Comparative Performance Evaluation of Pixel-Level and Decision-Level Data Fusion of Landsat 8 OLI, Landsat 7 ETM+ and Sentinel-2 MSI for Crop Ensemble Classification

Abstract

Crops mapping unequivocally becomes a daunting task in humid, tropical, or subtropical regions due to unattainability of adequate cloud-free optical imagery. Objective of this study is to evaluate the comparative performance between decision- and pixel-levels data fusion ensemble classified maps using Landsat 8, Landsat 7, and Sentinel-2 data. This research implements parallel and concatenation approach to ensemble classify the images. The multiclassifier system comprises of Maximum Likelihood, Support Vector Machines, and Spectral Information Divergence as base classifiers. Decision-level fusion is achieved by implementing plurality voting method. Pixel-level fusion is achieved by implementing fusion by mosaicking approach, thus appending cloud-free pixels from either Sentinel-2 or Landsat 7. The comparison is based on the assessment of classification accuracy. Overall accuracy results show that decision-level fusion achieved an accuracy of 85.4%, whereas pixel-level fusion classification attained 82.5%, but their respective kappa coefficients of 0.84 and 0.80 but are not significantly different according to Z-test at $\alpha = {\text{0.05}}$ . F1-score values reveal that decision-level performed better on most individual classes than pixel-level. Regression coefficient between planted areas from both approaches is 0.99. However, Support Vector Machines performed the best of the three classifiers. The conclusion is that both decision-level and pixel-level fusion approaches produced comparable classification results. Therefore, either of the procedures can be adopted in areas with inescapable cloud problems for updating crop inventories and acreage estimation at regional scales. Future work can focus on performing more comparison tests on different areas, run tests using different multiclassifier systems, and use different imagery.