Field Boundary Identification using Convolutional Neural Network and GIS on High Resolution Satellite Observations | IEEE Conference Publication | IEEE Xplore

Field Boundary Identification using Convolutional Neural Network and GIS on High Resolution Satellite Observations


Abstract:

With advent of constellation of high-resolution micro satellites, the acquisition rate of earth observation has surpassed the rate of data processing. In semi-arid rainfe...Show More

Abstract:

With advent of constellation of high-resolution micro satellites, the acquisition rate of earth observation has surpassed the rate of data processing. In semi-arid rainfed agricultural ecosystems area under crop varies across seasons. Due to small land holdings it’s often difficult to identify the field level crop cultivation information. Also, field boundaries are important for identification of crop extent, crop insurance, crop loan, carbon credit and to establish the credit score for the farm. Studies have reported effective edge detection using Deep Learning based classifiers. Geographic Information System (GIS) based topology operations for vector geometry are effective in correction of vector geometries. This study describes field boundary identification approach using Convolutional Neural Network (CNN) on high resolution satellite observations. A Holistically-nested Edge Detection algorithm is used to identify the edge raster images. The pixel error rate of 19% was obtained with 200 epochs and 131 training images. Finally, the edge raster images were geo-referenced and converted into vector polygon geometry. Topology operations such as sliver polygon removal, overshoot and undershoot error removal were applied to refine the field boundary output. The accuracy assessment of identified field boundaries was performed with manually drawn field boundaries. Key features such as area of the polygon and centroid shift were compared between actual and identified field boundaries. We observed mean of difference in area of 216 sq. meter and chentroid shift of 1.12 meter. We plan to train proposed architecture for different spatial resolutions and cropping conditions. Additional GIS based accuracy matrices like percent overlap will be used during the operational use.
Date of Conference: 26-29 July 2021
Date Added to IEEE Xplore: 08 September 2021
ISBN Information:
Conference Location: Shenzhen, China

Contact IEEE to Subscribe

References

References is not available for this document.