Detect, Consolidate, Delineate: Scalable Mapping of Field Boundaries Using Satellite Images

Abstract

Digital agriculture services can greatly assist growers to monitor their fields and optimize their use throughout the growing season. Thus, knowing the exact location of fields and their boundaries is a prerequisite. Unlike property boundaries, which are recorded in local council or title records, field boundaries are not historically recorded. As a result, digital services currently ask their users to manually draw their field, which is time-consuming and creates disincentives. Here, we present a generalized method, hereafter referred to as DECODE (DEtect, COnsolidate, and DElinetate), that automatically extracts accurate field boundary data from satellite imagery using deep learning based on spatial, spectral, and temporal cues. We introduce a new convolutional neural network (FracTAL ResUNet) as well as two uncertainty metrics to characterize the confidence of the field detection and field delineation processes. We finally propose a new methodology to compare and summarize field-based accuracy metrics. To demonstrate the performance and scalability of our method, we extracted fields across the Australian grains zone with a pixel-based accuracy of 0.87 and a field-based accuracy of up to 0.88 depending on the metric. We also trained a model on data from South Africa instead of Australia and found it transferred well to unseen Australian landscapes. We conclude that the accuracy, scalability and transferability of DECODE shows that large-scale field boundary extraction based on deep learning has reached operational maturity. This opens the door to new agricultural services that provide routine, near-real time field-based analytics.

Unlocking large-scale crop field delineation in smallholder farming systems with transfer learning and weak supervision

“…In this study, we assembled a dataset of 10,000 fields in India and used high resolution satellite imagery, transfer learning, and weak supervision with partial labels to automatically delineate smallholder fields. Building upon prior work [23,31], our methods achieve high performance (best model MCC = 0.65, median IoU = 0.86) through (1) access to very high resolution satellite imagery and (2) the use of partial field labels that relieve the labeling burden while enabling a large number of locations across the country to be sampled. In fact, taking a model naively trained in France and applied as-is in India as a benchmark, our method improves field-level identification by over two-fold.…”

“…We briefly describe the FracTAL-ResUNet architecture and loss function here and refer the reader to Diakogiannis et al [35] and Waldner et al [31] for more details. A FracTAL-ResUNet has three main features, reflected in its name: (1) a self-attention layer called a FracTAL unit that is inserted into standard residual blocks, (2) skip-connections that combine the inputs and outputs of residual blocks (similar to in the canonical ResNet), and (3) an encoder-decoder architecture (similar to a U-Net).…”

“…The first step of our field delineation pipeline is to use an attention-based deep neural network architecture to detect field edges. The architecture, named FracTAL-ResUNet, was first proposed in Waldner et al [31] and used to create production-grade field boundaries in Australia. We experimented with both the FracTAL-ResUNet, ResUNet-a [40], and regular U-Net [41] architectures and found that the FracTAL-ResUNet had the best performance in France and India.…”

“…High-resolution satellite imagery and recent advances in computer vision offer opportunities for automated field boundary delineation at low cost. Researchers have demonstrated that delineation can be automated in industrial agricultural systems like those in North America [2,26,27], Europe [28,29,30], and Australia [31] using publicly-available Landsat or Sentinel-2 imagery. In some cases, previous methods used large quantities of historical field boundaries and supervised machine learning to prove automated delineation is possible.…”

Looking for Change? Roll the Dice and Demand Attention

“…In this work, we did not experiment with deeper architectures, which would surely improve performance (e.g., D7nf32 models usually perform better), or with hyperparameter tuning. Finally, we should point out that the methods presented here have been successfully applied recently for the task of semantic segmentation (field boundary detection [59]). All Building Blocks use kernel size k = 3, padding p = 1 (SAME), and stride s = 1.…”