Abstract
This article investigates the effectiveness of multiclassifier fusion technique on domain adaptation for remote sensing image classification. Since it is impossible to find a domain adaptation method that is optimal for different datasets, and it is also difficult to select the best base classifier for domain-invariant features, multiple domain adaptation fusion (MDAF) method and the multiple base classifier fusion (MBCF) method are proposed to achieve a more stable and superior classification performance. The most crucial step of the weighted fusion approach is to assign weights for classifiers. It is known that different classifiers have varied performances on different subsets of data, and therefore a samplewise adaptive weight is more desirable than a fixed one. For each sample, a desired weight should be able to characterize the reliability of a classifier, so that the advantages of different classifiers can be exploited. We propose a neighborhood consistency based adaptive weighting method, which assigns a large weight to a classifier on a sample if the prediction of the sample is consistent to the predictions of its local neighbors. Experiments with three remote sensing images demonstrate the efficiency of the proposed weighting strategy in the proposed MDAF and MBCF methods.
Highlights
R EMOTE sensing images are important and increasingly available for earth observation [1]
Aiming to solve the two aforementioned problems, we propose two multiclassifier fusion based algorithms for domain adaptation: one focuses on multiple domain adaptation fusion (MDAF), and the other aims to multiple base classifier fusion (MBCF) on domain invariant features
It is worth noting that the BOT and Kennedy Space Center (KSC) images were obtained from upland and wetland areas, which are not easy to visualize the differences in the whole classification maps
Summary
R EMOTE sensing images are important and increasingly available for earth observation [1]. Classification of remote sensing images faces the problem of limited labeled data. Semisupervised learning [2] and dimensionality reduction [3], [4] are popular methods to decrease the labeling cost. “borrowing” labeled information from some temporally or spatially separate image is an attractive strategy for classifying image with few labels or even without labels. Directly reusing the labeled data may not perform well on the target image due to the spectral drift between the two images [5], [6].
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have