Land cover intelligent mapping using transfer learning and semantic segmentation

Monitoring the extension of irrigated areas at different stages of an agricultural cycle is of great interest for land and water resources management. With a limited training dataset (50 three bands Landsat 8 images), we rely on a transfer learning architecture combining UNet and ResNet50 as backbone to intelligently map irrigated areas. Three optimization methods are evaluated: Adam and two variants of Stochastic Gradient Descent (SGD) associated with two techniques (Cyclical Learning Rate and Warm Restart). We also assess the impact of data augmentation and temporal generalization of the model learned by predicting the location of irrigated areas at specific periods of the same agricultural cycle.