Deep learning on edge: extracting field boundaries from satellite images with a convolutional neural network (1910.12023v2)

Abstract: 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 error-prone, 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 address the problem of field boundary extraction from satellite images as a multitask 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 assign three labels to each pixel: 1) the probability of belonging to a field; 2) the probability of being part of a boundary; and 3) the distance to the closest boundary. These labels can then be combined to obtain closed field boundaries. Using a single composite image from Sentinel-2, the 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 only marginally decreased accuracy. We then showed in a series of experiments that our model generalised well across resolutions, sensors, space and time without recalibration. Building consensus by averaging model predictions from at least four images acquired across the season is the key to coping with the temporal variations of accuracy. By minimising image preprocessing requirements and replacing local arbitrary decisions by data-driven ones, our approach is expected to facilitate the extraction of individual crop fields at scale.

• The paper introduces ResUNet-a, a CNN-based multi-task method that automatically extracts field boundaries from satellite images.
• It employs residual blocks, atrous convolutions, and pyramid scene parsing pooling to effectively capture multiscale contextual features.
• Results on Sentinel-2 imagery from South Africa achieved an MCC of 0.82 and 99.4% field detection, demonstrating strong performance and scalability.

Deep Learning Approach for Field Boundary Extraction from Satellite Imagery

The paper entitled "Deep learning on edge: extracting field boundaries from satellite images with a convolutional neural network" proposes a method for automatic extraction of field boundaries using convolutional neural networks (CNNs), specifically ResUNet-a, which is designed to perform multi-task semantic segmentation. This approach seeks to alleviate the manual and time-intensive process traditionally required for delineating field boundaries in digital agriculture applications, while also aiming to enhance the precision of agricultural monitoring and assessment processes.

Overview of Methodology

The authors frame the field boundary extraction problem as a multi-task semantic segmentation challenge, where the goal is to output several correlated classification masks. ResUNet-a is adapted to yield three main outputs: the probability map for field extent, the boundary delineation, and the distance from each pixel to the nearest boundary. Additionally, the network simultaneously reconstructs the input image to facilitate model stability during training. This multitasking paradigm is hypothesized to enhance generalization abilities by exploiting correlated spatial features present in field images.

The architecture leverages the UNet backbone with layers composed of residual blocks and atrous convolutions to capture multiscale features efficiently. Pyramid Scene Parsing Pooling helps incorporate context, aiding in effectively distinguishing field boundaries from other land cover types. Post-processing techniques, namely thresholding and watershed segmentation, are introduced to refine the outputs and achieve instance segmentation, thereby enabling the delineation of individual fields with closed boundaries.

Results and Performance Evaluation

The model was trained on Sentinel-2 monthly composites of South Africa, reaching a high level of accuracy with a Matthew's Correlation Coefficient (MCC) of 0.82 for extent detection and successfully identifying 99.4% of fields in the test area. Compared to conventional edge detection methods, ResUNet-a notably outperformed by providing clearer and more accurate delineations, showing its advantages in detecting meaningful contextual boundaries in complex agricultural landscapes.

The real strength of the proposed method lies in its ability to generalize across various complexities. Experiments demonstrated successful application not only to single-date imagery but also across different sensors like Landsat-8, implying robust multispectral compatibility and spatial resolution adaptability. Moreover, by averaging predictions from multiple dates, the authors suggest building consensus, which effectively reduces temporal variability in accuracy and enhances robustness of field boundary extraction.

Implications and Future Directions

The results suggest that ResUNet-a's ability to generalize well across resolutions, sensors, time, and geographical areas could significantly lower preprocessing requirements and facilitate scalable field boundary extraction. The reduced reliance on multitemporal features points to the importance of hierarchical contextual features learned through CNNs, opening avenues for improved models for temporal analyses.

For practical applications, training on large blocks of data using cloud-free composites may best exploit spatial information, while inference on single dates with consensus is efficient for operational implementations. Future work could explore further integration of temporal dynamics, not fully leveraged in this paper, alongside advances in instance segmentation models like Faster R-CNN and Mask R-CNN, which promise enhancements in closed boundary accuracy.

Ultimately, this deep learning approach marks progress toward automated, scalable field boundary delineation, promising improvements in precision agriculture, crop mapping, and resource management.