Legume content estimation from UAV image in grass-legume meadows: comparison methods based on the UAV coverage vs. field biomass

Legume content (LC) in grass-legume mixtures is important for assessing forage quality and optimizing fertilizer application in meadow fields. This study focuses on differences in LC measurements obtained from unmanned aerial vehicle (UAV) images and ground surveys based on dry matter assessments in seven meadow fields in Hokkaido, Japan. We propose a UAV-based LC (LC_UAV) estimation and mapping method using a land cover map from a simple linear iterative clustering (SLIC) algorithm and a random forest (RF) classifier. The SLIC-RF classification achieved a high accuracy level for four different ground cover types (grasses, legumes, weeds, and background) in seven distinct meadows with an overall accuracy of 91.4% and an F score of 91.5%. By applying SLIC-RF to eliminate plots with low classification accuracy, we demonstrate the necessity of achieving a minimum classification accuracy of 0.82 for precise LC estimation. A non-linear relationship was revealed between the LC_UAV and LC_BM influenced by surface sward height (SSH, height of plant canopy). The results indicate a higher accuracy of the LC_BM estimation when SSH levels were lower, particularly when recommending SSH levels below 40 cm for optimal LC_BM estimation. This highlights the effectiveness of UAV-based remote sensing for assessing early growth or grazing in pastures with low SSH.