Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit

Yang Zhao; Shusheng Li; Xueshang Feng

doi:10.1145/3746027.3755609

Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit

Yang Zhao
, Shusheng Li
, Xueshang Feng

Harbin Institute of Technology Shenzhen

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

As the development of lightweight deep learning algorithms, various deep neural network (DNN) models have been proposed for the remote sensing scene classification (RSSC) application. However, it is still challenging for these RSSC models to achieve optimal performance among model accuracy, inference latency, and energy consumption on resource-constrained edge devices. In this paper, we propose a lightweight RSSC framework, which includes a distilled global filter network (GFNet) model and an early-exit mechanism designed for edge devices to achieve state-of-the-art performance. Specifically, we first apply frequency domain distillation on the GFNet model to reduce model size. Then we design a dynamic early-exit model tailored for DNN models on edge devices to further improve model inference efficiency. We evaluate our E3C model on three edge devices across four datasets. Extensive experimental results show that it achieves an average of 1.3x speedup on model inference and over 40% improvement on energy efficiency, while maintaining high classification accuracy.

Original language	English
Title of host publication	MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025
Publisher	Association for Computing Machinery, Inc
Pages	11862-11870
Number of pages	9
ISBN (Electronic)	9798400720352
DOIs	https://doi.org/10.1145/3746027.3755609
State	Published - 27 Oct 2025
Externally published	Yes
Event	33rd ACM International Conference on Multimedia, MM 2025 - Dublin, Ireland Duration: 27 Oct 2025 → 31 Oct 2025

Publication series

Name	MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025

Conference

Conference	33rd ACM International Conference on Multimedia, MM 2025
Country/Territory	Ireland
City	Dublin
Period	27/10/25 → 31/10/25

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

SDG 7 Affordable and Clean Energy

Keywords

dnn inference
early-exit
edge computing
knowledge distillation
remote sensing scene classification

Access to Document

10.1145/3746027.3755609

Cite this

Zhao, Y., Li, S., & Feng, X. (2025). Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit. In MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025 (pp. 11862-11870). (MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025). Association for Computing Machinery, Inc. https://doi.org/10.1145/3746027.3755609

Zhao, Yang ; Li, Shusheng ; Feng, Xueshang. / Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit. MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025. Association for Computing Machinery, Inc, 2025. pp. 11862-11870 (MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025).

@inproceedings{647cb526a353411796d27b60d61e3eee,

title = "Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit",

abstract = "As the development of lightweight deep learning algorithms, various deep neural network (DNN) models have been proposed for the remote sensing scene classification (RSSC) application. However, it is still challenging for these RSSC models to achieve optimal performance among model accuracy, inference latency, and energy consumption on resource-constrained edge devices. In this paper, we propose a lightweight RSSC framework, which includes a distilled global filter network (GFNet) model and an early-exit mechanism designed for edge devices to achieve state-of-the-art performance. Specifically, we first apply frequency domain distillation on the GFNet model to reduce model size. Then we design a dynamic early-exit model tailored for DNN models on edge devices to further improve model inference efficiency. We evaluate our E3C model on three edge devices across four datasets. Extensive experimental results show that it achieves an average of 1.3x speedup on model inference and over 40\% improvement on energy efficiency, while maintaining high classification accuracy.",

keywords = "dnn inference, early-exit, edge computing, knowledge distillation, remote sensing scene classification",

author = "Yang Zhao and Shusheng Li and Xueshang Feng",

note = "Publisher Copyright: {\textcopyright} 2025 ACM.; 33rd ACM International Conference on Multimedia, MM 2025 ; Conference date: 27-10-2025 Through 31-10-2025",

year = "2025",

month = oct,

day = "27",

doi = "10.1145/3746027.3755609",

language = "英语",

series = "MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025",

publisher = "Association for Computing Machinery, Inc",

pages = "11862--11870",

booktitle = "MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025",

}

Zhao, Y, Li, S & Feng, X 2025, Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit. in MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025. MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025, Association for Computing Machinery, Inc, pp. 11862-11870, 33rd ACM International Conference on Multimedia, MM 2025, Dublin, Ireland, 27/10/25. https://doi.org/10.1145/3746027.3755609

Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit. / Zhao, Yang; Li, Shusheng; Feng, Xueshang.
MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025. Association for Computing Machinery, Inc, 2025. p. 11862-11870 (MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit

AU - Zhao, Yang

AU - Li, Shusheng

AU - Feng, Xueshang

PY - 2025/10/27

Y1 - 2025/10/27

N2 - As the development of lightweight deep learning algorithms, various deep neural network (DNN) models have been proposed for the remote sensing scene classification (RSSC) application. However, it is still challenging for these RSSC models to achieve optimal performance among model accuracy, inference latency, and energy consumption on resource-constrained edge devices. In this paper, we propose a lightweight RSSC framework, which includes a distilled global filter network (GFNet) model and an early-exit mechanism designed for edge devices to achieve state-of-the-art performance. Specifically, we first apply frequency domain distillation on the GFNet model to reduce model size. Then we design a dynamic early-exit model tailored for DNN models on edge devices to further improve model inference efficiency. We evaluate our E3C model on three edge devices across four datasets. Extensive experimental results show that it achieves an average of 1.3x speedup on model inference and over 40% improvement on energy efficiency, while maintaining high classification accuracy.

AB - As the development of lightweight deep learning algorithms, various deep neural network (DNN) models have been proposed for the remote sensing scene classification (RSSC) application. However, it is still challenging for these RSSC models to achieve optimal performance among model accuracy, inference latency, and energy consumption on resource-constrained edge devices. In this paper, we propose a lightweight RSSC framework, which includes a distilled global filter network (GFNet) model and an early-exit mechanism designed for edge devices to achieve state-of-the-art performance. Specifically, we first apply frequency domain distillation on the GFNet model to reduce model size. Then we design a dynamic early-exit model tailored for DNN models on edge devices to further improve model inference efficiency. We evaluate our E3C model on three edge devices across four datasets. Extensive experimental results show that it achieves an average of 1.3x speedup on model inference and over 40% improvement on energy efficiency, while maintaining high classification accuracy.

KW - dnn inference

KW - early-exit

KW - edge computing

KW - knowledge distillation

KW - remote sensing scene classification

UR - https://www.scopus.com/pages/publications/105024064101

U2 - 10.1145/3746027.3755609

DO - 10.1145/3746027.3755609

M3 - 会议稿件

AN - SCOPUS:105024064101

T3 - MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025

SP - 11862

EP - 11870

BT - MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025

PB - Association for Computing Machinery, Inc

T2 - 33rd ACM International Conference on Multimedia, MM 2025

Y2 - 27 October 2025 through 31 October 2025

ER -

Zhao Y, Li S, Feng X. Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit. In MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025. Association for Computing Machinery, Inc. 2025. p. 11862-11870. (MM 2025 - Proceedings of the 33rd ACM International Conference on Multimedia, Co-Located with MM 2025). doi: 10.1145/3746027.3755609

Lightweight Remote Sensing Scene Classification on Edge Devices via Knowledge Distillation and Early-exit

Abstract

Publication series

Conference

UN SDGs

Keywords

Access to Document

Other files and links

Fingerprint

Cite this