Accurate leaf segmentation is a critical task in plant biology and agriculture, enabling applications such as plant species recognition, disease detection, and health monitoring. However, segmenting leaves in natural environments remains challenging due to complex backgrounds, overlapping foliage, soil artifacts, shadows, and lighting variations. This thesis addresses these challenges by proposing an innovative semi-supervised framework that combines multi-layer graph-based propagation with deep learning techniques. The method begins by spatially localizing the leaf to identify a foreground template representing the central leaf region. The image is then decomposed into multiple levels of homogeneous regions to capture both fine and coarse details at various scales. A graph is constructed to model relationships between these regions, with weighted edges based on shared boundaries or overlapping areas. This graph enables the propagation of ranking scores from the foreground template to the image boundaries, effectively distinguishing the leaf from its background. In this vein, a saliency map is generated to isolate the leaf, and the resulting binary mask is refined using random forests to ensure optimal separation between the leaf and its surroundings. Extensive experiments on a widely used dataset demonstrate that our method outperforms several state-of-the-art segmentation techniques in terms of accuracy, robustness, and adaptability to complex natural scenes. By reducing the reliance on manual annotations and improving segmentation precision, this research contributes significantly to the field of computer vision applied to plant biology. The outcomes of this work have practical applications in areas such as automated crop monitoring, precision agriculture, plant growth analysis, and biodiversity studies, providing a reliable tool for researchers and practitioners in these fields.