Improving the Accuracy of Cotton Water Stress Detection on DroneCaptured Data

Abstract

This dissertation addresses limitations in cotton water stress detection using UAV-captured RGB imagery and advanced deep learning techniques. Building upon Niu et al. (2024), who achieved 91% accuracy in cotton irrigation classification, this study develops an enhanced framework using over 21,000 UAV images collected across four sampling dates from experimental plots in Lubbock, Texas. We implemented an InceptionV3-based CNN with transfer learning and fine-tuning methodologies to distinguish between four irrigation management strategies: rainfed, fully irrigated, percent deficit irrigation, and time delay irrigation. Our two-phase training strategy combined ImageNet pre-trained features with selective layer unfreezing for domain-specific optimization. Results demonstrate exceptional performance improvements, achieving 96% overall accuracy across all temporal datasets compared to the 91% baseline. The model maintained temporal consistency with zero performance variance across sampling dates, indicating successful capture of phenology-invariant spectral signatures. Class-specific analysis revealed superior discrimination capabilities with rainfed conditions achieving 95-99% F1-scores and fully irrigated areas reaching 97-100% precision.