Day 4

Detailed paper information

Back to list

Paper title Monitoring three decades of center-pivot field dynamics in Saudi Arabia using a hybrid machine learning framework
  1. Ting Li King Abdullah University of Science and Technology Speaker
  2. Kasper Johansen King Abdullah University for Science and Technology (KAUST)
  3. Matthew McCabe King Abdullah University of Science and Technology
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Agricultural field masks or boundaries provide a basis for obtaining object-based agroinfomatics like crop type, crop yield, and crop water usage. Machine learning techniques offer an effective means of masking fields or delineating field boundaries using satellite data. Unfortunately, field boundary information can be difficult to obtain when trying to collect ground truth to train a machine learning model, since such information is not routinely available for many regions around the world. Manually creating field masks is an obvious solution to address this data gap, but this can consume a considerable amount of time, or simply be impractical when confronted with large mapping tasks (e.g. national scale). Here, we propose a hybrid machine learning framework that combines clustering algorithms and convolutional neural networks to identify and delineate center-pivot agricultural fields. Using a multi-temporal sequence of Landsat-based normalized vegetation index collected over one of the major agricultural regions in Saudi Arabia as input, a training dataset was produced by identifying field shape (circle, fan, or neither) and establishing whether it consisted of multiple fields. When evaluated against 4,099 manually identified center-pivot fields, the framework showed high accuracy in terms of identifying the fields, achieving 97.4% producer and 98.0% user accuracies on an object basis. The intersection over union accuracy was 96.5%. Based on the framework, the field dynamics across the study region from 1988 to 2020 were obtained, including the number and acreage of fields, the spatial and temporal dynamics of field expansion and contraction, and the number of years a field was detected as active. Our work presents the first long-term assessment of such dynamics in Saudi Arabia, and the resulting agroinformatic data correlated well with government-driven policy initiatives to reduce water consumption. Overall, the framework was trained using a dataset that was easy and efficient to produce and relied on limited in-situ records. It demonstrated stable performance when applied to different periods, and has the potential to be applied at the national scale, providing agroinformatic data that may assist food and water security-related concerns.