A major challenge in object-based remote sensing is developing computationally efficient methods to locate individual organisms and observe their characteristics. Here we present a computational framework to detect individual flowering trees and to track their flowering status across arbitrarily large geographic regions using near-daily satellite image data. Our approach includes: 1) Gaussian spot detection to identify candidate flowering trees; 2) a Random Forest (RF) classifier trained to discriminate candidates as flowering trees or other objects; and 3) spatial linking of detected trees across time using the Crocker-Grier algorithm to account for image alignment errors. We implemented this on 7,700 4-band images (1.3 TB) from Planet CubeSats covering a 625 km2 region of Central Panama from 2017-2020. We trained a RF model to identify one canopy tree species, Handroanthus guayacan.
Results/Conclusions
Tracking individual tree flowering status through time recovers a repeated signal of flowering detections that matches the known phenology for H. guayacan in central Panama. A remaining challenge for implementation on larger areas is to distinguish flowering time variation among individual organisms from commission errors that have several possible sources. Our results highlight the potential of object-based satellite remote sensing and the value of high temporal resolution imagery for observing individual-level variation in forest tree phenology, as well as population dynamics, at spatial scales that cannot be attained by other methods.
