Explainable Multimodal Deep Learning for Predicting Apple (Malus domestica Borkh.) Graft Performance in Ethiopian Highland Nurseries: Cross-Method Validation of Empirical Agronomic Relationships
Year: 2026
Type: preprint
Abstract: Background Apple production is growing in the highlands of Ethiopia. People who grow apples do not have good tools to help them pick the right apple variety or to predict how well a graft will work. This study aimed to create and test a deep-learning system that can explain its predictions about apple graft performance and to check if the explanations match rules that farmers use. Objective The goal of this study was to build a deep-learning framework that can predict the performance level of apple grafts and to see whether the explanations given by the model agree with simulated farming rules. Methods A simulator that uses the climate and soil conditions of Injibara District, Ethiopia (altitude 2,500–2,700 m and 400–800 chill hours) was made. The simulator produced 1,500 apple trees from five apple cultivars. The data set contained tables of numbers a 52-week series of plant growth and a spot for a picture. Each tree was put into one of five performance levels. After checking for and fixing any data-leak problems a neural network called GraftNet was tested with five random starts. Its results were compared with those from machine-learning methods. Two ways of looking at the model’s explanations—SHAP and grouped permutation importance—were used. Results GraftNet gave a weighted kappa of 0.954 ± 0.003 an accuracy of 0.799 ± 0.011 and a balanced accuracy of 0.775 ± 0.018. However GraftNet did not do better than a forest or a logistic regression. Adding the 52-week growth series did not help the classification. A simpler method, ridge regression predicted how fast the trees would grow accurately (RMSE 1.7 cm versus 6.1 cm). The important factors were the chilling requirement and the chill deficit. Variables that were not part of the simulator’s rule set were almost unimportant. GraftNet worked well on the simulated data but it did not give an edge over traditional machine-learning models. The work shows that explainable machine learning can help explore apple graft performance. Still more tests are needed with nursery and field data, from many apple varieties, rootstocks and different farming areas.
Source: Research Square
Author: Simachew Melaku Ayalew
Institution: Bahir Dar University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225640307