论文信息

论文解读

AI 辅助整理
方法
色彩直方图与距离度量,2172张无人机影像,28个南美地点验证。
发现
低数据下Top-1准确率81%—85%,ESP32部署达82.8%,但精度逊于NDVI。
机会点
可探索融合色彩与NDVI的轻量模型,并开展田间验证以提升边缘部署可靠性。

论文摘要

轻量可解释的植被健康分类算法,每类仅需5张图、无需GPU,以色彩直方图与距离度量在28个南美地点2172张无人机影像验证。低数据下Top-1准确率81%—85%,部署ESP32达82.8%。属预印本;纯色彩精度逊于NDVI,未田间验证。

全文

arXiv (2026)

Vegetation Health Classification for Low-Data and Edge Hardware Deployment
Yifei Cheng
Canton High School
Canton Township, MI
yifeichengmi@gmail.com
Haoyu Li
Grand Valley State University
Allendale, MI
lihao@gvsu.edu
Abstract
Unmanned aerial vehicles (UAVs) equipped with cameras
are popular, viable tools for vegetation and environmen-
tal health monitoring. However, existing image classifica-
tion methods are incompatible with UAV implementation for
independent practitioners in remote, resource-constrained
environments due to large dataset requirements, GPU hard-
ware, inflexible retraining, and low interpretability. This
paper presents a lightweight, interpretable vegetation clas-
sification algorithm operating with as few as five refer-
ence images per class, without GPUs or heavy machine
learning frameworks. It uses a 2D hue-saturation his-
togram feature representation and Earth Mover’s Distance
or Quadratic Form Distance, paired with a log-likelihood
probability score, to classify images into three vegetation
health categories. We evaluated six systems: three cus-
tom pipelines and three established models across low-
data scenarios (n= 5,10,25) and a full-data regime of
2,172 drone images from 28 South American locations.
Our custom pipelines achieve competitive Top-1 accuracy
(81–85%) under low-data constraints, maintaining sub-1ms
inference times that enable real-time classification and sub-
sequently tasks such as seeding and watering. The al-
gorithms also support incremental retraining without full
model reconstruction. We validate edge feasibility by de-
ploying the QFD variant on an ESP32 microcontroller,
achieving 82.8% accuracy with 1.01% heap memory uti-
lization (∼45kb), demonstrating that effective vegetation
classification is achievable on low-cost, low-power embed-
ded hardware without cellular connectivity.

  1. Introduction
    Unmanned aerial vehicles (UA Vs) with cameras and video
    recorders have emerged as a powerful tool for environmen-
    tal monitoring due to their cost effectiveness and versatil-
    ity. The role of drone imagery as a vegetation health as-
    sessment tool stretches across numerous domains and ap-
    plications, such as wildfire risk monitoring, agriculture, and
    biodiversity conservation, especially when tools such as in-
    frared imaging aren’t available [4, 10, 11, 21, 36].
    Compared to the Normalized Difference Vegetation In-
    dex (NDVI), which is based on infrared imaging, pure color
    detection systems are less accurate in terms of measur-
    ing plant health since it actually measures physiological
    changes in chlorophyll density and plant stress, which are
    only visible in more severe stages in pure color. This makes
    NDVI a great choice for large scale environmental moni-
    toring. However, in many practical scenarios independent
    practitioners or small organizations working on remote con-
    servation efforts do not have access to infrared imaging, ne-
    cessitating pure color classification [22, 23].
    This is especially true for commercial drones, which
    are increasingly being equipped with onboard cameras and
    used for exactly this kind of monitoring work. These plat-
    forms are often operated in remote areas without cellular
    or network coverage, making it impractical or impossible
    to stream imagery back to a remote computer for process-
    ing. They also typically lack the dedicated GPU hardware
    found on higher-end research-funded platforms, and may
    rely on legacy microcontrollers or other low-power embed-
    ded hardware for onboard computation.
    In addition, many tasks done by drones reqiure real-time
    classification to make decisions under field conditions, such
    as reforestation seeding and watering [2, 29]. However, far
    fewer of these models simultaneously satisfy the computa-
    tional and logistical constraints faced by resource-limited
    practitioners operating in remote field environments [3].
    To address the aforementioned practical issues with pre-
    vious approaches, this paper presents a lightweight, inter-
    pretable vegetation classification algorithm designed to op-
    erate with as few as five reference images per class, with-
    out requiring GPU hardware, extensive model retraining,
    or heavy external machine learning frameworks. The al-
    gorithm is designed with the out-of-the-box practitioner
    in mind: rather than assuming access to hyper-optimized
    pipelines or high-performance infrastructure, it targets the
    hardware and workflows that independent field operators
    are most likely to actually have available.
    1
    arXiv:2611.19495v1 [cs.CV] 14 Nov 2026

Because our algorithm runs its inference entirely on-
device, it is well suited to deployment directly onboard
commercial drones and low-power embedded hardware, in-
cluding in areas where cellular or other network signals can-
not reach, removing the need to transmit imagery back to a
remote computer for processing and also increasing opera-
tional range. To validate this directly, we deploy and bench-
mark the algorithm on an ESP32 microcontroller as a rep-
resentative example of the low-cost, low-power hardware
a practitioner might have on hand that has been proven to
possess capabilities for computer vision processing [8, 9].
2. Related Work
This section provides an overview of existing literature,
highlighting capabilities of existing small classification
models while also noting constraints in these methods.
2.1. Aerial Vegetation Monitoring and Color Spaces
The field of real-time automated vegetation classification
has grown alongside both drone and machine learning tech-
nology. In particular, supervised approaches such as Sup-
port Vector Machines (SVMs) and CNNs have demon-
strated great, deployable accuracy scores for identifying
vegetation health [16, 25, 26]. Neural networks have also
gained massive popularity recently. With the help of Im-
ageNet, architectures such as Ultralytics’ YOLOv8 are
adaptable enough to fit into this domain, achieving ex-
tremely high accuracy with massive datasets [17].
HSV was selected as the color representation method
over RGB because prior research has consistently demon-
strated that luminance functions as a confounding factor
in RGB-based image classification [15, 28]. Furthermore,
HSV facilitates more straightforward distance calculations
through eliminating the need for normalization and en-
abling efficient value isolation. This methodological choice
is well-established within the literature [13, 30, 32, 38, 39].
2.2. Deep Learning and Few-Shot Baselines in Data-
Scarce Regimes
Additionally, Few Shot Learning (FSL) methods have been
proposed as a solution to low-data situations. By learning
generalizable feature embeddings from related tasks or by
employing metric-based similarity frameworks, FSL meth-
ods have been proven to be robust, accurately classifying
vegetation scenes with as few as five labeled examples per
class [27, 35, 40].
While these advancements represent significant steps
forward, they present meaningful practical limitations that
motivate the approach taken in this study.
• Massive data requirements: The performance of super-
vised deep learning and Support Vector Machines comes
at the cost of a massive data requirement. They often re-
quire hundreds to thousands of annotated images to be
robust, requiring many hours by experts to sort through
and categorize them for training.
• High computational expenses during training time, resul-
tant inflexible training cycles: SVMs and neural networks
require a complete retraining cycle to be completed when
even minor amounts of new data is added [20]. In field
conditions where an operator might want to continuously
add newly collected data to a model, this restraint is in-
credibly difficult to work around with.
• Black-box neural networks: Neural networks and FSL
methods essentially act as black boxes [19]. After train-
ing, the embeddings and internal representations of im-
ages are high-dimensional and unintuitive, and as a result
essentially inaccessible to any independent practitioner
trying to apply these technologies.
• Low interpretability in SVMs: SVMs seem like a solu-
tion to this problem at first; they are more transparent
and offer a more intuitive and well-understood geomet-
ric representation of images, but in most cases that in-
volve vegetation classification, feature extraction must be
programmed specifically for this domain, and a complex
kernel has to be applied. For the vast majority of users ap-
plying SVMs, they essentially become a black box after
applying a kernel trick due to the program now introduc-
ing infinite dimension space and no directly interpretable
feature weights [34].
2.3. Edge Deployment Constraints
Although FSL methods largely solve the data issue, they
almost always rely on deep feature extraction based off
of frameworks such as ResNet and EfficientNet [14, 33].
While highly effective and easy to deploy, they are compu-
tationally expensive both during training and on inference
time, especially when deployed on devices without dedi-
cated GPUs, older field laptops, the processors on commer-
cial drones, and legacy microcontrollers. For real-time as-
sessment and analysis, this computational overhead is a se-
rious challenge and delays conservation efforts.
3. Methodology
This section outlines how we organized the comparative
study to test our lightweight algorithm against traditional
and deep learning / neural baselines.
3.1. Dataset Sourcing
The dataset consists of drone-captured vegetation imagery
sorted into three classes: Dry, mixed, and dense. Dry veg-
etation consists of essentially no healthy vegetation and is
usually either completely bare or dried out. Mixed vegeta-
tion images have some mixture of healthy green vegetation
together with dried out vegetation or bare dirt. Dense vege-
tation images are dominated by green, lush vegetation.
2

Figure 1. Archetypical examples of the three vegetation classes
used in this study. Left: Dense. Center: Mixed. Right: Dry.
There are a total of 2172 images in the dataset. 783
of them are dense vegetation, 625 are dry, and 764 are
mixed. The drone imagery is taken from 28 different lo-
cations within South America from OpenAerialMap [1]. 21
of these locations were sorted for all three classes, and an
additional 7 locations were sorted specifically to add to the
dry class due to a severe dataset size mismatch in compari-
son to the mixed and dense classes. Out of the dry images,
330 come from the original 21 locations, and the additional
295 were supplemented from the additional locations.
All 2172 images were manually classified by the first au-
thor to establish ground truth. Ambiguous, low-quality, or
nondescript images were excluded.
3.2. Feature Representation and Extraction
All systems except for the Siamese Network, utilize the
same standardized feature representation layer. This shared
foundation isolates algorithm-level differences in perfor-
mance and accuracy from variation based on feature extrac-
tion.
The 2D histogram employed by four of the five systems
evaluated comprises 15 hue columns and 8 saturation rows.
Each hue column covers 24 degrees of the color wheel,
while each saturation row covers 12.5% of the saturation
spectrum.
When constructing the histogram, pixels with values ex-
ceeding 240 or falling below 20 are excluded from consid-
eration. These extreme values are perceptually ambiguous
and usually just come from camera glare, deep shadows, or
image processing artifacts. As a result, they contain very lit-
tle real information about the hue composition of an image
and are disregarded.
Finally, each pixel is assigned to the corresponding his-
togram cell, and the resulting histogram is normalized to ac-
count for variations introduced by the value threshold and
potential differences in image dimensions. This feature his-
togram is shared across all systems.
Prior to inference, both EMD algorithms prune the his-
togram for values under 0.5% and renormalize the his-
togram in order to avoid slowing down the algorithm with
more bins with tiny values that largely do not end up influ-
encing the final perceptual composition of the image.
The Siamese Network purely utilizes a CNN as a feature
extraction method since it is the industry standard feature
extraction pairing for Siamese networks (Although tech-
nologies such as Vision Transformers are emerging as more
popular, state-of-the-art alternative for CNNs that require
more data and compute power) and CNNs were specifically
made for the purpose of image analyzation.
3.3. Similarity Metrics and Mathematical Frame-
work
Earth Movers Distance (Abbreviated as EMD) and
Quadratic Form Distance (Abbreviated as QFD) are the two
cross-bin distance metrics of choice for the custom algo-
rithm, both respecting the shape of the histogram distribu-
tions rather than calculating distance based on isolated, bin-
by-bin comparisons [12, 24].
EM D(P, Q) =
minF∈F
P
i
P
j fijdijP
i
P
j fij
,whereF= [−f ij]is the optimal flow between bins of
histogramsPandQ, andd ij is the perceptually weighted
ground distance.
EMD treats each histogram as a distribution of ”mass”
spread across bins and computes the minimum cost required
to reshape or move one distribution into the other via a lin-
ear programming problem, calculating cross-bin perceptual
distance where cost is a function of how far and how much
mass travels [18].
However, although EMD is incredibly accurate at cap-
turing cross-bin distribution similarity, the linear program-
ming approach scales cubicly and requires a high memory
footprint since the optimization needs to search over many
possible reassignments of mass, which becomes expensive
as histogram resolution increases.
QF D(P, Q) =
q
(P−Q) T A(P−Q)
,whereAis a similarity matrix with entriesa ij = 1−
(dij/dmax).
QFD approximates this same cross-bin sensitivity with-
out an iterative solver. Instead of optimizing a flow, it di-
rectly computes a weighted distance between the two his-
togram vectors, where the weighting matrix encodes how
similar any given pair of bins is expected to be. Because this
reduces to a fixed matrix multiplication rather than a search
process, QFD is far cheaper to compute and far more pre-
dictable in memory usage, at the cost of using a fixed, pre-
computed notion of bin-to-bin similarity rather than solving
for the true optimal transport each time. Although EMD
is more robust at handling variable distribution sizes and
dramatic feature shifts, QFD is an efficient alternative for
ESP32 testing.
Within the distance calculation, a polar mapping ensures
that hue differences in lower saturation bins are penalized
less than in highly saturated bins, reflecting the perceptual
3

ambiguity of low-saturation colors. For histogram bini
with hue indexh i and saturation indexs i (out ofN H hue
bins andN S saturation bins), we first map the bin into a
2D polar coordinate. The ground distanced ij between bins
iandjused in both EMD and QFD is then the Euclidean
distance between their polar-mapped coordinates:
dij =
q
(xi −x j)2 + (yi −y j)2
Sincer i scales linearly with saturation, bins with low
saturation map close to the origin regardless of their hue
angleθ i, compressing the effective distance between hues
at low saturation. This distance is finally normalized by the
maximum pairwise distanced max observed across all bin
pairs, giving the similarity weightinga ij = 1−(d ij/dmax)
used in the QFD similarity matrixA.
Figure 2. Visualization of radial saturation calculation. Bins with
identical relative hue placements are penalized less with lower sat-
uration due to being closer to the origin, reflecting real-life percep-
tual ambiguity
Alongside the geometric distance metrics, the pipeline
design incorporates a log-likelihood term to help disam-
biguate images that may sit at similar geometric distances
from class archetype histograms but differ in how their
probability mass is distributed across bins.
The averaged reference histogram for each class is
treated as an empirical probability distribution over hue-
saturation bins. Taking its natural log (with a small epsilon
to prevent log(0)) converts it into per-bin log-probabilities.
The test histogram is then scored against each class via a
dot product.
Since both histograms are normalized, this is mathemat-
ically equivalent to negative cross-entropy, which is asking
how likely the test image’s color distribution would be if
drawn from each class. This score is min-max normalized
and combined with the distance score, making the two terms
act as independent votes. The distance metric captures dis-
tributional shape difference, while the log-likelihood cap-
tures class fit. During ablation testing, combining both was
found to stabilize predictions against the mixed class, whose
intermediate color profile can otherwise pull single-metric
classifiers toward it, raising Top 1 Accuracy by around5%.
4. Implementation & Evaluation
This section details the implementation of each system as
used in this research and how each will be evaluated. To
keep the comparative analysis fair, all six systems are tested
using the same image data across three low data constraints
(n=5, n=10, and n=25 reference images per class) and one
full data situation. These constraints help simulate early
conditions faced by independent practitioners when they
might not have access to extensive training data. Addition-
ally, to evaluate real-world feasibility on edge hardware, all
systems are tested across two test sessions: one on Google
Colab and one on an ESP32 microcontroller.
4.1. Tested Systems
We evaluate six system configurations using identical sam-
ple sizes: three custom histogram distance pipelines and
three standard machine learning baselines. To reflect
field conditions for environmental practitioners without
high-performance computing infrastructure, baselines uti-
lize standard out-of-the-box configurations. Additionally,
all systems perform unbatched, single-image inference to
simulate real-time on-the-fly streaming over low-bandwidth
connections.
•EMD-Average (Custom):Computes an average 2D ref-
erence histogram acrossnsample images per class, clas-
sifying test images using Earth Mover’s Distance (EMD).
•EMD-Median (Custom):Identical pipeline to EMD-
Average, but uses a bin-wise median reference histogram
to provide outlier resistance against class distribution
skew.
•QFD Variant (Custom):Shares the shared 2D histogram
representation but replaces EMD with Quadratic Form
Distance (QFD) for hardware efficiency.
•Siamese Network Baseline:A few-shot neural net-
work using a four-block CNN encoder backbone (32,
64, and 128 channels) projecting into a 64-dimensional
L2-normalized embedding space [7]. Trained with con-
trastive loss (margin 1.0, Adamlr= 10 −3, batch size 32,
30 epochs) on balanced 50/50 positive and negative pairs.
At inference, it acts as a prototypical classifier, assign-
ing test image embeddings to the nearest class prototype
calculated from average reference embeddings.
•Random Forest Baseline:Ascikit-learnforest
of 200 estimators with balanced class weighting trained
directly on the flattened 120-bin histogram [6]. Tex-
ture channels (LBP) were omitted as preliminary test-
ing showed marginal accuracy improvements (+0.71%)
at prohibitive latency costs.
•SVM Baseline:A standardscikit-learnSVM with
an RBF kernel and standard feature scaling, trained on
the shared 120-element flattened histogram with balanced
class weighting [31].
4

4.2. Preliminary Development & Design Rationale
Iterative development guided the algorithm’s architecture
and feature selection. Initial tests using a coarse8×4
hue-saturation histogram suffered from high classification
bias toward the intermediate mixed class. Increasing res-
olution to15×8resolved this bias by establishing tighter
class boundaries.
While EMD-Average and EMD-Median leverage cross-
bin properties, preliminary hardware profiling revealed crit-
ical edge deployment barriers. EMD’s iterative linear pro-
gramming and lack of vectorizability create severe mem-
ory overhead and sequential computation bottlenecks on
the ESP32 microcontroller, compounded by the hardware’s
lack of a floating-point divider. This necessitated the QFD
Variant, which computes all distances simultaneously via
matrix operations.
In evaluating feature sets, adding Local Binary Pattern
(LBP) texture channels (sampled at radii 1, 3, and 5) to
the Random Forest and SVM yielded a negligible top-1 ac-
curacy gain (+0.71%) while increasing end-to-end latency
by 100 ms, prompting its exclusion. A naive RGB base-
line withL 2 distance achieved only 50.84% top-1 accuracy
atn= 5and plateaued at 39% mixed-class accuracy at
n= 25(σ= 8.1%), confirming the inadequacy of basic
color averaging.
To resolve ambiguous geometric distances among
classes with distinct probability distributions, a log-
likelihood scoring term was integrated into the custom
pipelines, yielding a 5% top-1 accuracy improvement in ab-
lation testing with minimal latency. Finally, profiling re-
vealed that sparse, near-zero histogram bins severely bot-
tlenecked OpenCV’s EMD solver (O(K 3 logK)complex-
ity). Implementing a 0.5% bin pruning threshold reduced
per-image inference time from 16 ms to sub-1 ms without
negatively altering baseline metrics.
4.3. Experimental Procedure and Edge Validation
To simulate real-world field or data-scarce conditions, we
extract a permanent, isolated testing set, made up of 20%
of the images from each class. The test set was a strat-
ified sample to ensure that the slight class size mismatch
wouldn’t cause any issues. The remaining images form the
training pool. For each test run, we sample exactly n im-
ages per class to act as the training/reference data. Every
experiment configuration is mapped and executed during
our primary benchmarking platform (Google Colab), and
the ESP32-optimized QFD Variant on an ESP32 microcon-
troller.
Because low-data training is highly sensitive to outlier
images and variation, each system for each data constraint
is tested 20 times with n randomly chosen images. The fi-
nal metrics for these systems represent the average across
all runs. For the full-data scenarios, each system was only
tested on one round since additional rounds would contain
the same training data and would therefore be redundant.
4.3.1. Cloud Benchmarking Platform (Google Colab)
All training and inference testing in Google Colab is done
on the CPU for fairness and to match domain needs.
4.3.2. Edge Device Deployment (ESP32 Microcontroller)
To validate that our algorithm runs on low-end hardware
without GPU, we test directly on an ESP32 microcontroller,
which will serve as a representative platform for legacy,
low-power hardware found on commercial drone flight con-
trollers. Preliminary evaluation showed EMD’s iterative lin-
ear programming is incompatible with ESP32 resources, so
edge validation will test the QFD Variant, which shares the
feature extraction pipeline but uses predictable matrix op-
erations that conform to the ESP32’s hardware limitations
(See section Preliminary Development).
Only the QFD Variant is tested on ESP32 at n=5 to
demonstrate feasibility. Since QFD stores exactly three ref-
erence histograms regardless of n (one per class), memory
usage is invariant to training sample count, making higher-
n testing redundant. Accuracy scaling is already character-
ized by Google Colab results.
Training occurs in Google Colab; only inference is eval-
uated on ESP32. Reference histograms are compiled and
flashed onto the device to simulate real-world deployment.
The 15 training images (3 groups of 5) were randomly cho-
sen for one trial.
4.4. Evaluation Metrics
Each system is benchmarked against nine core performance
metrics on our main hardware platform. Of these, only the
metrics related to inference (Total Accuracy, Partial Ac-
curacy, Inference Time, Compute Power during Inference
Time, and RAM Usage during Inference Time) are also
measured during the ESP32 session.
• Total Accuracy (Top-1): The percentage of test images
placed into exactly the correct class.
• Partial Accuracy: The percentage of test images placed
into either the exact right class or an adjacent class. Be-
cause our classes are linear (Dry, Mixed, Dense), mis-
classifying a dense image as mixed is treated as a partial
success, whereas misclassifying dense as dry is a total
failure.
• Inference Time per Image: The average time (in millisec-
onds) required to process and classify a single input im-
age after the model is established.
• Training time per round (Only tested on main platform)
• Memory usage during inference (Only tested on ESP32)
Feature extraction time and data ingestion time is also iso-
lated and measured in Google Colab. Since all except for
the Siamese Network utilize the same feature extraction
method and ingests data at the same rate, they will not be
5

taken into consideration during comparative benchmarking,
and will be noted in results independently.
5. Results
5.1. Accuracy Under Low Data Constraints
Across low-data regimes, the custom pipelines (EMD-
Average, EMD-Median, and QFD Variant) consistently
achieve competitive or superior Top-1 accuracy over neural
and kernel baselines, demonstrating small sample efficiency
atn= 5andn= 10.
This low-data advantage narrows asnincreases, with
Random Forest eventually surpassing all custom variants
atn= 25. EMD-Average and the QFD Variant exhibit
nearly identical performance at the smallest sample size
before plateauing near 85% Top-1 accuracy. Something
interesting is that EMD-Average consistently suffers from
severe class imbalance, with dry and dense classes having
significantly ( 25%) more accurate placements compared to
mixed.
The SVM struggles most severely under extreme data
scarcity atn= 5, confirming that kernel-based classifiers
require broader training support to generalize, while its low
partial accuracy highlights a susceptibility to catastrophic
misclassification. The Siamese Network displays similar
low-nfragility before recovering strongly atn= 25. Ran-
dom Forest proves to be the most robust baseline, overtak-
ing the custom pipelines at higher sample counts. In ad-
dition, partial accuracy remains uniformly high across all
six systems and data regimes. They reach generally 95%
and reaching 100% for custom pipelines at highern, which
hows that high-nmisclassifications reflect benign bound-
ary shifts into neighboring classes rather than severe model
failures.
n = 5 n = 10 n = 25
EMD Averaging 81.1% ± 4.6%85.0% ± 2.8% 85.1% ± 1.8%100% 100% 100%
EMD Median 74.1% ± 6.0% 80.9% ± 2.9% 81.4% ± 2.6%100% 100% 100%
QFD Variant 81.4% ± 4.6%82.7% ± 2.3% 83.4% ± 1.4%99.97%100% 100%
Siamese Network 67.3% ± 9.7% 76.3% ± 5.8% 84.3% ± 5.5%98.56% 99.69% 99.96%
Random Forest 79.3% ± 5.2% 85.3% ± 4.6% 89.0% ± 2.3%94.00% 98.00%100%
SVM 61.8% ± 6.1% 70.5% ± 3.8% 77.3% ± 2.4%81.00% 82.00% 90%
Table 1. Top 1 macro accuracy ± standard deviation (Above) and
partial accuracy (Below)
5.2. Inference and Training Time Under Low Data
Constraints
With the exception of the Siamese Network baseline, all
systems utilize the identical 2D Hue-Saturation histogram
feature extractor, resulting in a consistent feature extrac-
tion time per image of ˜26 ± 7 ms. For the EMD and QFD
pipelines, this feature extraction time is the primary bottle-
neck. The Siamese Network utilizes a traditional CNN fea-
ture extractor, among the most common methods of feature
extraction paired with this baseline.
Forn= 5,10, and25, all three custom pipelines achieve
the shortest inference and training time in comparison to the
three baselines.
For the Siamese network, the flat inference time across
n values is a direct consequence of how the CNN oper-
ates: at inference, the network performs a single forward
pass through a fixed convolutional architecture to produce
an embedding vector, and then computes a similarity score
against all of the reference embeddings. Since the forward
pass cost depends only on image resolution and network
depth instead of how many reference images were used dur-
ing training, adding more reference samples does not in-
crease per-image inference latency. This contrasts with the
histogram-based systems, where inference involves com-
puting EMD or QFD distances against aggregated reference
histograms, though in practice these operations remain fast
due to the compact histogram representation.
n = 5 n = 10 n = 25
EMD Averaging 0.939 ± 0.508 1.009 ± 0.530 1.013 ± 0.519
EMD Median 0.741 ± 0.400 0.769 ± 0.325 0.734 ± 0.309
QFD Variant0.315 ± 0.074 0.316 ± 0.084 0.314 ± 0.086
Siamese Network 19.896 ± 4.139 19.896 ± 4.139 20.067 ± 4.281
Random Forest 16.196 ± 5.132 15.919 ± 4.384 15.990 ± 4.701
SVM 1.089 ± 0.351 1.076 ± 0.234 1.092 ± 0.241
Table 2. Inference time per image ± Standard deviation (ms)
For all three custom pipeline variants, training time
scales linearly in accordance withnsince additional im-
ages take the same amount of time to aggregate into the
histogram for training, yet remain largely manageable. The
SVM keeps pace with the custom pipeline variants, and all
four systems remain consistently the fastest to train across
allnvalues.
The Random Forest is the second slowest system for
training, scaling from 1.110 ± 0.201 seconds atn= 5to
4.035 ± 0.434 seconds atn= 25, slightly above the custom
pipelines and SVM at comparablenvalues.
The Siamese Network is the slowest by a substantial
margin, reaching 173.9 ± 4.372 seconds at justn= 5and
6

rapidly scaling up asnincreases. This reflects the cost
of fine-tuning or re-training the convolutional architecture
on CPU without GPU acceleration, where backpropagation
through the network layers grows proportionally with the
number of training samples. The result is that even at small
values ofn, the Siamese Network’s training overhead ex-
ceeds that of every other system by roughly two orders of
magnitude.
n = 5 n = 10 n = 25
EMD Averaging 0.721 ± 0.152 1.483 ± 0.224 3.514 ± 0.457
EMD Median0.702 ± 0.121 1.397 ± 0.2183.505 ± 0.452
QFD Variant 0.735 ± 0.114 1.539 ± 0.256 3.667 ± 0.495
Siamese Network 173.9 ± 4.372 347.5 ± 4.012 892.8 ± 7.684
Random Forest 1.110 ± 0.201 1.824 ± 0.274 4.035 ± 0.434
SVM 0.763 ± 0.174 1.443 ± 0.1973.480 ± 0.272
Table 3. Training time per group ofn± standard deviation (sec)
5.3. Full Data Environment
In the full-data setting, the accuracies observed at lown
reverses: all three baselines now outperform the custom
pipelines on Top-1 accuracy, with Random Forest (94.67%)
and the Siamese Network (93.25%) achieving the high-
est accuracy, followed by SVM (90.84%). The custom
pipelines plateau in a tight band around 83–85% regard-
less of the distance metric used, indicating that the 2D hue-
saturation histogram representation itself, rather than the
choice of EMD-Averaging, EMD-Median, or QFD, is the
limiting factor once enough labeled data is available to fully
exploit the higher-capacity baselines.
In the full data scenario, although all systems maintain a
100% partial accuracy, all baselines perform better in terms
of Top-1 accuracy compared to our custom systems. Cus-
tom systems plateau around 85% Top-1 accuracy.
Notably, partial accuracy remains at 100% for every sys-
tem in the full-data regime, so none of the added baseline
accuracy comes at the expense of severe misclassifications.
Top1 Partial
EMD Averaging 86.05% 100%
EMD Median 82.82% 100%
QFD Variant 85.29% 100%
Siamese Net. 93.25% 100%
Random Forest94.67%100%
SVM 90.84% 100%
Table 4. Model accuracy results
Training time diverges sharply: the Siamese Network re-
quires 20014.5 seconds (over 5.5 hours), compared to under
110 seconds for every other system. QFD and both EMD
systems keep the trend of maintaining the fastest inference
out of hte six systems. Both EMD-Averaging and EMD-
Median also achieve sub minute training times per group,
the fastest between all six groups.
In terms of average inference time for all images, SVM
performs the fastest out of the comparison baselines. Al-
though both Siamese Network and Random Forest perform
the slowest in terms of average inference time, even when
taking into account feature extraction ( ˜26ms), on compara-
ble hardware both can theoretically achieve around 20 fps
in a real time practical scenario, still making them viable
for deployment.
The QFD Variant maintains the fastest inference
(¡0.4ms) in the all-data environment, and Systems EMD Av-
erage and EMD Median incur higher inference costs due to
EMD’s iterative optimization, even with the pruning sys-
tem. For all non-neural systems, training times remain
tightly clustered (under 5 seconds at n=25), enabling prac-
tical field retraining. Although performing accurately, the
5.5 hour retraining cycle on a CPU for the Siamese Net-
work indicates that full retraining cycles would be difficult
to integrate into low-end situations for practitioners.
Training Time (sec) Average Inf. (ms)
EMD Averaging 80.187 ± 1.225 1.003 ± 0.528
EMD Median 79.937 ± 0.577 0.735 ± 0.311
QFD Variant 81.865 ± 0.5670.341 ± 0.261
Siamese Net. 20014.526 19.984 ± 3.908
Random Forest67.892 ± 0.62016.004 ± 4.226
SVM 91.013 ± 1.457 1.280 ± 0.338
Table 5. Training and average inference times
5.4. ESP32 Edge Deployment
To validate resource-constrained feasibility, the QFD Vari-
ant was deployed on an ESP32 at n=5, achieving 82.8%
overall accuracy while utilizing just 45,136 bytes of heap
memory (1.01% utilization) with a fragmentation rate of
6.41%.
Per image timing benchmarked on the ESP32 confirms
this efficiency, but also points out a critical real-life bottle-
neck for these types of applications. The vast majority of
the 326.09 ms end-to-end pipeline time is taken up by se-
rial data transfer and overhead, while inference takes only
around 20 ms.
End-to-End Latency 326.09ms
Inference 23.117ms
Serial data transfer + Overhead 302.97ms
Table 6. ESP32 Latency Breakdown
This breakdown demonstrates that the QFD Variant itself
7

is not the constraint on real-time performance; inference ac-
counts for only about 7% of the total pipeline latency. The
remaining 93% is attributable to the serial link between
the host computer and the microcontroller, suggesting that
future deployments using a faster interface (e.g., SPI or on-
board camera capture rather than serial streaming) or an on-
board camera module could reduce end-to-end latency sub-
stantially without any change to the underlying classifica-
tion algorithm. Even with this overhead included, the full
pipeline completes in roughly a third of a second per image
on hardware with only 520KB of internal SRAM, reinforc-
ing the feasibility of deploying this method on low-power,
low-cost field devices.
5.5. Retraining Efficiency
Unlike traditional machine learning models, the proposed
EMD-Average and QFD pipelines do not require full re-
training when new data is acquired, only requiring reag-
gregating the 2D hue-saturation reference histograms by
weighting the incoming images against the existing class
totals. Operators can perform these incremental updates di-
rectly in the field without storing raw dataset images or re-
running optimization passes, offering a distinct operational
advantage under time-stressed or data-constrained environ-
ments.
In contrast, the evaluation baselines introduce significant
retraining overhead:
•SVM:RBF-kernel SVMs must evaluate pairwise interac-
tions across all data points, requiring a complete retrain
from scratch whenever new samples are introduced.
•Random Forest:While new decision trees can techni-
cally be appended, balancing decision weights between
old and new trees is difficult, and modifying existing trees
without full retraining is impractical for field deployment.
•Siamese Network:Adding reference samples requires
fine-tuning or full end-to-end retraining to preserve the
global embedding space. Given the sub-optimal CPU
training latency observed in Section 4.3 (over 5.5 hours),
rolling field updates are not very viable..
6. Discussion & Future Work
We evaluated lightweight 2D hue-saturation histogram clas-
sification pipelines against three commonly used baselines
across low-data (n≤5) and full-data regimes. Our goal
was to assess whether cross-bin metrics paired with prob-
ability scoring are capable of real-time performance under
extreme data scarcity, zero GPU hardware, incremental up-
dates, and FPU-less microcontrollers.
Under low-data constraints (n= 5), the proposed EMD-
Averaging and QFD pipelines achieves 81% Top-1 ac-
curacy matching or exceeding all baselines with sub-1ms
CPU inference, enabling tasks that demand real-time clas-
sification such as forest seeding regeneration and watering
for further action. Their reaggregation scheme also allows
training data to be updated without retraining. Conversely,
SVM struggled at low sample sizes (61.8% atn= 5),
and the Siamese Network’s high training overhead ( 900s
atn= 25on CPU) prevents field updates.
In the full data scenario, high-capacity models (Ran-
dom Forest at 94.67%, Siamese at 93.25%) outperformed
our pipelines (83–86%). This plateau stems from the 2D
HSV representation, which trades fine discriminative reso-
lution and lack of depth for computational efficiency, a nec-
essary tradeoff under certain data constraints. ESP32 edge
deployment validated this approach. The QFD pipeline ap-
proach achieved 82.8% Top-1 accuracy while utilizing 45
KB of heap memory (1.01% utilization). Latency profiling
revealed that a primary real-time bottleneck stems from se-
rial communication rather than classification computation.
To address current representation limits in our compara-
tive analysis portion, future architectural investigations can
explore Prototypical networks as an alternative few-shot
baseline that unifies metric learning with direct class pro-
totype representation. We can also pursue class expansion
beyond the current three class system and explicitly mitigate
the mixed class imbalance caused by transitional intermedi-
ary ground types. Future work should also focus on refining
our methodology and testing by testing on separate, unseen
sources to more effectively identify model robustness.
To support real-world deployment, evaluation can also
expand to multi-region generalization and seasonal robust-
ness tests across diverse vegetative conditions. In addition,
future work can focus on transitioning from stand-alone
ESP32 benchmarking to full integration with production
drone Flight Control Systems, observing practical real-time
onboard inference directly and also potentially integration
with downstream field tasks, such as autonomous regenera-
tive seeding and targeted watering.
7. Conclusion
This paper presented a lightweight, interpretable vegeta-
tion classification pipeline combining 2D hue-saturation
histograms, polar ground-distance mapping, cross-bin met-
rics (EMD/QFD), and log-likelihood scoring. By achieving
81% Top-1 accuracy under extreme low-data regimes (n=
5) with sub-1 ms CPU inference and a 45 KB heap footprint
on ESP32 hardware, the framework can provide insight on
a potential approach on bridging the critical gap between
real-time performance and extreme resource constraints. In
adition, the reaggregation architecture enables instant field-
side updates without extensive retrain cycles, offering a re-
trainable alternative to heavy neural networks on FPU-less
edge devices. Future work should expand the class taxon-
omy, evaluate cross-region seasonal robustness, and inte-
grate the pipeline directly into drone flight control systems
for closed-loop, real-time tasks such as autonomous seeding
and targeted watering.
8

References
[1] Openaerialmap.https://openaerialmap.org. Ac-
cessed: 2026. 3
[2] Youssef Ahansal, Mourad Bouziani, Reda Yaagoubi, Imane
Sebari, Karima Sebari, and Lahcen Kenny. Towards smart
irrigation: A literature review on the use of geospatial tech-
nologies and machine learning in the management of water
resources in arboriculture.Agronomy, 12(2):297, 2022. 1
[3] Andrea Albanese, Matteo Nardello, and Davide Brunelli.
Low-power deep learning edge computing platform for re-
source constrained lightweight compact uavs.Sustainable
Computing: Informatics and Systems, 34:100725, 2022. 1
[4] Karen Anderson and Kevin J Gaston. Lightweight unmanned
aerial vehicles will revolutionize spatial ecology.Frontiers
in Ecology and the Environment, 11(3):138–146, 2013. 1
[5] Fadhillah Azmi, M Khalil Gibran, and Amir Saleh. Herbal
plant image retrieval using hsv color histogram and random
forest algorithm.Journal of Computer Science, Informa-
tion Technology and Telecommunication Engineering, 6(2):
1019–1027, 2025.
[6] Leo Breiman. Random forests.Machine learning, 45(1):
5–32, 2001. 4
[7] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard
S¨ackinger, and Roopak Shah. Signature verification using
a” siamese” time delay neural network.Advances in neural
information processing systems, 6, 1993. 4
[8] Yeong-Hwa Chang, Feng-Chou Wu, and Hung-Wei Lin. De-
sign and implementation of esp32-based edge computing for
object detection.Sensors, 25(6):1656, 2025. 2
[9] Espressif Systems.ESP32-WROVER-E & ESP32-WROVER-
IE Datasheet, 2024. Version 1.4. 2
[10] Rafik Ghali, Moulay A Akhloufi, and Wided Souidene
Mseddi. Deep learning and transformer approaches for uav-
based wildfire detection and segmentation.Sensors, 22(5):
1977, 2022. 1
[11] Luis F Gonzalez, Glen A Montes, Eduard Puig, Sandra John-
son, Kerrie Mengersen, and Kevin J Gaston. Unmanned
aerial vehicles (uavs) and artificial intelligence revolution-
izing wildlife monitoring and conservation.Sensors, 16(1):
97, 2016. 1
[12] James Hafner, Harpreet S. Sawhney, William Equitz, My-
ron Flickner, and Wayne Niblack. Efficient color histogram
indexing for quadratic form distance functions.IEEE trans-
actions on pattern analysis and machine intelligence, 17(7):
729–736, 1995. 3
[13] Mohamed Hassanein, Zahra Lari, and Naser El-Sheimy. A
new vegetation segmentation approach for cropped fields
based on threshold detection from hue histograms.Sensors,
18(4):1253, 2018. 2
[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Deep residual learning for image recognition. InProceed-
ings of the IEEE conference on computer vision and pattern
recognition, pages 770–778, 2016. 2
[15] Jos ´e Luis Hern ´andez-Hern´andez, Gin ´es Garc´ıa-Mateos, JM
Gonz´alez-Esquiva, D Escarabajal-Henarejos, Antonio Ruiz-
Canales, and Jos ´e Miguel Molina-Mart´ınez. Optimal color
space selection method for plant/soil segmentation in agri-
culture.Computers and Electronics in Agriculture, 122:124–
132, 2016. 2
[16] Thani Jintasuttisak, Patompong Chabplan, Sasitorn Issaro,
Orawan Saeung, and Thamasan Suwanroj. Accurate seg-
mentation of vegetation in uav desert imagery using hsv-
glcm features and svm classification.Journal of Imaging,
12(1):9, 2025. 2
[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics
yolo.Zenodo, 2023. 2
[18] Haibin Ling and Kazunori Okada. An efficient earth mover’s
distance algorithm for robust histogram comparison.IEEE
transactions on pattern analysis and machine intelligence,
29(5):840–853, 2007. 3
[19] Zachary C Lipton. The mythos of model interpretability.
Communications of the ACM, 61(10):36–43, 2018. 2
[20] Viktor Losing, Barbara Hammer, and Heiko Wersing. Incre-
mental on-line learning: A review and comparison of state of
the art algorithms.Neurocomputing, 275:1261–1274, 2018.
2
[21] Salvatore Manfreda, Matthew F McCabe, Pauline E Miller,
Richard Lucas, Victor Pajuelo Madrigal, Giorgos Malli-
nis, Eyal Ben Dor, David Helman, Lyndon Estes, Giuseppe
Ciraolo, et al. On the use of unmanned aerial systems for en-
vironmental monitoring.Remote sensing, 10(4):641, 2018.
1
[22] Nathalie Pettorelli, Jon Olav Vik, Atle Mysterud, Jean-
Michel Gaillard, Compton J Tucker, and Nils Chr Stenseth.
Using the satellite-derived ndvi to assess ecological re-
sponses to environmental change.Trends in ecology & evo-
lution, 20(9):503–510, 2005. 1
[23] John W Rouse Jr, R Hect Haas, DW Deering, JA Schell, and
James C Harlan. Monitoring the vernal advancement and ret-
rogradation (green wave effect) of natural vegetation. Tech-
nical report, 1974. 1
[24] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The
earth mover’s distance as a metric for image retrieval.In-
ternational journal of computer vision, 40(2):99–121, 2000.
3
[25] Dee Shi and Xiaojun Yang. Support vector machines for land
cover mapping from remote sensor imagery. InMonitoring
and Modeling of Global Changes: A Geomatics Perspective,
pages 265–279. Springer, 2015. 2
[26] Weibo Shi, Xiaohan Liao, Shaoqiang Wang, Huping Ye,
Dongliang Wang, Huanyin Yue, and Jianli Liu. Evaluation of
a cnn model to map vegetation classification in a subalpine
coniferous forest using uav imagery.Ecological informatics,
87:103111, 2025. 2
[27] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical
networks for few-shot learning.Advances in neural informa-
tion processing systems, 30, 2017. 2
[28] Zhenqi Song, Yuefeng Lu, Ziqi Ding, Dengkuo Sun, Yuanxin
Jia, and Weiwei Sun. A new remote sensing desert vegetation
detection index.Remote Sensing, 15(24):5742, 2023. 2
[29] I Stamatopoulos, TC Le, and Fugen Daver. Uav-assisted
seeding and monitoring of reforestation sites: a review.Aus-
tralian Forestry, 87(2):90–98, 2024. 1
9

[30] Shamik Sural, Gang Qian, and Sakti Pramanik. Segmenta-
tion and histogram generation using the hsv color space for
image retrieval. InProceedings. international conference on
image processing, pages II–II. IEEE, 2002. 2
[31] Shan Suthaharan. Support vector machine. InMachine
learning models and algorithms for big data classification:
thinking with examples for effective learning, pages 207–
235. Springer, 2016. 4
[32] Michael J Swain and Dana H Ballard. Color indexing.Inter-
national journal of computer vision, 7(1):11–32, 1991. 2
[33] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model
scaling for convolutional neural networks. InInternational
conference on machine learning, pages 6105–6114. PmLR,
2019. 2
[34] Alaa Tharwat. Parameter investigation of support vector ma-
chine classifier with kernel functions: A. tharwat.Knowl-
edge and Information Systems, 61(3):1269–1302, 2019. 2
[35] Aleksandr Vladimirovich Uzhinskiy, Gennadii Aleksee-
vich Ososkov, Pavel Vladimirovich Goncharov, An-
drei Vasil’evich Nechaevskiy, and Artem Alekseevich
Smetanin. One-shot learning with triplet loss for vegetation
classification tasks.Computer Optics (In Russian), 45(4):
608–614, 2021. 2
[36] Sheng Wang, Kexin Chen, Ziheng Wei, Linzhe Yang, Qiao
Wang, Mingzhi Liu, Kangjian Cao, Chaoyue Zhao, Runlong
Chang, Zhen Wang, et al. Uav-based deep learning for bio-
diversity monitoring: Advances, applications, and future di-
rections.Ecological Informatics, page 103710, 2026. 1
[37] Pete Warden and Daniel Situnayake.TinyML. O’Reilly Me-
dia Incorporated: Sebastopol, CA, USA, 2019.
[38] Yang Wenbo, Guan Peng, Shi Honglei, Zhang Wei, and Lei
Fuqiang. Identification of grassland vegetation coverage and
height based on vegetation index and hsv space.Journal of
Resources and Ecology, 16(4):933–945, 2025. 2
[39] Wenzhu Yang, Sile Wang, Xiaolan Zhao, Jingsi Zhang, and
Jiaqi Feng. Greenness identification based on hsv decision
tree.Information Processing in Agriculture, 2(3-4):149–160,
2015. 2
[40] Tao Zhang, Yuge Bi, and Chuanzhong Xuan. Convolu-
tional transformer attention network with few-shot learning
for grassland degradation monitoring using uav hyperspec-
tral imagery.International Journal of Remote Sensing, 45
(6):2109–2135, 2024. 2
10

论文原文与出版信息

arXiv (2026)
查看论文原文

继续探索