Domain adaptation for land use classification: A spatio-temporal knowledge reusing method

Land use classification requires a significant amount of labeled data, which may be difficult and time consuming to obtain. On the other hand, without a sufficient number of training samples, conventional classifiers are unable to produce satisfactory classification results. This paper aims to overcome this issue by proposing a new model, TrCbrBoost, which uses old domain data to successfully train a classifier for mapping the land use types of target domain when new labeled data are unavailable. TrCbrBoost adopts a fuzzy CBR (Case Based Reasoning) model to estimate the land use probabilities for the target (new) domain, which are subsequently used to estimate the classifier performance. Source (old) domain samples are used to train the classifiers of a revised TrAdaBoost algorithm in which the weight of each sample is adjusted according to the classifier's performance. This method is tested using time-series SPOT images for land use classification. Our experimental results indicate that TrCbrBoost is more effective than traditional classification models, provided that sufficient amount of old domain data is available. Under these conditions, the proposed method is 9.19% more accurate.