Tracking changes in extent and distribution of tropical coastal covers using simple semi-supervised classification

Abstract

New and more complex methods of classification and monitoring of land cover change are developed daily. But simpler strategies could be sufficient for the tasks, more accessible to geographic analysts, and more interpretable to end-users at local-regional scales. This article uses a sequence of simple thresholds of spectral indices to obtain a coarse classification of four coarse coastal covers predominant in many tropical regions: water, mangrove, tidal, and sand. This semi-supervised method was used to (1) obtain current and historical classifications of Iscuandé and Guapi River Mouths, Colombia, and (2) measure the change in extension and distribution of coastal covers across 36 years (1984–2019). The overall accuracy of simple thresholding was high (85%), but lower than four machine learning algorithms (RF, CART, SVM, and GTB, ranging from 94 to 96%). Vegetation and water were classified with higher accuracy (97%), while tidal and sandy bare soils have lower accuracy (87 and 79% respectively). The simplicity of the current method allows the detection of temporal and spatial changes in coastal covers. Moreover, tidal flats and sandy bare soil increased an overall 42% and 83% respectively representing 743 ha of new bare soils during 36 years. Larger growth occurred in the early nineties and while changes were heterogeneous in magnitude, most of the localities studied show an overall increase in tidal and sand. Performance of simple thresholding could be improved using an alternative combination of indices, especially for bare soils, optimal adjustment of thresholds, object-oriented classification, or different strategies for building image mosaics. However, we believe this simple approach could aid the exploration of changes in coastal landscapes with sparse coverage of optical imagery and a lack of ground surveys.