 11.3 %, respectively. These models revealed the best estimation results when compared with the rest of the machine learning models.

Conclusion

The fitted models effectively estimated petiole nitrate concentration and pre-harvest tuber yield in commercial potato fields, demonstrating their practical utility. Key features, such as the Canopy Chlorophyll Content Index via UAV and the red edge band via Sentinel-2, were crucial in the prediction process. These findings underscore the potential of AI algorithms to enhance agricultural productivity and precision in crop management.

Implications

The use of AI algorithms trained on high-resolution air-borne data can significantly improve the accuracy of estimating the petiole nitrate concentration and pre-harvest tuber yield. By leveraging key important features with the robust feature engineering techniques, farmers and researchers can make more informed decisions regarding crop management. This approach provides a valuable tool for advancing sustainable and efficient farming practices.