Effect of the Training Set Configuration on Sentinel-2-Based Urban Local Climate Zone Classification

Abstract

ABSTRACT:As any supervised classification procedure, also Local Climate Zone (LCZ) mapping requires reliable reference data. These are usually created manually and inevitably include label noise, caused by the complexity of the LCZ class scheme as well as variations in cultural and physical environmental factors. This study aims at evaluating the impact of the training set configuration, i.e. training sample number and imbalance, on the performance of Canonical Correlation Forests (CCFs) for a classification of the 11 urban LCZ classes. Experiments are carried out based on globally available Sentinel-2 imagery. Besides multi-spectral observations, different index measures extracted from the images as well as the Global Urban Footprint (GUF) and Open Street Map (OSM) layers are fed into the CCFs classifier. The results show that different LCZs favor different configurations in terms of training sample number and balance. Based on the findings, majority voting of different predictions from different configurations is proposed and performed. This way, a significant accuracy improvement can be achieved.

Urban Local Climate Zone Classification with a Residual Convolutional Neural Network and Multi-Seasonal Sentinel-2 Images

“…1, where both the majority voting-based result with multiseasonal Sentinel-2 images (left) and the average of the four singleseason results (right) are shown. The achieved classification accuracy is comparable to our previous work, where the same cross validation has been carried out with single-seasonal Sentinel-2 images and additional features (Global Urban Footprint and Open Street Map) (Qiu et al, 2018), using Canonical Correlation Forests as the classifier. Using the majority voting-based result, an exemplary LCZ map over the city of Munich, Germany, is shown in Fig.…”

“…Furthermore, it has to be highlighted that this multi-seasonal result relies only on Sentinel-2 data, and only on four independent, multi-seasonal predictions, which are fused in a simple decision fusion manner based on majority voting. This is in contrast to the more complex set-up that has been necessary in our previous work (Qiu et al, 2018), for which the majority voting-based fusion of 20 different CCF classifier results and external auxiliary data (the Global Urban Footprint and two Open Street Map layers) were necessary. Especially since these Open Street Map layers are not available globally, this shows the great potential of the multi-seasonal ResNet-based approach presented in this paper for global applicability and generalization.…”

Feature Importance Analysis for Local Climate Zone Classification Using a Residual Convolutional Neural Network with Multi-Source Datasets

“…However, there are also several small classes lacking the same amount of labeled samples: compact high-rise, compact low-rise, open high-rise, sparsely built, heavy industry, bush (scrub), bare rock or paved, bare soil or sand. It seems that balancing the training samples is necessary for LCZ mapping, as also investigated and shown by our previous work with a Canonical Correlation Forest (CCF) as the classifier [37]. Using the reflectance of Sentinel-2 over the same study area, Figure 9 illustrates the effect of different balancing methods for each LCZ.…”

Self-Training Classification Framework with Spatial-Contextual Information for Local Climate Zones

“…The range of the majority filter in the WUDAPT method was a 3 × 3 square window size, and the nonnegative constant λ in CRF was set as 0.5. It is well known that the configurations of the training set and testing set play important roles in the assessment of the LCZ classification [28,45,59]. Since the ground truth data was usually limited, 10 labeled samples were randomly selected in each class for simulating the insufficiency of labeled samples and the remaining samples were used as a testing set, which is widely used in many semi-supervised researches [60,61].…”