Search

GeoAI combines artificial intelligence (AI) with geospatial data, science, and technologies. In this paper, a successful case in utilizing GeoAI to improve the data quality in the national topographic database (TDB) of Finland was introduced. The project employed a GeoAI model to identify buildings from the input data of true orthophotos, digital elevation model (DTM), and digital surface model (DSM). The GeoAI-derived buildings served as reference data, enabling a comparison with building polygons from the topographic database (TDB) to reveal TDB building location deviations, missing structures, and demolished buildings. Throughout the project, algorithms were developed to match the TDB building vectors to the GeoAI-derived building polygons. The challenges include i) the differences between the GeoAI-derived building outlines and the TDB build footprints; ii) the reliability of the GeoAI model across data over different environments: urban, suburban, rural, and forest areas. Throughout the project, the GeoAI model was continuously improved by training massive new datasets: corrected vectors from the model prediction. Testing was conducted on datasets from twelve areas of Finland, covering 2204 km2 and including hundreds of thousands of buildings. The test areas covered urban, suburban, rural, and forest areas. The evaluation was conducted by the mapping team in the organization. The results showed that our methods greatly enhanced the quality of the TDB building footprints. The challenges and lessons of the project were addressed in the paper. [ABSTRACT FROM AUTHOR]

Automatizing the extraction of different objects from remote sensing data with deep learning methods has been a popular research topic. Buildings have been one of those popular objects to be extracted. Not only does the selection of neural network affect the results and accuracy of extracted buildings, but also the selection of different types of data for the task. Digital surface models (DSMs) are increasingly used in remote sensing and their demand has increased. Retrieving height information from surface models has proved helpful for accurate extraction of buildings. In this study was investigated, if the use of light detection and ranging (LiDAR) DSMs and DEMs with 25 cm pixel resolution will lead to more accurate building extraction results in comparison to the use of aerial DSMs. Results with UNet models trained with building vector labels, DSMs, DEMs and true orthophotos from multiple areas of Finland, were produced with different data combinations from two Finnish cities, Savonlinna and Pudasj¨arvi, to see, which combination would lead to the most accurate building detection results. Results were evaluated partly by visual inspection, and partly by quantitative assessment. Based on the tests carried out, combining the information from true orthophotos with LiDAR DSMs and 25 cm DEMs provided the most accurate results. In forest area, using LiDAR data increased the accuracy of building detection. However, in urban area, due to missing buildings from LiDAR data, its advantages were compromised. We suggest that the use of both imagery and LiDAR data should be the optimal solution. [ABSTRACT FROM AUTHOR]

Vector datasets of small watercourses, such as rivulets, streams, and ditches, are important for many visualization and analysis use cases. Mapping small watercourses with traditional methods is laborious and costly. Convolutional neural networks (CNNs) are state-of-the-art computer vision methods that have been shown to be effective for extracting geospatial features, including small watercourses, from LiDAR point clouds, digital elevation models (DEMs), and aerial images. However, the cause of the false predictions by machine-learning models is often not thoroughly explored, and thus the impact of the results on the process of producing accurate datasets is not well understood. We digitized a highly accurate and complete dataset of small watercourses from a study area in Finland. We then developed a process based on a CNN that can be used to extract small watercourses from DEMs. We tested and validated the performance of the network with different input data layers, and their combinations to determine the best-performing layer. We analyzed the false predictions to gain an understanding of their nature. We also trained models where watercourses with high levels of uncertainty were removed from the training sets and compared the results to training models with all watercourses in the training set. The results show that the DEM was the best-performing layer and that combinations of layers provided worse results. Major causes of false predictions were shown to be boundary errors with an offset between the prediction and labeled data, as well as errors of omission by watercourses with high levels of uncertainty. Removing features with the highest level of uncertainty from the labeled dataset increased the overall f1-score but reduced the recall of the remaining features. Additional research is required to determine if the results remain similar to other CNN methods. [ABSTRACT FROM AUTHOR]

National mapping authorities aim to increase automation in their mapping processes. Automation increases the rate of updating features in topographic databases (TDBs), improves the quality in terms of regional homogeneity and logical consistency between the mapped features, and lowers costs. Recently, the National Land Survey of Finland (NLS) has achieved countrywide coverage for its digital elevation model with a 2-meter resolution (NLS-DEM2m). During the production of the NLS-DEM2m, the NLS has initiated several pilot projects to find automatic methods for extraction of topographic features from the DEM, including, contour lines (Kettunen, Koski and Oksanen 2017), cliffs, and hydrographic features. While producing encouraging results, these projects have highlighted the inconsistency between automatically generated features from the NLS-DEM2m and manually mapped features in the NLS TDB. For example, this is apparent when overlaying automatically generated contour lines and the current manually mapped hydrographic features of the TDB (Figure 1). Increased computing power, more data, and better models have made machine learning (ML) popular for automation. In May 2021, NLS started piloting automated mapping of hydrographic features from the NLS-DEM2m with the use of ML. Using the NLS-DEM2m to map hydrographic features to the NLS TDB can improve their positional accuracy and completeness and make them consistent with automatically generated contour lines. Previous piloting of automated mapping of hydrographic features with non-ML methods from high-resolution DEMs has not yet achieved a high enough quality for them to be viable options for fully automated production of hydrographic features (Eraña-Beitia 2013, Oksanen 2014). The main challenges have been to correctly identify and map small streams and ditches. Therefore, in this project, we will focus mainly on ML-based extraction of hydrographic features. The objectives of automated mapping of hydrographic features in the project are as follows: 1. To assess the quality of the ditch and stream vectors that are derived from the results of ML models. 2. To identify best practices for producing test and validation data of ML models that detect ditches and streams from the NLS-DEM2m. 3. To identify the most critical areas of the stream and ditch extraction process that need to be improved. 4. To improve the used methods by developing new and innovative solutions for automated mapping of hydrographic features from DEMs. Recently, convolutional neural networks (CNNs) have been used to identify both streams (Xu et al. 2021) and ditches (Paul, Ågren, and Lidberg 2021) from DEMs. These CNN approaches have displayed promising results using DEMs with resolutions as high as 1 and 0.5 meters as input data. Results of the CNN models have been assessed based on recall and precision on the pixel level. We will apply these CNN approaches and assess results achieved with them, using the NLSDEM2m as input data as this is the highest resolution country-wide DEM that is available. The results will then be used to generate ditch and stream vectors, which will be assessed based on their completeness and positional accuracy. The quality of results produced by CNN models is dependent on high-quality training and validation data, as well as input data. The positional accuracy and completeness of the data are critical for producing high quality results. However, mapping ditches and streams, especially those in forested areas can be difficult, even when done manually from the available source datasets (Figure 2). Apart from the positional accuracy and completeness of the training data, we need to assess how much training data is needed, what types of test areas are needed considering different types of terrain in Finland, and how to produce the most accurate data possible considering available resources and source data. In the final steps of the project, we aim to identify areas that need improvements and develop methods to improve the outcome. Possible areas of improvement include quality of training data, resolution of input DEM, CNN model architecture, and automated vectorization of the CNN results. We will identify the most promising areas of improvement by testing higher resolution DEMs produced from denser point clouds, by assessing the impact of the quality of training and validation data on the results, by testing different CNN models, and by comparing different methods for vectorization of the CNN model results. [ABSTRACT FROM AUTHOR]