An exploratory journey of representation learning’s enhancement, adaptation and related intelligent methods

Wu, Jing

Permalink

https://hdl.handle.net/2142/124344 Copy

Description

Title

An exploratory journey of representation learning’s enhancement, adaptation and related intelligent methods

Author(s)

Wu, Jing

Issue Date

2024-04-24

Director of Research (if dissertation) or Advisor (if thesis)

Hovakimyan, Naira

Doctoral Committee Chair(s)

Hovakimyan, Naira

Committee Member(s)

• Salapaka, Srinivasa
• Martin, Nicolas Federico
• Wang, Yuxiong

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Representation Learning
• Machine Learning
• Intelligent Agriculture

Abstract

Representation learning models employing Siamese structures have consistently demonstrated exceptional performance across various fields, including deep learning, computer vision, and natural language processing. Furthermore, the applicability of representation learning has broadened to encompass wider domains such as agriculture, remote sensing, and earth observation, which are significantly challenged by data scarcity. This dissertation aims to enhance the quality and adaptability of learned representations across these diverse application domains. Meanwhile, we have also expanded the scope of our research to a broader area of intelligent agricultural systems. Initially, we delve into contrastive representation learning within the general computer vision domain and introduce a novel ``Hallucinator" module to reduce mutual information, increase the batch size of positive pairs, and improve representation quality. Subsequently, we extend the representation framework to agriculture and remote sensing, proposing spatial-temporal-aware architectures tailored to the unique characteristics of remote sensing data. Furthermore, we introduce the Extended Agriculture Vision dataset to address data scarcity issues and showcase the effectiveness of proposed representation frameworks. Furthermore, we demonstrate that the learned representations are powerful features for few-shot tasks in remote sensing and earth observation. We introduce GenCo, a generator-based representation learning framework that simultaneously pre-trains backbones and explores variants of feature samples. During fine-tuning, the auxiliary generator enriches the limited labeled data samples in the feature space. We validate the effectiveness of our method in enhancing few-shot learning performance on the Agriculture-Vision and EuroSAT datasets. Notably, our few-shot approach surpasses purely supervised training in both classification and semantic segmentation tasks trained over ten thousand images in the Agriculture-Vision Dataset. Lastly, we propose an intelligent nitrogen (N) management system utilizing deep reinforcement learning (RL) in conjunction with crop simulations through the Decision Support System for Agrotechnology Transfer (DSSAT). Initially, we framed the N management issue as an RL problem. Subsequently, we train management policies using deep Q-network and soft actor-critic algorithms, along with the Gym-DSSAT interface. This interface facilitates daily interactions between the simulated crop environment and RL agents. According to our experiments with maize crops in both Iowa and Florida, USA, the RL-trained policies surpass previous empirical methods.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Jing Wu

Owning Collections