Abstract: The plant leaf area can serve as an indicator of the plant's growth rate, nutrient uptake, and photosynthetic efficiency. However, some challenges are still remained to accurately segment the leaf areas and subsequent measurements, particularly with the leaves bearing some distinct characteristics, such as the serrated edges found on cucumber leaves. It is a high demand to promote the reliability of research outcomes and practical applications in the field of agriculture. In this study, an innovative "Marm" model was introduced to measure the cucumber leaf area from the realm of deep convolutional neural networks (CNN). The architecture of Mask R-CNN was also extended to incorporate the Sobel operator. The intricate contours of cucumber leaf edges were capture and detect using the Sobel operator with the enhanced edge details. A suite of mask was obtained to fully depicted the intricate topography inherent in these leaves. Among them, the edges were acted as a connective thread weaving precision into the process. The architecture of the model was enhanced to introduce a tailored component of edge loss. A strategic addition was then used to improve the segmentation accuracies. A weight factor was introduced into the new loss component, in order to balance among the existing loss functions and potentially computational costs. The better performance of the improved model was then effectively enhanced with the pragmatic feasibility. A comprehensive strategy was employed to leverage a synergistic blend of model-generated mask images and validated reference object labels. The morphological intricacies and contextual validation were integrated to facilitated the steadfast and reliable measurements of cucumber leaf area across various growth stages. A better resilience of data and validation was then established to deftly navigate the diverse terrain of leaf morphology from germination to maturity. The precision, recall, and Intersection over Union (IoU) scores reached 99.1%, 94.87%, and 92.18%, respectively, after the empirical validation, indicating over the benchmarks of the original Mask R-CNN architecture. The performance was improved by 1.28, 1.13, and 1.05 percentage points in the precision, recall, and IoU scores, respectively. There was also the substantial reduction of 1.43 percentage points in the error of area measurement, suitable for the complexities of leaf edge segmentation. Although there are various factors affecting the images, such as occlusion and shadows, the error rate in area measurement can be controlled to around 5.45%. In conclusion, the improved Marm model can be expected to use for the plant biometric measurements with the higher accuracy than before. The serrated leaf edges can also be accurately segmented for the fully understanding of the plant growth dynamics and informed agricultural practices. The intricate scenarios of leaf edge can be captured to extend the promising transformative applications in the diverse fields, including the computer vision in modern agriculture. As the technological advancements empowered agriculture, this finding can significantly contributed to the ongoing evolution of sustainable food production. The landscape of advanced techniques can be evolved for the plant phenotype. The improved model can also be utilized to navigate the complexities of leaf edge segmentation in the field of botanical sciences.