Lowe, Amy
(2018)
Early detection of stress in strawberry plants using hyperspectral image analysis.
PhD thesis, University of Nottingham.
Abstract
Strawberry plants produce one of the highest quantities of soft fruits in the UK. The plants are grown in fields and glass houses where the environment is hard to completely control. There are a variety of biotic and abiotic stresses that affect the plants' production rate, lifespan and the quality of the fruit. It is possible to mitigate these stresses, but first they must be detected as early as possible.
Drought, or water deficit, is an environmental stress which impacts the plants' productivity. The stress can be monitored by looking for visible signs, but detection by a person can occur too late. Using technology could improve the detection before a person can see the signs. One such technology is hyperspectral imaging. Hyperspectral imaging has the potential to detect certain features in plants by examining their reflectance spectrum. In this work, a spectral range from visible to near-infrared will be used to record reflectance changes from plants during drought experiments.
The data collected in this thesis comprises strawberry plants undergoing drought conditions. Initial inspection of the data suggests that a difference in reflectance may exist throughout this time period, but the data is noisy and changes subtle: using whole-plant measures may not be as accurate as selecting individual leaves. Also, particular leaves may be more suited to hyperspectral inspection than others. Therefore, the work in the rest of the thesis focuses on automatically selecting suitable quality leaves from which to derive hyperspectral measures.
Leaf locations are selected automatically by combining 2D spatial information and select wavelengths from the hyperspectral data to extract leaf features (leaf centres and vein patterns), and this information is used to initialise a level set shape model to segment the leaves. The results from the 2D segmentation are evaluated, and results suggest clearly visible leaves can be segmented well automatically. However, it is hypothesised that further information is required to select the best possible leaves for analysis from the set of segmented leaves, so 3D plant information is then incorporated into the approach.
The orientation of leaves, and shadows, are likely to influence the spectral signatures, and occluded leaves are undesirable. Therefore, the leaf orientation, position and height are taken into consideration when selecting leaves for measurement. To do this, 3D surface models are created from multiple digital images captured at the same time as the hyperspectral data. The hyperspectral data is then mapped onto the model and suitable leaves are identified based on their orientation, height and distance from the centre of the plant. This allows for selection of a subset of leaves as sources of hyperspectral measurement for each plant. This approach, from segmentation to leaf selection and measuring, is built into a fully-automated pipeline, and evaluated.
The resulting leaves are selected from the other time points in the time series data, and the spectral signatures are compared to the initial manual inspection. The automatically selected leaf data show similar spectral responses to the manual approach. The challenge of detecting changes in these spectral profiles is investigated and discussed.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)