A Method for Multitemporal Classification of PlanetScope Images for Detailed Land Cover Analysis

Abstract

Multitemporal classification faces challenges such as varying temporal conditions and the limitations of using the same model over different periods or requiring training samples selection for each time stamp. These challenges necessitate advanced strategies to improve efficiency and accuracy, even in relatively small areas. Existing literature has explored various approaches to address these issues, highlighting the importance of feature selection and algorithm choice. This paper presents a method for classifying PlanetScope images to monitor changes in land cover, specifically urban areas, grasslands, bare soil, and forest over a reduced area, covering a region of interest of approximately 4 km2 over the period 2021-2022. To achieve this goal, several machine learning algorithms were applied, resulting in the random forest as the best performing one with an overall accuracy of 100% in the training process. The Support Vector Machine model also showed a high accuracy with 94%. The models were trained using 30 training samples per class, selected by photointerpretation and over different features such as the Normalized Difference Vegetation Index, the spectral bands R, G, B, and NIR of the Planet images, a grayscale image obtained from converting a BGR false color composition to grayscale, and a Gabor filter. The use of these features helped in dealing with classification in areas smaller than 5 km2 and limited data availability for training a classification model.