DSCA-PSPNet: Dynamic spatial-channel attention pyramid scene parsing network for sugarcane field segmentation in satellite imagery

Yujian Yuan; Lina Yang; Kan Chang; Youju Huang; Haoyan Yang; Jiale Wang

doi:10.3389/fpls.2023.1324491

DSCA-PSPNet: Dynamic spatial-channel attention pyramid scene parsing network for sugarcane field segmentation in satellite imagery

Front Plant Sci. 2024 Jan 17:14:1324491. doi: 10.3389/fpls.2023.1324491. eCollection 2023.

Authors

Yujian Yuan^{1

2}, Lina Yang^{1

2}, Kan Chang¹, Youju Huang³, Haoyan Yang¹, Jiale Wang¹

Affiliations

¹ School of Computer, Electronics, and Information, Guangxi University, Nanning, China.
² Guangxi Key Laboratory of Multimedia Communications and Network Technology, School of Computer, Electronics, and Information, Guangxi University, Nanning, China.
³ Guangxi Institute of Remote Sensing of Natural Resources, Nanning, China.

Abstract

Sugarcane plays a vital role in many global economies, and its efficient cultivation is critical for sustainable development. A central challenge in sugarcane yield prediction and cultivation management is the precise segmentation of sugarcane fields from satellite imagery. This task is complicated by numerous factors, including varying environmental conditions, scale variability, and spectral similarities between crops and non-crop elements. To address these segmentation challenges, we introduce DSCA-PSPNet, a novel deep learning model with a unique architecture that combines a modified ResNet34 backbone, the Pyramid Scene Parsing Network (PSPNet), and newly proposed Dynamic Squeeze-and-Excitation Context (D-scSE) blocks. Our model effectively adapts to discern the importance of both spatial and channel-wise information, providing superior feature representation for sugarcane fields. We have also created a comprehensive high-resolution satellite imagery dataset from Guangxi's Fusui County, captured on December 17, 2017, which encompasses a broad spectrum of sugarcane field characteristics and environmental conditions. In comparative studies, DSCA-PSPNet outperforms other state-of-the-art models, achieving an Intersection over Union (IoU) of 87.58%, an accuracy of 92.34%, a precision of 93.80%, a recall of 93.21%, and an F1-Score of 92.38%. Application tests on an RTX 3090 GPU, with input image resolutions of 512 × 512, yielded a prediction time of 4.57ms, a parameter size of 22.57MB, GFLOPs of 11.41, and a memory size of 84.47MB. An ablation study emphasized the vital role of the D-scSE module in enhancing DSCA-PSPNet's performance. Our contributions in dataset generation and model development open new avenues for tackling the complexities of sugarcane field segmentation, thus contributing to advances in precision agriculture. The source code and dataset will be available on the GitHub repository https://github.com/JulioYuan/DSCA-PSPNet/tree/main.

Keywords: D-scSE; PSPNet; deep learning; precision agriculture; remote sensing; satellite imagery; sugarcane field segmentation.

Grants and funding

The author(s) declare financial support was received for the research, authorship, and/or publication of this article. The authors gratefully acknowledge the financial support provided by the National Natural Science Foundation of China (62171145)(61966003)(62371144) and Natural Science Foundation of Guangxi Province (2020GXNSFAA159171).