Crop type mapping by using transfer learning

Crop type mapping currently represents an important problem in remote sensing. Accurate information on the extent and types of crops derived from sensing can help managing improving agriculture especially for developing countries where such is scarce. In this paper, high-resolution RGB drone images are input data classification performed using a transfer learning (TL) approach. VGG16 GoogLeNet, which pre-trained convolutional neural networks (CNNs) used tasks coming computer vision, considered crop types. Thanks to transferred knowledge, proposed models successfully classify studied with high overall accuracy two cases, achieving up almost 83% Malawi dataset 90% Mozambique dataset. Notably, these results comparable ones achieved by same deep CNN architectures many vision tasks. With regard analysis, application very limited so far due requirements number samples needed train complicated architectures. Our demonstrate that efficient way overcome take full advantage benefits drone-based mapping. Moreover, based experiments different TL approaches we show frozen layers parameter fine-tuning all weights significantly better performance than apply only some numbers last layers.