Intelligent mapping of irrigated areas from Landsat 8 images using transfer learning

The lack of reliable and up-to-date data in developing countries is a major obstacle to sustainable development. In Morocco, where groundwater withdrawals by farmers are very intensive informal, maps describing monitoring the extension irrigated areas scarce labor-intensive obtain. this paper novel transfer learning algorithm proposed map at different stages an agricultural cycle from Landsat 8 images. results obtained displays satisfactory performance over traditional machine algorithms. On small dataset, we initially tested three well known deep architectures (SegNet, DenseNet Unet). were not satisfactory. So, get high performance, rely on architecture combining UNet with ResNet50 backbone (trained 2012 ILSVRC ImageNet dataset) as baseline after phase configurations tested. first part study, compared use optimization methods: Adam two variants Stochastic Gradient Descent (SGD) associated techniques (Cyclical Learning Rate Warm Restart) find optimal rate then test impact augmentation overall accuracies. Data had improved accuracy for methods. based method 94% 97% mean IoU 0,79 (for all land cover classes) 0,86 class. For SGD methods, increased 91% 0,75 0,82 As interested having key periods cycle, also explored, second temporal generalization best model.