01 — Doctoral research · BINUS University

TANDAN

Technology for Affordable Near-field Drone-based Assessment of Oil Palm Plantations

Affordable eyes in the sky for oil palm plantations.

5tree conditions detected

17 FPSsustained engine throughput on a Jetson Nano

0.9694mAP@0.5:0.95 at 768 px (0.9648 at the deployed 704 px)

See the drone Read the research

The TANDAN drone flying between oil palm rows during field testing.

02/14The problem

Spotting a struggling palm among thousands

Indonesia is the world's largest palm oil producer. Oil palm contributes roughly 17% of Indonesia's agricultural GDP and employs up to 7.8 million people across its value chain (Purnomo et al., 2020), and palm oil supplies about 40% of global vegetable oil demand (Qaim et al., 2020). Indonesia's palm oil exports reached USD 22.97 billion in 2020 (Xin et al., 2022).

Keeping plantations productive means spotting problem trees early: dead palms, yellowing crowns, poorly maintained trees and stunted growth. Diseases such as basal stem rot (Ganoderma boninense) often become visible only at a late stage. Walking thousands of hectares to inspect trees by eye is slow, costly and inconsistent. Commercial drone-analytics services exist, but they are expensive and usually depend on cloud processing that remote plantations cannot rely on.

~17%of Indonesia's agricultural GDP (Purnomo et al., 2020)

7.8 Mpeople employed across the value chain, at most (Purnomo et al., 2020)

~40%of global vegetable oil demand (Qaim et al., 2020)

USD 22.97 BIndonesian palm oil exports in 2020 (Xin et al., 2022)

“The question was not whether AI can recognize a struggling palm — it was whether it can do so on hardware a plantation can actually afford.”

03/14What TANDAN is

An end-to-end system in three parts

A custom drone

Built on an S500 quadcopter frame, carrying an NVIDIA Jetson Nano as its onboard computer.

A lightweight detection model

YOLO11n, recognizing five tree conditions, optimized to run on low-power edge hardware.

TANDAN-V1

Monitoring software that tracks each tree, counts unique palms, computes a plantation health score and exports survey reports.

Two operating modes

Mode	How it works	Status
Onboard edge mode	Camera on the drone → Jetson Nano runs the optimized TensorRT engine on the aircraft	Engine validated on the Jetson Nano; in-flight integration in progress
Ground-station mode (TANDAN-V1)	Drone video streamed over RTMP → laptop runs detection, tracking and the live dashboard	Operational; used in field testing

04/14The drone

An open, custom-built airframe

TANDAN uses a custom-built UAV rather than a closed commercial drone, so that the flight controller, onboard computer and camera can be integrated openly for research.

**FIG. 4.1**The custom S500-based TANDAN drone. Select a numbered marker to label a component.

The build uses an open ArduPilot flight stack and off-the-shelf hobby-grade components. A closed agricultural flight controller (DJI N3-AG) was evaluated and not adopted, because its ecosystem does not expose the open interfaces that onboard AI integration needs.

Specifications
Frame	S500 quadcopter frame, 480–500 mm class
Flight controller	APM 2.6 (ArduPilot Mega)
Firmware	ArduPilot, configured with Mission Planner
Motors	2212 920 KV brushless (×4)
ESCs	20 A BLHeli-S (×4)
Propellers	1045 (10 × 4.5 in)
Battery	LiPo 3S 2200 mAh
Radio transmitter	FrSky Taranis X9D Plus
RF link	RadioMaster Ranger Micro 2.4 GHz ExpressLRS (ELRS) external TX module with receiver
Onboard computer	NVIDIA Jetson Nano 4 GB (JetPack 4.6.6, CUDA 10.2, TensorRT 8.2.1)
Camera	Arducam IMX219 CSI camera (8 MP; sensor modes up to 3264 × 2464 @ 21 fps and 1280 × 720 @ 120 fps)

Measured camera capture through the Jetson's hardware image pipeline: about 28.8 FPS at 1280 × 720 — well above the 15 FPS real-time threshold, so capture is not the bottleneck.

Jetson Nano and IMX219 camera mounted on the drone, with the camera feed on a bench monitor — **FIG. 4.2**Jetson Nano and IMX219 camera mounted on the airframe during bench testing of the camera pipeline.

TANDAN drone, front view on a plantation road — **FIG. 4.3**TANDAN drone, front view, on a plantation road.

FIG. 4.4TANDAN drone, top view while hovering.

05/14The detection model

Five conditions, one very uneven dataset

The model was trained on MOPAD (Multi-class Oil PAlm Detection dataset; Zheng et al., 2021), Site 2 subset.

1,803training images

500validation images

1024 × 1024pixels per image

151,583annotated trees

Five condition classes
Class (EN)	Class (ID)	Meaning	Training instances
healthy	normal	Healthy mature palm	89,958
smallish	kecil	Small or stunted growth	26,644
yellowish	menguning	Yellowing crown, possible nutrient or disease stress	1,473
mismanaged	tidak terawat	Poorly maintained, overgrown surroundings	431
dead	mati	Dead palm	231

The data is heavily imbalanced. The mismanaged class is both rare and hard, because it is recognized mainly from fine texture around the crown.

389 : 1healthy palms for every dead palm in training

Full class distribution
Set	dead	healthy	mismanaged	smallish	yellowish	Total
Training	231	89,958	431	26,644	1,473	118,737
Validation	61	24,884	95	7,395	411	32,846
Total	292	114,842	526	34,039	1,884	151,583

06/14The research journey

Nine ideas, then one change that mattered

We tested nine ways to make the model smarter. None was adopted. Then we changed one thing — the image resolution.

Step 1

Choosing a base model

Four nano-sized YOLO models were compared in the same screening run.

Model	mAP@0.5:0.95	Params (M)	GFLOPs	FPS on Jetson	Real-time?
YOLOv8n	0.9505 ± 0.0070	3.006	8.1	21.8	Yes
YOLO11n	0.9578 ± 0.0023	2.583	6.32	20.2	Yes — selected
YOLO12n	0.9696 ± 0.0026	2.558	6.32	9.8	No
YOLO26n	0.9394 ± 0.0074	—	—	export failed	No

YOLO12n was the most accurate but ran at only 9.8 FPS on the Jetson Nano, because its attention mechanism relies on hardware (FlashAttention / Tensor Cores) that the Nano's Maxwell GPU lacks. YOLO26n could not be exported to TensorRT. YOLO11n offered the best balance of accuracy, size and stability.

In the main experiments that followed, the YOLO11n baseline at 640 px measured 0.9619 ± 0.0023 — this is the reference used for every comparison below.

Step 2

Nine modifications, none adopted

Each was tested in isolation over the same baseline and judged against a statistical noise floor: the standard deviation of the baseline across three training seeds (42, 123, 456), ≈ 0.0023.

Group A — tested on YOLO12n, the accuracy leader
Modification	mAP@0.5:0.95	Change	Cost	Verdict
YOLO12n baseline	0.9696 ± 0.0026	—	2.558 M params, 6.32 GFLOPs	reference
SPD-Conv (P3–P5)	0.9692 ± 0.0048	−0.0004	3.995 M params, 9.86 GFLOPs (+56%)	neutral, expensive
Varifocal Loss	0.9756 (seed 42)	+0.0060	unchanged	not adopted — background false positives rose 2.2× (364 → 809)
A2C2fRep (attention removed)	0.9565 ± 0.0031	−0.0131	2.927 M params, 6.79 GFLOPs	declined

SPD-Conv is a small-object technique, but oil palm crowns are medium-sized objects (about 62 px), so it added 56% compute for no gain.

Group B — tested on YOLO11n (baseline 0.9619 ± 0.0023; seed-42 reference 0.9593 for single-seed rows)
Modification	Part changed	mAP@0.5:0.95	Verdict
Class-balanced focal loss (Cui et al., 2019)	Loss	0.9597 ± 0.0021	neutral
SCS-Head (shared-convolution head)	Head	0.9156	failed
SimAM parameter-free attention (Yang et al., 2021)	Backbone	0.9501	failed
RepC3k2 re-parameterization (after Ding et al., 2021)	Backbone	0.9509	declined
BiFPN weighted feature fusion (Tan et al., 2020)	Neck	0.9586	neutral
Network Slimming structured pruning (Liu et al., 2017)	Whole network	channels did not polarize — halted	failed

Pruning failure evidence
BN scale statistic	Before	After	Needed
Mean absolute value	1.1045	1.0644	≪ 1.0
Channels < 0.01	0.013%	0.013%	> 20%
Channels < 0.10	0.026%	0.026%	> 30%

The class-balanced loss shifted where false positives went (background → healthy dropped from 66.3% to 52.6%) but did not remove them — they moved to smallish (31.0% → 42.1%).

Step 3

What worked: resolution

Input resolution	mAP@0.5:0.95	GFLOPs	AP mismanaged	Seeds
640 px	0.9619 ± 0.0023	6.32	0.894	3
704 px	0.9648	7.64	0.908	1
768 px	0.9694 ± 0.0026	9.1	0.926	3

Raising resolution from 640 to 768 px improved mAP by 0.0075 — more than three times the noise floor — and brought YOLO11n level with YOLO12n (0.9696) while remaining deployable. The hardest class, mismanaged, gained over three points.

Try it — input resolution

640704768

mAP@0.5:0.95—

AP mismanaged—

GFLOPs—

Jetson FPS (TensorRT FP16)—

Step 4

Why it worked

Per-class accuracy across eight configurations. Spread of each class across configurations: healthy 0.021, dead 0.043, yellowish 0.048, smallish 0.050, mismanaged 0.125 — two to six times wider than any other class. Ranking configurations by mismanaged accuracy almost reproduces the ranking by overall mAP; the other four classes were already saturated.

Hover or focus a point to see its configuration. Gold = resolution 768.

AP per class
Configuration	dead	healthy	mismanaged	smallish	yellowish

Key insight

The bottleneck was pixels, not model capacity. Fine crown texture disappears when images are downsampled, and no module can recover information that is not in the input. The nine rejected modifications are not wasted effort — they are the evidence that ruled every other explanation out.

Error diagnosis. Most errors were not confusion between classes but background mistaken for trees — about 66% of background false positives were labeled healthy and 31% smallish.

Overall accuracy tracks the mismanaged class — **FIG. 6.1**Overall mAP against AP of the *mismanaged* class across configurations.

Accuracy versus compute trade-off — **FIG. 6.2**Accuracy versus compute on the Jetson Nano, marked by real-time feasibility.

07/14Edge deployment on Jetson Nano

Nine engines, one that clears the bar

PyTorch.pt
ONNXopset 12
TensorRT enginebuilt with trtexec
NVIDIA Jetson Nano 4 GBMaxwell GPU, compute capability 5.3

Throughput (queries per second) · three resolutions × three precisions

11.220.2 qps

INT8 gives no speedup on this GPU — within 0.3% of FP32 at every resolution. The Maxwell GPU lacks the DP4A instruction that accelerates 8-bit math (available only from compute capability 6.1). Measured, not assumed.
FP16 gives a consistent ~30% speedup at every resolution.
768 px fails the 15 FPS real-time threshold (14.52) despite being most accurate; 704 px passes with a 13% margin (17.02). 704 px FP16 is the deployed configuration.

Engine throughput against the real-time threshold — **FIG. 7.1**Engine throughput against the 15 FPS real-time threshold.

FP16 accuracy check

Measured on the Jetson, using PyTorch half precision as a numerically equivalent proxy because TensorRT's Python bindings are unavailable for Python 3.8 on JetPack 4.6.

Class	FP32	FP16	Δ
dead	0.979	0.981	+0.002
healthy	0.993	0.993	0.000
mismanaged	0.931	0.927	−0.004
smallish	0.977	0.971	−0.006
yellowish	0.988	0.986	−0.002
all	0.974	0.971	−0.003

Thermal stress test

313 s of continuous inference.

Metric	Idle	Under load
GPU temperature	23.1 °C	21.5 → 30.5 °C, plateau, max 35.0 °C
CPU temperature	21.4 °C	21.5 → 31.5 °C, max 36.5 °C
GPU utilization	~0%	stable 89–98%
Throughput	—	17.0025 qps sustained vs 17.02 instantaneous (Δ 0.1%)

No thermal throttling: temperatures plateaued more than 60 °C below the ~97 °C throttling point, and sustained throughput matched the instantaneous value.

Final configuration

Model: YOLO11n
Input: 704 px
Engine: TensorRT FP16
Throughput: 17 FPS sustained
Parameters: 2,583,127
Output: 1 × 9 × 10,164

08/14TANDAN-V1 monitoring software

From detections to a survey a manager can use

TANDAN-V1 turns raw detections into survey information a plantation manager can use.

Drone flight
Videolive RTMP stream via mediamtx, or recorded file
YOLO detection
ByteTrack multi-object tracking
Per-tree classificationby majority vote
Live dashboard
Automatic session report

VID. 8.1Demo of TANDAN-V1 tracking palms: annotated video on the left, analytics panel on the right. Figures visible in the recording belong to a demo session and are not reported results.

Split-screen dashboard: annotated video (about 65% of the screen) and analytics panel (about 35%).
Box labels show track ID · class · confidence.
Unique palms counting: each tree keeps a stable track ID; when it leaves view for 15 frames it is retired and its final class is decided by majority vote across all its observations. The dashboard shows both detections (every frame) and unique palms (each tree once) — unique counts are what matter for inventory.
Plantation health score (0–100) with a traffic-light gauge: green ≥ 75, yellow 50–75, red < 50.
Alerts: a flashing red border appears when a dead or mismanaged palm is in view.
Live FPS and detection sparklines, confidence histogram, per-class counts.
Exports per session: annotated MP4, CSV (frame, timestamp, elapsed seconds, track ID, class, confidence, box coordinates, inference time, FPS) and a JSON summary.
Post-flight analytics page: loads a session CSV and shows class distribution, timeline, confidence and FPS charts, and a table of at-risk palms.

TANDAN-V1 live monitoring dashboard shown on a tablet in the field — **FIG. 8.2**TANDAN-V1 live monitoring dashboard on a tablet.

Post-flight analytics of a survey session — **FIG. 8.3**Post-flight analytics of a demo session.

Health score

good  = healthy + 0.7 × smallish
bad   = dead + 0.8 × mismanaged + 0.4 × yellowish
score = good / (good + bad) × 100
        (computed on unique palms)

Adjust the five class counts to see how the score and gauge colour respond.

healthy

smallish

yellowish

mismanaged

dead

—

09/14Field testing

In the plantation

TANDAN was tested at an oil palm plantation in Kabupaten Tanjung Jabung Barat, Jambi, Indonesia. Aerial imagery was processed into orthomosaics with Agisoft Metashape, and the detector — trained on data from a different region — was applied to the new site's imagery.

Research team during field testing, with the TANDAN drone and ground-station laptop — Research team during field testing.

Mission planning on a handheld controller at the plantation — Planning an aerial survey at the test site.

Drone take-off at the test site.

10/14Results at a glance

What the study established

0.9694

mAP@0.5:0.95 at 768 px — level with a newer, heavier model

0.9648

mAP at the deployed 704 px

+0.0075

mAP from resolution alone (3× the noise floor)

modifications tested, none adopted

17 FPS

sustained on a Jetson Nano, 13% above the real-time threshold

~30%

speedup from FP16; none from INT8

−0.003

mAP cost of FP16

60 °C+

thermal headroom under continuous load

2.58 M

parameters

11/14Network architecture

Stock YOLO11n, deliberately

The deployed network is stock YOLO11n. The contribution is the resolution choice and the deployment pipeline, not a new architecture.

BackboneLayers 0–10: Conv → Conv → C3k2 → Conv → C3k2 → Conv → C3k2 → Conv → C3k2 → SPPF → C2PSA; channel widths 16 → 256.
NeckLayers 11–22: PAN-FPN with upsampling and concatenation skip connections from layers 4, 6 and 10.
HeadLayer 23: detection at three scales — P3/8 (88 × 88 grid at 704 px), P4/16 (44 × 44), P5/32 (22 × 22) — decoded to 10,164 candidate boxes (4 box coordinates + 5 class scores).
Size2,583,127 parameters · 101 fused layers · 6.3 GFLOPs at 640 px

**FIG. 11.1**Deployed YOLO11n network architecture. Select to enlarge.

**FIG. 11.2**Methodology from dataset to deployment-ready engine. Select to enlarge.

12/14Status

Where the project stands

Dataset preparation and model trainingDone
Nine-modification study and resolution studyDone
Jetson Nano engine benchmark, FP16 validation, thermal testDone
TANDAN-V1 monitoring softwareDone
Custom S500 drone buildDone
Field testingDone
Live onboard camera pipeline on the JetsonIn progress
In-flight onboard detectionPlanned

13/14Team

People

Researcher: Muhammad Alfhi Saputra, S.Kom., M.Kom.
Promotor: Prof. Dr. Ir. Widodo Budiharto, S.Si., M.Kom., IPM., SMIEEE
Co-promotors: Dr. Ir. Haryono Soeparno, M.Sc.
Dr. Ir. Yulyani Arifin, S.Kom., M.M., IPP
Institution: BINUS University
Dataset credit: MOPAD dataset by Zheng et al. (2021)

Dissertation

Development of an Efficient Real-Time UAV-Based Detection System for Oil Palm Condition Monitoring

Pengembangan Sistem Deteksi Berbasis Real-Time-UAV yang Efisien untuk Monitoring Kondisi Kelapa Sawit

Doctoral dissertation, Doctor of Computer Science.

References

Sources

Zheng, J. et al. (2021). Growing status observation for oil palm trees using UAV images. ISPRS J. Photogramm. Remote Sens. 173, 95–121. (MOPAD dataset)
Purnomo, H. et al. (2020). Reconciling oil palm economic development and environmental conservation in Indonesia. For. Policy Econ. 111, 102089.
Qaim, M. et al. (2020). Environmental, economic, and social consequences of the oil palm boom. Annu. Rev. Resour. Econ. 12, 321–344.
Xin, Y., Sun, L., Hansen, M.C. (2022). Oil palm reconciliation in Indonesia. J. Clean. Prod. 380, 135087.
Cui, Y. et al. (2019). Class-balanced loss based on effective number of samples. CVPR.
Yang, L. et al. (2021). SimAM: a simple, parameter-free attention module. ICML.
Tan, M., Pang, R., Le, Q.V. (2020). EfficientDet: scalable and efficient object detection. CVPR.
Ding, X. et al. (2021). RepVGG: making VGG-style ConvNets great again. CVPR.
Liu, Z. et al. (2017). Learning efficient convolutional networks through network slimming. ICCV.
Jocher, G., Qiu, J. (2024). Ultralytics YOLO11. github.com/ultralytics/ultralytics