[1]田子昊,刘振军,张玉娟,等.基于改进SegFormer模型和多分辨率影像融合的耕地精细化提取方法[J].江苏农业学报,2026,42(08):1597-1607.[doi:doi:10.3969/j.issn.1000-4440.2026.08.009]
　TIAN Zihao,LIU Zhenjun,ZHANG Yujuan,et al.Detailed farmland extraction method based on improved SegFormer model and multi-resolution image fusion[J].,2026,42(08):1597-1607.[doi:doi:10.3969/j.issn.1000-4440.2026.08.009]
点击复制

基于改进SegFormer模型和多分辨率影像融合的耕地精细化提取方法()

江苏农业学报[ISSN:1006-6977/CN:61-1281/TN]

卷:
42
期数:
2026年08期
页码:
1597-1607
栏目:
农业信息工程
出版日期:
2026-08-31

文章信息/Info

Title:
Detailed farmland extraction method based on improved SegFormer model and multi-resolution image fusion
作者:
田子昊1; 刘振军2; 张玉娟1; 焦英华1; 刘耀辉1; 赵展1
(1.山东建筑大学测绘地理信息学院,山东济南250101；2.山东省国土测绘院,山东济南250102)
Author(s):
TIAN Zihao1; LIU Zhenjun2; ZHANG Yujuan1; JIAO Yinghua1; LIU Yaohui1; ZHAO Zhan1
(1.School of Surveying and Geo-Informatics, Shandong Jianzhu University, Jinan 250101, China;2.Shandong Institute of Land Surveying and Mapping, Jinan 250102, China)
关键词:
多分辨率影像; 影像融合; 耕地提取; 深度学习; SegFormer
Keywords:
multi-resolution image; image fusion; farmland extraction; deep learning; SegFormer
分类号:
S127
DOI:
doi:10.3969/j.issn.1000-4440.2026.08.009
文献标志码:
A
摘要:
耕地的精细化提取对于农业资源管理与粮食安全保障具有重要意义。为提高地块破碎区域耕地的提取精度和准确率, 本研究首先以资源一号卫星高分辨率数据和哨兵2号卫星中分辨率数据构建不同的数据集:资源卫星红光、绿光、蓝光3波段数据构成的数据集1，哨兵2号卫星影像上采样至资源卫星分辨率后的12波段数据构成的数据集2，融合资源一号卫星影像后的哨兵2号12波段数据构成的数据集3，数据集1+数据集2构成的数据集4，数据集1+数据集3构成的数据集5。然后，对原始SegFormer模型进行优化:采用自适应输入层设计, 使编码器灵活适配多源遥感影像的光谱通道数, 并通过增加特征提取深度, 增强网络在高维空间的表征能力；利用可学习的转置卷积(Transposed convolution)替代原始解码器中的双线性插值上采样, 在还原空间分辨率的同时保留更多边缘细节信息, 有效防止耕地边界在上采样过程中出现模糊、断裂或粘连等现象。最后，以山东省淄博市高青县为研究区, 明确不同数据集下原始SegFormer模型和改进SegFormer模型对耕地提取精度的影响。结果表明, 利用融合影像能够显著提高耕地分割精度。其中, 利用融合数据集5, 改进SegFormer模型对耕地的识别精度平均交并比(mIoU)、平均F1分数(mF1)、平均像素准确率(mAcc)分别比数据集1提升3.32个百分点、1.73个百分点、1.77个百分点, 比数据集2分别提高3.50个百分点、1.83个百分点、1.86个百分点, 效果最好。利用数据集5, 改进SegFormer模型对耕地的识别精度mIoU、mF1、mAcc分别比原始模型提高0.48个百分点、0.25个百分点、0.25个百分点。总体来看, 哨兵影像先融合再与资源一号影像叠加的数据预处理方案能较大程度地提高改进SegFormer模型的耕地提取精度。本研究结果为高精度的耕地提取与农业遥感监测提供了可行的技术方法。
Abstract:
The fine extraction of cultivated land is of great significance for agricultural resource management and food security safeguarding. In order to improve the extraction accuracy and precision of cultivated land in the broken area of the plot, this study first constructed different datasets based on the high-resolution data of ZY-1 satellite and the medium-resolution data of Sentinel-2 satellite. Dataset 1 was composed of red band, green band and blue band data of ZY-1 satellite; Dataset 2 was composed of 12 band data upsampled from Sentinel-2 satellite image to YZ-1 satellite resolution; Dataset 3 was composed of 12 band data of Sentinel-2 after fusion with ZY-1 satellite image; Dataset 4 was composed of Dataset 1 and Dataset 2; Dataset 5 was composed of Dataset 1 and Dataset 3. Then, the original SegFormer model was optimized: the adaptive input layer design was adopted to make the encoder flexibly adapt to the number of spectral channels of multi-source remote sensing images, and the representation ability of the network in high-dimensional space was enhanced by increasing the depth of feature extraction. Instead of bilinear interpolation upsampling in the original decoder, the learnable transposed convolution was used to restore the spatial resolution while retaining more edge detail information, which could effectively prevent the blur, fracture or adhesion of the cultivated land boundary during the upsampling process. Finally, Gaoqing County of Zibo City of Shandong province was taken as the research area to clarify the influence of the original SegFormer model and the improved SegFormer model on the extraction accuracy of cultivated land under different datasets. The results showed that the use of fusion images could significantly improve the accuracy of cultivated land segmentation. Among them, using the fusion Dataset 5, the mean intersection over union (mIoU), average F1 score (mF1), and average pixel accuracy (mAcc) of the improved SegFormer model for cultivated land recognition accuracy were 3.32 percentage points, 1.73 percentage points, and 1.77 percentage points higher than the Dataset 1, and were 3.50 percentage points, 1.83 percentage points, and 1.86 percentage points higher than the Dataset 2, respectively. The fusion Dataset 5 showed the best effect. Using the Dataset 5, the improved SegFormer model’s recognition accuracies of cultivated land mIoU, mF1, and mAcc were 0.48 percentage points, 0.25 percentage points, and 0.25 percentage points higher than the original model, respectively. In general, the data preprocessing scheme that first fused Sentinel image and then overlaid with ZY-1 image could greatly improve the cultivated land extraction accuracy of the improved SegFormer model. This study provides a feasible technical method for high-precision farmland extraction and agricultural remote sensing monitoring.

参考文献/References:

[1]邓艾丽,李平光,吴郁玲. 耕地转入主体的异质性对耕地“非粮化”的影响研究[J/OL].中国农业资源与区划：1-16
[2025-11-18]. https://link.cnki.net/urlid/11.3513.S.20250731.1506.016.
[2]林娜, 全海琳, 李双桃, 等. 基于SHAP可解释特征优选的撂荒耕地遥感提取[J]. 农业工程学报,2025,41(14):291-302.
[3]原昊. 基于图像分割的生产建设项目图斑解译[D]. 西安:西安电子科技大学,2022:20-30.
[4]常明会,李世华,彭帅峰,等. 集成渐进式多特征融合的复杂场景耕地提取方法[J]. 农业工程学报,2025,41(10):294-303.
[5]王耀磊,郑毅,张成程,等. 基于GF-2遥感影像的澳洲坚果林空间分布信息提取[J]. 南方农业学报,2025,56(1):74-86.
[6]Hotelling H. Analysis of a complex of statistical variables into principal components[J]. Journal of Educational Psychology,1933,24(6):417-441.
[7]Daubechies I. Orthonormal bases of compactly supported wavelets[J]. Communications on Pure and Applied Mathematics,1988,41(7):909-996.
[8]Hinton G E, Osindero S, Teh Y W. A fast learning algorithm for deep belief nets[J]. Neural Computation,2006,18(7):1527-1554.
[9]Badrinarayanan V, Kendall A, Cipolla R. SegNet:a deep convolutional encoder-decoder architecture for image segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(12):2481-2495.
[10]Ronneberger O, Fischer P, Brox T. U-net:convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer-Assisted Intervention Society:MICCAI 2015. Cham:Springer,2015:234-241.
[11]Chen L C, Zhu Y K, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Ferrari V, Hebert M, Sminchisescu C, et al. Computer Vision-ECCV 2018. Cham:Springer,2018:833-851.
[12]Xie E Z, Wang W H, Yu Z D, et al. SegFormer:simple and efficient design for semantic segmentation with transformers[EB/OL].
[2026-01-03]. 2021:arXiv:2105.15203. https://arxiv.org/abs/2105.15203
[13]郭乐乐,席瑞,齐家国,等. 基于高分辨率遥感影像的耕地提取方法研究[J]. 杭州师范大学学报(自然科学版),2025,24(1):35-44.
[14]周楠,杨鹏,魏春山,等. 地块尺度的山区耕地精准提取方法[J]. 农业工程学报,2021,37(19):260-266.
[15]Tao K J, Li H, Huang C, et al. Extraction of cropland based on multi-source remote sensing and an improved version of the deep learning-based segment anything model (SAM)[J]. Agronomy,2025,15(5):1139.
[16]陈启浩,李广潮,曹文静,等. 综合时序光学和SAR数据的南方多云雨地区耕地种植强度提取[J]. 测绘学报,2024,53(12):2361-2374.
[17]刘超兵,甘淑,袁希平,等. 级联改进DexiNed和DeepLabv3+网络的坡耕地提取[J]. 自然资源遥感,2025,37(2):49-55.
[18]顾余庆,李晓文,曹伟,等. 基于Segformer网络的地块尺度作物种植结构精细化识别与分类[J]. 江苏农业学报,2024,40(2):293-302.
[19]李倩楠,张杜娟,潘耀忠,等. MPSPNet和UNet网络下山东省高分辨耕地遥感提取[J]. 遥感学报,2023,27(2):471-491.
[20]董金玮,吴文斌,黄健熙,等. 农业土地利用遥感信息提取的研究进展与展望[J]. 地球信息科学学报,2020,22(4):772-783.
[21]刘巍,吴志峰,骆剑承,等. 深度学习支持下的丘陵山区耕地高分辨率遥感信息分区分层提取方法[J]. 测绘学报,2021,50(1):105-116.
[22]蔡玉林,刘照磊,孟祥磊,等. 基于HRNet和自注意力机制的多源遥感影像水稻提取[J]. 农业工程学报,2024,40(4):186-193.
[23]张新长,黄健锋,宁婷. 高分辨率遥感影像耕地提取研究进展与展望[J]. 武汉大学学报(信息科学版),2023,48(10):1582-1590.
[24]李旭青,吴冬雪,王玉博,等. 基于改进HRNet的遥感影像冬小麦语义分割方法[J]. 农业工程学报,2024,40(3):193-200.
[25]肖茂池,林娜,李双桃,等. 融合注意力机制的DeepLabV3+耕地遥感提取研究[J]. 遥感技术与应用,2025,40(5):1323-1332.
[26]冯权泷,江子航,牛博文,等. 东北黑土区侵蚀沟遥感识别的多尺度特征提取模型:以海伦市为例[J]. 遥感学报,2024,28(12):3147-3157.
[27]Lu H, Qiao D Y, Li Y X, et al. Fusion of China ZY-1 02D hyperspectral data and multispectral data:which methods should be used?[J]. Remote Sensing,2021,13(12):2354.
[28]Drusch M, Del Bello U, Carlier S, et al. Sentinel-2:ESA’s optical high-resolution mission for GMES operational services[J]. Remote sensing of Environment,2012,120:25-36.
[29]Brodu N. Super-resolving multiresolution images with band-independent geometry of multispectral pixels[J]. IEEE Transactions on Geoscience and Remote Sensing,2017,55(8):4610-4617.
[30]He K M, Zhang X Y, Ren S Q, et al. Delving deep into rectifiers:surpassing human-level performance on ImageNet classification[C]//IEEE:2015 IEEE International Conference on Computer Vision (ICCV). Santiago,Chile:IEEE,2016:1026-1034.
[31]张圣微,杨林,叶德成,等. 基于深度学习的内蒙古河套灌区2000—2021年种植结构提取与动态[J]. 农业工程学报,2023,39(18):142-150.
[32]张志博,赵西宁,姜海晨,等. 基于无人机影像和深度学习的渭北旱塬区土地利用精准分类[J]. 农业工程学报,2022,38(22):199-209.
[33]杨丽丽,陈炎,田伟泽,等. 田间道路改进UNet分割方法[J]. 农业工程学报,2021,37(9):185-191.

相似文献/References:

[1]金正婷,李卫国,景元书.基于影像融合的冬小麦种植面积提取适宜尺度研究[J].江苏农业学报,2015,(06):1312.[doi:doi:10.3969/j.issn.1000-4440.2015.06.018]
　JIN Zheng-ting,LI Wei-guo,JING Yuan-shu.Appropriate extraction scale of winter wheat planting area based on image fusion[J].,2015,(08):1312.[doi:doi:10.3969/j.issn.1000-4440.2015.06.018]
[2]廖靖,胡月明,赵理,等.结合数据融合算法的光能利用率模型反演水稻地上部生物量[J].江苏农业学报,2019,(03):594.[doi:doi:10.3969/j.issn.1000-4440.2019.03.013]
　LIAO Jing,HU Yue-ming,ZHAO Li,et al.Inversion of aboveground biomass of rice by combining light use efficiency model with data fusion algorithm[J].,2019,(08):594.[doi:doi:10.3969/j.issn.1000-4440.2019.03.013]

备注/Memo

备注/Memo:
收稿日期：2026-01-06基金项目：国家自然科学基金项目（42201077）作者简介：田子昊（2000-），男，山东淄博人，硕士研究生，主要研究方向为地块精细化提取。(E-mail)2224019348@qq.com通讯作者：赵展，(E-mail)zhaozhan19@sdjzu.edu.cn
更新日期/Last Update: 2026-09-16