Silage Bale Detection for the «Cultivable Area» Update of the Cantonal Agricultural Office, Thurgau

EasyChair Preprint 8817

Abstract

In Switzerland direct subsidies are paid to farms for sustainable agricultural practice. The cultivable agricultural area layer (German: Landwirtschaftliche Nutzfläche, LN) serves as an annual basis for the calculation of these contributions at the Swiss cantonal agricultural offices. Material deposits like silage bale stacks are usually excluded from the LN. Therefore, the canton of Thurgau could profit from a spatial vector layer indicating locations and area consumption extent of silage bale stacks intersecting with the LN perimeter.

To ease the monitoring process, we propose a Mask-RCNN based prototypical Deep Learning framework which was trained on 10cm SWISSIMAGE orthophoto datasets (swisstopo, Bern). Embedded in an efficient python-based geodata workflow the model boasts a high F1-Score of 92% on evaluation data. This approach allows robust and accurate inference detections over the whole cantonal area. Having the silage bale stack detections at hand reduces the manual workload of the responsible official by directing the eyes to the relevant hotspots.

Keyphrases: Administration, Agriculture, Mask RCNN, Subsidy Payments, aerial imagery, cadastral, monitoring, object detection, remote sensing

@booklet{EasyChair:8817,
  author    = {Adrian Meyer and Denis Jordan},
  title     = {Silage Bale Detection for the «Cultivable Area» Update of the Cantonal Agricultural Office, Thurgau},
  howpublished = {EasyChair Preprint 8817},
  year      = {EasyChair, 2022}}