Hedgerow object detection in very high-resolution satellite images using convolutional neural networks

Abstract

Hedgerows are one of the few remaining natural landscape features within European agricultural areas. To facilitate hedgerow monitoring, cost-effective and accurate mapping of hedgerows across large spatial scales is required. Current methods used for automatic hedgerow detection are overly complicated and generalize poorly to larger areas. We examine the application of transfer learning using two neural networks (Mask R-CNN and DeepLab v3+) for hedgerow mapping in south-eastern Germany using IKONOS imagery. We demonstrate the potential of such networks for hedgerow monitoring by investigating performances across varying input image bands, seasonal imagery, and image augmentation strategies. Both networks successfully detected hedgerows across a large spatial scale (562 km 2 ), with DeepLab v3+ (75% F1-score) outperforming Mask R-CNN. Differences between band combinations were minimal, implying hedgerow detection could be achieved using RGB sensors. Results suggested that using all available training images across seasons is preferred and should have the same model generalizing effects as data augmentation. Experiments with varying data augmentations found augmentations effecting object geometries to greatly increase performance for both networks while results using augmentations modifying pixel spectral values showed concerning effects. Overall, our study finds that transfer learning in neural networks offers a simplified approach that outperforms previously established methods.

High-Throughput Plot-Level Quantitative Phenotyping Using Convolutional Neural Networks on Very High-Resolution Satellite Images

“…The superpixel and per-pixel methods require that precise plot boundaries are known, whereas the centred method only requires the centre of the plot, making the centred method the most flexible. Additionally, most existing works at a similar scale pose per-pixel problems [22,24]. Thus, our per-pixel models are most appropriate for transfer learning between our task and other remote sensing tasks.…”

“…There now exist works which resolved counties at hundreds of metres per pixel [18], fields at tens of metres per pixel [19], and individual trees at <1 m per pixel [20,21]. Operating at a sub-metre pixel resolution has created new opportunities for detecting and describing crops, fields, and farm infrastructure with unprecedented precision [22][23][24]. In particular for this work, the plots used in crop trials are visually distinct in these highest-resolution satellite images.…”

Deep Learning-Based Defect Detection Framework for Ultra High Resolution Images of Tunnels

“…There are several preprocessing methods for using UHR images as input for DCNN (n convolutional neural Network)-based complex detection models, including resizing the image itself [13,14,24]. In this study, a method of splitting UHR images into patches of appropriate size was used to minimize information loss.…”

Detection and Monitoring of Woody Vegetation Landscape Features Using Periodic Aerial Photography

“…The monitoring of such woody vegetation landscape features requires higher precision to capture changes in narrow linear structures and small patches of non-forest greenery, which can be only a few meters in width or diameter. Using very high spatial resolution (VHR) satellite images can provide a necessary solution [11,12], but is less attractive due to the higher costs involved in periodic updates. Satellite data were, therefore, not used in this study.…”