We Can’t Have It Both Ways - Accepting the Trade-off of Detection Time and Accuracy in Multi-source, Near Real-time Deforestation Monitoring
Abstract
Monitoring deforestation in near real-time continues to be a high priority for remote sensing science. The proliferation of Earth-observing satellites, and the emergence of time series approaches, promises higher accuracy detections with lower latency. But the potential trade-off between optimizing detection accuracy and time has been underexplored, and combining disparate information from multiple sensors remains challenging. We devised a straightforward approach to monitoring deforestation using multiple satellite image time series, and used multiobjective optimization to answer the following questions':' Do multi-source approaches give higher accuracy, lower latency change detections compared to single-source? And, can a single approach optimize detection time and accuracy, or are there trade-offs? We used PlanetScope image sequences to create a validation set of deforestation events with high temporal accuracy in north-central Myanmar (n=159). Similar to CCDC, we estimated time series models from Landsat 8 and Sentinel-2 NDVI time series, and VV and VH backscatter from Sentinel-1. Standardized residuals were combined across sensors, and aggregated using a temporal weighting scheme before converting to a deforestation probability using logistic regression. We optimized for detection time, alongside Type 1 and II errors across a range of weighting parameters. Multi-source approaches were superior to any single-source approach, with both detection times and error rates being twice as good. We also show clear trade-offs between detection time and error rate':' weighting scenarios with the fastest detections (median value of 2 days) had higher false positive (0.015) and false negative rates (0.23). Thus, we recommend that further developments focus on using multi-source time series to increase accuracy and reduce detection time. Users should also be aware of the inherent trade-offs in detection latency and accuracy, and be able to prioritize these differently across applications.
