Enhancing large-scale chlorophyll-a estimation across U.S. Lakes through spectral and spatio-temporal feature integration: a machine learning approach using AquaMatch

Harmful algal blooms (HABs) in inland waters threaten public health, drinking water security, and aquatic ecosystems. Chlorophyll-a (Chl-a) concentration is a critical indicator for bloom risk, yet its continental-scale satellite predictions remain challenging due to the optical complexity and environmental heterogeneity of inland waters. Despite the growing availability of large Landsat satellite–field matchup datasets across U.S. inland waters, Chl-a models trained on these heterogeneous datasets have generally achieved only moderate predictive performance. These large-scale models have relied largely on broadband spectral predictors, although geographic and seasonal variability can also influence phytoplankton dynamics and the relationship between reflectance and Chl-a. We therefore investigated whether integrating spatial and temporal information with Landsat spectral predictors can improve Chl-a estimation using a continental-scale dataset of U.S. inland-water matchups. Using 14,125 quality-controlled matchups from the AquaMatch dataset within a 2-hour satellite–field window and 22 predictors, we compared six machine learning models spanning three broad families: latent-variable regression, neural networks, and tree-based ensembles. The results indicate XGBoost performed best, achieving R 2 = 0.718 and RMSE = 16.02 µg/L on the held-out test set, with stable performance across 35 repeated random train–test splits (R 2 = 0.700 ± 0.018). Its performance remained moderate under repeated spatial tests conducted at the monitoring-station level stratified by hydrologic region (R 2 = 0.577 ± 0.023), which provided a more stringent test of spatial generalization. Complementary predictor-group ablation experiments showed that adding spatial and temporal context to the spectral-only model produced relative R 2 gains of 28.7% under repeated random validation and 65.7% under hydrologic-region spatial validation, with spatial information contributing approximately twice the gain provided by temporal information. SHAP analysis identified latitude as the most influential predictor overall, followed by the Triangular Greenness Index, Red-Blue Index, Surface Water Index, longitude, and cyclical day-of-year terms, highlighting collective contributions from spectral, geographic, and seasonal information. Independent validation in western Lake Erie demonstrated moderate transferability without local recalibration (R 2 = 0.532 for ≤1-hour matchups), although high Chl-a concentrations were increasingly underestimated. These results show substantial relative gains in predictive performance from incorporating readily available geographic and seasonal predictors alongside broadband Landsat spectral information, while highlighting remaining challenges in spatial transfer and accurate retrieval of extreme Chl-a concentrations.