Detecting Nitrogen and Water status of Potato Crops with Machine Learning

This study investigates the potential of using a random forest approach to detect the nitrogen and water status of potato crops based on multispectral reflectance. For this purpose, images from 2 different experimental fields in the Netherlands were used. Each field had plots with 6 different water and nitrogen treatments and 3 different potato varieties, with one of them being the same in both. The images consist of 9 bands in the visible - near infrared region and one band in the Long-Wave Infrared.

The results showed that the random forest algorithm was able to achieve an accuracy of 85% for detecting the 6 different treatments, and 90% for detecting water or nitrogen status after local calibration. The results, however, were different if the algorithm was trained in one field and tested in another. By applying the algorithm only to the variety that was common in both the fields, the 6 treatments were detected with accuracy of approximately 48% while it was able to detect water or nitrogen status with accuracy of 63% and 79% respectively.