Hybrid Neural Networks with Attention-based Multiple Instance Learning for Improved Grain and Yield Predictions
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3070213Utgivelsesdato
2022Metadata
Vis full innførselSamlinger
Sammendrag
Agriculture is a critical part of the world’s food production, being a vital aspect of all societies.Procedures need to be adjusted to their specific environment because of their climate and fieldcondition disparity. Existing research has demonstrated the potential of grain yield predictions onNorwegian farms. However, this research is limited to regional analytics, which is unable to acquiresufficient plant growth factors influenced by field conditions and farmers’ decisions. One factorcritical for yield prediction is the crop type planted on a per-field basis.This research effort proposes a novel approach for improving crop yield predictions using a hybriddeep neural network utilizing temporal satellite imagery from a remote sensing system. Additionally, We apply a variety of data, including grain production, meteorological data, and geographicaldata. The crop yield prediction system is supported by a field-based crop type classification model,which supplies features related to crop type and field area. Our crop classification system takesadvantage of both raw satellite images as well as carefully chosen vegetation indices. Further, wepropose a multi-class attention-based deep multiple instance learning model to utilize semi-labeleddatasets, fully benefiting Norwegian data acquisition.Our best crop classification model, which consists of a time distributed network and a gated recurrent unit, classifies crop types with an accuracy of 70% and is currently state-of-the-art forcountry-wide crop type mapping in Norway. Lastly, our yield prediction system enables realisticin-season early predictions that could benefit actors in real-life scenarios.