Thu, Aug 18, 2022: 10:15 AM-10:30 AM
515A
Background/Question/MethodsThe proliferation of satellite imagery has made analyzing deforestation trends easier using multi-source, time-series approaches. In addition to lower temporal latencies, the combination of optical and synthetic aperture radar (SAR) data can be crucial for near real-time (NRT) monitoring in tropical regions with extensive cloud cover, as SAR is able to “see” through clouds. Despite recent progress, there remains little discussion about trade-offs in NRT monitoring and how these should be leveraged for resource managers. In other words, what is the potential for multi-source imaging in an NRT system, and how useful can NRT monitoring be for enforcement?To address these questions, we developed a new NRT method for monitoring deforestation. We identified the extent and timing of 230 disturbances using PlanetScope imagery in Myanmar, before obtaining time-series of Landsat-8 (L8), Sentinel-2 (S2), and Sentinel-1 SAR (S1) surface reflectances or backscatter. We fit time-series models to an assumed stable period, and then combined each sensor’s standardized residuals via a 30-day exponentially weighted moving average, with weights determined by multiobjective optimization. After transforming the weighted values to a daily disturbance probability, we assessed the algorithm by calculating detection latency and true/false positive rates using a probability threshold of 90%.
Results/ConclusionsOur results were dominated by trade-offs between accuracy and latency. For example, prioritizing fast detections (median 1-4 days) led to relatively high false positive rates (median 0.006) and low false negative rates (median 0.2759), whereas prioritizing slow detections gave the opposite (9-12 days, 0.0031, and 0.3103, respectively). We also found that while including S1 decreased latency, it increased the errors by 100% compared to S2 and L8 alone. However, sampling bias towards disturbances during the dry season, when less-noisy L8 and S2 observations are abundant, likely affected this result.Our contributions to ecology are two-fold. First, our multi-source method clearly illustrates how NRT monitoring can be useful for enforcement due to low latencies and our use of probabilities. Instead of relying on binary alerts over weeks, our method allows managers to quickly and efficiently prioritize limited resources. Second, we show the potential of a multi-source NRT system in the context of trade-offs. Though the low detection latencies support better enforcement, we demonstrate the cost from the higher false positive or false negative rates, depending on the management goal. Overall, our findings demonstrate a clear improvement in NRT monitoring of forest disturbances.
Results/ConclusionsOur results were dominated by trade-offs between accuracy and latency. For example, prioritizing fast detections (median 1-4 days) led to relatively high false positive rates (median 0.006) and low false negative rates (median 0.2759), whereas prioritizing slow detections gave the opposite (9-12 days, 0.0031, and 0.3103, respectively). We also found that while including S1 decreased latency, it increased the errors by 100% compared to S2 and L8 alone. However, sampling bias towards disturbances during the dry season, when less-noisy L8 and S2 observations are abundant, likely affected this result.Our contributions to ecology are two-fold. First, our multi-source method clearly illustrates how NRT monitoring can be useful for enforcement due to low latencies and our use of probabilities. Instead of relying on binary alerts over weeks, our method allows managers to quickly and efficiently prioritize limited resources. Second, we show the potential of a multi-source NRT system in the context of trade-offs. Though the low detection latencies support better enforcement, we demonstrate the cost from the higher false positive or false negative rates, depending on the management goal. Overall, our findings demonstrate a clear improvement in NRT monitoring of forest disturbances.