Segmentação de imagens para análise do crescimento de alface em sistema hidropônico por aprendizado profundo

Abstract

Lettuce is a herbaceous plant that, in addition to its health benefits, is easy to cultivate yearround when grown in a protected environment. Among the cultivation methods, the hydroponic system stands out, offering advantages over soil-based cultivation. However, even under controlled conditions, issues may arise in plant development, such as poor formation, water stress, and nutrient deficiency. In this context, this study aimed to develop and apply image segmentation techniques for lettuce grown in a hydroponic system, focusing on the analysis and monitoring of growth throughout the development cycle. To achieve this, a dataset consisting of 1,570 annotated images of lettuce was created. Based on this dataset, different segmentation models were evaluated in order to compare their performance and suitability for the problem. Among the analyzed models, SAM 2.1 and SAM 3 stood out, along with YOLOv11 and YOLOv12, which were used for real-time detection and segmentation tasks. The T-Rex tool was also employed as a supporting resource. As a result, the SAM 2.1 model showed the best performance after supervised training, achieving Dice, Precision, and Recall values of 0.90, 0.98, and 0.84, respectively. From the segmented areas, it was possible to quantify the area occupied by the plants in the images in pixels and track its variation over time, enabling the monitoring of plant growth. These data were used to model growth using the Verhulst logistic model. The fitted model showed an average coefficient of determination of 0.978 and an average relative error below 7%, indicating a good fit to the observed data. It is concluded that the proposed approach is effective for monitoring plant growth in hydroponic systems, as it allows the automatic extraction of leaf area in pixels over time, in addition to using metrics such as Dice, Precision, and Recall to evaluate model performance, contributing to a more accurate assessment of plant development.