Rethinking reference data quality: the role of mixed pixels in remote sensing classification

Current recommendations for supervised machine learning classification in Remote Sensing advocate for using only high-quality reference data, which is often associated to pure pixels or endmembers. This focus, however, overlooks a fundamental challenge: real-world environmental images are rarely composed primarily of pure pixels. As a result, using only pure pixels as training data can leave a significant portion of the image unrepresented and hinder adequate classification. To demonstrate this effect, we introduce the Reference Sample Selection (RSS) approach. RSS systematically varies pixel purity in the training dataset to assess its impact on classification results. Pixel purity is determined based on a finer spatial resolution image. In this study, we present a case study within a select region of the Brazilian Amazon Rainforest. We applied RSS to the commonly used medium spatial resolution data from the Sentinel-2 MultiSpectral Imager (MSI) to target four land cover types: forest, water, grassland, and bare soil. The analysis used three common shallow classifiers: K-Nearest Neighbours (KNN), Support Vector Machines (SVM), and Random Forests (RFR). Our results demonstrate that including pixels with varied purity levels can significantly alter classification accuracy, depending on the land cover class. This finding challenges the conventional definition of reference data quality and highlights the need for using training samples that represent the entire image, not just its purest components. This method is readily applicable to a wide range of Remote Sensing studies. The source code and used data are available at https://github.com/paeslemesa/rss_approach/ and https://www.kaggle.com/datasets/sabrinacorra/uruara-s2msi-samples.