Prediction Estimation of Surface Area and Number of Leaves on Duckweed Focused on Leaf-Shape and Leaf-Color Features

In this paper, we propose a method to estimate the frond area and frond number of a floating macrophyte, duckweed (Lemna minor) using photographs as input. Duckweed is a fast-growing plant that holds potential for water purification and biomass production. Accurate recognition of plant growth is essential for its effective utilization. The frond area and the frond number of duckweed are used as indicators of duckweed growth. The proposed method consists of three parts: (1) colony segmentation, (2) colony health estimation, and (3) leaf number estimation within the colony. The term “colony” is a state where one or more leaves of duckweed are gathered together. For (1) colony segmentation, we use a technique called Mask R-CNN, which is an instance segmentation method. For (2) colony health estimation, we constructed a classifier using ResNet152 and a classification layer, taking the image regions recognized as colonies as input. Using these methods, the frond area is estimated from the regions believed to be healthy colonies. Furthermore, for the recognized healthy colonies, we performed (3) the frond number estimation using ResNet152 and a regression layer. During the feature extraction, we made efforts to focus on the shape of the colonies. To facilitate training and testing, we cultivated duckweed and created a dataset through photo collection and annotation. For evaluation, we used Mean Absolute Percentage Error (MAPE). Regarding the performance of frond area estimation, MAPE was 8.8% for existing methods, while our proposed method achieved 3.1%. As for frond number estimation, MAPE was 18.6% without incorporating the shape-based feature extraction, but it improved to 17.2% with the proposed technique.