One of the challenges of training artificial intelligence models for classifying satellite images is the presence of label noise in the datasets that are sometimes crowd-source labeled and as a result, somewhat error prone. In our work, we have utilized three labeled satellite image datasets namely, SAT-6, SAT-4, and EuroSAT. The combined dataset consists of over 900,000 image patches that belong to a land cover class. We have applied some standard pixel-based feature extraction algorithms to extract features from the images and then trained those features with various machine learning algorithms. In our experiment, three types of artificial label noises are injected - Noise Completely at Random (NCAR), Noise at Random (NAR) and Noise Not at Random (NNAR) to the training datasets. The noisy data are used to train the algorithms, and the effect of noise on the algorithm performance are compared with noise-free test sets. From our study, the Random Forest and the Back-propagation Neural Network classifiers are found to be the least sensitive to label noises. As label noise is a common scenario in human-labeled image datasets, the current research initiative will help the development of noise robust classification methods for various relevant applications.
展开▼
