Multisensor Land Cover Classification With Sparsely Annotated Data Based on Convolutional Neural Networks and Self-Distillation

Extensive research studies have been conductedin recent years to exploit the complementarity among multi-sensor (or multi-modal) remote sensing data for prominentapplications such as land cover mapping. In order make a step further with respect previous which investigate multi-temporal SAR and optical or multi-temporal/multi-scale combinations, here we propose deep learning framework that simultaneously combine all these input sources, specifically SAR/optical fine scale information. Our proposal relies on patch-based multi-branch convolutional neural network (CNN) exploits different per source CNN encoders deal specificity of signals. Additionally, our is equipped self-distillation strategy boost analyses. This new leverages final prediction multi-source guide supporting learn from itself.Experiments are carried out two real world benchmarks, namely Reunion island (a french overseas department) Dordogne study site southwest department in France)where annotated reference was collected under operational constraints (sparsely ground truth data). Obtained results, providing an overall classification accuracyof about 94% (resp. 88%) theReunion island(resp. theDordogne) site. These findings highlight effectivenessof based Convolutional Neural Networks andself-distillation heterogeneous remotesensing confirm benefit multi-modal analysisfor downstream tasks