Articles
YIELD PREDICTION IN FRUIT CROPS USING IMAGE ANALYSIS
Article number
903_155
Pages
1115 – 1119
Language
English
Abstract
Fluctuations in the EU (EU-25) apple harvest are about 20% with fluctuations from country to country and from year to year.
Yield prediction is a relevant pre-requisite for the EU and US commissions, shipper, retailer, fruit trade, marketing organisations and fruit grower.
The objective of this joint project was to develop an alternative to the present yield prediction method based on manual visual fruit counting.
The new scheme utilizes digital images with a resolution of at least 1280 × 980 pixel. Digital images were taken in July of fully-grown apple trees on M.9 at Klein-Altendorf near Bonn, Germany and other sites in Europe with different crop loads.
Image analysis was employed to predict yield on a net acreage, defined as 93% of gross orchard acreage for the purpose of this study.
Harvest data were compared with yield prediction and presented as harvest prediction index, i.e. the ratio of predicted to harvested fruit yield.
This harvest prediction index ranged between 0.7 and 1.2 over 3 years for three major apple cultivars Golden Delicious, Braeburn and Elstar. The image analysis software was optimised over the three years for the majority of fruit orchards with best results in the 25-50 t/ha yield range.
The upper limit for the current image analysis software is 200 fruit/tree.
Larger yields than 50 t/ha are under-estimated, because larger fruit loads on the tree resulted in overlapping fruit segments in the images with under-estimation of fruit number and yield; similarly, yields lower than 25 t/ha are over-estimated.
Yield prediction was more accurate for early ripening than late-ripening apple cultivars due to the longer period between yield prediction and harvest and weather in between.
Higher image resolution than 1.5 megapixel (Mpx) did not improve the accuracy of the yield prediction, but required more than 1 min for the fully automated image analysis of a single image.
The accuracy of the yield prediction can be further improved by a larger number of photos per tree and fine tuning of the customised software, but not by higher resolution of the digital images vs. photos, which only slowed the process down.
The vision or possible future of this informal initiative is a low-cost, still slow but efficient, fully (or largely) automated scheme, in which any fruit farmer in the world with an ordinary digital camera could participate at low cost with reasonable orchard records, if he follows simple instructions for the photography.
The digital images could be posted or mailed via an FTP server for image analysis.
Yield prediction is a relevant pre-requisite for the EU and US commissions, shipper, retailer, fruit trade, marketing organisations and fruit grower.
The objective of this joint project was to develop an alternative to the present yield prediction method based on manual visual fruit counting.
The new scheme utilizes digital images with a resolution of at least 1280 × 980 pixel. Digital images were taken in July of fully-grown apple trees on M.9 at Klein-Altendorf near Bonn, Germany and other sites in Europe with different crop loads.
Image analysis was employed to predict yield on a net acreage, defined as 93% of gross orchard acreage for the purpose of this study.
Harvest data were compared with yield prediction and presented as harvest prediction index, i.e. the ratio of predicted to harvested fruit yield.
This harvest prediction index ranged between 0.7 and 1.2 over 3 years for three major apple cultivars Golden Delicious, Braeburn and Elstar. The image analysis software was optimised over the three years for the majority of fruit orchards with best results in the 25-50 t/ha yield range.
The upper limit for the current image analysis software is 200 fruit/tree.
Larger yields than 50 t/ha are under-estimated, because larger fruit loads on the tree resulted in overlapping fruit segments in the images with under-estimation of fruit number and yield; similarly, yields lower than 25 t/ha are over-estimated.
Yield prediction was more accurate for early ripening than late-ripening apple cultivars due to the longer period between yield prediction and harvest and weather in between.
Higher image resolution than 1.5 megapixel (Mpx) did not improve the accuracy of the yield prediction, but required more than 1 min for the fully automated image analysis of a single image.
The accuracy of the yield prediction can be further improved by a larger number of photos per tree and fine tuning of the customised software, but not by higher resolution of the digital images vs. photos, which only slowed the process down.
The vision or possible future of this informal initiative is a low-cost, still slow but efficient, fully (or largely) automated scheme, in which any fruit farmer in the world with an ordinary digital camera could participate at low cost with reasonable orchard records, if he follows simple instructions for the photography.
The digital images could be posted or mailed via an FTP server for image analysis.
Publication
Authors
D. Stajnko, M.M. Blanke
Keywords
Malus × domestica, apple, food chain, fruit growth, harvest, image analysis, marketing, precision agriculture, precision horticulture, yield prediction
Online Articles (170)