Land-Use Classification in a Tropical Wetland: A Comparison of MLC and Machine-Learning Algorithms

Historical monitoring of land-cover and land-use change provides a means to quantify anthropogenic and atmospheric processes (e.g., floods and droughts) that affect lagoon systems. The Chaschoc-Sejá lagoon system (SLCh-S), located in Tabasco, Mexico, is a natural complex dominated by interconnected lagoons and noted for its high biodiversity, including endemic species. The SLCh-S exhibits strong seasonal dynamics. During the rainy season, it behaves as an interconnected network of water bodies linked by meandering tributaries; extensive flooding occurs, vegetation cover declines, and water bodies display striking color variability. In contrast, during the dry season, interconnections disappear, sediments become exposed, and wet soils and flood-tolerant vegetation emerge along lagoon margins.

 

Although the SLCh-S is undergoing anthropogenic and environmental pressures, the magnitude of these impacts at the regional scale remains poorly understood. Land-use maps derived from remote sensing offer a key first step for large-scale monitoring. However, multiple mapping methods are available, and their performance depends strongly on the characteristics of each study area; therefore, testing is required to identify the most suitable approach.

 

The objective of this study is to evaluate and compare supervised classifiers applied to high-resolution (3 m) satellite imagery to determine which performs best in the region. Two PlanetScope images were analyzed, one from the dry season (March 2024) and one from the rainy season (September 2024). We implemented the traditional Maximum Likelihood Classification (MLC) method and three machine-learning classifiers: Random Forest (RF), Support Vector Machine (SVM), and Random Trees (RT). Classification accuracy was assessed using the Kappa index.

 

Kappa scores were 0.82 for MLC, 0.77 for RF, 0.68 for SVM, and 0.70 for RT. Results indicate that in flat terrain with homogeneous vegetation, agricultural areas, and well-defined water bodies, MLC can effectively classify land use and vegetation, outperforming the tested machine-learning algorithms. Nevertheless, all methods showed limitations in discriminating vegetation with high intra-class spectral variability. The moderate accuracies also highlight the need for post-classification refinement to improve final maps, a step that can be labor-intensive in high temporal-resolution monitoring. Integrating derived variables (e.g., NDVI/NDWI, texture) and complementing accuracy assessment with per-class ROC/AUC metrics (one-vs-rest) is recommended to better characterize class separability.

 

Overall, the study clarifies the strengths and limitations of common classifiers for high-resolution monitoring of tropical wetlands.