Edge-Based Predictive Data Reduction for Smart Agriculture: A Lightweight Approach to Efficient IoT Communication

Dora Kreković 1, * , Mario Kušek 1 , Ivana Podnar Žarko 1 , Danh Le-Phuoc 2 · 2025

Read Edge-Based Predictive Data Reduction for Smart Agriculture: A Lightweight Approach to Efficient IoT Communication by Dora Kreković 1, * , Mario Kušek 1 , Ivana Podnar Žarko 1 , Danh Le-Phuoc 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-19103.txt.

Introduction

Dora Krekovic and three co-authors cut the data traffic from farm sensors, where consecutive readings often barely change yet are sent to the cloud anyway, wasting bandwidth and battery in remote fields. Their approach runs a predictive filter at the network edge, on site.

The filter forecasts the next reading and transmits only when the real value deviates beyond a set tolerance, while a cloud model guards overall integrity. Simulations show lower communication load and better energy use. Models also transfer across sites without retraining, and combining local and satellite observations strengthens them, making the method scalable for smart agriculture.

Catalog record

CREATORDora Kreković 1, * , Mario Kušek 1 , Ivana Podnar Žarko 1 , Danh Le-Phuoc 2
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-19103

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Edge-Based Predictive Data Reduction for Smart Agriculture: A Lightweight Approach to Efficient IoT Communication Thanks: This work has been supported by the Horizon Europe WIDERA program under the grant agreement No. 101079214. * Corresponding author: Dora Kreković ([email protected]) Dora Kreković 1, * , Mario Kušek 1 , Ivana Podnar Žarko 1 , Danh Le-Phuoc 2 Abstract The rapid growth of IoT devices has led to an enormous amount of sensor data that requires transmission to cloud servers for processing, resulting in excessive network congestion, increased latency and high energy consumption. This is particularly problematic in resource-constrained and remote environments where bandwidth is limited, and battery-dependent devices further emphasize the problem. Moreover, in domains such as agriculture, consecutive sensor readings often have minimal variation, making continuous data transmission inefficient and unnecessarily resource intensive. To overcome these challenges, we propose an analytical prediction algorithm designed for edge computing environments and validated through simulation. The proposed solution utilizes a predictive filter at the network edge that forecasts the next sensor data point and triggers data transmission only when the deviation from the predicted value exceeds a predefined tolerance. A complementary cloud-based model ensures data integrity and overall system consistency. This dual-model strategy effectively reduces communication overhead and demonstrates potential for improving energy efficiency by minimizing redundant transmissions. In addition to reducing communication load, our approach leverages both in situ and satellite observations from the same locations to enhance model robustness. It also supports cross-site generalization, enabling models trained in one region to be effectively deployed elsewhere without retraining. This makes our solution highly scalable, energy-aware, and well-suited for optimizing sensor data transmission in remote and bandwidth-constrained IoT environments. I Introduction The demand for sustainable and data-driven agriculture is driving the adoption of intelligent monitoring systems based on the Internet of Things (IoT) and Artificial Intelligence (AI) [ 1 ] . These technologies promise improved yield productivity, efficient resource use, and better adaptation to climate variability. However, deploying such systems in remote rural farms introduces critical challenges related to connectivity, power consumption, and data availability [ 2 ] . Agro-IoT deployments typically consist of distributed low-power sensors that measure environmental parameters such as air and soil temperature, humidity, and precipitation. These devices operate on limited battery reserves and rely on wireless communication protocols, ranging from short-range (Bluetooth, Zigbee, WiFi [ 3 , 4 , 5 ] ) to long-range (LoRa, Sigfox, NB-IoT [ 6 , 7 , 8 ] ), to transmit data to the cloud. While cloud-based architectures enable centralized data storage and analytics, they also create bottlenecks: 1. Dependence on stable Internet connectivity, often lacking in rural areas. 2. Continuous transmission of redundant or slowly changing sensor readings, leading to energy drain and network congestion. To address these constraints, edge computing has emerged as a promising paradigm [ 9 ] . It offers a compelling alternative to centralized cloud-based solutions by bringing intelligence closer to the data source. In agricultural contexts where sensor readings change gradually, transmitting every measurement to the cloud is often unnecessary and inefficient [ 10 , 11 ] . Processing data locally on a sensor or gateway node reduces bandwidth usage, lowers latency, and preserves battery life [ 12 ] . In the proposed use case, the edge device functions as a local gateway that collects readings from nearby environmental sensors and performs lightweight predictive inference, transmitting only significant deviations to the cloud. Yet, a common limitation in deploying predictive models to the edge is the lack of historical labeled data required for training. This is particularly problematic in newly instrumented or under-monitored regions. To address this, we propose an edge-based prediction architecture that leverages pretrained models on publicly available satellite-derived climate datasets and transfers them to in-situ stations without retraining. This cross-site inference strategy enables rapid deployment while maintaining high prediction accuracy, eliminating the need to train and update separate models for each location. Our approach employs a LSTM model suitable for the edge deployment. Each device forecasts future measurements and transmits data only when the observed value deviates from the prediction by more than a configurable error threshold. This predictive filtering significantly reduces the volume of transmitted data, conserving both energy and bandwidth without sacrificing data integrity. The primary contributions of this paper are summarized as follows: • Edge-Based Predictive Filtering: We implement a compact LSTM model optimized for edge deployment, enabling local forecasting and reducing communication frequency. • Cross-Site Generalization: We demonstrate that models trained on satellite-derived temperature data from one region generalize well to other in-situ locations, supporting scalable deployment in data-scarce environments. • Real-World Evaluation: Using agrometeorological datasets from Croatia, we demonstrate substantial reductions in communication load (up to 94%) while maintaining high prediction accuracy under realistic deployment conditions. The remainder of this paper is organized as follows: Section II reviews related work. Section III details the prediction and transmission algorithm as well as model architecture. Section IV describes the datasets and deployment strategy. Section V presents experimental results. Finally, Section VI concludes the paper and outlines directions for future research. II Related Work Reducing data transmission in IoT networks is a well-studied problem, especially in domains like smart agriculture where energy efficiency and intermittent connectivity are critical constraints [ 13 ] . Data reduction strategies are typically categorized as compression, aggregation, or prediction-based [ 14 , 15 ] , each with its own trade-offs in complexity, energy usage, and fidelity. II-A Data Compression and Aggregation Data compression reduces payload size by encoding readings at the source and decoding at the destination [ 16 ] . Techniques such as DASA [ 17 ] and Huffman-based encoding [ 18 ] show effectiveness in edge environments. In Agricultural IoT, combining compression with secure protocols like MQTT has been explored to ensure scalability and data integrity [ 19 ] . However, these techniques often underperform in slowly changing environments, typical in agriculture. Aggregation methods reduce data by summarizing multiple values (e.g., averages or statistics) before transmission [ 20 , 21 ] . These methods are often implemented at gateways or cluster heads to alleviate network congestion. While scalable, they can obscure temporal trends important for fine-grained decision-making and result in loss of detail. II-B Prediction-Based Data Reduction Prediction-based methods are an effective class of techniques for reducing data transmission in IoT systems. These methods forecast future sensor values and suppress transmission when the predicted value falls within an acceptable error range, thus minimizing communication overhead. Traditional approaches rely on statistical models such as autoregression or exponential smoothing. More recent work leverages machine learning and deep learning models, MLPs, LSTMs, CNNs, and attention-based architectures, for improved accuracy in capturing complex temporal dependencies [ 22 , 23 , 24 ] . Adaptive frameworks further enhance flexibility. For instance, [ 25 ] proposed an adaptive dual prediction model that adjusts in real time to environmental dynamics. Prediction can occur centrally [ 26 ] or locally at the sensor node [ 27 , 28 ] . Local prediction reduces dependency on constant connectivity but requires computationally efficient models. Dual prediction schemes (DPS) [ 29 , 30 , 31 , 32 ] synchronize models at both the sensor and gateway, transmitting only when both predictions fail. While this improves robustness, it

Read the complete text