Landcover and crop type classification with intra-annual times series of sentinel-2 and machine learning at central Portugal

Abstract

Land cover and crop type mapping have benefited from a daily revisiting period of sensors such as MODIS, SPOT-VGT, NOAA-AVHRR that contains long time-series archive. However, they have low accuracy in an Area of Interest (ROI) due to their coarse spatial resolution (i.e., pixel size > 250m). The Copernicus Sentinel-2 mission from the European Spatial Agency (ESA) provides free data access for Sentinel 2-A(S2a) and B (S2b). This satellite constellation guarantees a high temporal (5-day revisit cycle) and high spatial resolution (10m), allowing frequent updates on land cover products through supervised classification. Nevertheless, this requires training samples that are traditionally collected manually via fieldwork or image interpretation. This thesis aims to implement an automatic workflow to classify land cover and crop types at 10m resolution in central Portugal using existing databases, intra-annual time series of S2a and S2b, and Random Forest, a supervised machine learning algorithm. The agricultural classes such as temporary and permanent crops as well as agricultural grasslands were extracted from the Portuguese Land Parcel Identification System (LPIS) of the Instituto de Financiamento da Agricultura e Pescas (IFAP); land cover classes like urban, forest and water were trained from the Carta de Ocupação do Solo (COS) that is the national Land Use and Land Cover (LULC) map of Portugal; and lastly, the burned areas are identified from the corresponding national map of the Instituto da Conservação da Natureza e das Florestas (ICNF). Also, a set of preprocessing steps were defined based on the implementation of ancillary data allowing to avoid the inclusion of mislabeled pixels to the classifier. Mislabeling of pixels can occur due to errors in digitalization, generalization, and differences in the Minimum Mapping Unit (MMU) between datasets. An inner buffer was applied to all datasets to reduce border overlap among classes; the mask from the ICNF was applied to remove burned areas, and NDVI rule based on Landsat 8 allowed to erase recent clear-cuts in the forest. Also, the Copernicus High-Resolution Layers (HRL) datasets from 2015 (latest available), namely Dominant Leaf Type (DLT) and Tree Cover Density (TCD) are used to distinguish between forest with more than 60% coverage (coniferous and broadleaf) such as Holm Oak and Stone Pine and between 10 and 60% (coniferous) for instance Open Maritime Pine. Next, temporal gap-filled monthly composites were created for the agricultural period in Portugal, ranging from October 2017 till September 2018. The composites provided data free of missing values in opposition to single date acquisition images. Finally, a pixel-based approach classification was carried out in the “Tejo and Sado” region of Portugal using Random Forest (RF). The resulting map achieves a 76% overall accuracy for 31 classes (17 land cover and 14 crop types). The RF algorithm captured the most relevant features for the classification from the cloud-free composites, mainly during the spring and summer and in the bands on the Red Edge, NIR and SWIR. Overall, the classification was more successful on the irrigated temporary crops whereas the grasslands presented the most complexity to classify as they were confused with other rainfed crops and burned areas.