Evaluation of the Potential of Convolutional Neural Networks and Random Forests for Multi-Class Segmentation of Sentinel-2 Imagery

Abstract

Motivated by the increasing availability of open and free Earth observation data through the Copernicus Sentinel missions, this study investigates the capacity of advanced computational models to automatically generate thematic layers, which in turn contribute to and facilitate the creation of land cover products. In concrete terms, we assess the practical and computational aspects of multi-class Sentinel-2 image segmentation based on a convolutional neural network and random forest approaches. The annotated learning set derives from data that is made available as result of the implementation of European Union’s INSPIRE Directive. Since this network of data sets remains incomplete in regard to some geographic areas, another objective of this work was to provide consistent and reproducible ways for machine-driven mapping of these gaps and a potential update of the existing ones. Finally, the performance analysis identifies the most important hyper-parameters, and provides hints on the models’ deployment and their transferability.

Crop identification by massive processing of multiannual satellite imagery for EU common agriculture policy subsidy control

“…It shows the results generated in the models by gro uping years. As mentioned, there are several studies that use ANNs for the recognition of crop patterns (Ji et al, 2018;Kamilaris & Prenafeta-Boldú, 2018;Nevavuori et al, 2019;Syrris et al, 2019) and employ several satellite images of an area, studying some indexes such as the NDVI to learn the phenological pattern of a crop. In our study, we have chosen to use all available information fr om satellite sensors (the 12 bands) without any calculation of indexes that combine these bands.…”

“…Recently, several studies have used Convolutional Neural Networks (CNN), a DL algorithm, to identify different land covers. In this respect, Syrris et al (2019) experimented with four CNN variants to detect land cover in general (with a general class concerning crops): a standard one, a fully convolutional network (FCN) a U-net and a SegNet with four groups of 11 Sentinel-2 images (44 images all together). Other examples are the study by Nevavuori et al (2019) that used CNN to detect different types of crops, but with UAV images, and the work of Phartiyal and Singh (2018) that used 5 SAR and optical images.…”

Application of the Symbolic Machine Learning to Copernicus VHR Imagery: The European Settlement Map

“…Ongoing experiments are also exploring different quantization approaches, other than uniform, which could improve the results of the SML. Future work will also explore the added value of convolutional neural networks for the multiclass segmentation of the input imagery [15]. The acquisition of VHR imagery every three years by the Copernicus program provides an opportunity for regular monitoring and change detection starting with this layer.…”

Joint Supervised Classification and Reconstruction of Irregularly Sampled Satellite Image Times Series

“…Classification of land cover or land use using machine learning techniques has emerged as a major application of SITS. Deep learning was intensively studied either using only one or few temporal acquisitions with convolutional neural networks (CNN) [23], [24], or using the full temporal stack of acquisitions using recurrent neural networks (RNN) [25], [26]. Combination of RNN applied on SITS with CNN applied on SPOT-6 image was also investigated for land cover mapping over Reunion Island [27].…”