NJ-Unet: A novel Unet-based greenhouse tomato leaf segmentation neural network and its application

Accurate estimation of leaf area index (LAI) is fundamental for understanding canopy dynamics and yield formation in greenhouse-grown tomato, yet remains hindered by severe occlusion and structural complexity. However, conventional approaches-such as extinction coefficient-based methods that rely on top–bottom measurements and threshold-based image segmentation with fixed cutoffs-struggle to delineate occluded regions and thus show poor generalization. To overcome these limitations, this study developed an interpretable deep learning framework that integrates porosity-based structural modeling with an enhanced segmentation network, NJ-Unet, for precise and scalable LAI estimation. NJ-Unet incorporates a Smooth Maximum Unit (SMU) activation function and a Spatial-Channel Self-Attention (SCSA) module into the Unet backbone, substantially improving segmentation accuracy and efficiency under complex canopy conditions. Based on segmented leaf masks, a concentric ring-based porosity analysis method was employed to estimate LAI, achieving high estimation efficiency and outperforming optical sensor and destructive sampling methods (R 2 = 0.72, RMSE = 0.73, MAE = 0.54). Furthermore, LAI values were used to predict tomato yield across developmental stages, with the strongest correlation observed during the fruiting stage (R 2 = 0.90). Altogether, this study demonstrates how coupling interpretable deep learning with porosity-based modeling to enhance the accuracy, efficiency, and applicability of LAI retrieval and yield estimation in protected horticultural systems.