A First National Seagrass Map of Venezuela

Kurzfassung

Coastal wetland habitats including seagrass meadows are important for their ecosystem services. Yet, there are still many knowledge gaps in the global map of seagrasses. Some countries do not even have a first baseline. One such country is Venezuela, which has extensive seagrass meadows extending along its entire Caribbean Sea coast, but no national seagrass map or systematic in situ monitoring on seagrass ecosystems. The limited understanding of spatial and temporal trends does not help to develop informed national conservation and restoration strategies, or national blue carbon accounting. Here, we describe results from a new remote sensing study to evaluate the first national seagrass map along the whole Venezuelan Caribbean coast, derived from remote sensing techniques. One of the strategies to produce an initial regional seagrass map is via a multitemporal composite approach. This has been done in Greece, the Mediterranean Sea, East Africa, Bahamas and Seychelles. Often, issues of image quality, cloud cover, sun glint, and atmospheric and water turbidity reduces the pool of viable images for composition. Even when the cloud cover metadata is used to filter out excessively cloudy images, some cloudy pixels remain in the pool of filtered images, causing cloud artefacts in the composite image and compromising its quality. To improve the composite quality of the composite image and its derived seagrass map, we tested a new Google Earth Engine product: the Cloud Score+, which provides a per-pixel quality assessment (QA) band based on an atmospheric similarity model and a space-time context network model. The Cloud Score+ product has two bands, the Cloud Score probability score (CS) or the cumulative distribution function of this cloud score QA band (CDF). We compare the performance of Cloud Score+ derived products against previously established multi-temporal image composites acquired in different time ranges, and the more conservative ACOLITE-processed single image composite using Sentinel-2 (S2) Level-1C (L1C) imagery in the whole Venezuelan coastline. The S2 L1C imagery was processed following three different approaches: 1) using a multi-temporal composition of the full S2 L1C archive available and processed in GEE; 2) integrating Cloud Score+ dataset into the previous approach; and 3) using a single-image offline approach applying ACOLITE atmospheric correction which has been widely used for water applications. All images were further processed from L1C to L2A remote sensing reflectance (Rrs), for the purpose of comparability. Additional image features such as Gray Level Co-occurrence Matrix (GLCM) and Principal Component Analysis (PCA) were generated. The training data were randomly split into roughly 70% and 30% for training and test, respectively. This was bootstrapped 20 times to produce 20 sets of training and test data for the classification and validation. Per bootstrap, a two-step classification was performed. A first classification was trained on the 70% training dataset with Random Forest in GEE. A variable selection was performed on GEE using their native ee.Classifier.explain function, and only the top ten features were retained. The second classification was trained using these top 10 features on the 70% training dataset. We defined five classes for the classification: sand, seagrass, turbid, deep waters, and coral. For the training and test design, the point data were obtained along the coast and intertidal areas of the whole nation through existing literature, data banks and visual interpretation. We found that the performances across the different thresholds within the CS or CDF composites were largely similar, with small differences in their confidence intervals. In terms of temporal range for the multi-temporal processing, the full archive seven–year composite had the most consistent quantitative performance over the two different optical water types, achieving an F1 Score of seagrass class of 0.664 with a 95% Confidence Interval (CI) [0.634, 0.695] and 0.631 [0.588, 0.675] for coastal and open waters, respectively. For coastal waters, the ACOLITE composite had a very competitive F1 Score and best Overall Accuracy (OA) performance at 0.668 [0.649, 0.688] and 0.781 [0.649, 0.688], respectively. For open waters, the full archive seven–year composite performed best, with both the CS and CDF product having a comparable performance. The ACOLITE composite had the weakest quantitative performance in open waters, although its confidence intervals do overlap with all the other three products. Qualitatively, the ACOLITE composite was deemed to instead perform better than its competitors. In optically clear waters such as the open reef waters, where the main concerns were clouds and cloud shadows, the simpler Cloud Score+ products provided a pragmatic alternative to both the full archive and the ACOLITE products. For optically complex waters, it was better to rely on either a larger temporal interval or the ACOLITE atmospheric processor. Based on this comparison of the various products, the full archive seven–year composite forms a good initial baseline for the baseline national seagrass map for Venezuela.