The National Institute of Geographic and Forest Information (IGN) has developed Artificial Intelligence (AI) models that describe land cover at the pixel level from IGN aerial images. This is part of the production process for the Large Scale and Land Use Reference (OCS GE). This contribution is threefold:

Methodology: the training strategy and the use of these models will be reviewed by focusing on i) the selection of the task performed by the models, ii) the approach for choosing and producing learning samples and iii) the training strategy to generalize to the scale of Metropolitan France. The evaluation of the models using various metrics will also be discussed. Visuals will be provided to illustrate the quality of the results. Furthermore, we will explain how AI products are incorporated into the production of the OCS GE.

Continuous improvement: the models are continuously improved, particularly through the implementation of FLAIR (French Land cover from Aerospace ImageRy) challenges towards the scientific community. The challenges FLAIR#1 and FLAIR#2 dealt with model generalization and domain adaptation as well as data fusion, i.e., how to develop an AI model that can process very high spatial resolution images (e.g., IGN aerial acquisitions) and satellite image time series (Sentinel-2 images) as input. We will both review the challenges implementation and the obtained results, leveraging convolutional and attention-based models, ensembling methods and pseudo-labelling. As the AI model for land cover goes far beyond the context of OCS GE production, additional experiments outside of the challenges will be discussed, allowing the development of additional AI models to process other modalities (very high spatial resolution satellite images, historical images, etc.).

Open access: all source code and data, including AI land cover predictions maps, are openly distributed. These resources are distributed via the challenges and as products (CoSIA: Land Cover by Artificial Intelligence) by a dedicated platform, which is of interest for AI users and non-specialists including users from the geoscience and remote sensing community.

How to cite: Garioud, A.: Artificial intelligence for country-scale land cover description., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6477, https://doi.org/10.5194/egusphere-egu24-6477, 2024.

Landform mapping (also referred to as geomorphology or geomorphometry) can be divided into two domains: general and specific (Evans 2012). Whereas general landform mapping categorizes all elements of the study area into landform classes, such as ridges, valleys, peaks, and depressions, the mapping of specific landforms requires the delineation (even if fuzzy) of individual landforms. The former is mainly driven by physical properties such as elevation, slope, and curvature.  The latter, however, must consider the cognitive (human) reasoning that discriminates individual landforms in addition to these physical properties (Arundel and Sinha 2018).

Both mapping forms are important. General geomorphometry is needed to understand geological and ecological processes and as boundary layer input to climate and environmental models. Specific geomorphometry supports such activities as disaster management and recovery, emergency response, transportation, and navigation.

In the United States, individual landforms of interest are named in the U.S. Geological Survey (USGS) Geographic Names Information System, a point dataset captured specifically to digitize geographic names from the USGS Historical Topographic Map Collection (HTMC). Named landform extent is represented only by the name placement in the HTMC.

Recent work has investigated CNN-based deep learning methods to capture these extents in machine-readable form. These studies first relied on physical properties (Arundel et al. 2020) and then included the HTMC as a band in RGB images in limited testing (Arundel et al. 2023). Results from the HTMC dataset surpassed those using just physical properties. The HTMC alone performed best due to the hillshading and elevation (contour) data incorporated into the topographic maps. However, results fell short of an operational capacity to map all named landforms in the United States. Thus, our current work expands upon past research by focusing on the HTMC and physical information as inputs and the named landform label extents.

Specifically, we propose to leverage pre-trained foundation models for segmentation and optical character recognition (OCR) models to jointly map landforms in the United States. Our approach aims to bridge the disparities among independent information sources to facilitate informed decision-making. The modeling pipeline performs (1) segmentation using the physical information and (2) information extraction using OCR in parallel. Then, a computer vision approach merges the two branches into a labeled segmentation. 

References

Arundel, Samantha T., Wenwen Li, and Sizhe Wang. 2020. “GeoNat v1.0: A Dataset for Natural Feature Mapping with Artificial Intelligence and Supervised Learning.” Transactions in GIS 24 (3): 556–72. https://doi.org/10.1111/tgis.12633.

Arundel, Samantha T, and Gaurav Sinha. 2018. “Validating GEOBIA Based Terrain Segmentation and Classification for Automated Delineation of Cognitively Salient Landforms In Proceedings of Workshops and Posters at the 13th International Conference on Spatial Information Theory (COSIT 2017), Lecture Notes in Geoinformation and Cartography, edited by Paolo Fogliaroni, Andrea Ballatore, and Eliseo Clementini, 9–14. Cham: Springer International Publishing.

Arundel, Samantha T., Gaurav Sinha, Wenwen Li, David P. Martin, Kevin G. McKeehan, and Philip T. Thiem. 2023. “Historical Maps Inform Landform Cognition in Machine Learning.” Abstracts of the ICA 6 (August): 1–2. https://doi.org/10.5194/ica-abs-6-10-2023.

Evans, Ian S. 2012. “Geomorphometry and Landform Mapping: What Is a Landform?” Geomorphology 137 (1): 94–106. https://doi.org/10.1016/j.geomorph.2010.09.029.

How to cite: Arundel, S., Von Pohle, M., Akbari Asanjan, A., Oza, N., and Lott, A.: GeoAI advances in specific landform mapping, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20985, https://doi.org/10.5194/egusphere-egu24-20985, 2024.

Canadian national air photo library (NAPL) comprises millions of historical airborne photographs dating over 100 years. Historical photographs are rich chronicles of countrywide geospatial information. They can be used for creating long-term time series and supporting various analytics such as monitoring expansion/shrinking rates of built areas, forest structure change measurement, measuring thinning and retreating rates of glaciers, and determining rates of erosion at coastlines. Various technical solutions are developed at Natural Resources Canada (NRCan) to generate analysis-ready mapping products from NAPL.

Photogrammetric Processing with a Focus on Automated Georeferencing of Historical Photos: The main technical challenge of photogrammetric processing is identifying reference observations, such as ground control points (GCP). Reference observations are crucial to accurately georeference historical photos and ensure the spatial alignment of historical and modern mapping products. This is critical for creating time series and performing multi-temporal change analytics. In our workflow, GCPs are identified by automatically matching historical images to modern optical satellite/airborne ortho-rectified images. In the matching process, first, we use convolutional neural networks (D2Net) for joint feature detection and description in the intensity space. Then, we convert intensity images to phase congruency maps, which show less sensitivity to nonlinear radiometric differences of the images, and we extract an additional set of features using the Fast detector and describe them using the radiation-invariant feature transform (RIFT). Feature-matching outliers are detected and removed via random sample consensus (Ransac), enforcing a homographic transformation between corresponding images. The remaining control points are manually verified through a graphical interface built as a QGIS plugin. The verified control points are then used in a bundle block adjustment, where external orientation parameters of the historical images and the intrinsic calibration parameters of the cameras are refined, followed by dense matching and generation of digital elevation models and ortho-rectified mosaics using conventional photogrammetric approaches. These solutions are implemented using our in-house libraries as well as MicMac open-source software. Through the presentation, examples of the generated products and their qualities will be demonstrated.

Deep Colourization, Super Resolution and Semantic Segmentation: Considering the fact that NAPL mostly contains grayscale photos, their visual appeal and interpretability are less than modern colour images. In addition, the automated extraction of colour-sensitive features from them, e.g. water bodies, is more complicated than colour images. With this regard, we have developed fully automated approaches to colourize historical ortho-rectified mosaics based on image-to-image translation models. Through the presentation, the performance of a variety of solutions like conditional generative adversarial networks (GAN), encoder-decoder networks, vision transformers, and probabilistic diffusion models will be compared. In addition, using a customized GAN, we improve the spatial resolution of historical images which are scanned from printed photos at low resolution (as opposed to being scanned directly from film rolls at high resolution). Our semantic segmentation models, trained initially on optical satellite and airborne imagery, are also adapted to historical air photos for extracting water bodies, road networks, building outlines, and forested areas. The performance of these models on historical photos will be demonstrated during the presentation.

How to cite: Shahbazi, M., Sokolov, M., Mahoro, E., Alhassan, V., Bousias Alexakis, E., Gravel, P., and Turgeon-Pelchat, M.: Applications of GeoAI in Extracting National Value-Added Products from Historical Airborne Photography, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2378, https://doi.org/10.5194/egusphere-egu24-2378, 2024.

The ability to detect changes occurring on the Earth's surface is essential for comprehensively monitoring and understanding evolving landscapes and environments.

To achieve a comprehensive understanding, it is imperative to employ methodologies capable of efficiently capturing and analyzing both two-dimensional (2D) and three-dimensional (3D) changes across various periods.

Artificial Intelligence (AI)  stands out as a primary resource for investigating these alterations, and when combined with Remote sensing (RS) data, it has demonstrated superior performance compared to conventional Change Detection (CD) algorithms.

The recent introduction of the MultiTask Bitemporal Images Transformer [1] (MTBIT) network has made it possible to simultaneously solve 2D and 3D CD tasks leveraging bi-temporal optical images.

However, this network presents certain limitations that necessitate being considered. These constraints encompass a tendency to overfit the training distribution and challenges in inferring extreme values [1]. To address these shortcomings, this work introduces a series of custom augmentations, including strategies like Random Crop, Crop or Resize, Mix up, Gaussian Noise on the 3D CD maps, and Radiometric Transformation. Applied individually or in specific combinations, these augmentations aim to bolster MTBIT's ability to discern intricate geometries and subtle structures that are otherwise difficult to detect.

Furthermore, the evaluation metrics used to assess MTBIT, such as Root Mean Squared Error (RMSE) and the change RMSE (cRMSE), have their limitations. As a response, the introduction of the true positive RMSE (tpRMSE) offers a more comprehensive evaluation, specifically focusing on MTBIT's efficacy in the 3D CD task by considering only the pixels affected by actual elevation changes.

The implementation of custom augmentations particularly when applied in synergy, like Crop or Resize with Gaussian Noise on the 3D map, yielded substantial improvements. These interventions led – through the best augmentation configuration – to the reduction of the cRMSE to 5.88 meters and the tpRMSE to 5.34 meters, compared to the baseline (standard MTBIT) values of 6.33 meters and 5.60 meters, respectively.

The proposed augmentations significantly bolster the practical usability and reliability of MTBIT in real-world applications, effectively addressing critical challenges within the realm of Remote Sensing CD. 

REFERENCES:

• [1] Marsocci, V., Coletta, V., Ravanelli, R., Scardapane, S., Crespi, M., 2023. Inferring 3D change detection from bitemporal optical images. ISPRS Journal of Photogrammetry and Remote Sensing, 196, 325-339

How to cite: Contu, R., Marsocci, V., Coletta, V., Ravanelli, R., and Scardapane, S.: Urban 3D Change Detection with Deep Learning: Custom Data Augmentation Techniques, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9729, https://doi.org/10.5194/egusphere-egu24-9729, 2024.

Urban and natural landscapes are distinguished by different built and vegetated elements with unique features, and their proper identification is crucial for many applications, from urban planning to forestry inventory or natural resources management. With the rapid evolution and deployment of high-resolution airborne and Unmanned Aerial Vehicle (UAV) technologies, large areas can be easily surveyed to create high-density point clouds. Photogrammetric cameras and LiDAR sensors can offer unprecedented high-quality 3D data (a few cm on the ground) that allows for discriminating and mapping even small objects. However, the semantic enrichment of these 3D data is still far from being a fully reliable, accurate, unsupervised, explainable and generalizable process deployable at large scale, on data acquired with any sensor, and at any possible spatial resolution.

This work reports the state-of-the-art and recent developments in urban and natural point cloud classification, with a particular focus on the:

• Standardization in defining the semantic classes through a multi-resolution and multi-scale approach: a multi-level concept is introduced to improve and optimize the learning process by means of a hierarchical concept to accommodate a large number of classes. 
• Instance segmentation in very dense areas: closely located and overlapping individual objects require precise segmentation to be accurately identified and classified. We are developing a hierarchical segmentation method specifically designed for urban furniture with small samples to enhance the comprehensiveness of dense urban areas.
• Generalization of the procedures and transferability of developed models from a fully-labelled domain to an unseen scenario.
• Handling of under-represented objects (e.g., pole-like objects, pedestrians, and other urban furniture): classifying under-represented objects presents a unique set of challenges due to their sparse occurrence and similar geometric characteristics. We introduce a new method that specifically targets the effective identification and extraction of these objects in combination with knowledge-based methods and deep learning.
• Available datasets and benchmarks to evaluate and compare learning-based methods and algorithms in 3D semantic segmentation: urban-level aerial 3D point cloud datasets can be classified according to the presence of color information, the number of classes, or the type of sensor used for data gathering. The ISPRS - Vaihingen, DublinCity, DALES, LASDU and CENAGIS-ALS datasets, although extensive in size, do not provide color-related information. Conversely, Campus3D, Swiss3DCities, and Hessigheim3D include color data but feature limited coverage and a few class labels. SensatUrban, STPLS3D, and HRHD-HK were collected across extensive urban regions, but they also present a reduced number of classes. YTU3D surpasses other datasets in terms of class diversity, but it encompasses less extensive areas than SensatUrban, STPLS3D, and HRHD-HK. Despite these differences, the common deficiency in all datasets is the presence of classes with under-represented objects, the limited generalization, and the low accuracy in classifying unbalanced categories, making using these models difficult for real-life scenarios.

The presentation will highlight the importance of semantic enrichment processes in the geospatial and mapping domain and for providing more understandable data to end-users and policy-makers. Available learning-based methods, open issues in point cloud classification and recent progress will be explored over urban and forestry scenarios.

How to cite: Bayrak, O. C., Zhenyu, M., Farella, E. M., and Remondino, F.: Operationalize large-scale point cloud classification: potentials and challenges, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10034, https://doi.org/10.5194/egusphere-egu24-10034, 2024.

Regulations relating to the disposal of the use of asbestos was introduced in Italy with Law no. 257 of 1992 and its implementation took place over time. The Regional Asbestos Plan was put in place in 1996 and is updated periodically.

Modern remote sensing techniques constitute an essential tool for studies over an environmental and territorial scale. These systems can detect for each pixel of the acquired image from tens to hundreds of bands of the electromagnetic spectrum. This is useful as any material has its own characteristic spectral signature that can be exploited for different types of investigation.

The work involved the experimentation of a neural network for the classification of airborne remotely sensed hyperspectral images to identify and map the asbestos-cement roofing existing in some Municipalities of the Autonomous Region of Friuli Venezia Giulia.

The Region covers an area of approximately 8,000 square kilometres. To detect the entire area, it was necessary to carry out flights on different directions, different days and with different solar exposure conditions and so, the radiometric quality of the images is not uniform. Moreover, the images have high geometric resolution (1 meter pixel) and radiometric resolution (over 180 bands), that required a particular attention in their management: more than 4,000 images, for a total size of 25-30 TB.

Starting from these hyperspectral images and using the information already available relating to the mapping of the asbestos roofs of 25 Municipalities of the Region, we generated an adequate ground truth to train, test and validate a neural network implemented using the Keras library.

Given the differences in the territories of the various Municipalities, in the first step of the processing we calculated 3 different models generated on different training datasets for each considered Municipality: a total and a partial one that are independent on the considered Municipality, and the last one adapted to the specific Municipality. The combination of these predictions allowed us to obtain a raster result which is supposed to better adapt to the characteristics of the considered Municipality.

Obtained the data, it was then necessary to move on from the raster results to vector data using a zonal analysis on the buildings available in the Regional Numerical Map. An initial automatic classification, determined through the definition of adequate thresholds, was then manually refined exploring it with additional tools, such as Google StreetView and the 10 cm regional orthophoto, to obtain a final refined classification.

The results obtained for the 5 pilot Municipalities represent a clear indication of the presence of asbestos material on some building roofs. This work emphasized an operational workflow using data at a regional scale and could also be easily extended to other territorial entities. It has the great advantage to allow the government authority to save at least an order of magnitude in term of costs with respect to traditional investigations. Finally, the automation of the neural network represents a useful tool for programming, planning and management of the territory also in terms of human health.

How to cite: Viero, E., Gubiani, D., Basso, M., Marin, M., and Sgrazzutti, G.: Identification of asbestos roofing from hyperspectral images, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15565, https://doi.org/10.5194/egusphere-egu24-15565, 2024.

The development of high resolution mapping models of forest attributes based on employing machine or deep learning techniques has increasingly accelerated in the last couple of years. The consequence of this is the widespread availability of multiple sources of information, which can either lead to a potential confusion, or to a possibility to get an "extended” insight into the state of our forests by interpreting these sources jointly. This contribution aims at addressing the latter, by relying on the Bayesian model averaging (BMA) approach.

BMA is a method that can be used in building a consensus from an ensemble of different model predictions. It can be seen as weighted mean of different predictions with weights reflecting the predictive performances of different models, or as a finite mixture model which estimates the probability that each observation from the independent validation dataset has been generated by one of the models belonging to the ensemble. BMA can thus be used to diagnose and understand the difference in the predictions and to possibly interpret them.

The predictions in our case are the forest canopy height estimations for the metropolitan France coming from 5 different AI models [1-5], while the independent validation dataset comes from the French National Forest Inventory (NFI) disposing with some 6000 plots per year, distributed across the territory of interest. For every plot we have several measurements/estimations of the forest canopy height out of which the following two are considered in this study: h_m – the maximum total height (from the tree's base level to the terminal bud of the tree's main stem) measured within the plot, and h_dom – the average height of the seven largest dominant trees per hectare.

In this contribution we present for every considered plot the dominant model with respect to both references i.e. the model having the highest probability to be the one generating measurements/estimations at NFI plot (h_m and h_dom). We present as well as the respective inter-model and the intra-model variance estimations, allowing us to propose a series of hypotheses concerning the established differences between predictions of individual models in function of their specificities.

[1] Schwartz, M., et al.: FORMS: Forest Multiple Source height, wood volume, and biomass maps in France at 10 to 30 m resolution based on Sentinel-1, Sentinel-2, and Global Ecosystem Dynamics Investigation (GEDI) data with a deep learning approach, Earth Syst. Sci. Data, 15, 4927–4945, 2023, https://doi.org/10.5194/essd-15-4927-2023

[2] Lang, N., et al.: A high-resolution canopy height model of the Earth, Nat Ecol Evol 7, 1778–1789, 2023. https://doi.org/10.1038/s41559-023-02206-6

[3] Morin, D. et al.: Improving Heterogeneous Forest Height Maps by Integrating GEDI-Based Forest Height Information in a Multi-Sensor Mapping Process, Remote Sens., 14, 2079. 2022, https://doi.org/10.3390/rs14092079

[4] Potapov, P., et al.: Mapping global forest canopy height through integration of GEDI and Landsat data, Remote Sensing of Environment, 253, 2021, https://doi.org/10.1016/j.rse.2020.112165.

[5] Liu, S. et al.: The overlooked contribution of trees outside forests to tree cover and woody biomass across Europe, Sci. Adv. 9, eadh4097, 2023, 10.1126/sciadv.adh4097.

How to cite: Besic, N., Picard, N., Vega, C., Renaud, J.-P., Schwartz, M., Planells, M., and Ciais, P.: Bayesian model averaging of AI models for the high resolution mapping of the forest canopy height, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8420, https://doi.org/10.5194/egusphere-egu24-8420, 2024.

Deadwood is a vital part of the habitat for many threatened species of animals, plants and fungi. Thus, presence of deadwood is an important indicator for the probability that a given site harbors threatened species. Nowadays, field work is the most common method for monitoring dead trees. However, it is time consuming, costly and labor-intensive. Therefore, there is a need for an automatic method for mapping and monitoring deadwood. The combination of fine-resolution remote sensing and deep learning techniques have a potential to provide exactly this. Unfortunately, due to the typical location of lying deadwood under the canopy, this is a challenging task as the visibility of the lying trees is limited notably with optical remote sensing techniques. Therefore, laser scanning data seems to be the most appropriate for this purpose as it can penetrate the canopy to some extent and hence gather data from a forest floor.

In this work we aim at the development of methods enabling detection of lying deadwood at the national scale in protected forests and we focus on the presence of deadwood in 15-meter-radius circular plots. To achieve this goal, we use Airborne Laser Scanning (ALS) data that is publicly available for the whole Denmark and, as a reference, almost 6000 forestry plots acquired as a part of the Danish national habitats monitoring program. The binary classification into plots that contain deadwood and the ones that do not is performed using SparseCNN deep neural network. In this study we showed that it is possible to detect plots having deadwood with an overall accuracy of around 61%. However, the accuracy of the classifier depends on the volume of the deadwood present in a plot.  

How to cite: Walicka, A., Bladt, J., and Moeslund, J. E.: Remote sensing techniques for habitat condition mapping: deadwood monitoring using airborne laser scanning data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13166, https://doi.org/10.5194/egusphere-egu24-13166, 2024.

The study addresses the need for flood risk anticipation and planning, through the development of a flood zone mapping approach for different return periods, in order to best prevent and protect populations. Today, traditional methods are too costly, too slow or require too many requirements to be applied over large areas. As part of a project funded by the Canadian Space Agency, Geosapiens and the Institut National de la Recherche Scientifique set themselves the goal of designing an automatic process to generate water presence maps for different return periods at a resolution of 30 m, based on the historical database of Landsat missions from 1982 to the present day. This involved the design, implementation and training of a deep learning algorithm model based on the U-Net architecture for the detection of water pixels in Landsat imagery. The resulting maps were used as the basis for applying a frequency analysis model to fit a probability of occurrence function for the presence of water at each pixel. The frequency analysis data were then used to obtain maps of water occurrence at different return preiods such as 2, 5 and 20 years. 

How to cite: Chokmani, K., Zidi, H., El Alem, A., and Gill-Fortin, J.: Development of an approach based on historical Landsat data for delineating Canadian flood zones at different return periods, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1396, https://doi.org/10.5194/egusphere-egu24-1396, 2024.