Year: 2026
Type: dissertation
Abstract: Accurate characterization of future environmental conditions is important for agricultural planning and crop performance prediction. This thesis develops a machine-learning framework that converts daily environmental records into compact annual embeddings and predicts future embeddings from recent years. The analysis uses 1981--2022 records from 923 locations across 23 countries. Three monthly representation methods are compared using a self-supervised contrastive annual-encoder architecture. The autoencoder-based representation gives the strongest prediction performance and is used for future-embedding prediction. The average of the three preceding annual embeddings predicts the future embedding more accurately than learned predictors. A separate downstream evaluation examines future-year yield prediction without weather from that year. The direct embedding gives the lowest test error for both crops. The predicted embedding also improves grain-maize yield prediction over historical yield information alone, with a corresponding within-learner improvement for maize. The results show that annual environmental embeddings can support longer-term crop performance prediction.
Author: Haque, Asif Ihtemadul
Institution: Western University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDGs: Climate action, Zero hunger
Open Access status: green
OpenAlex ID: W7214880267
