Bulletin of Surveying and Mapping ›› 2026, Vol. 0 ›› Issue (8): 137-144.doi: 10.13474/j.cnki.11-2246.2026.0820

Previous Articles    

Intelligent cultivated land extraction from ZY1E imagery via dual-backbone deep learning network

Zhou Hong1, Zhang Guohe1, You Siqi2, Wang Ran1, Ding Pengfei1   

  1. 1. Military-Civilian Integration Geological Survey Center, China Geological Survey, Chengdu 610036, China;
    2. School of Geography and Planning, Chengdu University of Technology, Chengdu 610059, China
  • Received:2025-11-14 Published:2026-09-12

PDF

Abstract: [Purposes] Cultivated land is a fundamental resource crucial for ensuring national food security and ecological stability.With the rapid development of high-resolution satellite data,utilizing deep learning technology for high-precision automated identification of cultivated land has become a significant direction in agricultural remote sensing research. [Methods] This study takes the Suqian area of Jiangsu province as the research area.ZY1E images underwent preprocessing operations including radiometric calibration,atmospheric correction,fusion,and geometric registration.A high-quality training sample set was constructed through a multi-type data augmentation strategy.On this basis,a DB-DeepLabV3+semantic segmentation model integrating the dual-backbone networks of EfficientNet-B3 and ResNet-50 is proposed.In this model,the EfficientNet-B3 backbone network is used for efficient multi-scale feature extraction,while the ResNet-50 backbone enhances deep semantic representation capability； the two achieve structural complementarity and boundary optimization within the DeepLabV3+framework. [Findings] Results indicate that the DB-DeepLabV3+model demonstrates high accuracy and stability in cultivated land identification,achieving an overall accuracy (OA)of 90.91%, a Kappa coefficient of 0.818,and an intersection over union (IoU)of 0.832,outperforming the single-backbone DeepLabV3+,EfficientNet-B3+DeepLabV3+,and ResNet-50+DeepLabV3+models.The model maintained good performance consistency under different data split ratios,verifying its strong generalization ability. [Conclusions] The research findings demonstrate that the dual-backbone structure holds significant advantages for cultivated land extraction from high-resolution remote sensing imagery and provides effective technical support for the application of the domestic ZY1E satellite in agricultural intelligent recognition and land monitoring．

Key words: cultivated land identification, high-resolution remote sensing, deep learning, ResNet-50, EfficientNet, DeepLabV3+

CLC Number: