The technical challenge is deceptively subtle. A fresh betel leaf shows saturated green pigmentation with a moist, high-specular surface; a moderate leaf begins to yellow and desiccate; a stale leaf presents brown discoloration, wrinkled texture, and partial tissue breakdown. These transitions are continuous rather than discrete, producing large intra-class variance and small inter-class margins that confound threshold-based color analysis and handcrafted feature descriptors. Deep convolutional neural networks sidestep this limitation by learning hierarchical feature representations directly from pixel data, capturing spectral, textural, and morphological cues simultaneously. But standard CNNs that achieve strong accuracy on general benchmarks carry parameter counts in the tens of millions, which is impractical for edge deployment on farm-level embedded hardware.

BeLFG-Net’s core computational trick is the depthwise separable convolution, a factorization popularized by MobileNet and refined in MobileNetV2 and Xception. Instead of applying a single k-by-k spatial convolution across all input and output channels at once, the separable form splits the operation in two: a per-channel depthwise convolution applies one small spatial filter independently to each channel, and a subsequent pointwise 1-by-1 convolution mixes information across channels. The combined computational cost of this two-step approach relative to a standard convolution reduces to roughly 1/Cout + 1/k², where Cout is the number of output channels and k the kernel size. For the 3-by-3 kernels and 128-channel layers used in BeLFG-Net’s second block, the authors calculate a ratio of approximately 0.119—meaning an 8.4-fold reduction in computation relative to the equivalent standard convolution, with the saving approaching 8 to 9 times as channel counts grow.

But BeLFG-Net does not simply copy MobileNet’s playbook. The authors describe a hybrid block design that begins each of its four encoding stages with a full standard 3-by-3 convolution to establish a rich channel-wise feature space at the original spatial resolution, and only then applies the depthwise separable pair for efficient cross-channel mixing. MobileNet and MobileNetV2, by contrast, apply separable or inverted-residual blocks from the very first layer, which the authors argue fails to establish the rich initial feature representation that their hybrid approach provides. Each block follows a consistent internal structure: a standard convolution with batch normalization and ReLU activation, a depthwise convolution, a pointwise convolution with batch normalization and ReLU, a 2-by-2 max-pooling operation that halves spatial resolution, and a dropout layer. Filter counts progress through the blocks as 64, 128, 256, and 128, progressively reducing spatial resolution while building channel depth.

After the four encoding blocks, global average pooling aggregates the final 14-by-14-by-128 feature map into a compact 128-dimensional vector by averaging each channel’s spatial values. This step provides implicit regularization by replacing spatially indexed feature maps with class-discriminative spatial averages, and it dramatically reduces the parameter count of the fully connected classification head relative to standard flattening. The classification head then produces class logits through a three-layer fully connected network that progressively narrows from 512 to 256 to 3 units, enabling the network to construct class boundaries in a graduated dimensionality reduction. The entire model totals 979,651 parameters—roughly 0.98 million, or more than 58 times fewer than AlexNet and more than 26 times fewer than ResNet-50.

Training the network required a purpose-built dataset, since publicly available betel leaf imagery is dominated by labeled disease categories rather than freshness grades. The authors collected 766 original RGB photographs from betel farms in the Cox’s Bazar district of Bangladesh using a handheld consumer smartphone under ambient daylight, with no studio lighting, flash, or controlled backdrop. Domain experts assigned freshness labels at the time of capture according to agronomic criteria: fresh leaves exhibit intact, saturated green tissue with no visible discoloration; moderate leaves display the onset of yellowing or mild surface wrinkling consistent with 24 to 48 hours of post-harvest storage; stale leaves show brown discoloration, significant wrinkling, or partial tissue necrosis consistent with storage beyond 48 hours. The resulting class distribution comprised 292 fresh, 181 moderate, and 293 stale images.

A crucial methodological safeguard addresses a common pitfall in augmented deep learning datasets. The 766 original photographs were assigned unique identifiers and split 80/20 with stratification on the three freshness labels before any augmentation, yielding 613 training identifiers and 153 held-out test identifiers. Offline augmentation—random rotation within ±15 degrees, horizontal and vertical flips, shear, zoom, translation, brightness jitter, and contrast modification—was then applied only within each partition, so every augmented view inherits the split of its parent image and no leaf can contribute to both training and test sets. This identifier-level split eliminates train-test leakage, a necessary condition for an honest accuracy estimate. The augmented dataset grew to 6,262 views: 5,010 training views from the 613 training identifiers and 1,252 held-out views from the 153 test identifiers. The membership lists, the manifest of all views with their transform parameters, and a verification script accompany the paper so the split can be audited independently.

Trained with the Adam optimizer at an initial learning rate of 10⁻⁴ for 100 epochs with a batch size of 32, BeLFG-Net reached 90% training accuracy by epoch 30 and converged to 97.8% training accuracy and 95.72% validation accuracy at epoch 100, with validation loss descending smoothly from 1.02 to 0.2647. The generalization gap of roughly 2.1 percentage points between training and test accuracy is modest and consistent with the regularization provided by progressive dropout and batch normalization. Per-class performance was remarkably consistent, with F1-scores of 0.96 for fresh, 0.96 for moderate, and 0.95 for stale. The moderate category presented the highest classification difficulty due to its transitional nature—moderate specimens share visual characteristics with both fresh and stale classes—yet the model still achieved 95% recall and 97% precision for this category. Of the 1,252 test samples, the model misclassified just 54, with most errors occurring at fresh-moderate and moderate-stale boundaries, consistent with the continuous rather than discrete nature of the freshness transition.

Comparative analysis against established baselines trained from scratch on the same dataset and protocol shows BeLFG-Net outperforming every alternative. AlexNet achieved 86.34% accuracy with 57 million parameters; VGG-16 reached 88.57% with 138 million; MobileNet 90.21% with 4.2 million; MobileNetV2 91.65% with 3.4 million; InceptionV3 92.40% with 23.8 million; and ResNet-50 93.77% with 25.6 million. BeLFG-Net topped them all at 95.72% with just 0.98 million parameters. The authors attribute this counterintuitive result—fewer parameters yielding higher accuracy—to three architectural factors: the hybrid block design that combines the representational benefits of standard convolution with the parameter efficiency of separable factorization; a progressive dropout schedule with rates of 0.25, 0.25, 0.25, and 0.50 across the four blocks that imposes graduated regularization strongest at the point of highest channel density; and global average pooling that provides a channel-wise spatial summary naturally aligned with the class-discriminative regions of the betel leaf surface. Larger models trained from scratch on small datasets, the authors note, fail to generalize due to overfitting.

An ablation study dissects the contribution of each design component. Starting from a minimal CNN baseline with standard convolutions only, fixed dropout, and flattening, which achieved 88.96% accuracy with 1.82 million parameters, the introduction of depthwise separable blocks provided the largest single-component improvement of 2.69 percentage points, reducing parameter count to 0.98 million while simultaneously improving accuracy. This reflects the regularization effect inherent in the separable factorization: by constraining spatial and channel-wise operations to be orthogonal, the model is less likely to overfit spurious correlations in the training set. Progressive dropout added a further 1.75 percentage points, and replacing flattening with global average pooling contributed the remaining 2.32 percentage points, confirming that spatial averaging provides better generalization than position-specific flattening on a dataset of this size.

The authors are candid about the boundaries of their evidence. All 766 original photographs were collected from a single growing district during one harvest window between January and March 2023, with a single handheld smartphone, making the dataset homogeneous in cultivar, agronomic practice, soil and microclimate, seasonal stage, and imaging sensor. The 95.72% held-out accuracy should therefore be read as an estimate of within-distribution performance for this locality, sensor, and season—not as evidence of cross-region, cross-cultivar, or cross-season accuracy. The authors also note that augmentation multiplies views but adds no new leaf morphology, no new senescence trajectory, and no new disease state; the effective number of independent biological samples remains 766. They anticipate measurable degradation under different sensors with different white-balance or saturation responses, under indoor tungsten or mixed-spectrum lighting that introduces a color cast mimicking the onset of yellowing, and under presentation conditions such as wet leaves with strong specular highlights or stacked overlapping leaves encountered in bulk handling. Disease lesions can also produce brown discoloration that visually resembles staleness, and because the dataset is labeled for freshness only, the network has no mechanism to separate senescence from pathology.

Senescence itself is continuous, whereas the three grades are discrete bins imposed on it, and the authors acknowledge that leaves near a bin boundary are intrinsically ambiguous. Part of the reported error is therefore attributable to the labeling scheme rather than to the model. Because grades were assigned by expert judgment at the time of capture without repeated independent gradings, inter-rater agreement for this dataset is not quantified, and label noise near the boundaries remains unmeasured. The authors recommend that BeLFG-Net be calibrated on locally collected images before use at any new site, rather than deployed on the strength of the accuracy reported here, and they propose leave-one-farm-out and leave-one-session-out evaluation to measure cross-domain transfer directly.

Looking ahead, the research team outlines several directions: on-device and packing-house deployment with latency and energy measurement on target hardware; collection of additional original identifiers from independent farms in other districts, cultivars, and seasons; replacement of the three-class scheme with a continuous freshness index together with repeated expert gradings to quantify label noise; density-adaptive feature extraction to handle illumination and maturity heterogeneity in open-set field conditions; and semi-supervised pretraining strategies to leverage unlabeled betel leaf imagery from farm monitoring systems. For now, BeLFG-Net stands as a demonstration that careful architectural design—hybrid convolutions, progressive dropout, and global average pooling—can outperform larger standard models when the training dataset is compact, offering a glimpse of a future where the freshness of a betel leaf is judged not by a weary human eye but by a network small enough to live on a chip in the field.

Subject of Research: Deep learning

Article Title: BeLFG-Net: A lightweight betel leaf freshness grading network via hybrid depthwise separable convolutions

Article References: Haque, A. U., & Ur Rahman, F. (2026). BeLFG-Net: A lightweight betel leaf freshness grading network via hybrid depthwise separable convolutions. Smart Agricultural Technology, 15, Article 102597. https://doi.org/10.1016/j.atech.2026.102597

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102597

Keywords: BeLFG-Net, lightweight, betel, leaf, freshness, grading, network, hybrid, depthwise, separable, convolutions, scientific research