A Comparative Assessment of Multi-view fusion learning for Crop Classification

With a rapidly increasing amount and diversity of remote sensing (RS) datasources, there is a strong need for multi-view learning modeling. This is acomplex task when considering the differences in resolution, magnitude, andnoise of RS data. The typical approach for merging multiple RS sources has beeninput-level fusion, but other - more advanced - fusion strategies mayoutperform this traditional approach. This work assesses different fusionstrategies for crop classification in the CropHarvest dataset. The fusionmethods proposed in this work outperform models based on individual views andprevious fusion methods. We do not find one single fusion method thatconsistently outperforms all other approaches. Instead, we present a comparisonof multi-view fusion methods for three different datasets and show that,depending on the test region, different methods obtain the best performance.Despite this, we suggest a preliminary criterion for the selection of fusionmethods.