Evaluation and comparison of the earth observing sensors in land cover/land use studies using machine learning algorithms

The rapid transformation of land cover/land use (LCLU) is a strong indication of global environmental change. In order to monitor LCLU through maps, a significant dataset and robust technique are necessary. Thus, the primary objective of the current research is to evaluate and compare the efficiency of several notable satellite sensors including Landsat-8 (L-8), Sentinel-2 (S-2), Sentinel-1 (S-1), combined Sentinel-1 and Sentinel-2 (S-1-2), LISS III (L-3), and LISS IV (L-4) for LCLU mapping applying random forest (RF), logit boost (LB), stochastic gradient boosting (SGB), artificial neural network (ANN), and K-nearest neighbor (KNN) models. For this purpose, 300 samples for each of the six LCLU classes have been selected based on field survey and high resolution Cartosat-3 images. The classification accuracy namely producer accuracy (PA), user accuracy (UA), overall accuracy (OA) and kappa coefficient have been calculated from the confusion matrix of the applied models. This results show the highest accuracy has been derived from the integration of S-1-2 datasets followed by S-2, L-8, L-3, L-4, and S1. On the other hand, LB model is the most consistent and efficient in comparison with other models for all the datasets. Regarding importance of variable, SWIR band is repeatedly the most crucial factor while blue band is the least significant variable. From this comparative assessment of sensors, it has been found that high spatial and spectral resolutions along with combination of satellite datasets are required to get better accuracy rather than only high spatial resolution in regional scale mapping. The present study strongly advocates the use of combined S-1-2 data together with the application of LB model for LCLU classification.