Journal of Geo-information Science >
Prior Knowledge Guided Deep Learning for Monitoring Buildings and Greenhouses within Cultivated Land
Received date: 2023-06-06
Revised date: 2023-09-07
Online published: 2023-11-02
Supported by
Fundamental Research Funds for the Central Universities(2021YCPY0113)
National Natural Science Foundation of China(52204190)
National Natural Science Foundation of China(42271368)
Accurate and automated monitoring of the non-agriculturalization of cultivated land has important implications for upholding the arable land red line and ensuring sustainable socio-economic development. This paper proposes a deep learning method guided by prior knowledge to achieve precise monitoring of the illegal occupation of cultivated land by buildings and greenhouses, with the ultimate goal of issuing timely warnings. Firstly, the vector range and category attributes are obtained from the third national land survey database, serving as prior knowledge. Then, parcel-level segmentation of high-resolution remote sensing images is performed using the front-phase vector data of the cultivated land to locate detection areas. Next, the SRAM-SegFormer model, integrated with prior knowledge, is employed to extract potential non-agricultural patches. Finally, post-processing operations such as mosaic, reclassification, and overlay are performed to obtain final monitoring results of arable land non-agriculturalization. Taking the Peixian County in Xu Zhou City as the study area, the performance of common sematic segmentation networks, including Deeplabv3+, PSPNet, U-Net, HRNet, SegFormer, and SRAM-SegFormer, in extracting potential non-agricultural patches are compared. The results show that Deeplabv3+ and PSPNet are prone to omissions and false detections in complex areas; U-Net tends to miss large-scale buildings; HRNet exhibits irregular boundaries in extracted non-agricultural patches; SegFormer has poor extraction ability for small-scale buildings and greenhouses, and tends to merge buildings and greenhouses in densely populated areas; SRAM-SegFormer shows the best results, with the highest accuracy rate for Mean Pixel Accuracy (MPA) (84.30%), Mean Intersection-Over-Union (MIoU) (73.76%), and Overall Accuracy (OA) (97.91%), especially in extracting small-scale buildings and greenhouses. Therefore, the proposed method in this paper can achieve more efficient and automated monitoring of arable land non-agriculturalization.
TAN Min , LIN Huijing , HAO Ming . Prior Knowledge Guided Deep Learning for Monitoring Buildings and Greenhouses within Cultivated Land[J]. Journal of Geo-information Science, 2023 , 25(11) : 2293 -2302 . DOI: 10.12082/dqxxkx.2023.230315
表1 不同方法的潜在非农化图斑提取结果精度评价Tab. 1 Accuracy evaluation of the potential non-agricultural land extraction results by different methods (%) |
方法 | MPA | MIoU | OA |
---|---|---|---|
Deeplabv3+ | 74.46 | 61.05 | 95.61 |
PSPNet | 76.54 | 64.53 | 96.00 |
U-Net | 76.56 | 66.38 | 96.84 |
HRNet | 79.19 | 64.33 | 96.26 |
SegFormer | 79.71 | 68.42 | 96.99 |
SRAM-SegFormer | 84.30 | 73.76 | 97.91 |
注:表中加粗数值表示表格中每一行的最佳值。 |
表2 消融实验精度评价结果Tab. 2 Accuracy evaluation results of ablation studies (%) |
方法 | MPA | MIoU | OA | ||
---|---|---|---|---|---|
S1+SRAM | S2+SRAM | S3+SRAM | |||
√ | √ | √ | 84.30 | 73.76 | 97.91 |
× | × | √ | 82.44 | 72.00 | 97.85 |
× | √ | × | 83.98 | 71.79 | 97.76 |
√ | × | × | 80.94 | 70.50 | 97.73 |
× | × | × | 79.71 | 68.42 | 96.99 |
注:表中加粗数值表示表格中每一行的最佳值。 |
[1] |
黄祖辉, 李懿芸, 毛晓红. 我国耕地“非农化”“非粮化”的现状与对策[J]. 江淮论坛, 2022(4):13-21.
|
[2] |
李宏鹏, 田达睿, 谭静斌. 延安市2000—2020年耕地非农化时空格局演变及其影响因素[J]. 水土保持通报, 2022, 42(4):330-337,372.
|
[3] |
张兆鹏, 田昕, 朱新杰, 等. 基于GF-1影像的兰州新区土地利用类型时空变化分析[J]. 地理信息世界, 2022, 29(3):58-65.
|
[4] |
闫利, 张毅, 杨见兵, 等. 自然资源典型地类遥感智能主动调查监测技术及应用[J]. 地理信息世界, 2022, 29(5):66-73.
|
[5] |
|
[6] |
|
[7] |
李刚, 宁晓刚, 张翰超, 等. “三调”成果数据引导的耕地“非农化”遥感监测[J]. 测绘科学, 2022, 47(7):149-159.
|
[8] |
|
[9] |
高鸣, 周鑫鑫, 刘琦, 等. 深度学习网络支持下的农房侵占耕地自动化监测[J]. 测绘通报, 2022(3):47-53.
|
[10] |
徐志红, 关元秀, 王善华, 等. 融合对象影像分析和OCNN的耕地变化检测[J]. 遥感信息, 2022, 37(5):15-22.
|
[11] |
曹州, 刘士彬, 马勇, 等. 一种联合光谱-对象-时间特征的遥感影像变化检测方法[J]. 中国科学院大学学报, 2022, 39(1):102-109.
|
[12] |
|
[13] |
国务院办公厅. 国务院办公厅关于坚决制止耕地“非农化”行为的通知[J]. 中华人民共和国国务院公报, 2020(27):36-37.
[General Office of the State Council of the People’s Republic of China. Circular of the general office of the state council on resolutely stopping the “non-agriculturalization” of cultivated land[J]. Gazette of the State Council of the People's Republic of China, 2020(27):36-37. ]
|
[14] |
农业农村部自然资源部印发《关于开展“大棚房”问题专项清理整治行动坚决遏制农地非农化的方案》的通知农农发[2018]3号[J]. 中华人民共和国农业农村部公报, 2018(10):49-52.
[ Notice on the plan to carry out the special clean-up and rectification action of “greenhouse” to resolutely curb the non-agriculturalization of agricultural landJ. Gazette of the Ministry of Agriculture and Affairs of the People’s Republic of China, 2018(10):49-52. ]
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
/
〈 | 〉 |