NLA: Natural Light Augmentation Method for Crop Image Datasets

Data augmentation is a key factor for improving the performance of object detection models in computer vision.Agricultural imagery is captured in open fields and greenhouses, and lighting varies widely with shooting time, weather conditions, and the influence of structures.These characteristics destabilize the distribution of training data, and when conventional augmentations generate data that differ from real environments, they can degrade a model's generalization.To address this problem, we propose Natural Light Augmentation (NLA), a lighting-aware augmentation that reflects illumination changes in agricultural images.NLA simulates variations in brightness, color temperature, and contrast so that conditions likely to occur in practice are applied to the training data.We validated the effectiveness of the proposed method on Oriental melon and strawberry datasets.On the Oriental melon dataset, NLA improved overall performance compared with existing augmentations, with the largest gain observed on mAP50-95 (4-6%).On the strawberry dataset, the Baseline model showed a decline in validation performance in the later training phase-suggesting overfitting-whereas NLA maintained stable performance as epochs progressed.We additionally conducted experiments with PatchSwap, which exchanges patches within the same image, but the performance improvement was limited.These results demonstrate that augmentation reflecting real lighting conditions in agricultural datasets effectively improves model performance and increases training stability.As an augmentation tailored to agricultural imagery, NLA should be extended to a wider range of crops and environmental conditions in future work.