Land cover is one of the main environmental climate variables (ECVs) as it is highly correlated with climate change. In this context, in the framework of the Climate Change Initiative (CCI) of ESA , the High Resolution Land Cover (HRLC) project is aimed to study the role of the spatial resolution in the mapping of land cover and land-cover changes to support climate modelling research . Land cover and related changes are indeed both cause and consequence of human-induced or natural climate changes. This has been demonstrated by the previous phase of the CCI program, focused on the generation of Medium Resolution (MR) Land Cover maps at global scale. Differently from the MR land cover CCI, which provided annual land cover maps at 300m resolution in the period 1992-2020 , the HRLC project produces regional maps characterized by a spatial resolution of 10m/30m. Moving from 300m to 30m requires the definition of new data analysis methods, reframing the perspective with respect to the MR project both from the theoretical and the operational viewpoints. Although HR potentially increases the capability of a detailed analysis of spatial patterns in the land cover, many challenges are introduced with respect to the MR case and limitations in the available data make the development of products at very large scale very challenging.
This contribution presents the architecture and the methodologies developed for implementing the full processing chains that have been developed to process Earth Observation (EO) data and generate the HRLC products. The primary products of the project consist of: (i) HR land-cover maps at subcontinental scale at 10m as reference static input (generated for 2019 only) to the climate models, (ii) a long-term record of regional HR land cover maps at 30m in the regions identified for the historical analysis every 5 years (generated in the period 1990-2015), and (iii) change information at 30m at yearly scale consistent with historical HR land-cover maps.
The development of the proposed architecture was based on the observation that temporal availability of HR data in the past/current archives is much lower than that of the MR ones and strongly varies across the years. Differently from the MR case (e.g., SPOT-Vegetation archive), no daily acquisitions are available and only in the very recent years it was possible to get a quite dense temporal sampling due to Sentinel and Landsat-8 missions. Prior to them, the number of yearly-based images available in archives dramatically reduces (being Landsat Thematic Mapper, ASAR and ERS-1 and 2 the most relevant data sources), resulting in a much more challenging problem for the development of HRLC products. This scenario led to a complex process to produce historical time series of products. Moreover, it required a shift in the processing paradigm that moves from the analysis of many images per year acquired at MR to a few images (for some areas and years single or no images are available) characterized by high spatial resolution.
To produce the land-cover maps, two multisensor (optical and SAR) processing chains have been designed and implemented: one is based on the exploitation of Sentinel 1 (S1) and Sentinel 2 (S2) images for the generation of maps at 10 m resolution (used in the project for generating products in 2019) (Figure 1) and the other one generates historical maps every 5 years going back to 1990 by exploiting Landsat (Enhanced) Thematic Mapper images and ASAR and ERS-1/2 data (Figure 2). Both architectures share two pre-processing branches (one for optical and the other for SAR data) and a fusion module for the final map production. The main difference between the two processing chains is related to the pre-processing techniques (which consider the large differences in data quality and availability between Sentinel and previous missions) and in the paradigm exploited for the classification mechanism. The S1/S2 architecture classifies independently the time series of images acquired in the target year (in the project 2019) and generates the land-cover products by fusing the classification results obtained independently on the two branches (optical and SAR) by using consensus theory and Markov Random Field approaches . The historical processing chain assumes as baseline the classification results generated with S1/S2 data and exploits the cascade classification paradigm  to properly model the temporal correlation between images when producing the historical land-cover maps. This is done to mitigate the well-known problem of error propagation in multitemporal classification, which is extremely critical when independent classification of multitemporal data is performed. The cascade classification paradigm is robust and theoretically well founded as it is based on the Bayesian decision theory. The classification techniques included in the optical and SAR branches include “shallow” machine learning techniques (Support Vector Machines, Random Forest) and specific SAR detectors focused on built -up and water related classes . Both architectures provide in output uncertainty measures for the classification of each pixel in the map and also indications on the second alternative class for a better representation of the real complex conditions on the ground. These are crucial information to be given as input to the climate modelling task when using the generated products. Specific methodologies have also been devised to support the definition of the training sets to be used for the 2019 and the historical image classifications .
To produce land-cover change maps every year, a third architecture has been defined (figure 3) that is driven by the cascade classification output and is aimed at identifying the location in time of the changes on a yearly base. This allows to localize in time the changes occurred between 5 years maps. The change detection products have been generated by using optical data acquired by Landsat (Enhanced) Thematic Mapper. The main challenge is related to the very uneven distribution of data available in different areas and in different years. This was addressed by defining an architecture based on a feature extraction module, a time series regularization module (based on a “shallow” neural network) and an abrupt change detection module . The change detection products have associated reliability information for each pixel in terms of the probability of change.
3. Product generation and conclusion
The processing chain has been developed according to the use of dockers and with the requirement to be able to process big data volumes of optical and SAR images. Processors have been fully integrated in Python-based pipelines that automatically retrieve the needed products for the specific task and perform the processing. The production has been run on Amazon Web Services (AWS) cloud computing, even if the processing chain is flexible and can be run on DIAS and other cloud infrastructures.
The Climate User Group involved in this project defined three large regions of particular interest to study the climate/LC feedbacks in three continents involving climate (tropical, semi-arid, boreal) and complex surface atmosphere interactions that have significant impact not only on the regional climate but also on large-scale climate structures. The three regions are in Amazon basin, the Sahel band in Africa and in the in the northern high latitudes od Siberia.
The products generated on the three areas have accuracies that given the complexity of the task and of the legend of land covers classes (which includes also seasonal classes) are satisfactory (see the “ESA CCI High Resolution Land Cover Products” presentation).
 ESA – European Space Agency: ESA Climate Change Initiative description, EOP-SEP/TN/0030-09/SP, Technical Note – 30 September 2009, 15 pp., 2009.
 L. Bruzzone et al, "CCI Essential Climate Variables: High Resolution Land Cover,” ESA Living Planet Symposium, Milan, Italy, 2019.
 P. Defourny et al (2017). Land Cover CCI Product User Guide Version 2.0. [online] Available at: http://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdf
 D. Tuia, M. Volpi, G. Moser, “Decision fusion with multiple spatial supports by conditional random fields,” IEEE Trans. Geosci. Remote Sens., 56, 2018
 L. Bruzzone, R. Cossu, A multiple cascade-classifier system for a robust a partially unsupervised updating of land-cover maps, IEEE Trans. on Geosci. Remote Sens., Vol. 40, 2002.
 T. Sorriso, D. Marzi, P. Gamba, ”A General Land Cover Classification Framework for Sentinel-1 SAR Data.” in Proc. of the 2021 IEEE 6th Int. Forum on Research and Tech. for Society and Industry (RTSI), Napoli,. 2021.
 C. Paris, L. Orlandi, L. Bruzzone, “A Strategy for an Interactive Training Set Definition based on Active Self-Paced Learning,” IEEE Geosci. Remote Sens. Letters, Vol. 17, 2021.
 Y.T. Solano-Correa, K. Meshkini, F. Bovolo, L. Bruzzone, “A land cover-driven approach for fitting satellite image time series in a change detection context,” SPIE Conf. on Image and Signal Processing for Remote Sensing XXVI, 2020.
List of the other HRLC team members: C. Domingo (CREAF), L. Pesquer (CREAF), C. Lamarche (UCLouvain), P. Defourny (UCLouvain), L. Agrimano (Planetek), A. Amodio (Planetek), M. A. Brovelli (PoliMI), G. Bratic (PoliMI), M. Corsi (eGeos), C. Ottlé (LSCE), P. Peylin (LSCE), R. San Martin (LSCE), V. Bastrikov (LSCE), P. Pistillo (EGeos), M. Riffler (GeoVille), F. Ronci (eGeos), D. Kolitzus (GeoVille), Th. Castin (UCLouvain), R. San Martin (LSCE-IPSL), C. Ottlé, V. Bastrikov (LSCE-IPSL), Ph. Peylin (LSCE-IPSL).