Parcel-Based Active Learning for Large Extent Cultivated Area Mapping

Abstract

This paper focuses on agricultural land cover mapping at a high-resolution scale and over large areas from an operational point of view and from a high-resolution monodate image. In this context, training data are assumed to be collected by successive journeys of field surveys and, thus, are very limited. Supervised learning techniques are generally used, assuming that the classes distribution is constant over the whole image. However, in practice, a data shift often occurs on large areas due to various acquisition conditions. To alleviate these issues, active learning (AL) techniques define an efficient training set by iteratively adapting it through adding the most informative unlabeled instances. They can improve the classification process efficiency while keeping a limited training dataset. The novelty in this paper is the application of AL techniques on multispectral images for agricultural land cover mapping, using field sampling instead of pixel sampling, which is rarely done in the literature. Besides, we proposed a parcel-based AL scheme that is suitable for an operational land cover mapping in cultivated areas since the parcel is an agricultural unit and field observations are processed at parcel scale. Random forests classifier was used. Results were processed on a 6 m multispectral Spot6 image over a 35 km $^2$ Mediterranean cultivated area, in Lebna Catchment, north eastern Tunisia. The contribution of AL techniques was assessed with comparison to a random and stratified random strategies for sampling new instances. For iterative sample selection, two criteria are used and often coupled: uncertainty and diversity. For diversity metric, a new clustering-based metric was proposed based on a mean-shift clustering, which improved the classification accuracy. AL techniques showed to be efficient with complex data and fine land cover legend improving random-based selection up to 10%. Besides, the maximum of classification accuracy is reached using mean-shift breaking ties metric in just 5-day field survey, i.e., 30 days less compared to the random selection. Finally, results showed that the finer the definition of land cover classes, the more crucial is the choice of AL metrics.