BarkNet-Lite: A Lightweight Texture and Colour Network with the BarkBD Benchmark for Bark-Based Tree Species Recognition in Bangladesh

1. Introduction

• Bark recognition can support species-level records where trained observers are scarce, and unlike leaves, flowers, and fruit, bark is visible year-round at eye level.The introduction highlights Bangladesh, where many trees stand outside gazetted forests and inventories must be repeated as land use changes.Source: §1 · p. 1
• Fine bark texture is difficult to classify because it varies with age and diameter, overlaps across species, and changes with sun, cloud, and rain.Relevant cues include fissure depth and orientation, lenticel arrangement, and exfoliation geometry.Source: §1 · p. 1
• The work targets geographic and architectural gaps: public bark datasets lack Bengal delta species, while prior high accuracies generally rely on large ImageNet-pretrained backbones.The introduction frames the architectural question as how much accuracy independence from pre-training costs.Source: §1 · p. 2

1.1. Research questions •

• The research questions ask where a scratch-trained lightweight network sits on the accuracy–cost frontier, whether it transfers, which components matter, and whether it relies on bark rather than context.The component question covers multi-scale texture aggregation, attention, residual refinement, and the colour pathway.Source: §1.1 · p. 2

1.2. Contributions •

• BarkBD contributes 14,258 uncropped smartphone images of 20 native species from four districts, collected under three weather conditions.Its classes include English, Bengali, and scientific names, and the dataset uses a fixed stratified split.Source: §1.2 · p. 2
• BarkNet-Lite has 2.96 million parameters and combines multi-scale texture and parallel colour-aware pathways, training from scratch and reaching 96.64±0.66% accuracy over five seeds.The contributions also describe a matched comparison with nine ImageNet-fine-tuned backbones, cross-dataset evaluation, and Grad-CAM checks; finer component-level controls are left to future work.Source: §1.2 · p. 2

2.1. Bark as a diagnostic organ

• Bark identification relies chiefly on local texture cues, while ontogenetic change means results from mature trees need not generalize to saplings.Field cues include fissures, lenticels, exfoliation patterns, and, for palms, persistent leaf-base scars.Source: §2.1 · p. 3

2.2. Texture recognition, attention and explanation

• Multi-scale aggregation and dilation span bark structures, while attention can select informative channels and locations on uncropped images.The background contrasts learned descriptors with handcrafted ones, which are interpretable but fixed by their designers; it also notes depthwise separable convolution as a low-cost option for the colour pathway.Source: §2.2 · p. 3
• Grad-CAM uses late feature-map contributions to the predicted class score to locate evidence, distinguishing trunk-focused maps from foliage or sky context.The background presents it as a post-hoc, approximate check for context bias.Source: §2.2 · p. 3

3. Related work

• The related-work review organizes bark identification by representation type and sensing modality, then surveys datasets and the paper’s research gap.Table 1 summarizes studies quantitatively.Source: §3 · p. 4, Table 1 · p. 5

3.1. Handcrafted texture descriptors

• Handcrafted descriptors are compact and interpretable, but their fixed patterns can degrade when acquisition conditions shift or the species set changes.Earlier work found that adding colour helped each of the described handcrafted approaches.Source: §3.1 · p. 4

3.2. Deep learning on bark images

• Learned bark representations advanced through CNNs addressing scale, uncropped imagery, efficiency, and specialized tasks; BarkNet 1.0 emphasized sampling distinct trees for generalization.Deep BarkID instead reaches deployability by distilling a large pre-trained model.Source: §3.2 · p. 4

3.3. Three-dimensional and multi-organ approaches

• Terrestrial laser scans capture bark geometry but carry acquisition costs that confine them to research inventories, while multi-organ methods depend on other organs being visible.The latter condition does not hold for the off-season bark-only use case described here.Source: §3.3 · p. 4

3.4. Datasets and the geographic gap

• Public bark datasets differ in size, species coverage, and cropping; the cited South Asian corpus covers 13 species from two Indian states, not the Bengal delta.The text notes its restrictive CC BY-NC-ND licence and identifies several Bengal delta species absent from public bark benchmarks.Source: §3.4 · p. 4

3.5. Positioning of this work

• The paper addresses narrow geographic coverage, protocol-confounded accuracy, and post-hoc compactness with an uncropped Bengal delta corpus, single-image evaluation across five seeds, and a scratch-trained 2.96 M-parameter model.The model is evaluated against nine compact backbones under a matched protocol on the accuracy–cost frontier.Source: §3.5 · p. 4

4. The BarkBD Dataset

• BarkBD contains 14,258 uncropped smartphone photographs of 20 native species, collected across four districts and released with a fixed stratified split.The dataset was captured under field conditions without cropping or retouching.Source: §4 · p. 6, §4.4 · p. 8

4.1. Design Rationale

• Images were retained uncropped, sampled by individual tree and left unbalanced to better reflect field deployment, rather than maximize headline accuracy.The design rationale notes that cropping removes common background-related failure cases and that individual-tree sampling matters for generalisation.Source: §4.1 · p. 6

4.2. Study Area and Acquisition

• Sampling covered four districts to represent site variation, while three consumer phones, multiple viewpoints and heights, and sunny, cloudy and rainy sessions captured acquisition variation.The districts span the northern floodplain and south-eastern coast; the collection rationale cites soil, humidity, exposure and geographically restricted species.Source: §4.2 · p. 6

4.3. Annotation and Composition

• Field identifications were verified by an agroforestry specialist; the 20 classes include 15 hardwoods, four palms and one giant grass, with counts from 249 to 1,389.Table 2 lists each class with English, Bengali and scientific names and image counts.Source: §4.3 · p. 6, Table 2 · p. 7

4.4. Preparation Pipeline and Splits

• The preparation pipeline screens out blurred, out-of-focus or background-dominated frames, then assigns surviving images their source-tree labels and partitions them by class into fixed 70/15/15 splits.The shipped split contains 9,971 training, 2,138 validation and 2,149 test images, with no image appearing in more than one subset.Source: §4.4 · pp. 7–8, Figure 3 · p. 8
• Because photographs of one trunk can fall on both sides of the image-level split, reported accuracy concerns sampled individuals under unseen viewpoints, distances and illumination—not unseen individuals.The authors describe these scores as an upper bound on generalisation to individuals never seen in training; source-tree records allow a tree-grouped split to be constructed.Source: §4.4 · p. 8

4.5. Relation to Existing Bark Corpora

• BarkBD is larger than BarkTex and TRUNK12, comparable in size to CentralBark and BarkNet 1.0, and its species set overlaps none of those corpora or BarkVisionAI.The taxa dominating the Bengal delta remain absent from the public bark benchmarks discussed.Source: §4.5 · p. 8

5.2. Preprocessing and Augmentation; 5.4. Optimisation

• Images are resized to 224 × 224 without standardisation; training uses fresh geometric and colour augmentation, batch mixing, label-smoothed cross-entropy, and Adam with warmup and cosine decay.Validation and test images are not augmented; training stops when validation accuracy stalls for 40 epochs and restores the best-validation weights.Source: §5.2 · p. 9, §5.4 · p. 11

5.3. Architecture of BarkNet-Lite

• BarkNet-Lite combines four texture stages, an early-tapped colour pathway with two fusion points, and a dual-pooling head.The colour pathway is reinjected later so chromatic evidence reaches the classifier without passing through the full texture hierarchy.Source: §5.3 · p. 9

5.3.1. Stem; 5.3.6. Dual-Pooling Head; 5.3.7. Complexity

• A 5 × 5 stem kernel spans fissure width, and the head concatenates global average and max pooling to represent dense patterns and strong local responses.The model has 2,957,956 trainable parameters and occupies 11.32 MB on disk.Source: §5.3.1 · p. 9, §5.3.6 · p. 10, §5.3.7 · p. 10

5.3.2. Texture Inception Module

• Six parallel receptive-field branches target bark structures from lenticel detail to plate-scale patterns and palm leaf-base scars.The branches are concatenated and projected to 2F channels; dilated convolutions span wider windows without downsampling.Source: §5.3.2 · pp. 9–10

5.3.3. Texture Attention Module; 5.3.4. Texture Residual Block

• The attention module adds channel and spatial gates to the identity, while a fourfold-bottleneck residual block limits the cost of its 3×3 convolution.The additive attention form preserves strong evidence from one gate when the other is weak; the bottleneck keeps the parameter budget at 2.96 M.Source: §5.3.3 · p. 10, §5.3.4 · p. 10

5.3.5. Colour-Aware Module and Colour Skips

• The colour pathway branches off after the first stage, uses pointwise–depthwise–pointwise factorisation with squeeze-and-excitation, and is reinjected at two depths.It occupies under 4% of parameters; average pooling preserves mean chromaticity when matching the texture-pathway resolution.Source: §5.3.5 · p. 10

5.5. Experimental Protocol

• Evaluation uses strict single-image inference, compares nine identically trained ImageNet-pretrained backbones, tests three quantisation formats, and retrains the architecture on other datasets.The backbone comparison jointly tests compact design and no pretraining, while each baseline is trained once; cross-dataset runs change only classifier width.Source: §5.5 · p. 11

5.5. Experimental Protocol; 5.6. Evaluation Metrics

• Stage-removal ablations use fixed 100-epoch budgets, while Grad-CAM maps are checked with deletion/insertion measures and weight randomisation; metrics include macro and support-weighted scores.Ablation drops are interpreted by ordering because variants train for less than the full model; shallowest-stage removal is unresolved, and finer component controls are deferred.Source: §5.5 · p. 12, §5.6 · p. 12

6.1. Stability across seeds; 6.2. The reference run

• 96.64±0.66% test accuracy across five seeds; the single reference run scores 95.91% accuracy and 95.88% weighted F1.The standard deviation sets the comparison resolution: differences below roughly 1.3 points between single runs should not be interpreted.Source: §6.1 · p. 12, Table 4 · p. 13, §6.2 · p. 13

6.4. Per-species performance; 6.5. Error structure

• Sixteen species exceed 95% F1, but Litchi chinensis is lowest at 85.25%; errors cluster among similar bark, led by lychee-to-Monoon at 9.6%.The smallest class, Casuarina equisetifolia, reaches 97.44% F1, while the largest, Litchi chinensis, records the worst score; class size and F1 are uncorrelated.Source: §6.4 · pp. 13–14, Table 5 · p. 14, §6.5 · p. 14

6.6. Comparison with modern backbones

• 2.3 points below MobileNetV4-Hybrid-Medium (98.93% at 9.82 M), BarkNet-Lite ranks eighth by accuracy but is second-smallest at 2.96 M.The five-seed mean is within one standard deviation of MobileNetV4-Conv-Small (96.88%), the only other sub-3 M model; BarkNet-Lite is the only model trained from scratch.Source: §6.6 · pp. 14–15, Table 6 · p. 16

6.7. Deployment efficiency

• 15.34±0.39 ms per photograph on-device at 95.91% accuracy; quantised variants fall to 84.46% and 84.27%, so neither is deployed.The deployed single-precision graph uses 38.58 MB peak memory; its two-thread XNNPACK configuration is measured on a Snapdragon 7+ Gen 3.Source: §6.7 · pp. 15–16, Table 7 · p. 17
• 15.34±0.39 ms per uncropped photograph on a smartphone, with quantisation losses leaving single precision as the reported deployment.The quantised variants lose 11.45 and 11.63 points; standard remedies were not applied, so the results bound rather than measure achievable compression.Source: §6.7 · pp. 15–16, Table 7 · p. 17

6.8. Cross-dataset generalisation

• 95.86% on BarkVN-50 and 92.85% on BarkNet 1.0, versus 69.76% on TRUNK12’s 393 images across 12 classes.BarkVN-50 has no official partition, making its comparison indicative rather than like-for-like; BarkNet 1.0's 93.88% cited figure is single-image, while majority vote reaches 97.81%.Source: Table 8 · p. 17, §6.8 · p. 17

6.9. Ablation study

• 93.67% after deepest-stage removal, 2.97 points below the five-seed model; deeper removals hurt more, but fixed 100-epoch runs make each drop an upper bound.The first-stage removal is confounded with the colour pathway that taps from it, so its near-zero effect cannot be attributed to either alone.Source: Table 9 · p. 18, §6.9 · p. 18, Figure 9 · p. 18

6.10. Explainability; 6.11. Calibration

• Attributions concentrate on bark and separate from random controls; top-2 accuracy reaches 98.88%, while calibration ECE is 0.148.Grad-CAM is more faithful than Grad-CAM++ on every reported measure; cascade-randomising the head and final stage collapses maps to a near-uniform field.Source: Figure 10 · p. 19, Table 10 · p. 19, §6.10 · p. 19

7.1. Answering the research questions

• 96.64 ± 0.66% accuracy places BarkNet-Lite within one seed-level standard deviation of the smallest backbone, but 2.3 points behind the strongest fine-tuned backbones.It classifies one photograph in 15.34 ± 0.39 ms within 38.58 MB on a mid-range phone; it is not on the accuracy–compute frontier.Source: §7.1 · p. 20
• Retrained from scratch, the architecture reaches 95.86% on BarkVN-50 and 92.85% on BarkNet 1.0; TRUNK12 reaches 69.76%, where 33 images per class cannot support a texture representation learned from scratch.Source: §7.1 · p. 20
• Removing the deepest stage costs 2.97 points, but the single-seed, shorter-budget ablation cannot resolve the near-zero first-stage effect against noise.The first-stage removal is confounded with the colour pathway tapping that stage; separating their effects remains future work.Source: §7.1 · p. 20
• Grad-CAM localises fissures, lenticel fields, and leaf-base scars, while faithfulness and weight-randomisation checks support attribution to learned parameters rather than background.Failures are often confident species confusions despite localisation of the correct region; the authors point to finer input resolution as a remedy.Source: §7.1 · p. 20

7.2. Implications

• Regional datasets are the binding constraint, and compact design affords ImageNet independence at a cost; parameter count alone also misses compute and latency.The discussion further links uncropped, site-clustered data to a risk of learning location and argues that Grad-CAM verification supports the deployment claim; per-species F1 is uncorrelated with image count (r = −0.12).Source: §7.2 · p. 20

7.3. Limitations and threats to validity

• The image-level split places frames from the same tree on both sides, so the reported accuracies measure sampled-tree recognition rather than generalisation to unseen individuals.A tree-grouped split is the appropriate test for unseen individuals; the authors treat current figures as upper bounds, while noting that the split is shared by all nine baselines and leaves relative comparisons unaffected.Source: §7.3 · p. 20
• Cross-dataset results are one run per corpus, latency and memory come from one phone, and collection over four districts leaves regional, seasonal, and age generalisation untested.Quantised exports lose more than eleven points; the single-precision graph remains the deployable artefact, and the responsible weight layer for quantisation loss is unidentified.Source: §7.3 · p. 21

7.4. Future work

• Future work prioritises a tree-grouped BarkBD split and multi-seed ablations that separate the colour pathway from the first texture stage.Other proposed steps include recording trunk diameters, capturing close and wide views, aggregating images per tree, and recovering quantisation performance.Source: §7.4 · p. 21

8. Conclusion

• BarkNet-Lite reaches 96.64 ± 0.66% accuracy over five seeds on BarkBD under strict single-image inference, using 2.96 M parameters and random initialisation.BarkBD contains 14,258 uncropped smartphone photographs of 20 native species across four districts and three weather conditions, with a fixed stratified split.Source: §8 · p. 21
• 95.86% on BarkVN-50 and 92.85% on BarkNet 1.0 accompany 69.76% on TRUNK12, marking transfer as bounded by data regime.The paper also reports deployable single-precision inference at 15.34 ± 0.39 ms within 38.58 MB on a mid-range smartphone; quantisation costs over eleven points.Source: §8 · p. 21
• The image-level split bounds recognition of sampled trees rather than performance on unseen individuals, although the comparison among models shares that split.The authors state that this qualification sets what field deployment should expect.Source: §8 · p. 21
• The conclusion frames geographic concentration in bark research as a data problem before an algorithmic one, calling for regional datasets and device-sized models.It notes that Bangladeshi species were absent from available label spaces.Source: §8 · p. 21

Data and code availability

• The BarkBD dataset, training and evaluation code, trained BarkNet-Lite weights, and figure-generation notebooks are slated for release upon acceptance.The materials are also to be archived with a persistent identifier in a public research-data repository.Source: Data and code availability · p. 21
• The release includes the fixed stratified split, five-seed checkpoints, a reference-run checkpoint, and the signed expert botanical-validation certificate for BarkBD labels.Keeping the fixed split enables subsequent work to report directly comparable figures without re-splitting.Source: Data and code availability · p. 22

Funding

• This research did not receive a specific grant from funding agencies in the public, commercial, or not-for-profit sectors.Source: Funding · p. 22

• Aroshi Ali contributed to conceptualization, methodology, software, investigation, data curation, and writing—review and editing.Source: CRediT authorship contribution statement · p. 22
• Saad Ahmed contributed to conceptualization, supervision, project administration, and writing—review and editing.Source: CRediT authorship contribution statement · p. 22
• Md. Khalid Syfullah contributed to formal analysis, validation, visualization, and writing—the original draft and review and editing.Source: CRediT authorship contribution statement · p. 22

• LLMs were used to improve the manuscript's language and readability.Source: Declaration of generative AI and… · p. 22
• The authors reviewed and edited the content after using the tool and take full responsibility for the published article.Source: Declaration of generative AI and… · p. 22
• Generative AI was not used to produce, analyze, or interpret data, generate or alter figures, or derive reported numerical results.Source: Declaration of generative AI and… · p. 22