Per-pixel accuracy as a weighting criterion for combining ensemble of extreme learning machine classifiers for satellite image classification

Abstract

Reliable classification of satellite images is essential for various applications, including land cover and crop (LCC) mapping. In recent years, ensemble classifiers have shown remarkable success in satellite image classification as they provide solutions to combine and integrate multiple classifiers. This research presented a novel satellite image classification method called PAELM. The PAELM algorithm builds a diverse set of extreme learning machine (ELM) classifiers and combines them using the pixel-based LCC accuracy values in such a way that, for a given pixel, the most highly accurate ELM classifier among the ensemble of ELMs assigns the LCC class of that pixel. We validated PAELM on six experimental sites with varying geographical environments in the United States and compared it with three advanced machine learning classifiers, namely support vector machine, conventional ELM, and extreme gradient boosting and one advanced ensemble classifier. Our results showed that PAELM improved the accuracy of LCC mapping in comparison to the benchmark classifiers. The LCC maps generated by PAELM had an averaged overall accuracy and aggregated F1 score of 0.811 and 0.804, respectively, while these values for the most accurate benchmark classifier were 0.787 and 0.781, respectively. The results also implied that all the classifiers were sensitive to scene heterogeneity and LCC class composition, with PAELM being the least sensitive classifier to these factors. Overall, our findings suggested that PAELM was a promising approach for accurate LCC mapping, demonstrating the practicality of spatial accuracy as a weighting factor in the integration of ensemble of classifiers.

Keywords

• Ensemble classification
• Extreme learning machine
• Per-pixel accuracy
• Sentinel
• Cropland data layer