Deep learning on edge: Extracting field boundaries from satellite images with a convolutional neural network

Applications of digital agricultural services often require either farmers or their advisers to provide digital records of their field boundaries. Automatic extraction of field boundaries from satellite imagery would reduce the reliance on manual input of these records, which is time consuming, and would underpin the provision of remote products and services. The lack of current field boundary data sets seems to indicate low uptake of existing methods, presumably because of expensive image preprocessing requirements and local, often arbitrary, tuning. In this paper, we propose a data-driven, robust and general method to facilitate field boundary extraction from satellite images. We formulated this task as a multi-task semantic segmentation problem. We used ResUNet-a, a deep convolutional neural network with a fully connected UNet backbone that features dilated convolutions and conditioned inference to identify: 1) the extent of fields; 2) the field boundaries; and 3) the distance to the closest boundary. By asking the algorithm to reconstruct three correlated outputs, the model's performance and its ability to generalise greatly improve. Segmentation of individual fields was then achieved by post-processing the three model outputs, e.g., via thresholding or watershed segmentation. Using a single monthly composite image from Sentinel-2 as input, our model was highly accurate in mapping field extent, field boundaries and, consequently, individual fields. Replacing the monthly composite with a single-date image close to the compositing period marginally decreased accuracy. We then showed in a series of experiments that, without recalibration, the same model generalised well across resolutions (10 m to 30 m), sensors (Sentinel-2 to Landsat-8), space and time. Building consensus by averaging model predictions from at least four images acquired across the season is paramount to reducing the temporal variations of accuracy. Our convolutional neural network is capable of learning complex hierarchical contextual features from the image to accurately detect field boundaries and discard irrelevant boundaries, thereby outperforming conventional edge filters. By minimising over-fitting and image preprocessing requirements, and by replacing local arbitrary decisions by data-driven ones, our approach is expected to facilitate the extraction of individual crop fields at scale.