The paper presents an end-to-end framework for dense, pixel-wise classification of satellite imagery using convolutional neural networks (CNNs). The authors propose a fully convolutional architecture that directly produces classification maps from input images, addressing the challenge of imperfect training data through a two-step training approach: initial training on large amounts of potentially inaccurate reference data followed by refinement on a small amount of accurately labeled data. They also introduce a multi-scale neuron module to alleviate the trade-off between recognition and precise localization. Experiments on a Pleiades image dataset over France demonstrate the effectiveness of the proposed framework, showing that the networks can take into account a large amount of context to provide fine-grained classification maps. The contributions include a new fully convolutional architecture, a two-step training approach, and a multi-scale architecture, all aimed at improving the performance of CNNs for large-scale remote sensing image classification.The paper presents an end-to-end framework for dense, pixel-wise classification of satellite imagery using convolutional neural networks (CNNs). The authors propose a fully convolutional architecture that directly produces classification maps from input images, addressing the challenge of imperfect training data through a two-step training approach: initial training on large amounts of potentially inaccurate reference data followed by refinement on a small amount of accurately labeled data. They also introduce a multi-scale neuron module to alleviate the trade-off between recognition and precise localization. Experiments on a Pleiades image dataset over France demonstrate the effectiveness of the proposed framework, showing that the networks can take into account a large amount of context to provide fine-grained classification maps. The contributions include a new fully convolutional architecture, a two-step training approach, and a multi-scale architecture, all aimed at improving the performance of CNNs for large-scale remote sensing image classification.