Search

The capacity to accurately map field boundaries of smallholder farms is important for increasing food production and ensuring food security in these systems. Deep learning methods have shown effectiveness in learning high-level image features and detecting objects using satellite imagery. In this study, we used a deep-learning based method, Mask Region based Convolutional Neural Networks (Mask R-CNN) to detect and delineate individual smallholder fields with very-high-resolution satellite imagery (WorldView-3, 0.5 m spatial resolution) in Bihar, India. We compared the accuracies of models using different input image types including panchromatic, pan-sharpened multispectral, or edge-enhanced imagery. To reduce the amount of labor needed to collect training and validation data, we developed an iterative method to annotate field polygons, which relied on annotating incorrectly predicted field boundaries from an initial Mask RCNN model. Our final trained Mask R-CNN models reported mean average precisions larger than 0.67 and F1 Scores larger than 0.72 at Intersection over Union thresholds of 0.50, which demonstrates the potential of Mask R-CNN to map smallholder field boundaries. Our results suggest that the type of imagery used has limited effects on accuracy. We also find that our iterative approach to collect data improved accuracies while reducing polygon digitization effort.