AMGAN: A multimodal generative adversarial network for near-daily alfalfa multispectral image reconstruction

Abstract

Accurate and temporally consistent multispectral observations are essential for monitoring alfalfa yield and quality, given its frequent harvest cycles and rapid regrowth. However, optical satellite imagery is often constrained by cloud cover, revisit intervals, and sensor availability. To overcome these limitations, we propose a novel Alfalfa Multimodal Generative Adversarial Network (AMGAN) designed for near-daily multispectral image reconstruction. Unlike conventional image-to-image or spatiotemporal fusion methods that overlook crop-specific characteristics, are restricted to observed timestamps, or depend heavily on dense temporal series, AMGAN leverages multisource (Landsat-8/9, Sentinel-1, PlanetScope) and multimodal (climate, geographic, temporal) information within an adversarial learning paradigm. This enables high-quality image generation from minimal inputs. Extensive experiments across five major alfalfa-producing states in the United States (2022-2024) show that AMGAN consistently surpasses four state-of-the-art (SOTA) deep learning baselines. It achieves higher reconstruction accuracy across all spectral bands, with pronounced gains in red-edge and near-infrared (NIR) regions critical for vegetation assessment. Multisource integration and multimodal cues enhance robustness, ensuring reliable performance under diverse observation scenarios. The reconstructed imagery was subsequently evaluated in alfalfa yield and quality prediction tasks. Results demonstrated high predictive accuracy for dry matter yield (DM) in the cross validation (CV) experiment with a coefficient of determination (R²) of 0.80, and moderate correlations for selected quality traits such as crude protein (CP), non-fiber carbohydrates (NFC), and minerals, while nutritive value traits tied to complex biochemical processes remained more challenging. Overall, this study underscores the potential of multimodal adversarial learning to bridge observational gaps in alfalfa monitoring. The proposed framework provides a scalable, crop-specific approach for generating temporally dense imagery, supporting precision management for biomass-related and proximate quality traits, while performance for digestibility traits remains limited.

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.