10,000 Crop Field Boundaries across India

Release of dataset and neural network weights accompanying the paper "Unlocking large-scale crop field delineation in smallholder farming systems with transfer learning and weak supervision" (forthcoming in Remote Sensing). Ten thousand crop fields in India were delineated manually through inspection of high-resolution satellite imagery (Airbus SPOT). We also provide the weights of the highest performing neural network (FracTAL ResUNet architecture) pre-trained in France and fine-tuned on Airbus SPOT images in India. The model was trained in MXNet 1.6.0 and can be loaded with the "model.load_parameters()" function.