Local Climate Zones Classification Based On Remote Sensing With Small Sample Data

Global warming and urbanization are two major environmental issues in the 21 st century,which has aggravated the urban heat island effect(UHI)and seriously threatened the ecological environment and human health.Mitigating the urban heat island effect is of great significance for a sustainable city,while the Local Climate Zones(LCZ)is the most remarkable one in the recent study.Based on the extraction of temperature-sensitive indicators from land surface elements,LCZ simplifies the global urban area into 17 categories,providing a refined quantitative method of urban local thermal environment for urban heat island.It has become a benchmark framework for global urban research.The classification of LCZ is of great theoretical and practical significance to the related research of urban climate.The characteristics of wide coverage,short acquisition period,cost-effective,and convenience for remote sensing imagery make the remote sensing-based LCZ classification the most common method.With the support of the sufficient labeled samples,this method has great performance.However,LCZ samples reply to manual annotation with expert knowledge and field verification,requiring huge time and labor costs.The limited number of labeled samples and insufficient sample information available for training cause poor performance,which is called the problem of small samples.Driven by the application background of global-local climatic zone mapping,LCZ classification has developed from single-region to cross-region research,and the problem of small samples is the biggest challenge,mainly including:(1)The small samples of LCZ classification in a single region.In field-verifiable labeled areas,different LCZ has poor visual discrimination on remote sensing images,which makes it challenging to obtain the high-quality labels objectively,bringing about the poor prediction from traditional single-region LCZ classification methods.(2)The small samples of LCZ classification across regions.In areas unavailable for verifying,the global area can be regarded as the whole region,and the samples from the marked areas are used as the source domain data for training.In contrast,the unknown areas are used as the target domain data for prediction.However,the limited labeled samples severely restrain the generalization performance of cross-regional LCZ classification.For overcoming the challenge of small sample problems in single-region and cross-region LCZ classification,this thesis proposes two methods with full use of data level and feature level information,the conditional random fields(CRF)-based selftraining method for LCZ classification and the deep transfer learning-based domain adaptation method for LCZ classification.The main contents and contributions can be concluded as follows:(1)Systematically summarizing the basic theory and method of LCZ.Based on the complexity of the LCZ concept,this thesis introduces the background and the related works while analyzes the problem of limited samples in LCZ classification from a single region to across regions.Besides,the classic LCZ classification with WUDAPT is also provided.(2)In single-regional LCZ classification,the CRF-based self-training method for LCZ classification is proposed.Based on the self-training algorithm,a high-confidence sample selection strategy for LCZ combined with spatial information is carried out,which can automatically enrich the training dataset.At the same time,the traditional methods only rely on single-pixel information,which is prone to generate much noise,the CRF is introduced to supplement implicit spatial-contextual information through its flexible modeling ability,which can effectively mitigate the poor classification performance from the limited sample.(3)In cross-regional LCZ classification,the deep transfer learning-based domain adaptation method for LCZ classification is proposed.The deep convolutional neural network is used for automatic feature extraction,and the source domain data and target domain data are mapped to the feature space.While training the network on the source domain,the unlabeled information of the target domain is also considered to construct a domain adaptive loss function,which is beneficial for the alignment of feature space and reduction the difference between the feature distribution of the source and target domain,enhancing the generalized performance in the unlabeled target domain.This thesis researches remote sensing-based LCZ classification on the small sample problems in single-region and cross-region,the proposed CRF-based selftraining method,and the deep transfer learning-based domain adaptation method for LCZ classification can effectively improve LCZ classification accuracy under small sample conditions.It is promoting the construction of a global urban LCZ database,which has important scientific significance and social value for urban heat island,urban planning,and other fields.