Land cover information is fundamental for a wide range of fields, such as research and policymaking. Remote sensing has historically been a source of data on land cover and recognized as the only practical systematic and wall-to-wall source for crop mapping. The European Copernicus programme and its free data policy for Sentinel-2 made accessible large volumes of imagery for frequent mapping and updating, generating new challenges. One such challenge is timely mapping through supervised image classification. The need for a prompt classification workflow requires training to become automatic, which typically relies on samples collected manually via fieldwork or image interpretation. Another challenge is to map land cover classes that traditionally have been troublesome to identify when satellite observations were sparse. For instance, crops have a spectral response that changes substantially throughout the year or during narrow time windows, which cannot be observed with few image acquisitions. This paper presents research ongoing in Portugal to develop a methodology for automatic image classification using training samples labelling with no human intervention. Rather, auxiliary datasets are used to randomly extract labelled points from large training samples to produce a land cover and crop map in raster format at 10 m spatial resolution using 2018 Sentinel-2 images. The proposed methodology was tested with the Random Forest classifier achieving an overall accuracy of 76%. These results are promising and support the idea of refining the methodology to move towards an annual land cover map production at the national scale.

Remote sensing is also benefiting from the quick development of deep learning algorithms for image analysis and classification tasks. In this paper, we evaluate the classification performance of a well-known Convolutional Neural Network (CNN) models, such as ResNet50, using a transfer learning approach. We compare the performance when using vector-features acquired from general purpose data, such as the ImageNet [1], versus remote sensing data like BigEarthNet [2], UCMerced [3], RESISC45 [4] and So2Sat [5]. The results show that the model pre-trained on RESISC-45 data achieved the highest accuracy when classifying the Eurosat [6] testing dataset. This was followed by the model pre-trained on Imagenet with 95.94% and BigEarthNet with 95.93%. When presented with diverse remote sensing data, the classification improved in regards to large quantities of general-purpose data. The experiments carried out also show, that multi modal (co-registered synthetic aperture radar and multispectral) did not increase the classification rate with respect to using only multispectral data. The source codes of this work are available for reproducible research at https://github.com/itzahs/CNN-RS.

Current and upcoming Sun-Induced chlorophyll Fluorescence (SIF) satellite products (e.g., GOME, TROPOMI, OCO, FLEX) have medium-to-coarse spatial resolutions (i.e., 0.3–80 km) and integrate radiances from different sources into a single ground surface unit (i.e., pixel). However, intrapixel heterogeneity, i.e., different soil and vegetation fractional cover and/or different chlorophyll content or vegetation structure in a fluorescence pixel, increases the challenge in retrieving and quantifying SIF. High spatial resolution Sentinel-2 (S2) data (20 m) can be used to better characterize the intrapixel heterogeneity of SIF and potentially extend the application of satellite-derived SIF to heterogeneous areas. In the context of the COST Action Optical synergies for spatiotemporal SENsing of Scalable ECOphysiological traits (SENSECO), in which this study was conducted, we proposed direct (i.e., spatial heterogeneity coefficient, standard deviation, normalized entropy, ensemble decision trees) and patch mosaic (i.e., local Moran’s I) approaches to characterize the spatial heterogeneity of SIF collected at 760 and 687 nm (SIF₇₆₀ and SIF₆₈₇, respectively) and to correlate it with the spatial heterogeneity of selected S2 derivatives. We used HyPlant airborne imagery acquired over an agricultural area in Braccagni (Italy) to emulate S2-like top-of-the-canopy reflectance and SIF imagery at different spatial resolutions (i.e., 300, 20, and 5 m). The ensemble decision trees method characterized FLEX intrapixel heterogeneity best (R² > 0.9 for all predictors with respect to SIF₇₆₀ and SIF₆₈₇). Nevertheless, the standard deviation and spatial heterogeneity coefficient using k-means clustering scene classification also provided acceptable results. In particular, the near-infrared reflectance of terrestrial vegetation (NIRv) index accounted for most of the spatial heterogeneity of SIF₇₆₀ in all applied methods (R² = 0.76 with the standard deviation method; R² = 0.63 with the spatial heterogeneity coefficient method using a scene classification map with 15 classes). The models developed for SIF₆₈₇ did not perform as well as those for SIF₇₆₀, possibly due to the uncertainties in fluorescence retrieval at 687 nm and the low signal-to-noise ratio in the red spectral region. Our study shows the potential of the proposed methods to be implemented as part of the FLEX ground segment processing chain to quantify the intrapixel heterogeneity of a FLEX pixel and/or as a quality flag to determine the reliability of the retrieved fluorescence.

Earth data collection from satellites and aircraft has exponentially grown, but a substantial portion of it remains unlabeled. This has prompted the remote sensing community to explore effective methods for leveraging unlabeled data. In our prior investigation [1], we evaluated various deep semi-supervised learning algorithms on two very high-resolution (VHR) optical datasets (UCM [2] and AID [3]). Notably, the CoMatch [4] algorithm demonstrated the highest accuracy, motivating further exploration. This letter extends our earlier work by integrating the established Class-Aware Contrastive Semi-Supervised Learning framework (Comatch+CCSSL) [5] into CoMatch and introducing a new triplet metric learning loss (CoMatch+Triplet). CoMatch+Triplet excelled with 93.2% accuracy on UCM, while CoMatch led with 92.19% on AID. The addition of the triplet loss can produce a clearer separation of the samples from different classes in the embedding space at very early learning stages, being able to learn faster and getting maximum performance with few iterations. The exploration of diverse semi and self-supervised training methodologies presented in this work sheds light on the strengths and limitations of these approaches, enhancing our understanding of their applicability in remote sensing applications.