Convolutional Neural Networks (CNNs) are proposed for large-scale remote sensing image classification. The framework directly trains CNNs to produce classification maps from input images. A fully convolutional architecture is designed to address dense classification, and a two-step training approach is introduced to handle imperfect training data. The first step initializes the network with large, possibly inaccurate data, while the second step refines it with small, accurately labeled data. A multi-scale neuron module is designed to balance recognition and precise localization. Experiments show that the network considers a large amount of context to provide fine-grained classification maps. The framework is evaluated on a Pléiades image dataset over France, where OpenStreetMap data is inaccurate. The approach is end-to-end, with a fully convolutional network that avoids discontinuities at patch borders and improves accuracy and efficiency. The network is trained on a large dataset of Boston buildings and tested on a Pléiades image dataset. The results show that the fully convolutional network outperforms the patch-based approach in terms of accuracy and speed. The framework is also tested on a dataset with limited OpenStreetMap coverage, demonstrating its effectiveness in handling inaccurate training data. The multi-scale neuron module helps alleviate the trade-off between classification accuracy and the number of learnable parameters. The framework is evaluated on a Pléiades image over France, showing its effectiveness in large-scale satellite image classification.Convolutional Neural Networks (CNNs) are proposed for large-scale remote sensing image classification. The framework directly trains CNNs to produce classification maps from input images. A fully convolutional architecture is designed to address dense classification, and a two-step training approach is introduced to handle imperfect training data. The first step initializes the network with large, possibly inaccurate data, while the second step refines it with small, accurately labeled data. A multi-scale neuron module is designed to balance recognition and precise localization. Experiments show that the network considers a large amount of context to provide fine-grained classification maps. The framework is evaluated on a Pléiades image dataset over France, where OpenStreetMap data is inaccurate. The approach is end-to-end, with a fully convolutional network that avoids discontinuities at patch borders and improves accuracy and efficiency. The network is trained on a large dataset of Boston buildings and tested on a Pléiades image dataset. The results show that the fully convolutional network outperforms the patch-based approach in terms of accuracy and speed. The framework is also tested on a dataset with limited OpenStreetMap coverage, demonstrating its effectiveness in handling inaccurate training data. The multi-scale neuron module helps alleviate the trade-off between classification accuracy and the number of learnable parameters. The framework is evaluated on a Pléiades image over France, showing its effectiveness in large-scale satellite image classification.