基于特征融合的精细归一化植被指数生成方法

A Spatiotemporal Feature Fusion-Based Super-Resolution Method for Normalized Difference Vegetation Index

摘要

精细化、大范围的归一化植被指数（Normalized Difference Vegetation Index ， NDVI）是评价农作物生长状态的重要参数，获取高质量的NDVI数据对农业遥感具有重要意义。本文针对利用时序Landsat 8与Sentinel-2数据生成具有高时空分辨率的 NDVI 数据的需要，构建了一种时序超分辨率模型生成 NDVI 数据（Temporal Super-Resolution Network， TSRNet），模型以编码器-融合-解码器结构为基础，使用时序注意力机制计算各时间步的特征贡献，以捕捉多时相间的依赖关系；设计了融合通道注意力、空间注意力与全局注意力的空间增强模块，用于实现多层次、多尺度的特征增强。TSRNet通过权衡时序特征的贡献和空间特征的多维度增强，显著提高了时空特征提取的精度，能够生成与直接从高分辨率影像提取的NDVI质量一致的NDVI数据；较好保持了原始NDVI数据的变化趋势，有效保留作物物候信息。选取HighResNet、UTAE、TR-MISR三种模型作为对比，实验结果表明，TSRNet在R²、PSNR、SSIM等指标上均优于对比模型，表明TSRNet在时空特征提取和图像细节重建方面具有显著优势。

Abstract

High-resolution large-scale Normalized Difference Vegetation Index (NDVI) is an important parameter for assessing crop growth, and obtaining high-quality NDVI data is of great significance for agricultural remote sensing. To address the need for generating high spatiotemporal resolution NDVI data using time-series Landsat 8 and Sentinel-2 data, this study constructs a temporal super-resolution model for NDVI data generation (Temporal Super-Resolution Network, TSRNet). Based on an encoder-fusion-decoder architecture, the model employs a temporal attention mechanism to calculate the feature contribution of each time step, capturing dependencies across multiple temporal phases. Additionally, a spatial enhancement module integrating channel attention, spatial attention, and global attention is designed to achieve multi-level, multi-scale feature enhancement. By balancing the contribution of temporal features and multidimensional enhancement of spatial features, TSRNet significantly improves the accuracy of spatiotemporal feature extraction. It can generate NDVI data with quality consistent with that extracted directly from high-resolution images, better preserving the variation trend of the original NDVI data and effectively retaining crop phenology information. Three models-HighResNet, UTAE, and TR-MISR- are selected for comparison. Experimental results show that TSRNet outperforms the compared models in metrics such as R², PSNR, SSIM, indicating its significant advantages in spatiotemporal feature extraction and image detail reconstruction.

关键词

深度学习 / 农作物 / 归一化植被指数 / 超分辨率 / 物候信息

Key words

Deep Learning / crops / Normalized Difference Vegetation Index (NDVI) / Super-Resolution / temporal information

参考文献

[1] 孙 瑞, 张方敏, 翁升恒, 等. 2001-2021 年中国NDVI时空格局变化及对气候的响应[J]. 中国环境科学, 2023, 43(10): 5519-5528.

[2] 金 凯,王 飞, 韩剑桥, 等. 1982-2015 年中国气候变化和人类活动对植被NDVI变化的影响[J]. 地理学报, 2020, 75(05): 961-974.

[3] Andleeb T, Ali Z S, Mahmood Z, et al. Wheat varietal response to relative SPAD index (RSI) and relative normalized difference vegetation index (RNDVI) under variable nitrogen application and terminal heat stress along with yield repercussion[J]. Agronomy, 2022, 12(7): 1538.

[4] 鄂高阳,韩 芳, 秦秉希, 等. 遥感技术在农业资源与土壤环境综合监测上的应用[J]. 山东农业科学, 2024, 56(03): 163-170.

[5] 王来刚,曾 凯, 李文娟, 等. 黄淮海地区农业遥感研究进展、发展现状与未来展望[J]. 中国农业信息, 2023, 35(06): 60-70.

[6] Liu Y, Li H, Zhu L, et al. Revealing cropping intensity dynamics using high-resolution imagery: A case study in Shaanxi Province, China[J]. Remote Sensing, 2024, 16: 3832.

[7] Liu S, Liu J, Tan X, et al. A hybrid spatiotemporal fusion method for high spatial resolution imagery: Fusion of Gaofen-1 and Sentinel-2 over agricultural landscapes[J]. Journal of Remote Sensing, 2024, 4: 159.

[8] 王 帅,王 晓,梁 峰, 等. 无人机在我国农业监测与农事作业上的应用研究进展[J]. 山东农业科学, 2025, 57(06): 171-180.

[9] Lian Z, Zhan Y, Zhang W, et al. Recent advances in deep learning-based spatiotemporal fusion methods for remote sensing images [J]. Sensors, 2025, 25(4): 1093.

[10] Qiu B, Lin D, Chen C, et al. From cropland to cropped field: A robust algorithm for national-scale mapping by fusing time series of Sentinel-1 and Sentinel-2[J]. International Journal of Applied Earth Observation and Geoinformation, 2022, 113: 103006.

[11] Xiong C, Tang J, Chen T, et al. Improved global 250 m 8-day NDVI and EVI products from MODIS using LSTM networks[J]. Scientific Data, 2023, 10: 695.

[12] Xin Li, Li X, Xu J W, Jia Y Y, et al. Spatio-temporal dynamics of vegetation over cloudy areas in Southwest China retrieved from four NDVI products [J]. Ecological Informatics, 2024, 81: 102630.

[13] 孙玉洁, 王欣哲, 王文焯, 等. 基于光谱尺度注意力的任意倍率遥感图像超分辨率[J]. 火力与指挥控制, 2024, 49(11): 78-86+ 94.

[14] 陆婉芸, 王继周. 遥感影像超分辨率处理方法与研究进展[J]. 测绘科学, 2016, 41(12): 53-58+ 69.

[15] Li J, Pei Z, Li W, et al. A systematic survey of deep learning-based single image super resolution [J]. ACM Computing Surveys, 2024, 56(10): 1-40.

[16] Makwana Y, Swadas P B, Patel P S. Single image super resolution using deep learning: A survey[J]. International Journal of Advance Engineering and Research Development, 2020, 7(4): 22-27.

[17] Hu S, Li Y, Xu Y, et al. A review of deep-learning-based super-resolution: From methods to applications [J]. Pattern Recognition, 2025, 157: 110935.

[18] 赵利军,王 可, 张晋京, 等. 深度图超分辨率重建研究综述[J]. 计算机应用研究, 2023, 40(06): 25-31.

[19] 唐晓天,杨 雪,李 峰, 等. 自注意力映射超时相遥感图像超分辨率重建[J]. 遥感学报, 2024, 28(06): 1551-1559.

[20] Al-Mekhlafi H, Liu S. Single-image super-resolution: a comprehensive review and recent insight[J]. Frontiers of Computer Science, 2024, 18(1): 181702.

[21] Qin P, Huang H, Chen P, et al. Reconstructing NDVI time series in cloud-prone regions: A fusion-and-fit approach with deep learning residual constraint[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2024, 218(Part A): 170-186.

[22] 王睿琪. 图像超分辨率重建综述[J]. 计算机科学与应用, 2024, 14(02): 350-359.

[23] Salvetti F, Masi G, Bruna A, et al. Multi-image super resolution of remotely sensed images: exploiting spatiotemporal correlations to outperform SISR[J]. Remote Sensing, 2020, 12(14): 2207.

[24] Zhang Z, Hao X, Li J, et al. ESC-MISR: Enhancing spatial correlations for multi-image super-resolution in remote sensing[C]. Proceedings of the 31st International Conference on Multimedia Modeling (MMM 2025). Singapore: Springer, 2025: 373-387.

[25] Deudon M, Kalaitzis A, Goytom I, et al. Highres-net: recursive fusion for multi-frame super-resolution of satellite imagery [J]. arXiv Preprint arXiv: 2002.06460, 2020.

[26] Garnot V S F, Landrieu L. Panoptic segmentation of satellite image time series with convolutional temporal attention networks[C]. Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021: 4872-4881.

[27] Salvetti Francesco, Mazzia Vittorio, Khaliq Aamir, et al. Multi-image super resolution of remotely sensed images using residual attention deep neural networks [J]. Remote Sensing, 2020, 12(14): 2207.

[28] An T, Zhang X, Huo C Y, et al. TR-MISR: Multi-image super-resolution based on feature fusion with transformers[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2022, 15: 1373-1388.

[29] Ebel P, Meraner A, Schmitt M, et al. Multisensor data fusion for cloud removal in global and all-season Sentinel-2 imagery[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59(7): 5866-5878.

[30] Jing R, Duan F, Lu F, et al. Denoising diffusion probabilistic feature-based network for cloud removal in sentinel-2 imagery[J]. Remote Sensing, 2023, 15(9): 2217.

[31] 张金懿, 郝晋珉, 王 楠, 等. 山东省耕地质量空间自相关分析与耕地保护分区研究[J]. 土壤通报, 2023, 54(04): 757-767.

基金资助

风云卫星应用先行计划(FY-APP-ZX-2023.01)

山东省气象局创新团队专项(2024sdcxtd04)

长岛国家气候观象台开放基金项目(2024cdkfz04)

山东省自然科学基金(ZR2021MD097)