COMMITTEE CHAIR: Dr. Lijun Qian
TITLE: VEGETATION HEALTH PREDICTION USING RGB IMAGES FROM UAV WITH DEEP LEARNING
ABSTRACT: Vegetation health prediction is one of the most important research questions in precision agriculture. The Normalized Difference Vegetation Index (NDVI) is the most widely adopted indicator of vegetation health, such as crop vigor and stress. Traditional NDVI estimation relies on multispectral or hyperspectral imaging, which is expensive and inaccessible for small-scale farming. Recent advances in deep learning enable the prediction of NDVI from low-cost RGB imagery; however, the limited availability of paired RGB-NDVI datasets and domain discrepancies between existing sources hinder robust generalization. In this thesis, we construct UAV-based RGB-NDVI paired datasets and investigate multiple deep learning architectures, including a Fully Connected Neural Network (FCNN), an RGB-Split Convolutional Neural Network (RSCNN), and an Autoencoder with Dual Loss (ADL). Our experimental results show that the RSCNN consistently delivers the strongest regression performance across datasets, achieving R2 values of 82.43% and 89.53%, and the lowest MAE of 1.87% and 1.57% on Dataset 1 and Dataset 2, respectively. Furthermore, we studied transfer learning and proposed a self-learning approach that relied on pseudo-labeled target samples. The results demonstrate improved performance that varied with the chosen confidence threshold and the transfer direction.
Room Location: Electrical Engineering Conference Room 315D