Morphological characterization of Liberica coffee seedlings via machine vision

Abstract

A crux component in crop health monitoring is the morphological characterization that directly manifests the plant response to the environment and farm inputs like fertilizers and pesticides. A healthy plant presents firm leaves, well-developed flowers, fruits, and a root system. Leaf count, color, and canopy coverage can help characterize general crop health. In this study, an aerial drone is allowed to traverse an optimal path plan while positioned near an area of interest where several overhead images of crops are captured. The total collected aerial images count to 1,918 which is divided into training and testing sets with a 70:30 distribution ratio. From these aerial images, coffee seedlings are detected from the background, weeds, and other crops in the field using the VGG-16 model trained to recognize coffee seedlings. Once localized, leaf counting is performed using segmentation, while the canopy coverage estimation uses a patch-based DNN model to calculate the relative coverage concerning the overhead leaf area. Since leaf color extracted from an RGB image is very much affected by ambient light, normalization using an enhanced Triangular Greenness Index (eTGI) is implemented. The estimation results of the system reached up to 91.48% accuracy, 92.52% precision, and 93.82% recall for detecting coffee seedlings while the Mean Absolute Percentage Error (MAPE) for leaf count and canopy coverage of 11.61% and 15.67% respectively. For future work, the leaf color can be correlated to chlorophyll and percent nitrogen measurements which will require specialized instruments for validation. Estimation of chlorophyll and percent nitrogen is vital in identifying the amount and type of fertilizers to be applied.

__________________________________________

*Corresponding author

Email address: antonlouise.deocampo@g.batstate-u.edu.ph