Regional early warning systems for landslides rely on historic data to forecast future events and to verify and improve alarms. However, databases of landslide events are often spatially biased towards roads or other infrastructure, with few reported in remote areas. In this study, we demonstrate how Google Earth Engine can be used to create multi-temporal change detection image composites with freely available Sentinel-1 and -2 satellite images, in order to improve landslide visibility and facilitate landslide detection. First, multispectral Sentinel-2 images were used to map landslides triggered by a summer rainstorm in Jølster (Norway), based on changes in the normalised difference vegetation index (NDVI) between pre- and post-event images. Pre- and post-event multi-temporal images were then created by reducing across all available images within one month before and after the landslide events, from which final change detection image composites were produced. We used the mean of backscatter intensity in co- (VV) and cross-polarisations (VH) for Sentinel-1 synthetic aperture radar (SAR) data and maximum NDVI for Sentinel-2. The NDVI-based mapping increased the number of registered events from 14 to 120, while spatial bias was decreased, from 100% of events located within 500 m of a road to 30% close to roads in the new inventory. Of the 120 landslides, 43% were also detectable in the multi-temporal SAR image composite in VV polarisation, while only the east-facing landslides were clearly visible in VH. Noise, from clouds and agriculture in Sentinel-2, and speckle in Sentinel-1, was reduced using the multi-temporal composite approaches, improving landslide visibility without compromising spatial resolution. Our results indicate that manual or automated landslide detection could be significantly improved with multi-temporal image composites using freely available earth observation images and Google Earth Engine, with valuable potential for improving spatial bias in landslide inventories. Using the multi-temporal satellite image composites, we observed significant improvements in landslide visibility in Jølster, compared with conventional bi-temporal change detection methods, and applied this for the first time using VV-polarised SAR data. The GEE scripts allow this procedure to be quickly repeated in new areas, which can be helpful for reducing spatial bias in landslide databases.

Landslide risk mitigation is limited by data scarcity; however, this could be improved using continuous landslide detection systems. To investigate which image types and machine learning models are most useful for landslide detection in a Norwegian setting, we compared the performance of five different machine learning models, for the Jølster case study (30 July 2019), in Western Norway. These included three globally pre-trained models; (i) the continuous change detection and classification (CCDC) algorithm, (ii) a combined k-means clustering and random forest classification model, and (iii) a convolutional neural network (CNN), and two locally trained models, including; (iv) classification and regression Trees and (v) a U-net CNN model. Images used included Sentinel-1, Sentinel-2, as well as digital elevation model (DEM) and slope. The globally trained models performed poorly in shadowed areas and were all outperformed by the locally trained models. A maximum Matthew’s correlation coefficient (MCC) score of 89% was achieved with a CNN U-net deep learning model, using combined Sentinel-1 and -2 images as input. This is one of the first attempts to apply deep learning to detect landslides with both Sentinel-1 and -2 images. Using Sentinel-1 images only, the locally-trained deep-learning model significantly outperformed the conventional machine learning model. These findings contribute to developing a national continuous monitoring system for landslides.

<p>Using synthetic aperture radar (SAR) backscatter imagery can enable faster detection of landslides compared to optical images, particularly where there is persistent cloud cover or shadows. However, SAR images are underutilised for this purpose. This is partly due to the more complicated pre-processing requirements, and also due to the less intuitive interpretation of landslide signatures in SAR, relative to optical images. The problem of landslide identification in SAR backscatter imagery is complex. Landslides can occur in almost any land cover type and their expression in the environment can vary significantly depending on the material type and failure mechanisms. How this affects the expression of landslides in SAR backscatter data has so far not been well understood. In this study, we attempt to reduce this knowledge gap by investigating the physical basis for the expression of landslides in SAR backscatter data.</p> <p>This involved identifying trends in the spatial and temporal signatures of landslides in 30 case studies around the world, representing diverse physiographical and landslide types. Morphometric features of landslides (scarp, transport and deposition zone) were mapped separately, and quantitative analysis of their pixel values in multi-temporal Sentinel-1 SAR backscatter images was performed. The role of environmental factors including the orientation of the landslide with respect to the sensor (local incidence angle), land cover, seasonal variations, and water content were also analysed.</p> <p>The terrain influenced whether or not landslides were detectable, while the presence or absence of woody vegetation determined if there would be an increase or decrease in backscatter intensity. Landslides in non-forested areas that produce an increase in surface roughness, are best observed using VV polarisation and show increased backscatter intensity. Deposit zones also tend to show increased backscatter intensity, unless very fine material was deposited as a smooth flat surface (e.g. from non-turbulent mudflows). Removal of the forest is best viewed in VH polarisation, and produces a recognisable pattern of both decreased (due to radar shadow, and change from volumetric to surface scattering) and increased (due to direct and double bounce reflection from vertical tree trunks and scarp surface) backscatter intensity. Landslides that occur in mixed vegetation types, and those that do not significantly change the scattering properties of the ground surface, did not produce a detectable change in the C-band SAR images.  </p> <p>The findings were summarised in a conceptual model, based on SAR theory and empirical evidence. This can be used to help interpret landslides in SAR backscatter change images, and to design representative or synthetic datasets for training automatic landslide detection models. </p>

<p>Although Norway is a country with rough terrain and a high frequency instable steep slopes, there is a scarcity of landslide data available. This limits the accuracy of thresholds for early warning systems, and hazard maps, both of which rely on historic event data. There is great potential to supplement existing ground-based observations with automated landslide detection, using satellite imagery and deep learning. In working towards an automated system for landslide detection in Norway, we investigated which imagery types and machine-learning models performed best for detecting landslides in a formerly glaciated landscape.</p><p>We locally trained a deep learning model with the use of Keras, TensorFlow 2 and U-net architecture. As input data, we used multi temporal composites with Sentinel-1 and -2 image stacks of all available images from one month pre- and post-event. Processed bands included: dNDVI (difference in maximum normalised difference vegetation index) from Sentinel-2, and pre- and post-event Synthetic Aperture Radar (SAR) data (terrain-corrected, mean of multi-temporal ascending descending images, in VV polarisation) from Sentinel-1. Training and evaluation were performed with a well-verified landslide inventory of 120 manually mapped rainfall-triggered landslides from Jølster (30-July-2019), in Western Norway. We tested the model with four input data settings using different bands and various polarization for the pre- and post-event SAR data, including: 1) full version (all 13 bands) 2) dNDVI (Sentinel-2), preVV, postVV (Sentinel-1), 3) preVV, postVV (Sentinel-1), and 4) post-R, post-G, post-B, post-NIR, dNDVI (Sentinel-2). The results were compared to the results of a pixel-based conventional machine learning model (Classification and Regression Tree) using the same input data. The second input data setting provides the best results. The performance scores show precision results for all four input data settings between 80-85%, with Matthews corelation coefficient values from 51-89%. Moreover, the deep-learning model significantly outperforms the conventional machine learning model in the input data setting #3. We see that the patch-based classification method far out-performs the pixel-classification due to the ability to differentiate the landslide signal from random noise produced from speckle in undisturbed areas. In addition, this represents one of the first attempts to fuse SAR and optical data for landslide detection, and we show there is an advantage in doing so in this case.</p><p> </p>

<p>The Norwegian mass movements database contains over 33,000 registered snow avalanche and landslide events from the past 500 years and is used as an input for The Norwegian Landslide Early Warning System (LEWS). However, the usability of the database is limited by factors including a spatial bias towards transport systems and incomplete or missing information on landslide characteristics (including precise date, time or location). This has serious consequences for the definition of triggering thresholds. Sentinel-2 optical satellite data, with its frequent return period in Norway (up to three days) and relatively high resolution (10 m), could provide an alternative source of data on landslide occurrence to supplement ground-based observations and improve the information in the database.</p><p>This study examined the potential for using Sentinel-2 data to detect landslides with two approaches, using (i) a national-, and (ii) a local-survey. Both used the change in the vegetation index (denoted dNDVI) between pre- and post-event images, to identify a loss of vegetation as an indicator of landslide occurrence. Firstly, 30 well-documented landslides with a minimum volume of 1000m<sup>3</sup> were extracted from the national database. The selected landslides occurred across all Norway between 2015 to 2017. They were searched for in Sentinel 2 images to give insight into how factors including season, slope angle, aspect ratio, land cover, landslide size influenced landslide detection using the dNDVI-method. Secondly, the same approach was applied to the Jølster area in Western Norway, where an extreme short intense rainfall event in the summer of 2019 (30 July 2019) triggered multiple landslides. For Jølster, landslides were mapped and then verified by field and helicopter observations.</p><p>For the national survey, the season was found to have the greatest effect on detectability. For spring and summer events the percentage of landslides detected was 70-75%, while for winter and autumn this dropped to 14-20%. The main reasons for non-detection were clouds, shadows, snow, and lack of green vegetation. The average acquisition window for detected events was 43.3 days. The Jølster case study represented ideal conditions for using the dNDVI-method, with a five-day acquisition window (almost cloud-free images available from two days pre-, three days post-event), low shadow, and green summer vegetation. The mapping process produced an inventory of 99 events, giving a significant increase from the 14 events registered in the database.</p><p>The results indicate that the dNDVI-method has good potential for landslide detection in late-spring and summer in Norway, however, it is not recommended later in autumn and winter. We believe that the dNDVI-method provides an option for gaining more information on the size and location of landslides, which at the present, are only registered as points in the database. For the Jølster case, this method showed a great improvement with respect to the current practice, both in terms of an increased number of landslides and spatial distribution. This suggests good potential for improving inventories of landslides, necessary in landslide hazard analyses and definition of landslide thresholds.</p>