Grass-legume mixed swards are beneficial for feed nutrients and nitrogen fixation by rhizobia in legume roots. Since information on the proportion of legume is crucial, it is typically evaluated through yield surveys or expert visual assessments. However, yield surveys require significant labor, and visual assessments are subjective. Therefore, an efficient and objective evaluation method is required.
In this study, we developed CoverageMapMaker, a mapping system for estimating legume forage coverage in large-scale mixed cropping fields using drones and deep learning. To create a high-accuracy CNN model, we evaluated different training patch sizes and found that a 128×128 pixel patch size was more suitable. Data were collected over seven different days, resulting in eight datasets (each daily dataset and a combined dataset). Using these eight datasets under 20 different training conditions, a total of 160 CNN models were created to select high-accuracy CNN models. The model trained on the combined dataset demonstrated the highest generalization performance.
