Journal of Deep Learning, Computer Vision and Digital Image Processing

Development of Android-Based Smart Learning Media for the Operating Systems Course Using the ADDIE Model

Kurnia Wahyu Prima — 2026-06-09

Purpose – The rapid advancement of digital technology has encouraged higher education institutions to integrate innovative learning media to enhance the quality of the teaching and learning process. However, learning activities in Operating Systems courses are still predominantly conducted using conventional methods, causing students to experience difficulties in understanding abstract concepts such as process management, memory management, CPU scheduling, and file systems. This study aims to develop Android-Based Smart Learning media for the Operating Systems course and to determine the feasibility level of the developed media as an interactive learning tool.
Method – This study employed a Research and Development (R&D) approach using the ADDIE model, which consists of five stages: Analysis, Design, Development, Implementation, and Evaluation. The developed product was validated by subject-matter experts and media experts before being implemented with students of the Informatics and Computer Engineering Education Program who were enrolled in the Operating Systems course. Data were collected through validation sheets and student response questionnaires using a five-point Likert scale and were analyzed using descriptive quantitative methods.
Results – The findings indicate that the Android-Based Smart Learning media was successfully developed by integrating learning materials, instructional videos, interactive quizzes, and automated feedback features into a single Android application. The material expert validation yielded a score of 89.00%, while the media expert validation achieved a score of 90.00%, both categorized as highly feasible. Furthermore, student responses obtained an average percentage of 90.27%, classified as very good. Therefore, the developed media was considered suitable for supporting the learning process in the Operating Systems course.
Research Implications – This study was limited to a single study program and Android devices; therefore, the generalizability of the findings remains limited.
Originality – This research integrates the concepts of mobile learning and smart learning into a single interactive learning medium specifically designed to support Operating Systems education in higher education institutions.

An Intelligent IoT-Based Waste Bin System Utilizing Nearest Neighbor Algorithms for Optimized Waste Collection Routes

Dinan Yulianto — 2026-06-18

Purpose – Despite advances in IoT-enabled waste monitoring, existing solutions generally fail to integrate real-time bin status information with adaptive route optimization, resulting in inefficient collection operations. This study aims to design and implement an integrated system that leverages real-time waste data to facilitate intelligent, data-driven route optimization for improved waste collection operations.
Methods –This study presents an ESP32-based smart waste system using reed switch event-driven control and deep-sleep mode for energy efficiency. Waste levels were estimated using the arithmetic mean fusion of four VL53L0X sensors. A cloud-based MQTT-over-TLS architecture enables secure real-time communication, whereas a priority-based nearest-neighbor routing algorithm is evaluated across 150 nodes.
Findings – The results demonstrate that the proposed system provides accurate waste-level estimation with a mean error of 1.98%, significantly reduces energy consumption by 90.9% through deep-sleep operation, and supports near-real-time communication with an average latency of 4.66 s. Moreover, the priority-based route optimization strategy decreased the travel distance by 42.7%, ensured the immediate servicing of all full-status bins, and maintained operational feasibility within a fleet capacity of 2,700 L.
Research implications – The evaluation results demonstrate the feasibility of integrating real-time monitoring and adaptive route optimization for smart waste management. Future research should extend the validation to large-scale real-world deployments and incorporate road network-based routing models to enhance operational realism and optimization accuracy.
Originality – This study proposes an integrated smart waste platform that combines energy-efficient event-driven sensing, dynamic priority-based nearest-neighbor routing, and hardware-assisted digital twin validation for scalable and cost-effective waste management evaluation.

RT-DETR-Based Computer Vision System for Real-Time Detection and Classification of Oil Palm Fruit Maturity Levels in Plantations

Nur Hafiqah Rambe — 2026-06-18

Purpose – This study aimed to develop an automated oil palm fruit maturity level detection system using the real-time detection transformer (RT-DETR) algorithm to overcome the limitations of conventional visual inspection methods, which are often subjective and inconsistent. This study evaluated the effectiveness of the RT-DETR in detecting and classifying oil palm fruit maturity levels to support quality control processes in plantation operations.
Method – A computer vision-based approach was implemented using the RT-DETR-L object detection model. The dataset consisted of 14,620 annotated oil palm fruit images categorized into four maturity levels: unripe, underripe, ripe, and overripe. The research process included data collection, image annotation, preprocessing, model training, and evaluation of the model. The model performance was assessed using precision, recall, mean Average Precision (mAP@50), and inference speed metrics.
Findings – The experimental results show that the RT-DETR-L model achieved a precision of 93.2%, 95.6%, and mAP@50 of 96.9%, respectively. The model successfully detected and classified oil palm fruit maturity levels across all categories with high accuracy. Furthermore, the model achieved an inference time of 25–28 ms per image and a processing speed of 10–14 FPS on an NVIDIA RTX 3050 4GB GPU, demonstrating its capability for real-time applications.
Research Implications – The findings indicate that RT-DETR-L can improve the efficiency, consistency, and accuracy of oil palm fruit sorting and quality control processes. However, this study was limited to the available datasets and testing scenarios used. Future research should evaluate the model under diverse environmental conditions, lighting variations, and field deployment settings to improve its generalizability and robustness.
Originality – Unlike previous studies that primarily employed CNN-based detectors or focused on binary maturity classification, this study investigated the application of a transformer-based RT-DETR-L architecture for detecting four oil palm fruit maturity categories. The results demonstrate that RT-DETR-L can provide high detection accuracy while maintaining real-time performance in smart agriculture applications.

Comparative Analysis of IndoBERT and BiLSTM For Public Sentiment Classification Toward The Indonesian National Police on Youtube

Hardeva Satria Hazz — 2026-06-18

Purpose – This study aimed to compare the performance of IndoBERT and Bidirectional Long Short-Term Memory (BiLSTM) in classifying public sentiment toward the Indonesian National Police (INP) based on YouTube comments. This study sought to identify a robust sentiment classification model to support text-based public perception monitoring, particularly under a highly imbalanced sentiment distribution.
Method – YouTube comments were collected using the YouTube Data API. A total of 8,268 raw comments were obtained, and 7,197 comments were retained as the final dataset after preprocessing, automatic pseudo-labeling, and confidence filtering using a 0.5 threshold. To address concerns regarding threshold selection, an additional sensitivity analysis was conducted using confidence thresholds of 0.65 and 0.75. The experiment applied a dual-track preprocessing pipeline, cost-sensitive learning through class-weighted loss, bootstrap confidence interval analysis, and BiLSTM preprocessing ablation.
Findings – The results show that IndoBERT achieved stronger performance than BiLSTM. IndoBERT obtained an accuracy of 92.92% and a Macro-F1 Score of 0.8548, whereas BiLSTM achieved an accuracy of 76.11% and a Macro-F1 Score of 0.6124. Bootstrap analysis showed a Macro-F1 difference of 0.2424, with a 95% confidence interval of 0.1870 to 0.2959, indicating that IndoBERT’s advantage was statistically significant. Sensitivity analysis also confirmed that IndoBERT consistently outperformed BiLSTM across all the tested thresholds.
Research Implications – The findings indicate that IndoBERT is more suitable for Indonesian sentiment classification in public perception monitoring than other models. However, because the dataset labels were generated using a BERT-based classifier, the evaluation may contain architectural circularity that favors the IndoBERT model. Future studies should use human-annotated gold-standard data and broader cross-platform validations.
Originality – This study provides a comparative evaluation of transformer-based and recurrent models using sensitivity analysis, bootstrap testing, cost-sensitive learning, and pre-processing ablation under imbalanced sentiment conditions.

Design and Implementation of an AI Agent-Based Workflow Automation System for Scheduling and Information Dissemination in Oil Palm Plantations

Septianur Eka Amri — 2026-06-25

Purpose – Oil palm plantation operations in Indonesia require coordination across multiple divisions; however, meeting scheduling and information dissemination are often managed through separate, manual processes. This may cause communication delays, scheduling conflicts, and inconsistent information deliveries. This study aims to design and implement an AI-based workflow automation system that integrates meeting scheduling and information dissemination into a centralized platform to support the automated coordination and information management across organizational units.
Methods – This study employed the Design Science Research (DSR) approach, covering problem identification, literature review, system design, implementation, testing and evaluation. The proposed system integrates Gemini AI, Natural Language Processing (NLP), Telegram Bot, Zoom API, Google Calendar API, and Google Sheets to automate meeting scheduling, information dissemination, and document management.
Findings – The implemented system successfully automated meeting scheduling, calendar synchronization, information dissemination, and documentation management within an integrated platform. Functional testing confirmed that the core features operated as intended in the scenarios evaluated. The system also supports information classification based on public and private access.
Research implications – The system was evaluated through functional testing in a simulated oil palm plantation context and depends on third-party API services, which may limit its generalizability. User acceptance, organizational effectiveness, and efficiency were not evaluated. Nevertheless, the proposed architecture can be adapted to other organizational settings that require automated coordination and centralized information management.
Originality – This study proposes an AI-based workflow automation architecture that integrates communication and productivity services to support the end-to-end automation of meeting scheduling and information dissemination in oil palm plantation operations.

Comparative Analysis of Facial Feature Extraction in RGB and Near-Infrared Images Using YOLOv11 for Edge-Deployed Driver Monitoring

Ahmadil Barokah — 2026-06-25

Purpose – This study proposes a robust edge-computed Driver Monitoring System (DMS) using the YOLOv11 architecture to detect driver fatigue across daytime RGB and nighttime near-infrared (NIR) environments.
Methods – A lightweight YOLOv11 model was trained on an augmented multi-spectral dataset of 2,289 images containing two critical fatigue markers: drowsy_eye and open_mouth. For real-time deployment on a resource-constrained Raspberry Pi 4B, the model was compiled into an optimized ONNX format with a 240 × 320 pixel input matrix. A Temporal State Machine using strict logical conjunction (AND logic) was integrated to process sequential frame updates and reduce false-positive alerts caused by micro-blinking.
Findings – Under live multi-spectral stationary cabin hardware evaluation, the integrated prototype achieved real-time inference of 22.9–72.4 FPS in daytime RGB conditions and 20.7–28.7 FPS in nighttime NIR conditions. In total darkness, NIR feature extraction remained stable, with empirical confidence ranges of 0.70–0.82 for drowsy_eye and 0.93–0.94 for open_mouth. The state machine successfully confirmed microsleep events lasting more than two seconds and triggered synchronized voice alerts with a randomized LED array as a chaotic counter-fatigue sensory stimulus.
Research implications – The system demonstrates the feasibility of deploying advanced AI-based DMS models on low-power, standalone, cloudless edge hardware for automotive safety applications.
Originality – This study presents a multi-illumination RGB–NIR comparative evaluation of an ONNX-optimized YOLOv11 model integrated with an active randomized LED counter-fatigue intervention loop.

Artificial Intelligence in English Language Teaching (ELT) Classrooms: Using Hyperwrite as a Writing Tool to Enhance Writing Process

Alfi Hidayatu Miqawati — 2026-07-08

Purpose – AI-powered writing tools offer innovative breakthroughs by automating aspects of the writing process and providing helpful suggestions and feedback, thereby potentially improving writing outcomes and efficiency. This paper discusses the integration of HyperWrite in supporting and enhancing the writing process in ELT classrooms.
Methods – A qualitative research method using library study approach was deployed to formulate ideas about the implementation of HyperWrite in writing classes. The sources of data vary from journals, research reports, official documents, and curriculum. The data were analyzed through several steps that involved data reduction, data display, and conclusion drawing.
Findings – Several studies highlighted that HyperWrite has established itself as a potent tool among AI writing assistants, featuring a multifaceted platform that helps with both the efficiency and the quality of writing. It uses advanced machine learning algorithms to guide users in all aspects of writing, which include developing ideas and editing final drafts. This study also highlights the implementation of HyperWrite to guide students to write effectively.
Research implications – The use HyperWrite in ELT writing classes suggests a shift in the teacher’s role toward that of a facilitator who guides students in critically evaluating AI suggestions, while also fostering the development of critical thinking and self-confidence among students. Furthermore, these findings call for the integration of AI literacy into the ELT curriculum to develop students’ competence to use, evaluate, and edit AI-generated work ethically. This study was limited to only one AI tools and further studies about the uses of other tools for ELT classrooms are recommended.
Originality – This study offers a novel approach by conducting an integrated examination of the role of teachers, students’ critical thinking, and HyperWrite-assisted ELT writing classroom.

Web-Based Regional Map Repository System for Administrative Boundaries Survey Management

Muhammad Ramdani — 2026-07-08

Purpose – This study aims to design and implement a web-based regional map repository system to improve the management of administrative boundary survey archives at the Central Statistics Agency (BPS) of Bogor City. The study addresses the limitations of conventional physical archiving, including poor document traceability, potential data loss, damaged map records, and inefficient monitoring of map borrowing activities.
Methods – The system was developed using the Rapid Application Development (RAD) method through requirements planning, user design, construction, and cutover stages. Data were collected through semi-structured interviews, direct observation, and literature review. The system was built using Laravel 12, PHP, MySQL, and deployed through a local server environment. System design was modeled using UML, while evaluation was conducted using black-box testing and the System Usability Scale (SUS) involving 10 respondents.
Findings – The system integrates WA and WS map repositories, document borrowing history, role-based access, regional data management, and SLS merging and splitting features. Black-box testing showed a 100% success rate across all tested functional scenarios. The SUS evaluation produced an average score of 79 out of 100, categorized as “Good” and “Acceptable.”
Research implications – The findings are limited to one local BPS institution and a small usability sample. Further testing, including load and stress testing, is required to assess scalability and resilience.
Originality – This study offers an integrated repository model combining map archiving, administrative boundary management, and transaction-based tracking for local e-government archive transformation.

Attention Span Classification of Social Media Users Using Multi-Kernel Support Vector Machine Based on Survey Data

Reza Pahlevi — 2026-07-08

Purpose – This study aims to examine whether self-reported attention-related difficulty categories among social media users can be classified using a leakage-free machine learning framework. It addresses the risk of inflated performance in survey-based classification by excluding the same items used to construct the target label from the predictor set.
Methods – The study used the public Social Media and Mental Health (SMMH) Kaggle dataset with 478 valid respondents. A three-class label was constructed from Q10, Q12, and Q14 using percentile thresholds (P33 = 9.0; P66 = 12.0), producing High (n = 208), Medium (n = 152), and Low (n = 118) categories. These label-generating items were excluded from predictors. The remaining variables were processed in a scikit-learn Pipeline using MinMax scaling, ordinal encoding, and One-Hot Encoding. Multi-kernel SVM models and five baseline classifiers were evaluated using a stratified 70:30 split, cross-validation, F1 metrics, balanced accuracy, and permutation importance.
Findings – Random Forest achieved the highest performance, with 63.19% accuracy and 62.26% weighted F1. Linear SVM was the best SVM model, achieving 61.81% accuracy, 60.08% weighted F1, 58.99% macro F1, and 59.11% balanced accuracy. The strongest predictors were Restless Without Social Media, Use Without Purpose, and Interest Fluctuation.
Research implications – The findings are preliminary, dataset-specific, and based on a survey-derived composite label whose internal reliability still requires validation.
Originality – This study contributes a leakage-controlled classification approach for analyzing attention-related survey categories.

A Web-Based Sentiment Analysis System for IMDb Movie Reviews Using TF-IDF and Multinomial Naïve Bayes

Syahrur Ramadhan — 2026-07-08

Purpose – The rapid growth of online movie platforms has produced large volumes of user reviews that contain valuable audience opinions. However, manual review analysis is inefficient. This study aims to develop a web-based sentiment analysis application using TF-IDF feature representation and the Multinomial Naïve Bayes algorithm to classify movie reviews into positive and negative sentiments.
Methods – The model was trained and evaluated using the IMDb 50K Movie Reviews dataset with an 80:20 train–test split. An additional 600 reviews from six different movies were used to demonstrate application-level implementation. Text preprocessing included cleaning, lowercase normalization, tokenization, stopword filtering, and lemmatization using Natural Language Processing techniques. The processed texts were transformed into TF-IDF vectors and classified using Multinomial Naïve Bayes with the default smoothing parameter (α = 1.0). The trained model was deployed in a Flask-based web application for interactive sentiment prediction.
Findings – The model achieved an accuracy of 84.93%, with precision, recall, and F1-score showing relatively balanced performance across positive and negative classes. The web application successfully classified movie reviews and displayed sentiment distributions through an interactive interface.
Research implications – The findings indicate that lightweight machine learning methods can support practical web-based sentiment analysis with low computational demands. However, performance may decline when processing sarcasm, irony, or implicit contextual meaning.
Originality – This study combines benchmark evaluation with web-based validation using 600 additional real-world movie reviews, demonstrating practical applicability beyond dataset-level testing.

Design and Implementation of a Dual-LLM Prescriptive ESG Reporting System in the Indonesian Palm Oil Industry

Niko Firzi Anansyah — 2026-07-08

Purpose – This study aims to develop an automated Environmental, Social, and Governance (ESG) reporting information system based on Large Language Model (LLM) for the palm oil industry to overcome low efficiency, data inconsistencies, and analysis limitations inherent in manual reporting processes.
Methods – This applied research employs the Design Science Research (DSR) paradigm, encompassing needs analysis, system design, implementation, and black-box testing. The system was developed using the Laravel MVC framework and integrated a Dual-LLM API failover architecture (Groq Llama 3 as primary and Gemini as backup). The case study was conducted at PT Surya Mata Ie.
Findings – The developed system successfully automated ESG indicator extraction and prescriptive narrative generation. It utilizes a Strict Weighting Rule, a programmatic safeguard capping the ESG score at 50.0 (as a proof-of-concept testing constraint) if Ganoderma infection exceeds a 20% threshold (supported by agronomic research). During prototype evaluation, this rule intercepted an overly optimistic raw LLM score of 60.0 and corrected it to 50.0. This demonstrates the system's capability to function as a risk-control mechanism, mitigating potential hallucination-driven score inflation and supporting mathematically accountable outputs.
Research implications – The implementation of this system significantly accelerates reporting workflows and serves as an early warning instrument for environmental risks, thereby enhancing real-time managerial decision-making and corporate transparency in complying with global sustainability standards.
Originality – This study pioneers the integration of a Dual-LLM failover mechanism within a Laravel framework tailored for the palm oil sector. It introduces a novel programmatic constraint approach in JSON object parsing to maintain strict mathematical accountability in AI-generated ESG drafts.

Laravel Dashboard for Immature Oil Palm (TBM III) Monitoring Using XYZ Tiles and Large Language Models

Maghfirah — 2026-07-08

Purpose – This research addresses the strategic urgency of digitalizing plantation monitoring to achieve precision agriculture at PTPN IV Regional I. The monitoring of Immature Plants (TBM) III currently relies on fragmented manual spreadsheets, leading to data redundancy and delayed analysis.
Methods – A web-based data visualization dashboard was developed using the Laravel framework, integrating Geographic Information Systems (GIS) with Static Raster Tiling (XYZ Tiles) to optimize high-resolution map rendering. The system incorporates Large Language Model (LLM) API integration (Gemini 1.5 Flash and Llama 3) for prescriptive analytics, transforming biometric growth data into automated maintenance recommendations through prompt engineering.
Findings – Results indicate that the system achieves significant workflow simplification by transforming the fragmented, multi-stage manual reporting pipeline into an automated, single-step data ingestion process, successfully reducing administrative touchpoints. The Static Raster Tiling (XYZ Tiles) technique successfully reduced high-resolution orthophoto rendering latency from over 12,000 ms to an average of 180 ms. Validation using Fleiss' Kappa statistics yielded a score of 0.8105, categorized as "Almost Perfect Agreement," confirming that the AI-generated recommendations are highly consistent with expert agronomic standards.
Research implications – This system provides a comprehensive managerial evaluation tool, bridging the gap between raw field data and strategic decision-making in oil palm management.
Originality – The integration of spatial optimization and prescriptive AI analytics offers a novel approach compared to existing descriptive-only monitoring platforms.

Development of an Integrated Web-Based Palm Oil Production Monitoring Dashboard Using Laravel at PT Paluta Inti Sawit

Aditya Pratama — 2026-07-08

Purpose – Data management at PT Paluta Inti Sawit (PT PIS) currently faces efficiency constraints due to fragmented operational data for Fresh Fruit Bunches (FFB), Crude Palm Oil (CPO), and Kernel (PK) stored in scattered spreadsheets. This study aims to design and build an integrated web-based dashboard to centralize monitoring and automate production reporting.
Methods – The research employs an applied system development and functional validation design following the Waterfall System Development Life Cycle (SDLC). The system was developed using the Laravel framework and MySQL database, with ApexCharts for near-real-time interactive visualization after data ingestion.
Findings – The developed dashboard successfully integrates multi-departmental production data into a single source of truth. Key features include an automated Excel/CSV parser with conflict resolution (skip/overwrite), a two-level role-based verification system (Staff and Manager), and near-real-time KPI tracking for Oil Extraction Rate (OER) and Kernel Extraction Rate (KER) upon data approval.
Research implications – This system accelerates the daily reporting cycle and supports data integrity through a digital audit trail, assisting rapid, data-driven managerial decisions.
Originality – Unlike previous studies that focus on mobile-only reporting, this system provides a mill-level integrated prototype for synchronized FFB reception and processing yield analytics, minimizing manual recapitulation errors.

Automatic Detection of Toxic Content in Short Videos Using Deep Learning-Based Text and Audio Feature Integration

Heri Agus Supriyanto — 2026-07-08

Purpose – This study develops and evaluates a multimodal classification model combining audio and text features via configurable weighted late fusion to detect toxic content in keyword-retrieved Indonesian short videos containing profanity-related contexts, addressing text-only detection limitations.
Methods – The pipeline utilized FFmpeg for audio extraction, MFCC for audio features, and Google Speech Recognition for text. Three fusion configurations (Audio-Text: 40%–60%, 50%–50%, and 60%–40%) combining a DNN and BiLSTM were evaluated on 1,484 manually labeled Indonesian short videos.
Findings – The Audio 60%–Text 40% configuration achieved the numerically highest test accuracy of 93.94% with a 95% confidence interval of 90.62%–96.13%, using a decision threshold of 0.60 selected from the validation set. The model obtained an F1-score of 0.95 for the toxic class. Compared with unimodal baselines, all fusion models achieved higher accuracy, indicating the benefit of integrating audio and text features.
Research implications – The findings suggest that multimodal audio-text fusion can improve Indonesian short-form video toxicity detection compared with audio-only or text-only models. However, the differences among the three fusion weighting schemes were not statistically significant based on McNemar’s test, so the audio-dominant configuration should be interpreted as the numerically best configuration in this dataset rather than as a universally superior setting.
Originality – This study systematically compares unimodal baselines and configurable audio-text late-fusion weighting strategies for Indonesian short-form video toxicity detection. The study also applies validation-based threshold selection, confidence intervals, and pairwise McNemar testing to provide a more reliable evaluation of multimodal model performance.

YOLOv8-Based IoT System for Oil Palm Harvest Readiness Identification through Loose Fruit Detection with Real-Time Web Monitoring

Wira Jhohan Simatupang — 2026-07-08

Purpose - This study develops and evaluates an Internet of Things (IoT)-based monitoring system for automatically identifying oil palm harvest readiness through loose fruit detection using the YOLOv8 algorithm. The system addresses the subjectivity, labor intensity, and inefficiency of conventional manual observation in large-scale plantations.
Methods - An experimental design was applied using 2,000 images of oil palm loose fruits collected from the Indonesian Institute of Palm Oil Technology, North Sumatra. Images were captured using an ESP32-CAM, annotated through Roboflow, and used to train the YOLOv8x model for 120 epochs at a resolution of 640 × 640 pixels.
Findings - The prototype performed image acquisition, loose fruit detection, and dashboard-based monitoring in near real time. On the independent testing dataset, the model achieved a precision of 0.933, recall of 0.954, F1-score of 0.944, mAP@0.5 of 0.943, and mAP@0.5:0.95 of 0.495. These results demonstrate the system’s feasibility as a prototype for supporting harvest-readiness monitoring, although broader field validation is still required.
Research Implications - The dataset was obtained from a single plantation site and may not represent highly variable environmental conditions. Future studies should use larger and more diverse datasets and assess cloud-based deployment for improved scalability.
Originality - This study integrates ESP32-CAM, YOLOv8x, IoT communication, and web-based monitoring into a prototype architecture for data-driven oil palm harvest-readiness assessment through loose fruit detection.

Performance Evaluation and Comparative Analysis of Small Solar Panels for IoT Battery Charging Using ESP32 Microcontroller Platform

Saddam Husain Siregar — 2026-07-08

Purpose – Small photovoltaic panels are widely used to support low-cost Internet of Things (IoT) sensor nodes; however, many locally available panels provide only nominal voltage labels without complete electrical specifications. This study evaluates the charging performance of three small solar panel configurations for a single 18650 lithium-ion battery used in an IoT-based monitoring node.
Methods – The experimental system used Lolin D32 boards based on ESP32, TP4056 charging modules, and identical battery connections. Node 1 used a large 9 V solar panel connected to a DC-DC step-down module calibrated to 5 V before entering the TP4056 input. Node 2 used a medium 6 V panel, while Node 3 used a mini solar panel directly connected to the charging module. Battery voltage was recorded every five minutes, converted into apparent state of charge (SoC), and uploaded to Google Sheets. The main field test was conducted from 10:00 to 15:00 on 8 June 2026.
Findings – Node 1 produced the highest apparent charging performance, with a charging rate of 12.8 percentage points per hour and a total SoC increase of 63.8 percentage points. In contrast, Node 2 and Node 3 each achieved only 0.5 percentage points per hour during the same test period.
Research implications – The findings indicate that nominal voltage alone is insufficient for selecting solar panels for IoT battery charging. Panel area, current capability, charger compatibility, firmware measurement sequence, and direct field validation must be considered.
Originality – This study provides practical field-based evidence for selecting small photovoltaic panels in low-cost IoT battery charging applications.