Multisource Assessment of Machine Learning Models and Atmospheric Correction Products for Chlorophyll-a Retrieval in Meso-Eutrophic Waters: Application to Hamilton Harbour, an Area of Concern in Lake Ontario, Canada

Freely available multispectral satellite imagery enables frequent monitoring of lake Chlorophyll-a(Chl-a), but retrieval accuracy depends strongly on both atmospheric correction (AC) and model choice. Using more than 600 pixel–in-situmatchups (2000–2023) from more than 240 Landsat-5/7/8 and Sentinel-2 imagery in western Lake Ontario (±4-day window; 3×3 pixel means), we benchmarked four radiometric products (Level-1, Level-2, DOS, ACOLITE) with four machine learning models (LASSO, MDN, SVR, XGBoost). ACOLITE paired with XGBoost was consistently best on the held-out test sets: RMSLE = 0.34 (L5), 0.43 (L7), 0.56 (L8), and 0.42 (S2), with slopes ~0.7, 0.6, 0.4, and 0.7, respectively. Minimally corrected products (Level-1, DOS1) performed poorly (test RMSLE ≳ 0.6; slopes ≲ 0.1). For Landsat-8, the Provisional Aquatic Reflectance generalized better than LaSRC (test slope ≈ 0.6 vs 0.1; RMSLE 0.59 vs 0.68), but ACOLITE remained strongest overall. Against semi-empirical indices, ML—especially XGBoost—reduced errors and produced fits closer to unity across sensors. We then mapped seasonal Chl-ain Hamilton Harbour (2000–2024) to showcase model application. Monthly means rose from late spring to midsummer (May 13.2 μg L⁻¹; June 11.5; July 19.2; August 20.2), with persistent nearshore hotspots near the Royal Botanical Gardens and Windermere Basin. Mann–Kendall trend analyses during bloom months indicated little monotonic change: mean Sen's slopes were near zero (|mean| ≤ 2×10⁻⁴ μg L⁻¹ day⁻¹) andamapping, demonstrated here for a meso-eutrophic embayment of Western Lake Ontario.