AbstractClimate change is producing shifts in the distribution and abundance of marine species. Such is the case of kelp forests, important marine ecosystem-structuring species whose distributional range limits have been shifting worldwide. Synthesizing long-term time series of kelp forest observations is therefore vital for understanding the drivers shaping ecosystem dynamics and for predicting responses to ongoing and future climate changes. Traditional methods of mapping kelp from satellite imagery are time-consuming and expensive, as they require high amount of human effort for image processing and algorithm optimization. Here we propose the use of mask region-based convolutional neural networks (Mask R-CNN) to automatically assimilate data from open-source satellite imagery (Landsat Thematic Mapper) and detect kelp forest canopy cover. The analyses focused on the giant kelpMacrocystis pyriferaalong the shorelines of southern California and Baja California in the northeastern Pacific. Model hyper-parameterization was tuned through cross-validation procedures testing the effect of data augmentation, and different learning rates and anchor sizes. The optimal model detected kelp forests with high performance and low levels of overprediction (Jaccard’s index: 0.87 ± 0.07; Dice index: 0.93 ± 0.04; over prediction: 0.06) and allowed reconstructing a time series of 32 years in Baja California (Mexico), a region known for its high variability in kelp owing to El Niño events. The proposed framework based on Mask R-CNN now joins the list of cost-efficient tools for long-term marine ecological monitoring, facilitating well-informed biodiversity conservation, management and decision making.