Application of local binary patterns in digital images to estimate botanical composition in mixed alfalfa-grass fields

Botanical composition in mixed stands of alfalfa and grass is a critical parameter in equations estimating harvest fiber concentration for dairy rations. Composition is difficult to estimate by visual observation. Digital image analysis in mixed stands could reduce botanical composition uncertainty and improve spring harvest management decisions. Mixed stands were sampled (n = 168) in farmers’ fields in Tompkins County, New York in May 2011. A digital image was taken of standing samples at 5-Megapixels resolution using a Canon PowerShot A3100IS, and alfalfa and grass height relationships were recorded. After clipping representative samples at 10-cm above ground level, samples were manually separated into alfalfa (Medicago sativa L.) and timothy grass (Phleum pratense L.), and dried to calculate fractions on a dry matter basis. Uniform rotation invariant local binary patterns (LBP) were extracted from whole images and 64 × 64 pixel tiles, and were used to develop regression equations estimating grass fraction. Tiles were manually classified as alfalfa (0), grass (1) or unclassifiable. An iterative process selected most accurate local binary pattern operator settings. Grass fraction was estimated in three regression model development approaches: (1) using average tile LBP histogram bins from whole images and botanical height relationships, (2) developing a binary tile classification model from tile LBP histogram bins, and using tile model-predicted grass probability averaged for tiles in whole images (grass coverage estimate) and botanical height relationships as inputs in whole image models, and (3) using LBP histogram bins extracted directly from whole images (1024 by 1024 pixel square) and height relationships. Predictive accuracy in whole image models using tile LBP histogram averages was highest for models generated from LBP tile histogram bin means (R²_pred up to 0.847), followed closely by combined tile models and whole image models (R²_pred up to 0.807), with pairwise correlations between tile model-generated grass coverage estimates and sample grass fraction up to 0.895. Local binary patterns are effective in differentiating alfalfa and grass under field conditions, because the method is robust to changes in color and illumination. Furthermore, key LBP histogram bins (e.g., symmetric edges) strongly differentiate alfalfa and grass in tiles. The LBP method is promising based on this study, but further evaluation under diverse field conditions, including different cameras and grass species, is necessary to assess usefulness.