Land Cover and Land Use Detection using Semi-Supervised Learning

Abstract

There have been considerable advancements in semi-supervised learning in the remote sensing community. It is a technique that uses a small number of labeled data to train a model. Generally, deep learning networks learn from labeled data only. But since finding a huge corpus of a labeled dataset is rare and manually labeling datasets is timeconsuming and expensive. And labeling remote sensing satellite images is much more challenging than typical image datasets with good accuracy. Our proposed method aims to solve the problem for labelling unlabelled data with better accuracy. We use a SSL technique with a proper class-rebalancing technique to help solve the imbalanced dataset problem. We do it by creating “artificial” labels and training a model to gain reasonable accuracy. Moreover, it is a common occurrence that datasets are typically class-imbalanced. And if they are trained using it, with a high number of samples, the model becomes biased towards the majority classes and away from minority classes having few examples. This becomes a primary problem to the poor performance of an SSL model. We use a distribution alignment strategy to iteratively redistribute the classes through re-sampling. We showed that our proposed method improve a stateof- the-art SSL method with a tweaked augmentation strategy to generate high-quality pseudo-labels, updating the labeled set handling imbalanced data through re-sampling and also can reduce model bias. This is done on various class-imbalanced satellite image datasets. This method consistently outperforms other methods and greatly reduces the need for labeled data and also solves the issue of class imbalance in datasets.