Image enhancement and restoration for maize LAI estimation under low-light conditions

Leaf area index (LAI) is a key indicator of maize canopy structure, but unmanned aerial vehicle (UAV)-based LAI estimation is commonly constrained by narrow solar-noon observation windows. This study evaluated whether low-light image restoration can extend usable UAV observation periods for summer maize LAI estimation. Field-acquired low-light data collected in 2025 and degradation-simulated RGB data generated from 2024–2025 UAV imagery were combined with synchronous LAI measurements. The field-acquired dataset included 161 paired samples of midday RGB, evening low-light RGB (RGB-LL), nighttime thermal infrared (TIR-LL), and ground-measured LAI, while the simulated dataset included 2435 RGB low-light samples across V12, R1, R2, and R3 stages. A Relative Illumination Index was developed in HSV space to classify illumination conditions, and four restoration methods, CLAHE, Restormer, Retinexformer, and ZERO-IG, were compared. Partial least squares regression, support vector regression, random forest, and XGBoost were then used for LAI estimation. Restormer achieved the best restoration quality on the simulated dataset, with PSNR of 39.50–51.37 dB and SSIM of 0.85–0.97. By contrast, CLAHE performed best under field-acquired severe low-light conditions, reaching 28.07 dB/0.71 for restored RGB-LL images and 26.68 dB/0.88 for restored TIR-derived pseudo-color images. CLAHE also showed lower KL divergence and more stable feature-level RGB-channel distribution consistency. For LAI estimation, CLAHE-restored RGB images achieved the best field performance, with R 2 = 0.64 and RMSE = 0.42, and the highest simulated-stage accuracy occurred at R1 using CLAHE-SVR (R 2 = 0.87, RMSE = 0.37). These results indicate that appropriate low-light restoration can support extended-time-window UAV-based maize LAI estimation under specific illumination and data-source conditions.