MCVI-SANet: VI Saturation Aware Crop Monitoring

• The paper introduces a lightweight semi-supervised CNN that accurately estimates LAI and SPAD by addressing VI saturation challenges.
• It employs a VI-SABlock with dual attention mechanisms and inverted residual blocks to enhance feature representation in dense canopy conditions.
• The integration of VICReg-based self-supervised learning and vegetation height-guided stratification yields up to 8.95% improvement in R² over baseline models.

The Multi-Channel Vegetation Indices Saturation Aware Net (MCVI-SANet) is a lightweight semi-supervised convolutional network designed for precision estimation of leaf area index (LAI) and soil–plant analysis development (SPAD) values in winter wheat. It addresses critical challenges posed by vegetation index (VI) saturation in dense canopy conditions and sparse ground-truth annotation, deploying a combination of channel-spatial feature enhancement, domain-informed stratification, and modern contrastive self-supervision to deliver state-of-the-art accuracy with high efficiency (Zhang et al., 20 Dec 2025).

1. Network Architecture and Data Flow

MCVI-SANet processes multi-channel vegetation index imagery, specifically leveraging 11 spectral VIs, including NDVI, EVI, SAVI, GNDVI, NDRE, MCARI, OSAVI, RVI, DVI, CI$_{\text{green}}$, and VARI, each represented as 192×192 pixel spatial tiles. The architecture comprises three principal stages:

• Vegetation Index Saturation-Aware Block (VI-SABlock): Initial normalization and dual attention modulation to produce enhanced features resilient to VI saturation.
• Inverted Residual Backbone (IRBs): Twelve staged blocks (IRB$_{1-12}$) maintain high spatial resolution (96×96), escalate feature dimension to 96 channels, and enable efficient representation learning with minimal parameters.
• Regressor Heads: Adaptive average pooling yields 96-dimensional vectors, which are processed by bi-layer MLPs (96→32→1) for LAI or SPAD output per plot.

Major convolutional details and total parameterization (0.10M overall, see Table below) ensure MCVI-SANet remains highly efficient, enabling real-time inference on edge devices.

Network Component	Channels (in/out)	Parameter Order (count)
VI-SABlock	11 → 11	~1,000
IRB (1–4)	11/64 → 64	~13,000 each
IRB (5–12)	64/96 → 96	~16,000 each
Regression Head	96→32→1	3,100 + 33

2. Vegetation Index Saturation-Aware Block (VI-SABlock)

The VI-SABlock specifically addresses feature collapse caused by VI saturation, typical when NDVI exceeds 0.8 and traditional mean-based indices lose discriminability. Its key innovations include:

• Multi-Index Statistics Fusion: Both channel-wise means ($\mu_c$) and standard deviations ($\sigma_c$) are computed and concatenated, capturing first- and second-order distributional properties over input tiles.
• Dual Attention Mechanisms:
  • Channel Attention (FRE): Utilizing learned channel descriptors via Mish activation and excitation layers, generating adaptive weights for each VI channel.
  • Spatial Attention (DSAM): Depthwise 3×3 convolutions followed by tanh activations highlight spatial texture irregularities, crucial for distinguishing canopy states under saturation.
• Residual Modulation: Output features are recalibrated spatially and additively, followed by channel expansion and downsampling for efficient backbone integration.

When NDVI is saturated, channel-wise STD retains a coefficient of variation of 76.3%, whereas the mean drops to 4.9%, confirming the block's capacity to preserve feature discriminability in closed-canopy scenarios.

3. VICReg-Based Semi-Supervised Learning

Integration of VICReg-based self-supervised pretraining capitalizes on weakly labeled data, providing robust and domain-adaptive representations. Key elements:

• Contrastive Pretraining: 2,700 unlabeled 5-band images are augmented (flip, rotate) and fed through the encoder-projector stack, which enforces invariance ($L_{inv}$), non-collapse ($L_{var}$), and decorrelation ($L_{cov}$) constraints on projected embeddings.
• Loss Formulations:

  $$L_{\mathrm{VICReg}} = \lambda_{\mathrm{inv}} L_{\mathrm{inv}} + \lambda_{\mathrm{var}} L_{\mathrm{var}} + \lambda_{\mathrm{cov}} L_{\mathrm{cov}}$$

  with $\lambda_{inv}=25$, $\lambda_{var}=25$, $\lambda_{cov}=1$, $\epsilon=10^{-4}$.

• Fine-Tuning Regimen: Upon completion, the encoder is frozen, and regression heads are trained on 330 labeled samples under LR scheduling and early stopping. The LogME score selects the best checkpoint for transfer.

This semi-supervised approach enables MCVI-SANet to generalize across phenological stages and underlying domain shifts, outperforming purely supervised or conventional baselines.

4. Vegetation Height-Guided Data Partitioning

To ensure experimental rigor and maximize representativeness across growth stages, MCVI-SANet employs a vegetation height (VH)-informed stratified clustering process:

• Clustering Methodology: K-means (K=10, $N>5$ repetitions) is applied to normalized feature vectors ([LAI, SPAD, VH]), followed by majority vote cluster assignments.
• Train/Val/Test Split: Each cluster is partitioned (9:1:1), maintaining statistical parity across splits (mean, std for LAI/SPAD) and reduced distribution divergence metrics (MMD=1.3e−3, JS=0.557, $CV_{across stages}=0.220$).

This approach ensures unbiased evaluation, controls for phenological diversity, and mitigates distribution shift that impairs traditional ML and DL pipelines.

5. Experimental Evaluation and Quantitative Performance

Experiments leveraged Intel Core i9-14900K CPUs and dual NVIDIA RTX 3090 GPUs under Windows 11/Python 3.11/PyTorch, utilizing Optuna for baseline ML hyperparameter search. Comparative results across 10 repeated runs demonstrate MCVI-SANet's efficacy:

Model	LAI R²	LAI RMSE	SPAD R²	SPAD RMSE
SVR (VIs+TFs)	0.7278	0.5782	0.4739	3.1297
ResNet18	0.7456	0.5585	0.4995	3.0449
MCVI-SANet (supervised)	0.8070	0.4861	0.6520	2.5427
MCVI-SANet (semi-supervised)	0.8123	0.4796	0.6846	2.4222

Semi-supervised MCVI-SANet achieves an average improvement of 8.95% (LAI R²) and 8.17% (SPAD R²) over best-performing baselines, attaining maximum single-run scores of LAI R²=0.8619 (RMSE=0.4118) and SPAD R²=0.7756 (RMSE=2.0442).

MCVI-SANet's complexity (0.10M parameters, 0.46 MB, CPU throughput 58.64 samp/s) contrasts favorably against ResNet50 (23.6M, 90.33 MB, 31.04 samp/s).

6. Integration of Agronomic Priors and Limitations

Key agronomic domain priors are embedded throughout MCVI-SANet’s workflow:

• Statistical Exploitation: VI-SABlock’s use of channel-wise STD counteracts saturation-induced loss of mean discriminability.
• Phenological Representation: VH-guided partitioning ensures phenological spectra in winter wheat datasets are faithfully represented, repressing distribution shift.
• Self-Supervised Invariance: VICReg training yields non-collapsed, generalizable embeddings robust against seasonal and environmental variance.

Notable limitations remain: performance under stress conditions, other crop types, and modalities (thermal, SAR, time series) are unexplored. Future model improvements include advanced compression (NAS, distillation) and broader semi-supervised scaling to support deployment on resource-constrained UAV platforms.

The fusion of VI saturation-aware attention, semi-supervised VICReg representation learning, and phenology-informed stratification establishes MCVI-SANet as a domain-adaptive and resource-efficient system for remote sensing-driven precision agriculture (Zhang et al., 20 Dec 2025).