Overcoming difficulties in segmentation of hyperspectral plant images with small projection areas using machine learning.

Segmentation of hyperspectral image data is a well-established technique in remote sensing. While it is commonly applied to individual field crops, its use for individual trees is less prevalent. Conifers are crucial in forestry, and assessing physiological status, or genetic diversity is required for effective early-age treatment in nurseries and hyperspectral imaging (HSI) combined with high-throughput phenotyping (HTP) offers faster and non-destructive evaluation. NDVI-based thresholding is sufficient for detection of leaves with large projection areas, but needles of conifers present challenges due to spatial resolution constraints and increased proportion of border pixels. This study monitored the offspring of three locally adapted Scots pine (Pinus sylvestris L.) populations, representing distinct upland and lowland ecotypes. This study presents a hyperspectral image processing pipeline for segmenting and isolating individual Scots pine seedlings. Using a K-means algorithm, 23 hyperspectral centroids were successfully derived and subsequently classified into ten biologically distinct groups. Random forest classification model effectively differentiated Scots pine seedlings based on origin during water stress and recovery periods. This study highlights the potential of hyperspectral imaging and machine learning in evaluating the physiological state of conifer seedlings, demonstrating promising applications in forest tree physiology research and tree breeding.