National crop field delineation for the United States

Abstract

Comprehensive and accurate crop field boundary maps are crucial for digital agriculture, land management, and environmental monitoring. However, no high-quality field boundary dataset is publicly available in the United States. This thesis addresses this gap by creating a new, large dataset and training a deep learning model capable of mapping field boundaries. We built a dataset of over 15,000 image-mask pairs using high-resolution National Agriculture Imagery Program (NAIP) satellite imagery and curated field boundary labels. This dataset covers a variety of leading agricultural states and includes images taken at different scales to capture a wide variety of field sizes and layouts. We used this dataset to train an adapted ResUNet++ neural network model designed to segment crop fields. The trained model achieved around 0.8 for pixel-level accuracy, showing it can generally identify field areas well. However, its performance in matching predicted individual field instances with the ground truth instances (measured by mean instance Intersection over Union, or mIoU) was around 0.5. This lower instance score was largely due to the post-processing step, which converts the model’s probability predictions into separate field instances. Despite this, the field polygons produced by our approach are visually coherent with satellite field images and can be readily used with geospatial tools like Google Earth Engine. Our work provides a practical starting point for future research on mapping fields across the contiguous U.S. Potential directions for improvements may involve developing sharper boundary predictions, exploring direct instance segmentation models, refining post-processing methods, and expanding the dataset to include more challenging areas.