Lesquerella seed yield estimation using color image segmentation to track flowering dynamics in response to variable water and nitrogen management

Abstract

Seed oil from lesquerella (Physaria fendleri (Gray) O’Kane & Al-Shehbaz) can potentially supplement castor oil as a non-petroleum-based chemical feedstock in the production of many industrial products. However, before lesquerella will become commercially viable, further efforts are needed to address crop management challenges and to improve lesquerella varieties. Because lesquerella develops vibrant yellow flowers on top of the canopy, digital imaging can be used to track the dynamics of its indeterminate flowering period. The objective of this study was to investigate a digital image analysis approach to (1) assess differences in lesquerella flowering dynamics due to variable water and nitrogen (N) management and (2) estimate lesquerella seed yield from flowering data. During the winters of 2011–2012 and 2012–2013, field experiments tested lesquerella responses to two irrigation levels and six N fertilization rates at Maricopa, Arizona. Biomass was sampled within a 30cm×30cm area twice per month, and lesquerella flowers were manually counted. Twice per week, digital images were collected with a commercial digital camera at a nadir view angle approximately 2m above the canopy. To obtain the percentage of yellow flowers in each image, an analysis routine included (1) an image transformation to the hue, saturation, and intensity (HSI) color space and (2) a Monte Carlo approach to address uncertainty in HSI parameters used for image segmentation. The imposed irrigation and N fertilization treatments led to differences in both flower count and flower cover (p<0.05). However, the digital imaging approach permitted more frequent measurements, which revealed fine temporal changes in flowering patterns that could be explained by management factors. Due in part to the larger sampling area for the digital imaging approach, lesquerella seed yield was better estimated using flower cover percentage (r2≤0.84) from images as compared to manual flower counts (r2≤0.56). Overall, the digital imaging approach provided useful information on lesquerella flowering dynamics, which was affected by water and N management and highly correlated with seed yield.