Pseudo-Labeling Approach for Land Cover Classification Through Remote Sensing Observations With Noisy Labels
Islombek Mirpulatov, Svetlana Illarionova, Dmitrii Shadrin, Evgeny Burnaev
Abstract
Satellite data allows us to solve a wide range of challenging tasks remotely, including monitoring changing environmental conditions, assessing resources, and evaluating hazards. Computer vision algorithms such as convolutional neural networks have proven to be powerful tools for handling huge visual datasets. Although the number of satellite imagery is constantly growing and artificial intelligence is advancing, the present sticking point in remote sensing studies is the quality and amount of annotated datasets. Typically, manual labels have particular uncertainties and mismatches. Also, a lot of annotated datasets available in low resolution. Available visual representation of the observed objects can be more detailed than annotation. This causes the need for markup adjustment, which can be referred to as a pseudo-labeling task. The main contribution of this research is that we propose a pipeline for pseudo-labeling to address the problem of inaccurate and low-resolution markup improvement for solving land-cover and land-use segmentation task based on the data from the Sentinel-2 satellite. Our methodology takes advantages both of classical machine learning (ML) and deep learning (DL) algorithms. We examined random sampling, uniform sampling, and K-Means sampling and compared it with the full dataset usage. U-Net, DeepLab, and FPN models were trained on the adjusted dataset. The achieved findings show that a simple yet effective approach of data preliminary sampling and further markup refinement leads to significantly higher results than just using raw inaccurate data in a deep neural network pipeline. Moreover, the considered sampling technique allows to use less data for ML model training. The experiments involve markup adjustment and up-scaling from 30m to 10m. We verify the proposed approach in precise test area with manual annotation and show the improvement in F1-score from 0.792 to 0.816.