Fatih Bozkurt, Mustafa Teke, Fecir Duran
For high value-added aromatic plants such as lavender, accurate wet biomass estimation is critical in order to minimize essential oil losses caused by waiting time in post-harvest distillation processes and to optimize logistics operations. This study proposes a non-destructive machine learning model using Unmanned Aerial Vehicle (UAV) based RGB images to overcome the existing methodological limitations in the literature. Within the scope of the research, extended HSV segmentation and elliptic morphological filtering were applied on 224 images obtained from 64 different individual lavender plants. Logarithmic volume attributes were extracted from the pixel areas normalized according to the flight altitude and the Random Forest algorithm was trained on the dataset, which was expanded to 896 samples with data augmentation. The analyses revealed that the integration of traditional two-dimensional spectral indices such as CIVE and VARI into the model relatively weakened the predictive performance (explained variance of 58.1%). In contrast, the model based on the proposed three-dimensional volumetric features demonstrated a Mean Absolute Error (MAE) of 171.55 grams, explaining 59.74% (R2) of the variance. This model offers the potential of a decision support system for agricultural yield planning and industrial process management.