Day 4

Detailed paper information

Back to list

Paper title In season fields delineation from Sentinel-1 time series using Convolutional Neural Network for object-based crop monitoring system
Authors
  1. Quentin Deffense UCLouvain/Geomatics Speaker
  2. Diane Heymans UCLouvain/Geomatics
  3. Pierre Defourny UCLouvain-Geomatics Belgium
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Crop monitoring at field level depends upon the availability of consistent field boundaries. In Europe, each country has its own Land Parcel Information System (LPIS) used as reference parcels to calculate the maximum area eligible in the direct payment of the Common Agricultural Policy. The update of the parcels by the administration is time-consuming and often based on orthophotos not always up to date. An automated field delineation would greatly ease this process by detecting the new parcels and the changes of parcel boundaries from one season to another. On another hand, this delineation would allow the extraction of statistical features at field level without the need for manual intervention. This objective was successfully achieved by using ResUNet-a, a deep Convolutional Neural Network, on Sentinel-1 metrics based on the coherence time series at 10 meters spatial resolution. The use of Synthetic Aperture Radar (SAR) allows obtaining early in the season cloud-free composites with high contrast between different fields. ResUNet-a is a fully connected UNet that performs a multitask semantic segmentation by estimating three metrics for each pixel: the extent probability, (i.e., the probability of a pixel to belong to a field), the probability to be a boundary pixel and the distance to the closest boundary. This model is trained here on the LPIS of the year 2019 in Wallonia (Belgium) and applied on the year 2020. A watershed algorithm is then used on the three metrics to extract the predicted field polygons. The results validation compares these predictions to the LPIS of 2020 on one hand and to the LPIS of 2019 on the other hand to validate the detected changes. These results assessment obtained over more than 60 000 parcels demonstrates that the proposed method has very good accuracy for field delineation paving the way for in-season field delineation independent to manual inputs. On top of that, the method can detect the new parcels, the ones that are no longer exploited and the ones that have changed compared to the last season. While such a delineation was found really critical for near real time crop monitoring at field level, the approach is also very promising in the context of the LPIS management for the Common Agriculture Policy to point out which fields need to be updated or added.