Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Predicting crop yield in potatoes using sequential UAV imagery and machine learning

Kroné, Andreas LU (2026) In Student thesis series INES NGEM01 20251
Department of Earth and Environmental Sciences (MGeo)
Abstract
Accurate prediction of crop yield is crucial in precision agriculture for ensuring food security worldwide. This study used temporally distinct multispectral UAV (Unmanned Aerial Vehicle) imagery together with a RF (Random Forest) machine learning model to predict crop yield in Fontane potatoes, during the growth season of 2024 in an area in southern Sweden. The main aims were to explore how well an RF model can predict yield in potatoes, and when is the best time to do so. The multispectral UAV imagery consisted of radiometrically corrected orthomosaics with a spatial resolution of 4cm. From it five spectral bands (blue, green, red, red-edge (REG) and near-infrared (NIR)) and six vegetation indices (NDVI, GNDVI, NDRE, EVI, SAVI, OSAVI)... (More)
Accurate prediction of crop yield is crucial in precision agriculture for ensuring food security worldwide. This study used temporally distinct multispectral UAV (Unmanned Aerial Vehicle) imagery together with a RF (Random Forest) machine learning model to predict crop yield in Fontane potatoes, during the growth season of 2024 in an area in southern Sweden. The main aims were to explore how well an RF model can predict yield in potatoes, and when is the best time to do so. The multispectral UAV imagery consisted of radiometrically corrected orthomosaics with a spatial resolution of 4cm. From it five spectral bands (blue, green, red, red-edge (REG) and near-infrared (NIR)) and six vegetation indices (NDVI, GNDVI, NDRE, EVI, SAVI, OSAVI) were derived. Using yield data obtained from a harvester machine, yield buffers were created and the mean yield as well as pixel values for each spectral band and vegetation index inside each buffer were extracted, and this was then used to train a RF regression model. The model training process included feature importance analysis to optimize the selection of predictor variables, to achieve the best results. The results show that the best growth phase to predict yield is during the stem elongation stage (~38 days after planting), which achieves a moderate prediction accuracy (R2 = 0.66). However, the model suffered from significant overfitting, which indicated problems with the input data. This means that this specific method for predicting yield will likely not generalize well for other areas or varieties of potato. Improvements could include more and better predictor variables like plant- or soil specific data, such as LAI or soil moisture, which have been shown to improve yield predictions. In conclusion, the results demonstrate that an RF model trained on multispectral UAV data can predict yield in potatoes moderately well, but that careful selection of input variables is crucial to the prediction accuracy and ensuring generalization. (Less)
Popular Abstract
Can drones and machine learning forecast potato harvest?
Food security is a large problem all over the world. Crops can fail for any number of reasons such as drought, pests and lack of fertilisation. In areas of the world where these difficulties can lead to reduced availability of food, it is of particular importance to be able to know how much food will be produced in the fields. Knowing how large a harvest will be early on helps farmers plan fertilisation, irrigation and helps estimate yield. For potatoes this is difficult since the crop grows underground and thus cannot be seen during the growing season. This study investigated whether drone imagery combined with machine learning can be used to estimate potato yield, and when is the... (More)
Can drones and machine learning forecast potato harvest?
Food security is a large problem all over the world. Crops can fail for any number of reasons such as drought, pests and lack of fertilisation. In areas of the world where these difficulties can lead to reduced availability of food, it is of particular importance to be able to know how much food will be produced in the fields. Knowing how large a harvest will be early on helps farmers plan fertilisation, irrigation and helps estimate yield. For potatoes this is difficult since the crop grows underground and thus cannot be seen during the growing season. This study investigated whether drone imagery combined with machine learning can be used to estimate potato yield, and when is the best time during the season to do so. Drones are comparatively cheap and can produce high resolution images compared to those of satellites or planes. The study also investigated how well a machine learning model can predict yield when it is trained on the bare minimum of input data. This was in the interest of keeping the method as cheap as possible, since this would reduce the need to perform time-consuming manual collection of data.
The study was carried out in a 1.5 ha potato field in southern Sweden during the 2024 growing season. A drone equipped with a multispectral camera, which captures images in visible light as well as near-infrared, photographed the field on seven separate dates. The image data from the different bands of light in the images were used to calculate vegetation indices (VIs) that act as indicators of plant health. Healthy foliage looks different to that of a stressed plant, so this imagery can reveal how well the crops are doing. This data was then used together with the final yield data obtained from the potato harvester to train a Random Forest machine learning model, one for each date. The yield predictions were made by looking at how the patterns in the image data compared to areas with high or low yield.
The best model results came from the drone data obtained on June 24, which is about five and a half weeks after the potatoes were planted and when the plants are growing their main stems. At this date, the model could predict yield with approximately 66% accuracy, which is not very high. You might expect observations closer to the time of harvest to produce more accurate predictions; however, this was not the case found in this study.
The main problems with this method were that the limited amount of data was not sufficient to produce accurate predictions. It is likely that some manually collected data, such as soil moisture and fertiliser amount, is necessary to predict yield to a degree which is reliable enough for farmers to use. The method also resulted in something called overfitting, which is where the accuracy of the model is much higher on the data it was trained on when compared to data from other places or times. This is critical, as a model which only works on one field is of limited use to anyone. Further work should refine the model training process to ensure increased generalisation and thus application to other areas. (Less)
Please use this url to cite or link to this publication:
author
Kroné, Andreas LU
supervisor
organization
course
NGEM01 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
physical geography, crop yield prediction, machine learning, Random Forest, UAV imagery, growth stage, potatoes
publication/series
Student thesis series INES
report number
793
language
English
id
9253684
date added to LUP
2026-11-17 16:28:16
date last changed
2026-11-17 16:28:16
@misc{9253684,
  abstract     = {{Accurate prediction of crop yield is crucial in precision agriculture for ensuring food security worldwide. This study used temporally distinct multispectral UAV (Unmanned Aerial Vehicle) imagery together with a RF (Random Forest) machine learning model to predict crop yield in Fontane potatoes, during the growth season of 2024 in an area in southern Sweden. The main aims were to explore how well an RF model can predict yield in potatoes, and when is the best time to do so. The multispectral UAV imagery consisted of radiometrically corrected orthomosaics with a spatial resolution of 4cm. From it five spectral bands (blue, green, red, red-edge (REG) and near-infrared (NIR)) and six vegetation indices (NDVI, GNDVI, NDRE, EVI, SAVI, OSAVI) were derived. Using yield data obtained from a harvester machine, yield buffers were created and the mean yield as well as pixel values for each spectral band and vegetation index inside each buffer were extracted, and this was then used to train a RF regression model. The model training process included feature importance analysis to optimize the selection of predictor variables, to achieve the best results. The results show that the best growth phase to predict yield is during the stem elongation stage (~38 days after planting), which achieves a moderate prediction accuracy (R2 = 0.66). However, the model suffered from significant overfitting, which indicated problems with the input data. This means that this specific method for predicting yield will likely not generalize well for other areas or varieties of potato. Improvements could include more and better predictor variables like plant- or soil specific data, such as LAI or soil moisture, which have been shown to improve yield predictions. In conclusion, the results demonstrate that an RF model trained on multispectral UAV data can predict yield in potatoes moderately well, but that careful selection of input variables is crucial to the prediction accuracy and ensuring generalization.}},
  author       = {{Kroné, Andreas}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Student thesis series INES}},
  title        = {{Predicting crop yield in potatoes using sequential UAV imagery and machine learning}},
  year         = {{2026}},
}