Automated Object-based Remote Sensing of Individual Tree Phenology in Near-daily Cubesat Data

High-throughput detection and phenotyping of individual plants across natural populations is a frontier in satellite remote sensing with applications in biodiversity monitoring and research. Cubesats now acquire Earth observations with spatiotemporal resolution fine enough to resolve individual trees and characterize trait variation over time and across populations - at least for some model species. But, across disciplines, cubesats are rarely used for object detection and tracking due to radiometric inconsistencies, spatial errors, and the lack of computational frameworks to turn these data into organism-level observations. Here we develop a generalized approach to object detection and tracking in high-resolution image data. We demonstrate the algorithm using 11,582 scenes in a 2016 - 2022 time series from Planet cubesats in a 759 km2 area of central Panama, where we use the algorithm to detect individual flowering trees in the Neotropical canopy tree species Handroanthus guayacan. The algorithm requires four automated steps: (1) initial object detection; (2) removal of false-positive errors; (3) spatial alignment of unique objects detected on different dates; and (4) quantifying cloud obscuration. We use a Gaussian spot detector for initial object detection and remove false positive errors by training a Random Forest classifier to discriminate false-positive errors based on human-labeled examples. We use the Crocker-Grier algorithm to overcome geolocation error among scenes, and identify cloud obscuration using a band threshold. Our analysis classified 118,337 objects as flowering H. guayacan, corresponding to 16,853 unique individuals detected 2-24 times during the 5-year time series, observable without cloud obscuration every 3.8 days on average. Repeated detections of the same individuals demonstrate that the algorithm recovered the expected flowering phenology in this species. We show that although flowering synchrony is generally high, inter-annual variability exists and the propensity for synchronicity decreases with distance between individuals. The ability to census and phenotype trees using cubesats has implications for phenology monitoring, distribution modeling, demography and understanding the basis of trait variation.