Advances in remote sensing of the early Anthropocene in tropical wetlands: From biplanes to lidar and machine learning. (April 2023)

• Record Type: Journal Article
• Authors: Doyle, Colin; Luzzadder-Beach, Sheryl; Beach, Timothy
• Other Names: Beach Timothy guest-editor.; Luzzadder-Beach Sheryl guest-editor.; Cook Duncan guest-editor.; Turner Billie Lee guest-editor.
• Is Part Of: Progress in physical geography. Volume 47:Number 2(2023)
• Page Start: 293
• Page End: 312
• Publication Date: 2023-04
• Subjects: Anthropocene -- LiDAR -- machine learning -- Maya wetland canals and fields -- remote sensing; Physical geography -- Periodicals; 910.02
• DOI: 10.1177/03091333221134185
• Languages: English
• ISSNs: 0309-1333

Abstract

This paper reviews the history of remote sensing for mapping ancient Maya wetland fields in Central America and provides a new assessment using machine learning with LiDAR data. We evaluate past uses of radar, multispectral, and LiDAR datasets in Northwest Belize across well-studied wetland field complexes that occur under different vegetation conditions. Next, we compare topographic products derived from LiDAR data commonly used for archaeology and geomorphology. Lastly, we train a machine learning algorithm to detect ancient canals with LiDAR data and test the algorithm on a newly rediscovered field system. The spatial resolution of any dataset must be sufficiently high (2-m or finer resolution) to detect most of these canals reliably. High resolution multispectral sensors can detect canals in open areas, but most wetland complexes are under dense tropical forest impenetrable to multispectral instruments. LiDAR data were the most useful due to the high spatial resolution (0.5-m) and the ability to penetrate canopy, but still have limitations under certain conditions. The intensity of the LiDAR returns with multispectral LiDAR systems can reveal differences in soil and vegetation between ancient canals and fields in places leveled by modern farmers. The algorithm successfully maps ancient canals but has many false positives in natural depressions and drainages. Although the machine learning approach tested here cannot be used on its own, we used it with the LiDAR visualizations to refine the canal estimates from previous studies, mapping nearly 50% more canals in the Central Rio Bravo floodplain than originally published.