Accurate quality detection of fruits and vegetables is crucial for optimizing harvest timing, minimizing post-harvest losses, and reducing waste. This research aims to integrate remote-sensing and deep learning (DL) technologies to develop and evaluate object detection models employing a novel DL approach for precision agriculture through automated quality detection in fruits and vegetables. To achieve this, twelve state-of-the-art object detection models from the MMDetection framework were trained by utilizing a custom-created and annotated dataset that comprises 1535 images and 39 classes of fruits and vegetables categorized into unripe, ripe, and overripe qualities. To evaluate the performance of each model, metrics like loss, mean Average Precision (mAP), receiver operating characteristic (ROC) curve, area under the curve (AUC), and confusion matrix were employed. The results determined that the Detection Transformer with Improved Denoising Anchor Boxes (DINO) and Dynamic Denoising Query (DDQ) models outperformed the others, achieving a mAP of 0.65 and a loss of 1.8 and 1.9, respectively. These metrics demonstrate their ability to distinguish the quality of fruits and vegetables accurately. These findings highlight the potential of DL models for real-world agricultural applications, as they facilitate timely quality assessment and contribute to the development of intelligent solutions.

Lettuce (Lactuca sativa L.) is one of the most widely cultivated leafy vegetables worldwide, with growing adoption in greenhouse hydroponic systems where precise nutrient management is essential. Crop production in these controlled environments is highly sensitive to nitrogen, phosphorus, and potassium deficiencies. The resulting foliar symptoms, such as chlorosis, necrosis, and morphological alterations, evolve alongside the growth cycle, making early detection vital for optimized crop management. Yet, traditional diagnosis relies on specialized visual expertise or destructive chemical analyses, both of which are costly and incompatible with continuous monitoring. RGB-image classification with deep learning has emerged as a non-invasive alternative, with Convolutional Neural Network (CNN)-based studies reporting accuracies above 96 %. However, these figures come from random-split protocols that may place near-identical images of the same plant in both training and test sets, leaving open the question of whether such models generalize to later phenological stages unseen during training. We address this question with twocontributions. First, we introduce a chronological-split protocol that assigns early growth-stage images to training and late-stage images to testing, quantifying the temporal robustness that random splits fail to measure. Second, we introduce LettuceDeF, a public dataset of 7,297 RGB images spanning four cultivars and seven diagnostic classes, with ground-truth labels validated by Inductively Coupled Plasma Optical Emission Spectroscopy (ICP-OES). Using this protocol and dataset, we compare four vision foundation models (DINOv3, EVA-02, SigLIP, and BEiT-3) against nine representative CNNs from the literature. Each of the 26 model–protocol combinations is trained with 12 independent seeds, and comparisons use paired t-tests with Bonferroni correction. Under the random split, the three best vFMs and ConvNeXt Base were statistically indistinguishable (F1-macro ≈ 0.997). The chronological split broke this parity: EVA-02 led with F1 = 0.928 and the lowest degradation (7.03 percentage points, pp), while the nine CNNs degraded by 14.26 pp on average—1.6× more than the top three vFMs (8.71 pp). EfficientNet-B3 was the notable exception, degrading only 9.55 pp and rivalling SigLIP and DINOv3. The chronological split provides a more realistic estimate of deployment performance and reveals robustness differences between model families that random splits cannot detect.