Introduction and Motivation

This research tackles the complexities of remote sensing for moisture detection in plant canopies, with a focus on the sparse vegetation of the Atacama Desert, home to the unique plant species Tillandsia landbeckii. Central to this challenge is the underrepresentation of the shortwave infrared (SWIR) domain in prevalent satellite sensors like Sentinel-2 and Landsat. This spectral limitation significantly hinders large-scale monitoring, particularly for species like T. landbeckii, whose spectral characteristics are camouflaged within its inorganic surroundings yet are distinctly visible in the SWIR spectrum. The Environmental Mapping and Analysis Program (EnMAP) emerges as a promising solution to this 'spectral gap'. However, the effective utilization of EnMAP's capabilities is not without challenges, primarily due to the intricacies involved in interpreting and leveraging data of varying scales and resolutions. This is especially true when contending with lower resolution imagery, a common issue in remote sensing. Our study delves into advanced methodologies to harness EnMAP's potential, aiming to significantly advance our capability to monitor and understand vegetation moisture in arid landscapes like that of the Atacama Desert.

Methodology

In March 2023, during our comprehensive field campaign in the Atacama Desert, we utilized a variety of tools including UAV-LiDAR, UAV-SWIR, UAV-VIS-NIR, and ASD-FieldSpec-4, alongside special satellite tasking with Pléiades-Neo, WorldView3-SWIR, and EnMAP. This integrated approach, merging ground and satellite data, generated a validated multiscale dataset. Our unique SWIR camera system, equipped with custom filters, filled the spectral gaps of satellite imagery, providing high-resolution, ground truth-like data. To bridge the gap between UAV and larger-scale satellite data, we applied advanced Spectral Unmixing (SMA) and Linear Mixing Models (LMM). This technique, crucial for discerning smaller objects in low-resolution images, enabled us to dissect the 30 m EnMAP pixel data, identifying distinctive surface reflections of key endmembers in our study area.

Results and Discussion

Our analysis of over 200 spectra provided initial insights into the spectral fingerprints of T. landbeckii and its surroundings. Spectroradiometer measurements identified significant band-depth differences at 680, 1437, and 2090 nm wavelengths. Utilizing Spectral Mixing Analysis (LMA) with EnMAP imagery, we identified Tillandsia cover at both 30 m and 60 m pixel resolutions. By combining this with a Support Vector Machine (SVM) classification from UAV imagery for reference and validation, we assessed the fraction cover per pixel. The regression analysis between these methods showed an R² value of 0.51 at 30 m and 0.66 at 60 m, underlining a moderate correlation. Despite some overestimation by EnMAP, this indicates its potential for regional analysis of Tillandsia coverage. Our goal is to enhance our models to extend these local insights across the Atacama using both EnMAP's hyperspectral and WorldView-3's multispectral imagery, thereby effectively merging UAV and satellite data.

Accurate land use and land cover (LULC) mapping is crucial for environmental monitoring and sustainable development. Especially in regions like Amazônia Legal, in which major deforestation and degradation events happen all year around. In this context, supervised deep learning models for classification tasks demand high quality training labels. However, these training labels are not abundantly available as they usually require the expertise and time of human annotators. For this reason it is feasible to make use of already classified maps, such as those of the MapBiomas Brasil project. The yearly classification maps from MapBiomas are derived from 30m Landsat imagery through an ensemble of machine learning classification models. These models are in itself not accurate, especially when distinguishing between a high number LULC classes. Even more uncertainty is introduced through a spatial resolution mismatch when using higher-resolution satellite data, like the 10m Sentinel-2 data used in this study. This study focuses on the detection and filtering of noisy classification labels within LULC maps, specifically those provided by the MapBiomas project, which distinguishes 29 classes within 5 macro classes. Our approach involves a novel approach for label noise detection that leverages clustering in the spectral domain of pixels. This identifies and filters noisy labels, which should lead to improved reliability of training data for subsequent deep learning tasks. To address this the pixels are clustered through multiple self-organizing maps (SOMs), which build class-wise anchor points for further filtering. Notably, our filtering approach is independent of any learned model, providing a versatile and adaptable solution for various classification tasks. In addition to clustering-based label noise detection, we conduct experiments incorporating feature extractors that capture both spatial and spectral relationships of pixels and their neighbors. These extractors are designed to augment the clustering process, providing the SOMs with more information to improve clusterability. The effectiveness of our approach is evaluated through comprehensive experiments, comparing the performance of models trained with and without label filtering, as well as a qualitative visual assessment. In preliminary results the proposed methodology rejects approximately 19% of the labels. When reassigning these to the class within their cluster, a qualitative improvement in the map is already visible. Finer structures like roads and rivers, which were missing in the classification map, are detected and labeled properly. The final results of this study will be presented at the upcoming conference, providing valuable insights and practical solutions for addressing noisy classification labels in LULC mapping, with potential applications beyond the Amazon region.

Das Bundesamt für Kartographie und Geodäsie (BKG) arbeitet im Projekt „KI-basierte Analyse in der Fernerkundung“ an der Überführung von Anwendungen der Künstliche Intelligenz (KI) im Bereich Fernerkundung von der Forschung in die angewandte Praxis. Dies unterstützt die Rolle des BKG als Geodatenbroker und Berater für Geodaten sowie –methoden und soll zu einer schnelleren und genaueren Klassifizierung der Polygone des Landbedeckungsmodells führen. Dieses Projekt wird durch das Deutschen Zentrum für Luft- und Raumfahrt e.V. (DLR) und das BKG gemeinsam bearbeitet und vorangebracht.

Recent advances in remote sensing technology have increased the availability of high-quality image data, enabling the monitoring and understanding of Earth’s physical processes. Depending on the sensors’ characteristics, the collected data can provide complementary information with different characteristics for the same observed region. For example, aerial imagery can deliver textural information about the surface with a resolution in the range of decimetres, but usually with high revisit times. On the other hand, satellite systems allow for short revisit times, so that the resultant images can capture temporal changes and patterns, but usually at a coarser spatial resolution, e.g. with a ground sampling distance (GSD) of 10 m or more. The goal of this paper is to jointly use aerial images and multi-temporal information from satellite image time series (SITS) to improve land cover classification. This opens up challenges of how to effectively combine these complementary sources of information to leverage their potential.