Robust Wheat Residue Cover Quantification Under Moisture Variability from ASD Spectroscopy Using Conditional Autoencoder Normalization and Linear Unmixing

Crop residue cover (CRC) plays a crucial role in sustainable farming systems by improving soil structure, regulating water retention, and reducing soil erosion. Accurate CRC monitoring is therefore important for evaluating field management practices at scale. Existing field methods (e.g., line-transect and visual estimation) are labor-intensive and difficult to scale, while optical retrievals are often confounded by soil moisture. Moisture introduces nonlinear spectral distortions that can bias residue estimates, particularly in the shortwave infrared range. We propose a Deep Moisture-Invariant Autoencoder (DMIA) framework that performs conditional spectral normalization—referred to as moisture normalization (dry-equivalent spectral transformation)—before linear spectral unmixing. The workflow has two stages: (1) a conditional autoencoder that transforms moisture-affected spectra to dry-equivalent spectra, and (2) fully constrained linear unmixing on dry-equivalent spectra. The experiment included 63 controlled wheat-residue scenes at a semi-arid site in Morocco, spanning three moisture levels and seven residue proportions (0–100%) measured with ASD spectroscopy. Within this controlled experimental dataset, DMIA achieved a global coefficient of determination of R2 = 0.93, outperforming ordinary least squares (R2 = 0.65), fully constrained least squares (R2 = 0.68), and ELMM (R2 = 0.71), and matching the performance of MESMA (R2 = 0.93) while requiring only a single forward pass at inference rather than iterative library matching. Although both methods showed similar overall accuracy, a closer analysis reveals that DMIA’s advantage over MESMA widens under wetter, coarser-resolution conditions, which are highly representative of operational monitoring. This finding is further validated by a Monte Carlo uncertainty propagation, proving the results are unaffected by reference noise. Using spectrally resampled ground data to simulate satellite responses, performance remained robust for PRISMA (R2 = 0.93) and Sentinel-2 simulation (R2 = 0.87). Reconstruction diagnostics (mean SAM below 5°) support the physical plausibility of the learned transformation. These results suggest that conditional spectral normalization can reduce moisture-related distortions while preserving compositional signals under controlled experimental conditions; however, the use of three discrete moisture levels represents an experimental simplification; in open operational fields, soil moisture varies continuously and pixel-level states are unknown. This framework provides a proof-of-concept basis for further investigation across diverse soils, residue types, and operational sensor configurations.