Efficient Data Utilisation for Semantic Segmentation in Aerial Imagery using Convolutional Neural Networks
This research addresses the challenge of improving data efficiency in training convolutional neural networks (CNNs) for the semantic segmentation of buildings in aerial imagery. We evaluate efficient techniques for label acquisition, including the use of weakly supervised partially labelled data and active learning strategies. Additionally, a novel loss function is introduced to facilitate the reuse of existing labels with new aerial images, effectively mitigating the issues arising from lateral shifts between masks and images.
The proposed ”Mask Shift Loss” function specifically addresses the displacement challenges observed between historical and recent aerial imagery sets. This function accounts for minor lateral shifts in segmentation masks, enabling the effective reuse of older building outline datasets. By employing this technique, new models can be trained on contemporary imagery with existing building outline datasets, minimizing the necessity for new annotations, especially in urban areas where building alterations between surveys are minimal.
Weakly supervised learning with partially labelled data, such as point, scribble, and superpixel annotations are explored. Models trained using these techniques were evaluated against benchmarks from fully supervised learning, on the building outlines detection problem. When applied to a multiclass ground cover segmentation task lacking full mask annotations, these methods demonstrated robust segmentation, high classification accuracy, and stable performance over time, with results closely matching human-labelled ground cover proportions.
Several active learning query methods are examined using an active learning framework with point training on the building outlines problem. Active learning led to faster training and increased accuracy over randomly chosen point annotations. Epistemic model uncertainty estimated using Monte Carlo Dropout is found to be an effective active learning query method. However, a minimum kernel-smoothed entropy query method achieves the best model accuracy of query methods examined. The smoothed entropy based active learning point annotations with a large random sample size yielded a model accuracy comparable to some fully supervised models evaluated in prior work. While fully supervised models provided inference results with finer details and more defined building outlines, the proposed techniques demonstrate remarkable promise in terms of training efficiency, accuracy, and reduced annotation effort. This research enables more efficient model training in building outline detection enabling applications in areas like urban planning, disaster recovery, and demographic studies.
