Articles
Application of artificial intelligence in prediction of sunburn damage of fruit
Article number
1433_8
Pages
65 – 70
Language
English
Abstract
The TensorFlow Lite model has been developed to get utilized in a smart farming application.
Climate change induced heatwaves severely affect horticulture.
Fruit damage can scale from discoloration to necrosis.
While early symptoms decrease the aesthetic value, severely damaged fruit cannot be consumed or processed.
Weather station common parameters were monitored and linear discriminant analysis (LDA), support vector machine (SVM) and neural network (NN) models were tested to predict sunburn damage.
Models were built and validated using consecutive years for grape (2022 and 2023), while calibration of apple models was performed with one year (2023). Descriptive and cumulative parameters, as well as tendencies were calculated for temperature, solar radiation and relative humidity.
Ten parameters were finally selected and future damage within 72 h was estimated.
The NN model with 2 hidden layers was finally selected because of the highest performance in validation and the ability to improve with new data provided by the grower as feedback.
Additionally, we defined a set of warning threshold values for elevated risk of sunburn damage.
The model is freely available in TFLite format for developers as well as the mobile app named SHEET for farmers.
Climate change induced heatwaves severely affect horticulture.
Fruit damage can scale from discoloration to necrosis.
While early symptoms decrease the aesthetic value, severely damaged fruit cannot be consumed or processed.
Weather station common parameters were monitored and linear discriminant analysis (LDA), support vector machine (SVM) and neural network (NN) models were tested to predict sunburn damage.
Models were built and validated using consecutive years for grape (2022 and 2023), while calibration of apple models was performed with one year (2023). Descriptive and cumulative parameters, as well as tendencies were calculated for temperature, solar radiation and relative humidity.
Ten parameters were finally selected and future damage within 72 h was estimated.
The NN model with 2 hidden layers was finally selected because of the highest performance in validation and the ability to improve with new data provided by the grower as feedback.
Additionally, we defined a set of warning threshold values for elevated risk of sunburn damage.
The model is freely available in TFLite format for developers as well as the mobile app named SHEET for farmers.
Authors
L. Baranyai, L.L.P. Nguyen, T. Zsom, V. Zsom-Muha, Z. Gillay
Keywords
discriminant analysis, support vector machine, neural network, Keras, climate, weather
Online Articles (41)