Artificial Intelligence

Air temperature (Tair or T2m) is an important climatological variable for forest, biosphere processes and climate change research. Due to the low density and uneven distribution of weather stations, traditional ground-based observations cannot accurately capture the spatial distribution of Tair. In this study, Tair in Berlin estimated during day and night time over six land cover/land use (LC/LU) types by satellite remote sensing data over a large domain and a relatively long temporal period (7 years). Aqua and Terra MODIS (Moderate Resolution Imaging Spectroradiometer) data and meteorological data for the period from 2007 to 2013 were collected to estimate Tair. Twelve environmental variables (land surface temperature (LST), normalized difference vegetation index (NDVI), Julian day, latitude, longitude, Emis31, Emis32, altitude, albedo, wind speed, wind direction and air pressure) were selected as predictors. Moreover, a comparison between LST from MODIS Terra and Aqua with daytime and nighttime air temperature (Tday ,Tnight ) was done respectively and next to it, the spatial variability of LST and Tair relationship by applying a varying window size on the MODIS LST grid was examined. Analysis of the relationship between observed Tair and spatially averaged remotely sensed LST indicated that 3 × 3 and 1 × 1 pixel size was the optimal window size for the statistical model estimating Tair from MODIS data during day and night time, respectively. Three supervised learning methods (Adaptive Neuro Fuzzy Inference system (ANFIS), Artificial Neural Network (ANN) and Support vector machine (SVR)) were used to estimate Tair during day and nighttime, and their performances were validated by cross-validation for each LC/LU. Moreover, tuning the hyper parameters of some models like SVR and ANN were investigated. For tuning the hyper parameters of SVR, Simulated annealing (SA) was applied (SA-SVR model) and a multiple-layer feed-forward (MLF) neural networks with three layer and different nodes in hidden layer are used with LevenberMarquardt back-propagation (LM-BP) in order to achieve higher accuracy in estimation of Tair. Results indicated that the ANN model achieved better accuracy (RMSE= 2.16°C, MAE = 1.69°C, R2 = 0.95) than SA_SVR model (RMSE= 2.50°C, MAE = 1.92°C, R2 = 0.91) and ANFIS model (RMSE= 2.88°C, MAE = 2.20°C, R2 = 0.89) over six LC/LU during day and nighttime. The Q-Q diagram of SA-SVR, ANFIS and NN show that all three models slightly tend to underestimate and overestimate the extreme and low temperature for all LC/LU classes during day and night time. The weak performance in the extreme and low temperature are a consequence of small number of data in these temperatures. These satisfactory results indicate that this approach is proper for estimating air temperature and spatial window size is an important factor that should be considered in the estimation of air temperature.

UrbanAI, next Starlab spinoff, value proposition is to ease the access to complex data by levering all the value from Satellite, IoT network and Crowdsourced data to create a disruptive data exploitation capability based in Machine Learning processing. Nowadays, with the big data era and the IoT massive data collection, new datasets are becoming available every day at cities, but their exploitation has not yet reached full potential. With the full deployment of the sentinels, and also new players such as Planet similar things are happening in to the satellite ecosystem. Data access, mining, combination and exploitation is becoming the main bottleneck. UrbanAI provides information data and most important, enables data mining empowering the user with new capabilities that they didn’t had before due to the fact that all data (from different nature and sources) was scattered apart. UrbanAI will then enable data mining and easy the access to unreachable data before and empower the user with machine learning tools The first UrbanAI application purpose is to help cities maximise and demonstrate the environmental and socio-economic benefits related to having a healthy and resilient urban forest. Right now, UrbanAI is being accelatered by ESADE(Barcelona business school) and also ESA thanks to an OpenInnovation project: Street Health. The solution is being validated in the market through the completion of ongoing pilots in several cities such as Montreal, Singapore, Paris, Barcelona and Milton Keynes from different application lines that range from urban forest to urban sprawl, fires monitoring and also leak detection.

During the last years satellite data with high spatial and temporal resolution have become available under free and open licenses form Sentinel missions: Sentinel-1A/B and Sentinel-2A/B. The large volumes of these data allow providing classification maps at global, national and regional scale in operational procedures and update them every two weeks. At the same time, for crop identification at global or even at national scale effective algorithms of data storing and processing should be utilized. We propose a four-level deep learning architecture for crop mapping based on multi-temporal imagery from different satellites [1, 2]. These levels are pre-processing, supervised classification, post-processing and geospatial analysis. An ensemble of convolutional neural network is used for time series classification [3]. Also, effective data access algorithms have been implemented to avoid step with merging all images during vegetation period into single image cube. This methodology allows us update large scale crop maps every two weeks and use new acquired images and evaluate increasing accuracy of crop mapping with new data acquisition. Keywords: agriculture, image processing and data fusion, open data, Sentinel-1, Sentinel-2 1. A. Shelestov, M. Lavreniuk, N. Kussul, A. Novikov, and S. Skakun, “Exploring Google Earth Engine Platform for Big Data Processing: Classification of Multi-Temporal Satellite Imagery for Crop Mapping,” Front. Earth Sci., vol. 5, no. 17, pp. 1-10, 2017. doi: 10.3389/feart.2017.00017. 2. S. Skakun, N. Kussul, A. Y. Shelestov, M. Lavreniuk, and O. Kussul, “Efficiency Assessment of Multitemporal C-Band Radarsat-2 Intensity and Landsat-8 Surface Reflectance Satellite Imagery for Crop Classification in Ukraine,” IEEE Journal of Select. Topics in Applied Earth Observation and Remote Sensing, vol. 9, no. 8, pp. 3712-3719, 2016. 3. N. Kussul, M. Lavreniuk, S. Skakun, and A. Shelestov, “Deep Learning Classification of Land Cover and Crop Types Using Remote Sensing Data,” IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 5, pp. 778-782, 2017.

The arrival of sentinel 2 data provides a great opportunity for researchers and scientist to test their algorithms on a large scale with many variations, at least both in time and locations. Many of these methods originate from the computer vision and pattern vision communities and address detection, classification and recognition problems. The available resolutions (10 meter GDS for panchromatic image) forbid the search and monitoring of small objects such as cars for example. However, many challenges remain for agricultural and urban related problems. In this study, we have focused on the cloud detection, although it is partially addressed by the ground segment for level 1C product, we have found that was still an important matter of interest for the strengthening of the product quality, leading to more robust higher level object extractions. In this regard we also have focused on another recognition problem which deals with the building footprints recognition and extraction. Thus, this paper sums-up our experience with both of these problems using a few convolutional neural networks from the literature. Many companies have open sourced their framework, allowing at little cost the use of very powerful tools to enable the design, training and application of such networks both locally and in the cloud. From our experience, Tensor Flow with Keras is a good way to prototype networks. None the less, we emphasize that a huge work is required in building proper the data sets. As far as we are concerned, this task remains one of the most challenging phase in the whole process, since a high ground truth quality is often required to reach high quality detection rates or high quality segmentation masks. In this regard, we address some comments with respects to the available open ground truth data and elaborate on our results while broadening the discussion to the benefits of active learning and unsupervised learning.

Nowadays the Remote Sensing scenario is characterized by a huge availability of SAR data that offer the possibility to map the Earth surface deformation at a very large scale. In particular, starting from April 2014, we are collecting big data archives acquired by the new Sentinel 1-A sensor, which has been paired with its twin sensor Sentinel 1-B on April 2016, showing enhanced features in terms of revisit frequency and spatial coverage. Therefore the challenge is to maximize the exploitation of such data, and, in this direction, the use of distributed computing infrastructures and, in particular, of Cloud Computing platforms can play a crucial role. In this work we present a Cloud Computing solution for the advanced interferometric processing chain, based on the P-SBAS approach [1], [2], aimed at processing S1 Interferometric Wide Swath (IWS) data for the generation of deformation time series in efficient, automatic and systematic way. Such a DInSAR chain ingests Sentinel 1 SLC images and carries out several processing steps, such as SAR image coregistration, interferogram generation, interferometric phase unwrapping, in order to finally compute deformation time series and mean deformation velocity maps. Different parallel strategies have been designed ad hoc for each processing step of the P-SBAS S1 chain, encompassing both multi-core and multi-node programming techniques, in order to maximize the computational efficiency achieved within a Cloud environment and cut down the relevant processing times. The presented P-SBAS S1 processing chain has been implemented on the Amazon Web Services public cloud platform and a thorough analysis of the attained parallel performances has been performed in order to identify and overcome the major bottlenecks to the scalability. Some impressive results relevant to the national-scale DInSAR analyses performed over Italy, involving the processing of more than 1500 S1 IWS images, will be presented, with all the details about the processing times and costs. Such outcomes confirms the big advantage of exploiting Cloud Computing platforms in the context of massive SAR data processing, because of the large collection of computational and storage resources that they offer, which allows performing DInSAR analyses at unprecedented large scale. The presented Cloud Computing P-SBAS processing chain can be a precious tool within the EO Open Science scenario, allowing us to fully exploit the huge S1 data stream, also in the perspective of developing operational services disposable for the EO scientific community related to hazard monitoring and risk prevention and mitigation. [1] F. Casu et al., "SBAS-DInSAR Parallel Processing for Deformation Time-Series Computation," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 7, no. 8, pp. 3285-3296, 2014. [2] I. Zinno et al., "A Cloud Computing Solution for the Efficient Implementation of the P-SBAS DInSAR Approach," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 10, no. , pp. 802-817, 2017.