Skip to main content
eScholarship
Open Access Publications from the University of California

REMOTE SENSING OF WATER STRESS IN ALMOND TREES USING UNMANNED AERIAL VEHICLES

  • Author(s): Zhao, Tiebiao
  • Advisor(s): Chen, YangQuan
  • et al.
Abstract

The technologies of unmanned aerial vehicles (UAVs) and miniature cameras have been improved significantly in the last decade. It becomes more and more popular to use UAVs and onboard cameras to collect high resolution multispectral images with better operation flexibility. Although remote sensing using satellites or field scanners has been researched for many years, there is still lack of a workflow to fully explore the benefits of high resolution images from UAV-based remote sensing. In this dissertation, three key parts of the workflow are discussed: extraction of region of interest from high resolution images, extraction of related features for further classification or regression problems and optimization of UAV-based remote sensing practices, using almond tree water status detection as a case study.

Unlike satellites providing low resolution images, or field scanners with limited field of view, UAV-based remote sensing platforms can collect very high resolution images with flexible temporal resolution and spectral band configuration. To extract the region of interest more accurately, two types of methods are evaluated. One uses manual features such as color, texture and morphological features, and the other is based on the latest deep learning based instance-segmentation models.

After accurate extraction of region of interest, the next step is to extract application related information from these high resolution images. A methodology is proposed to convert the feature extraction problem to a dimensionality reduction problem. According to this methodology, moments, histograms and traditional dimensionality reduction methods are discussed based on the performances of irrigation treatment classification and almond tree variety classification, and stem water potential (SWP) regression. For SWP regression, different regression regularization methods are also experimented to extract more information from given tree canopy

pixels.

Finally, best remote sensing practices are discussed using irrigation treatment classification and variety classification as reference applications. Effects of spatial resolution, spectral configuration, band-to-band registration and image formats are evaluated in terms of classification accuracies.

Main Content
Current View