Enhanced wetland vegetation mapping via smallsat detection of phenological cycles

Accurate remote sensing of wetland vegetation is challenging due to highly heterogenous land cover, morphological similarity of target vegetation classes, phenological cycles and dynamic environmental conditions. The high spatiotemporal resolution of PlanetScope's (PS) constellation of Dove nanosatellites may circumvent some of these hurdles by enabling timely and robust monitoring of key dates of phenological separability and environmental disturbance. PS daily imaging generally results in multiple cloud-free scenes per month, enabling precision targeting of key phenological transition dates. We evaluate PS imagery effectiveness for detecting seven wetland vegetation classes in a complex riverine wetland in Rhode Island, USA over the course of a year based on a single set of training data. Four classification algorithms, Maximum Likelihood (MLC), Support Vector Machine (SVM), K-Means, and Artificial Neural Network (ANN) are tested on individual imagery, monthly-, and multi-seasonal stacked images. Temporally stacked imagery displayed an average increase in Overall Accuracy (OA) of ~9% across all classifiers relative to daily imagery. Multi-seasonal imagery performed best overall, with a October and December image composite outperforming single-month stacked images across most classes for both Producer's and User's Accuracies (PA and UA, respectively). Average PA and UA across all classifiers was generally highest in December, although only three classes, Phragmites, Forest and Marsh Grass, achieved their highest PA during this month, and only one class, Marsh Grass, its highest UA. Other classes achieved their highest PA during spring green-up (April and May) while the best UA were split between green-up and fall senescence. Unsupervised Classification (K-means) performed poorly relative to the supervised classifiers, but was able to accurately identify emergent vegetation over the period of summer green-up, which was missed by the supervised classifiers. These results highlight multi-temporal PS imagery's ability to capitalize on varying phenological cycles between wetland vegetation classes to optimize mapping accuracy.