Effective water management requires a large-scale understanding of agricultural irrigation systems and how they shift in response to various stressors. Here, we leveraged advances in Machine Learning and availability of very high resolution remote sensing imagery to help resolve this long-standing issue. To this end, we developed a deep learning model to classify irrigation systems at a regional scale using remote sensing imagery. After testing different model architectures, hyper parameters, class weights and image sizes, we selected a U-Net architecture with a Resnet-34 backbone for this purpose. We applied transfer learning to increase training efficiency and model performance. We considered four irrigation systems as well as urban and background areas as land use/cover classes, and applied the model to 8,600 very high resolution (1 m) images, labeled with ground-truth observations of irrigation types, in a case study in Idaho, USA. Images were obtained from the US Department of Agriculture's National Agriculture Imagery Program. Our model achieved state-of-the-art performance for segmentation of different classes on the train data (85% to 94%), validation data (72% to 86%), and test data (70% to 86%), which attests to the efficacy of the model for the segmentation of images based on spatial features. Aside from leveraging deep learning and remote sensing for resolving the standing real-world problem of multiple irrigation type segmentation, this study develops and publicly shares labeled data, as well as a trained deep learning model, for irrigation type segmentation that can be applied/transferred to other regions globally. Furthermore, this study offers novel information about the impacts of transfer learning, imbalanced training data, and efficacy of various model structures for multiple irrigation type segmentation.
ASJC Scopus subject areas