Su Ye , Zhe Zhu, Ji Won Suh

Near real time (NRT) monitoring of land disturbances holds great importance for delivering emergency aids, mitigating negative social and ecological impacts, and distributing resources for disaster recovery. Many past NRT techniques were built upon examining the overall change magnitude of a spectral anomaly with a pre-defined threshold, namely the unsupervised approach. However, their lack of fully considering spectral change direction, change date and pre-disturbance conditions often led to low detection sensitivity and high commission errors, especially when only a few satellite observations were available at the early disturbance stage, which could eventually result in a longer lag to produce a reliable disturbance map. For this study, we developed a novel supervised machine learning approach guided by historical disturbance datasets for accelerating land disturbance monitoring. This new approach first applied retrospective analysis based on historical Harmonized Landsat Sentinel-2 (HLS) datasets from 2015 to 2021 and several open disturbance products, in which various multifaceted change related predictors were extracted from satellite time series, followed by separate disturbance model construction for each consecutive anomaly number. Then, these models were applied for NRT prediction with a per-pixel disturbance probability with new observations (e.g., 2022 HLS images) ingested incrementally on a weekly basis. We developed this operational NRT system incorporating both unsupervised and supervised approaches. Latency and accuracy were evaluated against 3,000 samples randomly selected from five most influential disturbance events of United States in 2022 based on labels and disturbance dates interpreted from daily PlanetScope images. The evaluation showed that the supervised approach required 15 days (since the start of the disturbance event) to reach the plateau of its F_1 curve (where disturbances are detected with high confidence), seven days earlier with roughly 0.2 F_1 score improvement compared to the unsupervised approach (0.733 vs. 0.546 F_1 score). The further analysis showed the improvement was mainly due to the substantial decrease of commission errors (17.7% vs 44.4%). The latency component analysis indicated that the supervised approach only took an average of 4.1 days to yield the first disturbance alert at its fastest alerting speed when the NRT platform made a daily update. This finding highlighted the importance of past knowledge and machine learning for accelerating a NRT monitoring task.