Files in this item



(no description provided)PDF


Title:Learning-based crop management optimization using multi-stream convolutional neural networks
Author(s):Ormiga Galvao Barbosa, Alexandre
Director of Research:Hovakimyan, Naira; Martin, Nicolas
Doctoral Committee Chair(s):Hovakimyan, Naira
Doctoral Committee Member(s):Salapaka, Srinivasa; Voulgaris, Petros; Shi, Humphrey
Department / Program:Mechanical Sci & Engineering
Discipline:Mechanical Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Convolutional Neural Networks
Deep Learning
On-farm Research
Yield Modeling
Risk-averse Optimization
Deep Ensemble
Reinforcement Learning
Coverage Path Planning
Imitation Learning
Abstract:Improving crop management is an essential step towards solving the food security challenge. Despite the advances in precision agriculture, new methods are needed to create decision-support systems to help farmers increase productivity while accounting for environmental impacts and financial risks. This dissertation presents a class of learning-based optimization algorithms for spatial allocation of crop inputs, and a new framework for online coverage path planning with potential use in tasks such as planting and harvesting. The proposed algorithms use Multi-stream Convolutional Neural Networks (MSCNN) to learn relevant spatial features from the environment and use them to optimize the available control inputs. In the crop inputs optimization problem, an MSCNN combines five input variables as in a regression problem to better predict yield. The predictive model is then used as the base of a gradient-ascent algorithm to maximize a custom objective function. To leverage the applicability of this algorithm, a risk-aware version of this method is also proposed. The predictive uncertainty is measured and used as a constraint to comply with different levels of risk-aversion. Experiments with real crop fields demonstrate that this method significantly reduces the yield prediction errors when compared to the state of the art algorithms. Results from the optimization algorithm show an increase in the expected net revenue of up to 6.8% when compared with the status quo management while providing safety bounds. In the coverage path planning framework, an MSCNN agent learns a control policy from demonstrations of paths obtained offline through heuristic algorithms, by using imitation learning. The resulting control policy is further improved through policy-gradient reinforcement learning. Simulations show that the improved control policy outperforms the offline algorithms used during the imitation learning phase, and that the proposed framework can be easily adapted to different cost functions.
Issue Date:2020-07-06
Rights Information:Copyright 2020 Alexandre Ormiga Galvao Barbosa
Date Available in IDEALS:2020-10-07
Date Deposited:2020-08

This item appears in the following Collection(s)

Item Statistics