Adaptability of Machine Learning-based LULC Classification Techniques for Natural Resource Monitoring:

A Study on an Oxbow Lake in Manikganj, Bangladesh

Abstract

Accurate Land Use Land Cover (LULC) information can help with numerous research efforts relating to floods, droughts, migration, and climate change at several scales. In decision support systems for sustainable development and natural resource management, machine learning-based methods for detecting changes in land use and land cover using Remote Sensing (RS) data have become indispensable tools. However, Smaller areas like the selected study area may have issues regarding the spatial resolution of satellite imagery. Although the development of Machine Learning (ML) algorithms has boosted the accuracy and use of classification systems, the accuracy level of different classifiers in small areas is often not tested. In order to determine the most accurate classifier for a small-area LULC mapping, the study compares three classifiers in ArcGIS and six classifiers in Google Earth Engine (GEE) against manually classified high-resolution Google Earth data. According to the study, the Support Vector Machine (SVM) classifier outperformed the other one in terms of accuracy for both Sentinel and Landsat satellite imagery.  But because of the small number of training samples and spectral similarity between the classes, all six classifiers—Classification and Regression Tree (CART), Random Forest (RF), Support Vector Machine (SVM), Minimum Distance (MD), Naive Bayes (NB), and Maximum Likelihood Classifier (MLC)—had trouble correctly differentiating between bare land and built-up areas. The study suggests taking adequate samples for classification, using higher resolution imageries, reinforcement-based classification or indices for specific feature-oriented research. The study also urges a high-resolution earth resource satellite in Bangladesh to monitor the fragile environment and disaster situations better.