Comparative performance of machine learning algorithms for Forest Cover classification using ASI - PRISMA hyperspectral data

Remote sensing (RS) supports studies on land cover and land cover changes, agroforestry planning and helps in performing cross-checks in the framework of EU's common agricultural policy (CAP). Multispectral data (with limited number of spectral bands) and multi-temporal processing (sequence of scenes captured on different periods of time) were mainly used for RS. The use of hyperspectral imagery such as those provided by the PRISMA satellite, could enable further advances. This study aims at evaluating and comparing classification methods of the PRISMA product. For this purpose, the ASI pre-processed radiance level data (L2D) was used to investigate an area in the Lazio region (Italy). All the classification approaches tested are advanced techniques based on machine learning. Experimentation was carried out using K-nearest neighbors (KNN), Support Vector Machine (SVM) and Random Forest (RF) algorithms, which were compared in terms of accuracy, number of ground control points (GCP) required and ease of implementation. The study showed that the Random Forest approach presents the best compromise among the three proposed approaches, also considering its applicability, which is related to the wider literature and established experimentation in the GIS environment. The research is one of the first studies with the purpose of using the PRISMA data and presents some methodologies for the operational use of this data.