Use of digital images to disclose canopy architecture in olive tree

Abstract The use of digital cameras for monitoring natural vegetation and agricultural ecosystems is particularly attractive since it necessitates neither expensive equipment nor extensive skill. In this study we tested the use of digital images (DIs) to generate 3D plant reconstruction for retrieving the main plant architectural features (leaf area, leaf inclination, leaf azimuth) of olive tree branches. High resolution image of tree branches were firstly used to generate 3D reconstruction of plant structures using a Structure From Motion approach; we therefore answered the question whether these 3D models may be segmented to discriminate main plant structures (leaves and branch) proposing a simple classification algorithm (Random Forest, RF) with saliency features and color indices as predictors. Finally, on the good and robust performances of the proposed classification algorithm, the single leaves were analyzed to retrieve the relevant area, inclination and orientation and compared to the relevant observed data. The calibration of the RF model indicated that color indices better discriminated leaves and stem than the sole use of saliency features. The classification performances were further improved by tuning the scale at which saliency features were calculated and by filtering the final result to reduce misclassified points. A RF model calibrated on a single plant was successfully applied to 5 others, indicating the robustness of the calibration strategy. The analysis of single leaves, as segmented after the classification process, indicated that plant architecture was satisfactorily reproduced with strong correlations obtained between measured and calculated values of leaf inclination and azimuth, while biases were observed for leaf area. These results emphasize the effectiveness of SFM in reproducing complex arrangements of leaves like on an olive tree. The use of such a system can be therefore suggested as a first step towards an improved low cost plant phenotyping platform to speed up our understanding of plant responses to environment. Further experiments are required to test the effectiveness of the approach also under outdoor conditions.