Semantic Mapping-driven Temporal Color Consistency Correction: A Case Study on Lettuce Fresh Weight Monitoring

Fluctuating light environments severely degrade the color consistency of time-series image datasets, preventing accurate phenotypic color representation and compromising detection accuracy. Existing light-environment control schemes are either prohibitively expensive or ill-suited for facility agriculture scenarios. This study proposes an efficient, semantic mapping-based color transfer algorithm for facility agriculture to improve non-destructive lettuce fresh weight detection. The method employs RoMa (Robust Dense Feature Matching) to establish cross-frame semantic correspondences, then applies a multi-percentile averaging strategy to extract stable color reference features. Subsequently, a global linear transformation is applied to align the colors of the frames to be corrected with reference frames. Simultaneously, a historical offset weighted inheritance mechanism is introduced to achieve smooth consistency correction across image sequences while preserving the integrity of target structures and backgrounds. Experimental results demonstrate that after correction, the color standard deviation and coefficient of variation of crop temporal images significantly decreased from 7.542 and 0.100 to 4.980 and 0.077, respectively. In the lettuce biomass estimation task, the color compensation process suppressed color noise fluctuations in temporal images and enhanced the correlation between color features and fresh weight, thereby universally improving the performance of various regression models. Without relying on hardware aids such as standard color cards, this research effectively suppresses ambient light interference and enhances data reliability. These findings establish a high-quality data foundation for non-destructive lettuce weight sensing and provide a promising basis for deploying lightweight phenotyping equipment in smart agriculture.