LC-TSRNet: A Land-Cover-Guided Two-Stage Network for Annual MODIS Surface Reflectance Time Series Reconstruction

Wenjuan Zhang; Xuesong Ma; Zhen Li; Meixiang Wang; Caimeng Yi; Jie Pan; Nan Su; Bing Zhang

cs.CVOct 18, 2026

LC-TSRNet: A Land-Cover-Guided Two-Stage Network for Annual MODIS Surface Reflectance Time Series Reconstruction

Authors: Wenjuan Zhang, Xuesong Ma, Zhen Li, Meixiang Wang, Caimeng Yi, Jie Pan, Nan Su, Bing Zhang

Organizations: State Key Laboratory of Remote Sensing and Digital Earth, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100101, China · Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China · College of Resources and Environment, University of Chinese Academy of Sciences, Beijing 100049, China · School of Artificial Intelligence, Xidian University, Xi’an 710071, China ·

Abstract

MODIS surface reflectance (SR) products often suffer from data gaps due to cloud contamination and sensor malfunctions, limiting their applications in long-term and large-scale studies. Most existing methods rely on cloud-free observations from optical time series or introduce SAR time series as auxiliary information. However, due to the coarse spatial resolution of MODIS, cloud-free reference images are difficult to obtain, and SAR imagery is not well matched to the spatial resolution of MODIS, making it difficult to provide effective prior constraints. To address these issues, this study proposes LC-TSRNet for seamless reconstruction of annual MODIS SR products. In the first stage, an annual land cover map is introduced as prior knowledge, and CSPM is designed by leveraging the spectral similarity within identical land cover types, thereby providing prior spectral constraints for missing regions. In the second stage, to address large-area, persistent, and overlapping gaps in MODIS imagery, SMRM is designed to adaptively exploit and integrate multi-scale spatiotemporal contextual information. Meanwhile, to address the limited collaborative utilization of spatio-temporal-spectral information, 3DAM is designed to impose cross-dimensional joint constraints. Considering the lack of publicly available MODIS sample datasets, this study constructs the AIRMOD-TSR dataset. Extensive experimental results show that LC-TSRNet outperforms the comparison methods and achieves better reconstruction accuracy and spectral fidelity. Under the spatially adjacent scenario, LC-TSRNet improves PSNR by 0.56 dB and reduces SAM by 0.13°. The vegetation monitoring application further verifies the practical value of the reconstructed SR time series.

Figures & tables

Fig. 1: Cumulative frequency distributions of annual missing ratios in global MOD09A1 data during 2019–2023. (a) Regional annual mean missing ratio for 512 $\times$ 512-pixel regions. (b) Pixel-level annual missing ratio.

Fig. 2: Overall workflow of data preprocessing.

Fig. 3: Geographic locations and land cover composition of the AIRMOD-TSR ROIs. The dataset consists of 13 globally distributed ROIs (R1–R13). Red and yellow boxes denote high- and low-quality ROIs, respectively.

Region ID	Longitude range	Latitude range	Selected year(s)	Annual mean missing ratio
High-quality ROIs
R1	87.31 ∘ W– 89.61 ∘ W	32.17 ∘ N– 34.47 ∘ N	2019, 2022, 2023	20.78%
R2	80.83 ∘ W– 83.13 ∘ W	32.70 ∘ N– 35.00 ∘ N	2019, 2022, 2023	20.26%
R3	77.64 ∘ W– 79.94 ∘ W	35.32 ∘ N– 37.62 ∘ N	2019, 2020, 2022, 2023	18.70%
R4	93.07 ∘ W– 95.37 ∘ W	35.29 ∘ N– 37.59 ∘ N	2019, 2022, 2023	16.93%
R5	82.69 ∘ W– 84.99 ∘ W	30.83 ∘ N– 33.13 ∘ N	2019, 2022	22.85%

TABLE I: Detailed information on the AIRMOD-TSR ROIs.

Fig. 4: Overall architecture of the proposed LC-TSRNet, consisting of a land-cover-guided coarse representation stage and a spatio-temporal-spectral joint refinement stage.

Fig. 5: Illustration of (a) 3D convolution, (b) Conv block, and (c) GConv block.

Fig. 6: Schematic illustration of the CSPM.

Fig. 7: Illustration of the PFM structure.

Fig. 10: The regional temporal missing ratios and pixel-level annual missing ratios under different scenarios. (a) Spatially adjacent scenario (R9). (b) Cross-hemisphere scenario (R10). (c) Persistent and overlapping missing scenario (R11).

Fig. 11: Time series quantitative evaluation of the reconstruction performance of different methods under the spatially adjacent scenario (R9). The shaded area represents the missing ratio for each composite. (a) PSNR. (b) SSIM. (c) CC. (d) SAM.

Methods	PSNR (dB) $\uparrow$	SSIM $\uparrow$	CC $\uparrow$	SAM ( ∘ ) $\downarrow$
WLR (2013)	38.3745	0.9130	0.9865	1.8680
MNSPI (2012)	39.1155	0.9251	0.9884	1.5112
PSTCR (2020)	31.1337	0.8593	0.9607	2.1439
U-TILISE (2023)	32.4606	0.8711	0.9737	2.1982
TAMGAN (2025)	37.5036	0.9386	0.9891	1.4871
MS 2 TAN (2025)	38.9208	0.9401	0.9878	1.9728

TABLE II: Quantitative evaluation of the time series reconstruction performance of different methods under the spatially adjacent scenario (R9). Values are averaged across all 46 composites, with the best and second-best results indicated by boldface and underlining , respectively.

Fig. 12: Visual comparison of reconstruction results on three representative dates under the spatially adjacent scenario (R9). Standard false-color composites are shown using B2, B1, and B4. For each date, the complete scene is shown in the upper row, the 80 $\times$ 80-pixel subregion marked by the yellow box is enlarged in the middle row, and the corresponding RMSE map of the enlarged region is presented in the lower row.

Fig. 13: Band-wise RMSE analysis of reconstruction results in the spatially adjacent scenario (R9). (a) Annual mean RMSE of different methods across the seven bands. (b) Time series RMSE curves of LC-TSRNet for each band.

Fig. 14: Comparison of reconstructed SR curves for typical land cover types derived from the R9 image acquired on 20220906. The spatial locations of selected samples are also shown. (a) Deciduous broadleaf forests. (b) Mixed forests. (c) Croplands. (d) Urban and built-up lands. (e) Water bodies. (f) Spatial locations of selected samples.

Fig. 15: Comparison of reconstructed time series SR curves over vegetated areas. Blue shaded areas indicate dates with missing observations. (a) Deciduous broadleaf forests. (b) Croplands.

Methods	PSNR (dB) $\uparrow$	SSIM $\uparrow$	CC $\uparrow$	SAM ( ∘ ) $\downarrow$
WLR (2013)	35.8984	0.8808	0.9821	2.1624
MNSPI (2012)	37.0999	0.9097	0.9867	1.5321
PSTCR (2020)	28.8164	0.8243	0.9526	2.2271
U-TILISE (2023)	32.4131	0.8670	0.9776	2.1690
TAMGAN (2025)	34.9539	0.9196	0.9864	1.5806
MS 2 TAN (2025)	35.8663	0.9050	0.9840	2.1591

TABLE III: Quantitative evaluation of the time series reconstruction performance of different methods under the cross-hemisphere scenario (R10). Values are averaged across all 46 composites, with the best and second-best results indicated by boldface and underlining , respectively.

Fig. 16: Time series quantitative evaluation of the reconstruction performance of different methods under the cross-hemisphere scenario (R10). The shaded area represents the missing ratio for each composite. (a) PSNR. (b) SSIM. (c) CC. (d) SAM.

Fig. 17: Visual comparison of reconstruction results on three representative dates under the cross-hemisphere scenario (R10). Standard false-color composites are shown using B2, B1, and B4. For each date, the complete scene is shown in the upper row, the 80 $\times$ 80-pixel subregion marked by the yellow box is enlarged in the middle row, and the corresponding RMSE map of the enlarged region is presented in the lower row.

Methods	PSNR (dB) $\uparrow$	SSIM $\uparrow$	CC $\uparrow$	SAM ( ∘ ) $\downarrow$
WLR (2013)	30.5794	0.8541	0.9586	1.9029
MNSPI (2012)	32.1557	0.8524	0.9757	1.9190
PSTCR (2020)	24.4754	0.6779	0.8886	3.2217
U-TILISE (2023)	28.0706	0.7981	0.9528	2.9854
TAMGAN (2025)	30.8160	0.8734	0.9745	1.9762
MS 2 TAN (2025)	32.0039	0.8875	0.9749	1.8996

TABLE IV: Quantitative evaluation of the time series reconstruction performance of different methods under the persistent and overlapping missing scenario (R11). Values are averaged across 8 composites, with the best and second-best results indicated by boldface and underlining , respectively.

Fig. 18: Visual comparison of reconstruction results on three representative dates under the persistent and overlapping missing scenario (R11). Standard false-color composites are shown using B2, B1, and B4. For each date, the complete scene is shown in the upper row, the 80 $\times$ 80-pixel subregion marked by the yellow box is enlarged in the middle row, and the corresponding RMSE map of the enlarged region is presented in the lower row.

Methods				R9				R10
Land Cover	CSPM	SMRM	3DAM	PSNR (dB) $\uparrow$	SSIM $\uparrow$	CC $\uparrow$	SAM ( ∘ ) $\downarrow$	PSNR (dB) $\uparrow$	SSIM $\uparrow$	CC $\uparrow$	SAM ( ∘ ) $\downarrow$
$\times$	$\times$	✓	✓	38.6871	0.9465	0.9901	1.4707	31.2672	0.8976	0.9582	2.2841
✓	$\times$	✓	✓	39.0090	0.9490	0.9909	1.4496	33.3313	0.9079	0.9828	1.9829
✓	✓	$\times$	✓	39.4310	0.9474	0.9912	1.4902	37.1209	0.9310	0.9887	1.5843
✓	✓	✓	$\times$	38.9883	0.9456	0.9906	1.4884	37.0544	0.9331	0.9889	1.5165
✓	✓	✓	✓	39.6724	0.9519	0.9915	1.3548	37.4161	0.9348	0.9894	1.4371

TABLE V: Quantitative ablation study of LC-TSRNet under the spatially adjacent scenario (R9) and cross-hemisphere scenario (R10). Values are averaged across all 46 composites, with the best results indicated by boldface .

Fig. 19: Visual comparison of ablation results for the land cover prior and CSPM in the spatially adjacent scenario (R9) and cross-hemisphere scenario (R10). (a) R9-20220322. (b) R9-20221203. (c) R10-20220202. (d) R10-20221024.

Fig. 20: Visual comparison of masked real observation time series and LC-TSRNet reconstruction results in the real-world scenario (R12). (a) Masked real observation time series. (b) Reconstruction results of LC-TSRNet.

Fig. 21: Visual comparison of observed and reconstructed SR and NDVI time series in the application region (R13). (a) Observed time series. (b) Reconstruction results of LC-TSRNet.

Fig. 22: Time series NDVI curves for two sample points and their spatial locations. (a) Point A–croplands. (b) Point B–mixed forests. (c) Spatial locations and land cover information of two points.

Remote Sensing

Remote sensing uses satellite, aerial, and ground-based observations to measure and map Earth and other planetary surfaces.

See all 528 papers

HLS-GPT: A Generative Pretrained Transformer (GPT) for Continental-Scale NASA Harmonized Landsat and Sentinel-2 (HLS) Reflectance Reconstruction Across All Bands on Arbitrary Dates
Jun 16, 2026Junjie Li, Hankui K. Zhang, David P. RoyTransformer Remote Sensing
Geospatial Sciences Center of Excellence, Department of Geography and Geospatial Sciences, South Dakota State University, Brookings, SD 57007, USA · Department of Geography, Environment, and Spatial Sciences, & Center for Global Change and Earth Observations, Michigan State University, MI 48824, USA
Fast Fourier Convolutional GAN for 30 m Clear-Sky Land Surface Temperature Gap-Free Reconstruction
Jul 22, 2026Marwa Alfouly, Smajil Halilovic, Nils Bochow +3Remote Sensing Image Inpainting
Technical University of Munich, Chair of Renewable and Sustainable Energy Systems, Germany · Helmholtz Centre for Polar and Marine Research, Alfred Wegener Institute, Germany · Technical University of Munich, Professorship of Earth System Modelling, Germany +1
Global-Scale Self-Supervised Spatiotemporal Learning for NDVI Time-Series Reconstruction
Aug 3, 2026Ang Li, Menghui Jiang, Xiaobin Guan +2Time Series Imputation Self-Supervised Learning
School of Resource and Environmental Sciences, Wuhan University, Wuhan 430079, China · School of Geography and Tourism, Anhui Normal University, Wuhu 241002, China · Key Laboratory of Earth Surface Processes and Regional Response in the Yangtze River Basin, Wuhu 241002, China +2

Jun 16, 2026J/K move · Enter open · S save

HLS-GPT: A Generative Pretrained Transformer (GPT) for Continental-Scale NASA Harmonized Landsat and Sentinel-2 (HLS) Reflectance Reconstruction Across All Bands on Arbitrary Dates

Junjie Li, Hankui K. Zhang, David P. Roy

Geospatial Sciences Center of Excellence, Department of Geography and Geospatial Sciences, South Dakota State University, Brookings, SD 57007, USA · Department of Geography, Environment, and Spatial Sciences, & Center for Global Change and Earth Observations, Michigan State University, MI 48824, USA

Transformer Remote Sensing

Open paper AskPDF arXiv

Recent deep learning methods for Landsat and Sentinel-2 reflectance time series reconstruction remain limited by restricted spectral coverage, limited geographic scalability, or patch-based designs with short temporal contexts. We present HLS-GPT, a large-scale generative pretrained Transformer model for reconstructing NASA Harmonized Landsat Sentinel-2 30 m surface reflectance for all bands, any date, and any pixel location. HLS-GPT uses a hierarchical Transformer architecture to handle the different spectral band configurations of Landsat and Sentinel-2 and operates on single-pixel 12-month time series. To capture geographic and seasonal variability, the model was trained with nine years of HLS time series from more than 0.25 million training pixels across the conterminous United States. A random cropping and masking strategy extracts 12-month periods with varying start dates across epochs, masks 50% of valid observations, and trains the model to reconstruct the masked reflectance values from the remaining observations. Evaluation using more than 62,000 independent test pixels shows robust reconstruction under diverse land surface conditions, including complex crop phenology and sparse, irregular observations. Leave-one-observation-out evaluation achieved reconstruction RMSE below 0.026 for all HLS spectral bands, with relative RMSE below 35% for visible bands and below 13% for other bands. Red-edge band errors were comparable to red and near-infrared errors despite the absence of red-edge bands on Landsat. Sensitivity analyses that randomly masked 10% to 90% of test observations showed only modest degradation when 10% to 50% of observations were masked, with all-band RMSE below 0.028. Image reconstruction over nine independent 109 by 109 km CONUS HLS tiles further demonstrates that HLS-GPT outperforms two conventional methods and the NASA-IBM Prithvi model.

Abstract

Remote Sensing

Explore similar work

HLS-GPT: A Generative Pretrained Transformer (GPT) for Continental-Scale NASA Harmonized Landsat and Sentinel-2 (HLS) Reflectance Reconstruction Across All Bands on Arbitrary Dates

Fast Fourier Convolutional GAN for 30 m Clear-Sky Land Surface Temperature Gap-Free Reconstruction

Global-Scale Self-Supervised Spatiotemporal Learning for NDVI Time-Series Reconstruction