Land use mapping using Sentinel-1 and Sentinel-2 time series in a heterogeneous landscape in Niger, Sahel

Abstract

Land use maps describe the spatial distribution of natural resources, cultural landscapes, and human settlements, serving as an important planning tool for decision makers. In the Sahel area, such information is valuable for risk management and mitigation in challenging sectors like food security, flood control, and urban planning. Due to its uniform quality across large areas in regular time steps, remote sensing imageries are essential input for producing land use maps. However, spatially and temporally heterogeneous landscapes in Sahel make classification of landscape features difficult. Our overall goal is to create an accurate, high resolution land use map covering Niamey, the capital of Niger and its surroundings which represents the unique landscape features in the Sahel using Sentinel-1 and Sentinel-2 archives. We assessed the performance of three commonly used classifiers (i.e. Maximum Likelihood (ML), Support Vector Machine (SVM) and Random Forest (RF)) for land use classification. To understand the utility of different features from Sentinel-1 and Sentinel-2 imagery for classification, we performed feature selection and compared mapping accuracies with and without feature selection. To leverage the strength of each classifier, we developed a classifier ensemble (CE) map based on the mapping accuracy of each land use class and each classifier. The results of this study showed that the performance of individual classifiers depends on feature selection method and accuracies can be improved by combining different classifiers. The ensemble map had an overall accuracy of 72 ± 3.9% and it was found superior in terms of accuracy particularly with respect to built-up areas compared to the existing global land cover products in the study area. Our classification scheme also better characterized the regional environment in the Sahel. For example, we mapped rice and bare rocks that have important regional implication, which are not included in the existing products. Overall, our approach highlights the potentiality of combining multi-modal imageries and multiple classifiers for mapping a heterogenous environment such as the Sahel with high spatial resolution.