[预印本]视觉基础模型中的多分支集成头用于高效黄瓜物候阶段分类

Multibranch Ensemble Head in Vision Foundation Models for Efficient
Cucumber Phenological Stage Classification
Kostiantyn Khabarlak
Dnipro University of Technology, Ukraine
Khabarlak.K.S@nmu.one
Abstract
VisionFoundationModels(VFMs)pretrainedwithself-
supervised learning, such as DINOv2, provide highly
transferable features, yet adapting them to small, domain-
shifteddatasetsisstillchallenging. Weproposearegular-
izingmultibranchensemblehead: atrainingandinference
scheme that pairs low- and high-capacity branches, sam-
ples a single branch stochastically at each training step,
and averages both branches in probability space at in-
ference. The design adds negligible parameters and es-
sentially no inference overhead over a single heavy head,
while improving the accuracy-efficiency trade-off under
data-scarce conditions. Accurate classification of plant
growthstagesiscriticalforprecisionagriculture,enabling
timely decisions on fertilization, irrigation, and harvest-
ing. Therefore, as a challenging practical case we use cu-
cumber phenological stage classification, in which flow-
ersandfruitsarefrequentlyoccludedbyfoliage,andstage
transitions manifest primarily as subtle changes in leaf
morphology. TheNALEFFoodSafetydatasetcontainsim-
ages of five cucumber growth stages captured under di-
verse real-world conditions. On it we evaluate a broad
range of backbones, including DINOv2 and DINOv3 ViT
andConvNeXtvariants,againststronglysupervisedcoun-
terparts(ResNet,EfficientNet,ViT).TheproposedDINOv2
ViT-SwithmultibranchheadachievesaMacroF1of0.978
and Cohen’s Kappa of 0.971, matching the performance
of models more than ten times larger in parameters and
GFlops, while requiring only 12.2 GFlops and 22.3M pa-
rameters, making it deployable on smartphones and edge
devices for phenological stage classification.
Keywords:Plant growth stage classification, Vision
Foundation Models, Self-supervised learning, DINOv2,
DINOv3, Regularization, Transfer learning, Cucumber
phenology.

1. Introduction
   The timing and progression of plant developmental
   phasesarehighlydependentonenvironmentalfactors,par-
   ticularly temperature, precipitation, and photoperiod [1].
   Accurate classification of plant growth stages is critical
   for optimizing agricultural management, including deci-
   sions on fertilization, herbicide application, and harvest-
   ing. Automating this process reduces reliance on labor-
   intensive and subjective manual analysis, ultimately sup-
   porting higher and more consistent crop yields.
   Monitoring of plant growth can be performed using
   ground-level cameras, unmanned aerial vehicles (UAVs),
   or remote-sensing satellite imagery.
   This article has been published in the peer-reviewed confer-
   ence proceedings. Citation: K. Khabarlak, “Multibranch ensemble
   head in vision foundation models for efficient cucumber phenological
   stage classification,” in Proceedings of the International Scientific
   and Practical Workshop “Advanced Digital Technologies: Intelligent
   Systems, Modeling, and Data-Driven Control” (ADT-ISMDDC 2026)
   Co-Located with ICST 2026, in CEUR Workshop Proceedings, vol.
2. Odesa, Ukraine: CEUR-WS.org, 2026, pp. 251–262. doi:
   10.83095/ceurws_04300_018.
   Video presentation available on YouTube.
   1
   arXiv:2610.37118v1 [cs.CV] 27 Oct 2026

Remote-sensing satellite imagery can be used to es-
timate growth stages of large farmlands. For instance,
Sentinel-2satellitesprovidepublicaccesstomulti-spectral
imagery, including red-edge bands specifically designed
for vegetation analysis [2]. However, this approach is un-
suitable for use in cloudy conditions or in greenhouses,
anditslowspatialresolutionleadstopoorperformanceat
early growth stages when the canopy does not fully cover
the field, or for small fields.
UAVs and stationary cameras offer much greater flex-
ibility and precision. They can be used regardless of
weather conditions and inside greenhouses. Additionally,
smartphonesenablerapid,on-siteplantgrowthevaluation
bynon-experts. Camera-basedplantgrowthmonitoringis
often formulated as a classification task [3], [4].
In particular, automated classification of cucumber
growth stages from visual data remains an open prob-
lem despite the crop’s considerable agronomic signifi-
cance. While computer vision methods have been applied
to cucumbers across several related tasks including dis-
ease detection [5], fruit localization for robotic harvest-
ing [6], quality grading [7], and morphological phenotyp-
ing [8], the problem of growth stage classification has re-
ceived comparatively little attention in the literature. This
task poses unique challenges: flowers and fruits are fre-
quently occluded by foliage, and cucumber fruit shares
near-identical coloration with the surrounding canopy. As
a result, transitions between intermediate stages manifest
primarily as subtle changes in leaf morphology.
Vision-based plant monitoring is typically performed
by fine-tuning a backbone trained with strong supervi-
sion on a large labeled dataset. However, recent works
have shown that backbones trained with self-supervised
objectives offer far more universal features that can be
adapted to diverse downstream tasks using a simple linear
probe[9],[10]. TheseVisionFoundationModels(VFMs)
have demonstrated applicability across object classifica-
tion,detection,tracking,scenedepthestimation,andmany
other tasks; yet their use for plant growth stage classifica-
tion remains comparatively underexplored.
In this work we study how to adapt self-supervised
VFMstosmall,domain-shiftedimagedatasetsthroughthe
designoftheclassificationhead. Weintroducearegulariz-
ingmultibranchensemblehead: ageneraltrainingandin-
ference scheme in which a low-capacity linear branch and
a higher-capacity MLP branch are attached to the back-
bone. A single branch is sampled stochastically at each
training step, and the two branches are averaged in proba-
bility space at inference. This yields an in-model ensem-
blethatregularizesheadtrainingunderdatascarcitywhile
adding negligible parameters and essentially no inference
cost. Weusecucumberphenologicalstageclassificationas
a challenging fine-grained case study: on the recently in-
troduced NALEF Food Safety [11] dataset, the proposed
headsubstantiallyimprovestheaccuracy-efficiencytrade-
off over the DINOv2 ViT-L fine-tuning approach previ-
ously reported for this dataset [12].
Our main contributions are as follows:

1. We propose the regularizing multibranch ensemble
   head, a lightweight, backbone-agnostic scheme for
   training classification heads on top of Vision Mod-
   els under data-scarce conditions, combining stochas-
   tic per-step branch sampling with probability-space
   aggregation at inference.
2. Using a cucumber phenological stage classification
   dataset, we systematically compare a broad range of
   self-supervised (DINOv2, DINOv3) and strongly su-
   pervised backbones, and show that self-supervised
   pretraining is particularly beneficial in the small-
   model regime despite the large domain gap between
   its pre-training data and plant imagery.
3. TheresultingDINOv2ViT-Smodelsubstantiallyim-
   proves over existing solutions on the NALEF Food
   Safetydatasetatasmallfractionofitscomputational
   cost, requiring few GFlops and parameters and re-
   maining deployable on smartphones and edge de-
   vices.
4. Literature Overview
   Fine-tuning from a strongly pretrained backbone is the
   standard approach for training deep neural networks on
   2

downstreamtaskssuchasplantgrowthstageclassification.
Theauthorsof[3]collectastrawberrygrowthdatasetand
report that a fine-tuned EfficientNet-B7 achieves 83.7%
classification accuracy. In [13], the authors highlight the
importance of camera-based plant growth monitoring as
(1) such systems are low-cost and (2) the collected data is
crucial for nutrient management and harvesting schedul-
ing; they develop a YOLOv8-based detection system for
multi-plant basil growth tracking. In [4], a ResNet-50-
basedbutterheadlettucegrowthstageclassificationsystem
is presented.
Cucumber-specific research is an active area. [14] in-
vestigatetherelationshipbetweencucumberroottraitsand
cropyieldtosupportthebreedingofhigh-yielding,stress-
tolerant cultivars. Cucumber production further benefits
fromautomatedharvestingsystems. Accuratedetectionof
harvest-ready cucumbers remains challenging due to the
visualsimilaritybetweenfruitsandsurroundingfoliage,as
well as frequent occlusions. To address this problem, [6]
proposealightweightYOLO-baseddetectionnetworktai-
lored for cucumber harvesting scenarios. Quality assess-
ment has also been explored using deep learning, with [7]
introducing a CNN-based framework for classifying cu-
cumber fruits into defective, curved, and straight/sound
categories. For phenotyping applications, [8] present a
lightweight keypoint detection model that enables auto-
mated extraction of fruit phenotypic traits. [15] propose
a pest classification method applicable to different plant
species,includingcucumber. Closesttoourtask,[12]fine-
tune a single DINOv2 ViT-L backbone with progressive
layerunfreezingforcucumbergrowthstageclassification.
In contrast to strongly supervised approaches, large
language models (LLMs) have demonstrated remarkable
progress through self-supervised learning, typically by
training on large text corpora with a next-token predic-
tionobjective. Thisapproachenablesthenetworktolearn
strong linguistic representations from unlabeled data.
VisionTransformershavesuccessfullyadaptedthecon-
cept of tokens to images, often surpassing convolutional
neural networks. Initially, however, they were still trained
with strong supervision, as the design of effective self-
supervised objectives for vision is more challenging, and
the notion of a “next image token” is less self-evident.
To pretrain a vision backbone with self-supervision,
several pretext tasks have been proposed: image re-
colorization,predictionofimagetransformations,andoth-
ers. DINOv2usesmaskedimagemodelingcombinedwith
afeature-consistencylossbetweenteacherandstudentnet-
works observing different crops of the same image. A
large Vision Transformer (1.1B parameters [9]) is trained
andthendistilledintoseveralsmallerViTs. Thisapproach
was later extended by DINOv3 [10], which identified lim-
itations in DINOv2’s architecture scaling under extended
training and proposed a Gram loss, among other improve-
ments. The resulting 7B model achieves state-of-the-art
performance on many downstream tasks, and the authors
distill it into both ViT and ConvNeXt variants. Models
trained in this fashion are commonly referred to as Vi-
sion Foundation Models, as the rich features they produce
transfer broadly across diverse downstream tasks.
The self-supervised training objective allows the col-
lection of large unlabeled datasets for training. For in-
stance, the authors of DINOv2 assembled the LVD-142M
datasetcomprising142millionimageswithoutanylabels,
includingImageNet-22Kandmanyothersources. Mostof
these are not plant-focused; while the authors include the
small Flowers dataset, the larger nature-focused iNatural-
istdatasetisabsentfromthetrainingset[9]. TheDINOv3
authors introduce the LVD-1689M dataset (1,689 million
Instagram photos). While substantially larger, this dataset
likely entails a greater domain gap than DINOv2’s web-
sourced collection. Despite the domain gap, recent works
have shown that self-supervised methods can be success-
fully adapted to plant-related tasks [16], [17].
Despite these advances, most existing approaches to
plant growth stage classification still rely on strongly su-
pervised backbones, limiting their transferability to spe-
cialized agricultural domains with scarce labeled data.
Furthermore,despiteitsagronomicimportance,cucumber
growthstageclassificationhasreceivedonlylimitedatten-
tion in the literature, with VFM-based solutions for this
specifictaskexploredonlyinasinglelarge-backbonecon-
3

figuration [12]. Finally, the task stands to benefit from ad-
ditionalin-modelregularization,asclassificationmustrely
primarily on subtle leaf morphology cues due to frequent
occlusionofflowersandfruits;thepotentialoflightweight
ensembleheaddesignstoimprovetheaccuracy-efficiency
trade-off in this setting remains unexplored.
3. Materials and Methods
Toaddressplantgrowthstageclassification,wedevelop
an approach based on Vision Foundation Models paired
with a regularizing multibranch ensemble head. We com-
pare the proposed method against the architectures listed
in Table 1, spanning both convolutional neural networks
(CNNs) and Transformer architectures.
The classification head is an important component of
the proposed architecture. Given the rich features ex-
tracted by the backbone, the head must adapt them to the
downstream task. We develop three head variants. Fig. 1a
shows a lightweight head, also known as a linear probe,
often used with VFMs as recommended in [9], [10], and
frequently sufficient given the descriptive power of VFM
features. To address the domain shift between VFM pre-
training and downstream agricultural data, a head with a
larger number of parameters may be required. Fig. 1b is a
heavy head comprising two MLP layers with GELU ac-
tivation followed by Dropout regularization (probability
0.3). Fig. 1c shows the proposed multibranch ensemble
head, described in detail below.
(a) Light
(b) Heavy
(c) Multibranch Ensemble
Figure 1. Neural network heads proposed for fine-tuning.
To improve the performance of VFMs, an ensemble of
modelscanbeused;however,thisincreasesinferencetime
by a factor ofN. Several works instead build more robust
representations within a single network, effectively form-
ing an in-model ensemble. The prototypical example is
Dropout[23]: randomlymaskingactivationsduringtrain-
ingsamplesadifferentweight-sharingsubnetworkateach
step, and test-time weight scaling approximates averaging
the predictions of exponentially many such subnetworks.
The authors of PTA [24] make this idea explicit, build-
inganin-modelensemblefrommultipleparallelbranches
of different depths within the backbone, improving clas-
sification quality. StochasticDepth [25] demonstrates that
dynamically varying network depth during training acts
as a regularizer; this technique has been adopted by Ef-
ficientNet V1 and V2, DINOv2, and DINOv3. [26] pro-
poses collaborative networks where multiple branches are
trained simultaneously, each supervised by a combina-
tion of ground-truth labels and soft targets from the other
heads. [27] proposes a multi-head multi-loss classifier in
which each head is assigned higher weight to a subset of
classes, inducing specialization, with the final prediction
obtained by averaging across heads.
In this work we propose a multibranch ensemble head
(Fig. 1c). The two branches differ in depth and capac-
ity: the left branch is a simple linear probe, while the
right branch consists of several MLP layers with Dropout,
equivalent to the heavy head described above (Fig. 1b).
During training, the active branch is randomly selected
for each batch, so each branch learns to solve the task in-
dependently. The difference in capacity allows the light
and heavy branches to capture simpler and more complex
feature transformations, respectively. At inference, an in-
model ensemble is formed by aggregating the outputs of
both branches:
ˆy= log
 1
2

p(light) +p (heavy)

+ϵ

,(1)
where the branch outputp(k), k∈ {light,heavy}is de-
fined as:
p(k) = softmaxf (k)(z) .(2)
Heref (k) is the function computed by branchk, andz=
4

Table 1. Backbones considered.
Model Type Sizes
Strongly Supervised Backbones
ResNet [18] CNN 50, 101
ConvNeXt [19] CNN Tiny (T)
EfficientNet [20] CNN B3, B7
EfficientNetV2 [21] CNN Small (S), Large (L)
ViT [22] Transformer Base (B), Large (L)
Self-Supervised Backbones
DINOv2 ViT [9] Transformer Small (S), Large (L)
DINOv3 ViT [10] Transformer Small (S), Large (L)
DINOv3 ConvNeXt [10] CNN Tiny (T), Small (S), Large (L)
g(X)are the features produced by the backbonegfor the
input batchX. The vectorˆy∈R C holds the aggregated
log-probabilities,andϵisasmallpositiveconstantthaten-
sures numerical stability of the logarithm. Aggregation
is performed in probability space rather than logit space
(Eqs.1–2),sincelogitsfrombranchesofdifferentcapacity
may have substantially different scales. Probability-space
aggregationisconsistentwithstandardmixture-of-experts
averaging and corresponds to a uniform Bayesian model
averagetreatingeachheadasanequallyplausiblehypoth-
esis about the class distribution.
The branch selector shown in Fig. 1c activates exactly
one branch per training step. At each step a branch is
drawn independently from a Bernoulli distribution,k∼
Bernoulli(π),withπ= 0.5assigningequalprobabilityto
the two branches. Only the selected branch produces the
prediction for that step, and only its parameters, together
with those of the backbone, receive gradients from the
cross-entropy loss; the other branch is left unchanged. No
knowledge-distillationorcross-branchconsistencytermis
applied, so each branch learns to solve the task indepen-
dently. Algorithm 1 summarizes the procedure.
Models are trained using the AdamW optimizer with
a batch size of 32. The head is randomly initialized
while the backbone uses pretrained weights. To mitigate
catastrophic forgetting, training proceeds in two stages:
Algorithm 1:Training step with the multibranch
ensemble head
Input :mini-batch(X, Y); backboneg; branches
f (light), f(heavy); branch probabilityπ
1z←g(X)% backbone features
2drawk∼Bernoulli(π),k∈ {light,heavy}
3ℓ←CrossEntropy

f (k)(z), Y

4update parameters ofgandf (k) using∇ℓ
% other branch weights are not
updated
15 epochs of head-only fine-tuning at a learning rate of
3×10 −3, followed by 15 epochs of full-network fine-
tuning at a reduced learning rate of5×10−6.
4. Experiments
To evaluate the proposed approach we use the NALEF
Food Safety dataset [11], published in 2026. The dataset
contains images of cucumber crops captured under real-
worldconditions,annotatedwithfivegrowthstages: plan-
tula, early vegetative, vegetative growth, flowering, and
fruit setting. Overall there are 787 training and 239 test
images. As the public test set does not include ground-
truth annotations, we perform a stratified train/validation
split with an 80/20 ratio for all reported experiments. Im-
agesarecapturedusingdifferentdevices(professionaland
5

smartphonecameras)undervaryingenvironmentalcondi-
tions.
The dataset is strongly imbalanced across growth
stages,assummarizedinTable2. Thesmallestclass(plan-
tula)contributesanorderofmagnitudefewersamplesthan
thelargest(vegetativegrowth),leavingonlyabouttenval-
idationimagesfortheminoritystages. Accuracyisunsuit-
ableastheprimarymetricforhighlyimbalanceddatasets.
WethereforeevaluateallmodelsusingMacroF1scoreand
Cohen’s Kappa.
Table 2. Per-class image counts for the stratified 80/20 train/val-
idation split.
Growth stage Train Val Total
Plantula 40 10 50
Early vegetative 42 11 53
Vegetative growth 307 77 384
Flowering 70 17 87
Fruit setting 170 43 213
Total 629 158 787
To improve model generalization, the following aug-
mentations are applied: randomly resized crop, random
horizontal flip, and color jitter. All images are resized to
224×224.
5. Results
Experiment results are presented in Table 3. For each
model we report the head configuration that achieved the
best result, the type of backbone supervision, evaluation
metrics, and computational complexity. For strongly su-
pervised models we report the model name and size; for
self-supervised backbones we report the DINO version,
backbonearchitecture,andsize. Moststronglysupervised
backbones are outperformed by self-supervised ones of
comparable size. The exception is ViT-L, which is on
par with the best DINO-pretrained models and attains the
highest Macro F1 overall (0.979). Therefore, the advan-
tage of self-supervised pretraining is concentrated in the
small- and mid-sized models (Fig. 2).
Fig. 2 shows Macro F1 score as a function of compu-
tational complexity (GFlops). Strongly supervised mod-
0 20 40 60 80 100 120 140 160
GFlops
0.88
0.90
0.92
0.94
0.96
0.98F1
ConvNeXt
EfficientNet
EfficientNet V2
ResNet
ViT
DINOv2 ViT
DINOv3 ConvNeXt DINOv3 ViT
Strong Supervision
ViT
ConvNeXt
EfficientNet V2
ResNet
EfficientNet
Self-Supervised
DINOv2 ViT
DINOv3 ConvNeXt
DINOv3 ViT
Figure 2. Macro F1 score vs. model computational complexity
(GFlops). Self-supervised backbones (blue) hold a clear advan-
tage in the small- and mid-compute regime.
els are shown in red and self-supervised models in blue.
TheadvantageofDINO-pretrainedbackbonesismostpro-
nounced in the small-model regime. The best result for
a small model (Macro F1 of 0.978) is achieved by DI-
NOv2 ViT-S with the proposed multibranch ensemble
head(highlightedinredinTable3),matchingthestrongly
supervisedViT-Lmodel,whichisover10×larger. Asim-
ilarlysizedConvNeXt-Tinymodelachievesasubstantially
lower Macro F1 of 0.959; DINOv3-style pretraining of
ConvNeXt-Tiny raises this to 0.968.
DINOv2 models also outperform their DINOv3 ViT
and ConvNeXt counterparts. As discussed above, DI-
NOv3 models are trained on a larger dataset sourced pri-
marilyfromInstagram,andthegreaterdomainshiftisthe
likely cause of their slightly inferior performance.
6. Discussion
Table 4 details the incremental improvements to DI-
NOv2ViT-Sproposedinthiswork. Asabaselinewetrain
DINOv2 ViT-S with a light head (Fig. 1a) for 15 epochs.
ReplacingthelinearprobewithaheavyheadimprovesF1
to 0.955. Switching to the multibranch head further raises
F1 to 0.960. Subsequently fine-tuning the full backbone
yields the final F1 of 0.978. Importantly, the multibranch
head adds negligible parameters and essentially no addi-
6

Table3. Evaluationresultsoftheconsideredmodelsandheads. DINOv2ViT-Swiththemultibranchhead(highlightedinred)achieves
the best quality-efficiency trade-off.
Model Head B. Supervision F1 Kappa GFlops Params (M)
ViT-L Heavy Strong 0.979 0.972 123.108 303.831
DINOv2 ViT-L Heavy Self 0.978 0.971 162.025 304.898
DINOv2 ViT-SMultibranch Self 0.978 0.971 12.248 22.260
DINOv3 ConvNeXt-L Light Self 0.970 0.962 68.720 196.238
DINOv3 ViT-S Heavy Self 0.968 0.953 9.397 21.797
DINOv3 ConvNeXt-T Light Self 0.968 0.962 8.910 27.824
DINOv3 ViT-L Heavy Self 0.963 0.952 125.681 303.659
ConvNeXt-T Heavy Strong 0.959 0.952 8.910 28.218
DINOv3 ConvNeXt-S Light Self 0.955 0.943 17.366 49.459
ViT-B Heavy Strong 0.934 0.914 35.127 86.196
EfficientNet V2-L Heavy Strong 0.931 0.894 24.461 117.893
ResNet101 Multibranch Strong 0.913 0.875 15.601 43.562
ResNet50 Heavy Strong 0.897 0.874 8.176 24.560
EfficientNet-B7 Heavy Strong 0.883 0.874 10.337 65.101
EfficientNet V2-S Heavy Strong 0.875 0.845 5.699 20.836
EfficientNet-B3 Light Strong 0.871 0.810 1.922 10.704
Table 4. Ablation study of the proposed improvements over DINOv2 ViT-S.
Modification F1 Kappa GFlops Params (M)
DINOv2 ViT-S (baseline) 0.915 0.895 12.247 22.059

• multibranch head 0.960 0.944 12.248 22.260
• backbone fine-tuning 0.978 0.971 12.248 22.260
  tional inference overhead compared to the single-branch
  heavy head, since backbone computation dominates.
  As shown in Fig. 3, the proposed DINOv2 ViT-S with
  multibranch head is a robust classifier: it makes no er-
  rors between non-adjacent growth stages, and despite sig-
  nificant class imbalance it remains accurate on minority
  classes.
  The only misclassifications occur between the vegeta-
  tive growth and flowering stages. At first glance these
  stages should be visually distinct, separable by the pres-
  enceofflowers. However,closerinspectionoftheNALEF
  FoodSafetydatasetrevealsthatimagesprimarilyshowcu-
  cumberleaves,withflowersorfruitsoftenoccludedbyfo-
  liage. Themodelmustthereforediscriminatestagesbased
  on subtle differences in leaf and vine appearance alone,
  which is a considerably harder task.
  Limitations.The dataset is captured from a leaf-centric
  viewpoint in which flowers and fruits are frequently oc-
  cluded, which is also the main source of the residual
  vegetative-growthvs. floweringconfusion. Typically,field
  examination is performed during daytime and under good
  environmental conditions. We have not evaluated robust-
  ness to adverse conditions such as heavy shading, rain,
  night-time capture, or lens contamination on smartphone
  andedgedevices. InthisworkwetargettheNALEFFood
  7

Plantula
Early VegetativeVegetative Growth
Flowering
Fruit Setting
Predicted Class
Plantula
Early Vegetative
Vegetative Growth
Flowering
Fruit Setting
True Class
10 0 0 0 0
0 11 0 0 0
0 0 72 5 0
0 0 1 16 0
0 0 0 0 43
0
20
40
60
Count
Figure 3. Confusion matrix of DINOv2 ViT-S with multibranch
head.
Safety dataset. Therefore, retraining will be needed to
adapt the model to other cultivars or environmental con-
ditions. Also, the validation split is highly imbalanced,
so small differences in Macro F1 between the top models
should be interpreted with this in mind.
The multibranch ensemble head targets any fine-
grainedclassificationtaskwherealightweightheadontop
offrozenor fine-tunedfeaturesmustbetrainedon limited
data and is at risk of overfitting. Applying and evaluating
it on other fine-grained benchmarks, for instance, species
recognition or disease severity grading, is a natural next
step that would establish how far the regularization bene-
fit observed here carries over.
7. Conclusions
In this work we proposed a regularizing multibranch
ensemble head for adapting Vision Foundation Models
to small, domain-shifted datasets, and studied it on cu-
cumber growth stage classification as a fine-grained case
study. Experiments across a broad set of strongly su-
pervised and self-supervised backbones on the NALEF
Food Safety dataset indicate that self-supervised pretrain-
ing yields stronger feature representations for this task in
the small- and mid-sized model regime. DINOv2 and DI-
NOv3 backbones outperform ResNet, EfficientNet, and
stronglysupervisedViTvariantsofcomparablesize,while
at the largest scale the strongly supervised ViT-L remains
competitive; overall, VFMs transfer well to agricultural
imagerydespitesubstantialpre-trainingdomaingaps,with
the clearest benefit when model capacity is limited. No-
tably, DINOv2 ViT-S paired with the proposed multi-
branch ensemble head achieves a Macro F1 of 0.978 and
Cohen’sKappaof0.971,matchingmodelswithover300M
parameters while requiring only 12.2 GFlops and 22.3M
parameters, making it suitable for deployment on smart-
phones and edge devices.
The introduced multibranch ensemble head trains a
lightweight linear branch and a heavier MLP branch si-
multaneously via stochastic branch sampling. At infer-
ence, predictions are aggregated in probability space as a
uniform Bayesian model average. Ablation analysis con-
firms that the multibranch head improves over both the
linear probe and the heavy head, and that subsequent full-
network fine-tuning yields further substantial gains. The
methodaddsnegligibleparametersandessentiallyzeroin-
ferenceoverheadcomparedtoasingle-branchheavyhead.
Error analysis reveals that misclassifications are con-
finedtoadjacentphenologicalstages,specificallybetween
vegetative growth and flowering, a consequence of the
dataset’s leaf-centric imaging perspective in which flow-
ers and fruits are frequently occluded. The absence of er-
rorsbetweennon-adjacentstagesindicateswell-structured
internal representations.
Future work includes extending the approach to addi-
tionalcropspeciesandgrowthstagetaxonomies,aswellas
incorporating explainability by integrating large language
models.
Funding
This research was carried out as part of the scientific
project “Software and algorithmic solutions for agricultural in-
formation and analytical systems based on artificial intelligence
in variable agroclimatic conditions” (state registration number
0126U001122)fundedbytheMinistryofEducationandScience
of Ukraine at the expense of the state budget.
Declaration on Generative AI
During the preparation of this work, the author used
ChatGPT-5.6 for grammar and spelling checks. The author sub-
8

sequentlyreviewedandeditedthecontentasnecessaryandtakes
full responsibility for the content of this publication.
References
[1] N. Katal, M. Rzanny, P. Mäder, and J. Wäld-
chen, “Deep learning in plant phenological re-
search:Asystematicliteraturereview,”Frontiersin
Plant Science, vol. Volume 13 - 2022, 2022, ISSN:
1664-462X. DOI:10 . 3389 / fpls . 2022 .
805738[Online]. Available:https : / / www .
frontiersin . org / journals / plant -
science / articles / 10 . 3389 / fpls .
2022.805738
[2] “Sentinel-2 | Copernicus Data Space Ecosystem,”
Accessed: May 14, 2026. [Online]. Available:
https : / / dataspace . copernicus .
eu / data - collections / copernicus -
sentinel-missions/sentinel-2
[3] D. Darlan et al., “SmartBerry for AI-based growth
stageclassificationandprecisionnutritionmanage-
ment in strawberry cultivation,”Scientific Reports,
vol.15,no.1,p.14019,Apr.23,2025,ISSN:2045-
2322. DOI:10.1038/s41598-025-97168-
z[Online].Available:https://www.nature.
com/articles/s41598-025-97168-z
[4] J.-S. G. Kim, S. Chung, M. Ko, J. Song, and
S. H. Shin, “Comparison of Image Preprocessing
StrategiesforConvolutionalNeuralNetwork-Based
Growth Stage Classification of Butterhead Lettuce
in Industrial Plant Factories,”Applied Sciences,
vol. 15, no. 11, p. 6278, Jun. 3, 2025, ISSN: 2076-
3417. DOI:10.3390/app15116278[Online].
Available:https://www.mdpi.com/2076-
3417/15/11/6278
[5] J. Liu, X. Wang, Q. Chen, P. Yan, and X. Liu,
“Deeplearningmethodforcucumberdiseasedetec-
tion in complex environments for new agricultural
productivity,”BMC Plant Biology, vol. 25, no. 1,
p. 888, Jul. 7, 2025, ISSN: 1471-2229. DOI:10.
1186 / s12870 - 025 - 06841 - yAccessed:
May 25, 2026. [Online]. Available:https : / /
bmcplantbiol . biomedcentral . com /
articles / 10 . 1186 / s12870 - 025 -
06841-y
[6] B. Li, G. Zhong, and W. Ke, “LMS-Res-YOLO:
Lightweight and Multi-Scale Cucumber Detection
Model with Residual Blocks,”Sensors, vol. 25,
no. 23, p. 7305, Dec. 1, 2025, ISSN: 1424-8220.
DOI:10.3390/s25237305Accessed: May 25,
2026. [Online]. Available:https : / / www .
mdpi.com/1424-8220/25/23/7305
[7] S. Haseli Golzar, H. Bagherpour, and J. Amiri Par-
ian,“ANewMethodtoOptimizeDeepCNNModel
for Classification of Regular Cucumber Based on
GlobalAveragePooling,”JournalofFoodProcess-
ingandPreservation,vol.2024,no.1,p.5818803,
Jan. 2024, ISSN: 0145-8892, 1745-4549. DOI:
10 . 1155 / 2024 / 5818803Accessed:
May 25, 2026. [Online]. Available:https : / /
onlinelibrary . wiley . com / doi / 10 .
1155/2024/5818803
[8] H.Yu,X.Zhang,J.Yan,andX.Meng,“YOLO11m-
SCFPose: An Improved Detection Framework for
Keypoint Extraction in Cucumber Fruit Pheno-
typing,”Horticulturae, vol. 11, no. 7, p. 858,
Jul. 20, 2025, ISSN: 2311-7524. DOI:10.3390/
horticulturae11070858Accessed: May 25,
2026. [Online]. Available:https : / / www .
mdpi.com/2311-7524/11/7/858
[9] M. Oquab et al., “Dinov2: Learning robust vi-
sual features without supervision,”Trans. Mach.
Learn. Res., vol. 2024, 2024. [Online]. Available:
https : / / openreview . net / forum ? id =
a68SUt6zFt
[10] O. Siméoni et al., “Dinov3,”CoRR,
vol. abs/2508.10104, 2025. DOI:10 . 48550 /
ARXIV . 2508 . 10104arXiv:2508 . 10104.
[Online]. Available:https://doi.org/10.
48550/arXiv.2508.10104
9

[11] A. C. Menco Tovar, E. Puertas, J. C. Martinez-
Santos, and J. Serrano Castañeda,Nalef food
safety 2026, Zenodo, Mar. 2026. DOI:10 .
5281/zenodo.19192941[Online]. Available:
https : / / doi . org / 10 . 5281 / zenodo .
19192941
[12] K. Khabarlak, “Fine-tuned DINOv2 ViT-L for cu-
cumber growth stage classification: NALEF food
safety challenge solution,” inProceedings of the
NALEF 2026 Workshop, Cartagena, Colombia,
ser. CEUR Workshop Proceedings, In press, 2026.
[13] J.-S. G. Kim, S. H. Shin, and S. Chung, “Instance-
level phenotype-based growth stage classification
of basil in multi-plant environments,”Frontiers in
Plant Science, vol. 16, p. 1707985, Nov. 18, 2025,
ISSN: 1664-462X. DOI:10 . 3389 / fpls .
2025.1707985[Online].Available:https://
www . frontiersin . org / articles / 10 .
3389/fpls.2025.1707985/full
[14] C. Zhu, H. Yu, T. Lu, Y. Li, W. Jiang, and Q. Li,
“Deep learning-based association analysis of root
image data and cucumber yield,”The Plant Jour-
nal,vol. 118,no. 3,pp.696–716, May2024, ISSN:
0960-7412, 1365-313X. DOI:10 . 1111 / tpj .
16627Accessed: May 25, 2026. [Online]. Avail-
able:https : / / onlinelibrary . wiley .
com/doi/10.1111/tpj.16627
[15] K. Khabarlak, I. Laktionov, and G. Diachenko,
“Feature knowledge distillation using group con-
volutions for efficient plant pest recognition,” in
Proceedings of the 12th International Confer-
ence Information Control Systems & Technologies
(ICST 2024), Odesa, Ukraine, September 23–25,
ser.CEURWorkshopProceedings,vol.3790,2024,
pp. 377–387. [Online]. Available:https : / /
ceur-ws.org/Vol-3790/paper33.pdf
[16] R. Singh, R. Bidese-Puhl, K. Dhakal, and S. Sor-
napudi, “Few-shot adaptation of grounding DINO
for agricultural domain,” inIEEE/CVF Conference
onComputerVisionandPatternRecognitionWork-
shops,CVPRWorkshops2025,Nashville,TN,USA,
June 11-15, 2025, Computer Vision Foundation /
IEEE, 2025, pp. 5332–5342. [Online]. Available:
https : / / openaccess . thecvf . com /
content/CVPR2025W/V4A/html/Singh_
Few- Shot_Adaptation_of_Grounding_
DINO _ for _ Agricultural _ Domain _
CVPRW_2025_paper.html
[17] A. Picon et al.,Mitigating domain drift in multi
species segmentation with dinov2: A cross-domain
evaluationinherbicideresearchtrials,2026.arXiv:
2508 . 07514 [cs.CV]. [Online]. Available:
https://arxiv.org/abs/2508.07514
[18] K. He, X. Zhang, S. Ren, and J. Sun, “Deep resid-
ual learning for image recognition,” in2016 IEEE
ConferenceonComputerVisionandPatternRecog-
nition, CVPR 2016, Las Vegas, NV, USA, June 27-
30, 2016, IEEE Computer Society, 2016, pp. 770–
778. DOI:10.1109/CVPR.2016.90
[19] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T.
Darrell, and S. Xie, “A ConvNet for the 2020s,”
inIEEE/CVF Conference on Computer Vision and
PatternRecognition,CVPR2022,NewOrleans,LA,
USA, June 18-24, 2022, IEEE, 2022, pp. 11966–
11976. DOI:10 . 1109 / CVPR52688 . 2022 .
01167[Online]. Available:https : / / doi .
org/10.1109/CVPR52688.2022.01167
[20] M. Tan and Q. V. Le, “EfficientNet: Rethinking
model scaling for convolutional neural networks,”
inProceedingsofthe36thInternationalConference
onMachineLearning,ICML2019,9-15June2019,
Long Beach, California, USA, ser. Proceedings of
Machine Learning Research, vol. 97, PMLR, 2019,
pp. 6105–6114.
[21] M. Tan and Q. V. Le, “EfficientNetV2: Smaller
models and faster training,” inProceedings of the
38th International Conference on Machine Learn-
ing, ICML 2021, 18-24 July 2021, Virtual Event,
vol. 139, PMLR, 2021, pp. 10096–10106.
10

[22] A. Dosovitskiy et al., “An image is worth 16x16
words: Transformers for image recognition at
scale,”in9thInternationalConferenceonLearning
Representations, ICLR 2021, Virtual Event, Aus-
tria, May 3-7, 2021, OpenReview.net, 2021. [On-
line].Available:https://openreview.net/
forum?id=YicbFdNTTy
[23] G. E. Hinton, N. Srivastava, A. Krizhevsky, I.
Sutskever, and R. Salakhutdinov, “Improving neu-
ral networks by preventing co-adaptation of fea-
ture detectors,”CoRR, vol. abs/1207.0580, 2012.
arXiv:1207.0580. [Online]. Available:http:
//arxiv.org/abs/1207.0580
[24] K. Khabarlak, “Post-train adaptive MobileNet for
fast anti-spoofing,” inProceedings of the 3rd Inter-
nationalWorkshoponIntelligentInformationTech-
nologies&SystemsofInformationSecurity,Khmel-
nytskyi, Ukraine, March 23-25, 2022, ser. CEUR
Workshop Proceedings, vol. 3156, CEUR-WS.org,
2022, pp. 44–53. [Online]. Available:http : / /
ceur-ws.org/Vol-3156/keynote5.pdf
[25] G.Huang,Y.Sun,Z.Liu,D.Sedra,andK.Q.Wein-
berger, “Deep networks with stochastic depth,” in
Computer Vision - ECCV 2016 - 14th European
Conference, Amsterdam, the Netherlands, October
11-14, 2016, Proceedings, Part IV, ser. Lecture
Notes in Computer Science, vol. 9908, Springer,
2016, pp. 646–661. DOI:10 . 1007 / 978 - 3 -
319-46493-0_39
[26] G. Song and W. Chai, “Collaborative learning
for deep neural networks,” inProceedings of the
32nd International Conference on Neural Informa-
tion Processing Systems, ser. NIPS’18, Montréal,
Canada: Curran Associates Inc., 2018, pp. 1837–
1846.
[27] A. Galdran, J. W. Verjans, G. Carneiro, and M. A.
González Ballester, “Multi-head multi-loss model
calibration,” inMedical Image Computing and
Computer Assisted Intervention – MICCAI 2023,
Cham:SpringerNatureSwitzerland,2023,pp.108–
117, ISBN: 978-3-031-43898-1.
11