Leaf Area Index (LAI): Retrieval & Imputation

• LAI is defined as the one-sided area of green leaves per unit ground area, serving as a key indicator of canopy structure and photosynthetic capacity.
• Remote sensing using optical sensors and SAR, combined with advanced imputation models like BiLSTM, enables robust LAI estimation despite data gaps.
• Hybrid approaches using physical modeling and machine learning (e.g., VHGPR) improve retrieval accuracy and provide spatially explicit uncertainty measures.

The Leaf Area Index (LAI) is a foundational biophysical parameter in remote sensing, agronomy, and ecological modeling. It represents the one-sided green leaf area per unit ground surface area, typically expressed in $\mathrm{m}^2\,\mathrm{m}^{-2}$. LAI underpins quantitative characterization of crop canopy structure, photosynthetic capacity, and seasonal growth dynamics. For winter wheat and similar crops, timely and continuous LAI monitoring informs key phenological transitions—from emergence and rapid green-up through biomass maximization to senescence—and is critical for accurate crop-yield prediction and model assimilation. Persistent cloud cover or acquisition constraints in optical remote sensing result in gaps in LAI time series, which degrade the quality of yield forecasts and biophysical analyses. To address this, recent research has focused on multi-source data fusion and advanced imputation techniques, including bidirectional recurrent neural networks and kernel-based retrievals, that enable robust and reliable inference of LAI even under incomplete observational scenarios (Zhao et al., 2023, Estevez et al., 2020).

1. Definition and Biophysical Relevance

The Leaf Area Index is defined as the total one-sided area of green leaves per unit ground area, $\mathrm{m}^2\,\mathrm{m}^{-2}$. LAI serves as a primary quantitative measure of canopy architecture, influencing radiative transfer, photosynthetic uptake, light interception, energy, and water exchanges within terrestrial ecosystems. In agricultural applications, LAI time series precisely track crop development, providing critical input to crop-growth models such as DSSAT and WOFOST, and support frameworks for improved yield estimation by informing model states such as light interception and biomass accumulation (Zhao et al., 2023).

2. Remote Sensing Modalities for LAI Estimation

2.1 Optical Data

Sentinel-2 MSI provides high spatial (10/20 m) and temporal (5-day revisit) resolution optical imagery, which—following atmospheric correction—can be used to retrieve LAI via physically-based inversion of radiative-transfer models such as PROSAIL (Estevez et al., 2020). For LAI retrieval, reflection spectra after atmospheric effects are modeled using PROSAIL (which couples PROSPECT-4 and 4SAIL) and then mapped to LAI either directly (at bottom-of-atmosphere reflectance) or indirectly (via top-of-atmosphere radiance, using coupled radiative transfer models such as 6SV for atmospheric correction).

2.2 Synthetic Aperture Radar (SAR)

Sentinel-1 C-band SAR data provide all-weather, day-and-night imaging, unaffected by cloud cover. The cross-polarized (VH) to co-polarized (VV) ratio—after multitemporal speckle filtering—has demonstrated strong (often inverse) correlation with time series LAI in winter wheat, with Pearson’s $r$ typically $> 0.8$ during canopy green-up and early maturity periods. This enables VH/VV as an effective proxy for augmenting or imputing missing LAI values when optical data are unavailable (Zhao et al., 2023).

3. Data Processing and Time Series Imputation Strategies

3.1 Preprocessing Pipelines

• Sentinel-2: Top-of-atmosphere reflectances undergo atmospheric correction (e.g., SIAC), followed by LAI retrieval using inverse radiative-transfer emulators at 10 m resolution. Only scenes with low cloud-cover scores are retained.
• Sentinel-1: VH and VV GRD IW-mode images are subjected to multitemporal speckle reduction (Quegan & Yu, 2001) to obtain noise-free backscattered intensities, after which the VH/VV ratio is computed. Temporal alignment is enforced via interpolation for baselines, while LSTM/BiLSTM networks handle missingness using binary indicators, per Cao et al. (2018).

3.2 Sequence Imputation Using Deep Neural Networks

Bidirectional Long Short-Term Memory (BiLSTM) networks are used to model and impute LAI time series. At each timestep, input vectors consist of the normalized VH/VV ratio and the partially observed LAI. BiLSTM exploits both forward and backward temporal dependencies, with hidden state size 60 per direction and a fully connected layer (size 50) followed by dropout ($p=0.5$) for regularization. Final output predicts both LAI and VH/VV (primarily LAI). Missingness is explicit in the input encoding (Zhao et al., 2023).

The loss at each timestep is the half mean squared error (MSE): $$\mathcal{L}_t = \frac{1}{2} \sum_{j=1}^R (y_{t,j} - \hat{y}_{t,j})^2,$$ where $y_t$ is the true response and $\hat{y}_t$ is the network prediction, with $R=2$.

4. Retrieval via Physical Modeling and Bayesian Regression

A parallel stream of LAI retrieval research employs physical radiative transfer coupled with Gaussian Process Regression (GPR) and its heteroscedastic extension (VHGPR):

• Lookup Table (LUT) Generation: Coupled PROSAIL (for canopy) and 6SV (for atmosphere) models simulate a LUT of expected TOA radiances or BOA reflectances across sampled biophysical and atmospheric parameters.
• Hybrid Regression: The LUT provides the training set for GPR or VHGPR, mapping spectral features to LAI. VHGPR introduces an input-dependent noise process to more accurately capture retrieval uncertainty.
• Direct TOA Retrieval: GPR and VHGPR enable LAI estimation directly from Sentinel-2 L1C radiance data, circumventing the atmospheric correction step when using models trained to jointly span surface and atmosphere variability (Estevez et al., 2020).

5. Cross-Modal Imputation Performance and Comparisons

Performance is validated by root mean squared error (RMSE) against in-situ-calibrated Sentinel-2 LAI at independent field points. In Hengshui (North China Plain) March/May campaigns (Zhao et al., 2023):

Method	RMSE (March)	RMSE (May)
BiLSTM	0.68	1.7855
Unidirectional LSTM	0.5912	2.2886
Polynomial reg.	0.9339	2.4088
Exponential reg.	0.9361	2.6240

During green-up, all methods perform comparably, with BiLSTM slightly better. In senescence, BiLSTM surpasses unidirectional LSTM by $\sim$0.50 LAI units and regression models by $\mathrm{m}^2\,\mathrm{m}^{-2}$00.6–0.8 units. Although $\mathrm{m}^2\,\mathrm{m}^{-2}$1 is not explicitly reported, visual and RMSE evidence implies $\mathrm{m}^2\,\mathrm{m}^{-2}$2 for BiLSTM across the season.

For direct retrieval with GPR/VHGPR, validation against 114 Marchfeld field samples (Estevez et al., 2020) yields (approximate):

Model	RMSE ($\mathrm{m}^2\,\mathrm{m}^{-2}$3)	$\mathrm{m}^2\,\mathrm{m}^{-2}$4
GPR (TOA)	0.66	0.77
VHGPR (TOA)	0.62	0.80
GPR (BOA)	0.70	0.78
VHGPR (BOA)	0.63	0.80

Biases in all models are $\mathrm{m}^2\,\mathrm{m}^{-2}$5 $\mathrm{m}^2\,\mathrm{m}^{-2}$6.

6. Robustness, Generalization, and Uncertainty Quantification

BiLSTM demonstrates robustness to Sentinel-2 data sparsity, maintaining stable imputation when only 8–12 LAI observations are available, leveraging bidirectional temporal context. Unidirectional LSTM underperforms in late-season inference, lacking forward temporal information. Missing data is handled as learnable graph nodes, aligning with the approach in BRITS (Cao et al. 2018), enabling broad applicability to other multi-sensor or clinical/financial time series (Zhao et al., 2023).

VHGPR provides spatially-varying, per-pixel uncertainty maps by modeling input-dependent noise, quantified by standard deviation $\mathrm{m}^2\,\mathrm{m}^{-2}$7 and coefficient of variation $\mathrm{m}^2\,\mathrm{m}^{-2}$8. Uncertainty is low in mid-LAI croplands and increases for bare-soil or atypical surfaces, enabling user-directed confidence masking (Estevez et al., 2020).

7. Broader Implications and Applicability

The described approaches represent current state-of-the-art for LAI imputation and retrieval. BiLSTM methodology is generalizable to other time-series imputation tasks involving multi-modal or incomplete remote sensing datasets, and can be directly applied to gap-filling in NDVI and microwave indices or incomplete clinical and financial records. The hybrid LUT-driven machine learning paradigm for LAI retrieval is broadly applicable across variable atmospheric conditions, with the VHGPR variant yielding robust spatially-explicit uncertainty estimates suitable for automated processing pipelines (Zhao et al., 2023, Estevez et al., 2020).