Red clover is the second most valuable and intensively produced leguminous fodder crop in Hungary. Optimum harvesting should start when 75-80 per cent of seed-heads are brown and 90 per cent or more of the seeds are past the hard-dough stage of maturity. Harvesting must be done before an appreciable amount of seed-head deterioration begins. On fall plantings the first-year crop normally matures during the months of August and September. Seed from second-year stands, if not cut back, will mature somewhat earlier. The harvesting of red clover is divided into two distinct operations: (1) the curing or preparation and (2) the hulling and separation of the seed. The curing is by windrowing or contact desiccant spraying. It is the timing of each step that can make all the difference between a good yield and a mediocre one. The agronomists decide those actions with personal on-site inspections of the heads. Their decisions rely on random point sampling. A trend has developed with the recent interest in the harvesting time estimation of red clover in seed production to meet the demands of sustainable and precise agriculture. Unmanned aerial systems (UAS) are one of the current state-of-the-art tools when it comes to supporting farm-scale phenotyping trials. In this examination, we explored the red clover seed heads classification using deep learning on a UAS RGB aerial image. We assessed different ripening status of the flower heads with a single-shot detector (SSD) deep learning model. It is proposed that drone data collection was suggested to be optimum in this study, closer to the harvest date, but not later than the ageing stage. With SSD model we can calculate the number of heads on the vegetation surface but to differentiate them (e.g., purple-, light-brown-, dark-brown flower heads) with high confidence, we need more examinations, spectral bands or fine adjustments of the model with additional data.
Read full abstract