Marco Trombetti is a computer scientist, a serial entrepreneur and investor, but also an incurable optimist, who is in love with big ideas. At ESA-ESRIN Phi Week 2019, he will talk about his experience building Translated, an online translation service he co-founded with his wife Isabelle Andrieu in 1999, which pioneered the use of artificial intelligence to help professional translators worldwide. Translated is one of the most successful online translation companies in the world. In 2016, the company was named among the top 50 European fastest growing tech companies by Tech Tour. Through the profits of Translated and other tech companies he went on to co-found, he successively created a venture capital firm called Pi Campus. Pi Campus invests in early-stage technology ventures mainly in the fields of applied AI and Italian excellence, such as food, lifestyle, fashion and creativity, counting more than 40 investments from Europe and the US in its portfolio to date. It is a startup district with over 250 citizens focused on providing the best possible international work environment for talent, helping young entrepreneurs around the world start up great companies. Marco has also recognized the need to support top developers build solutions to real world problems provided by companies and he has therefore established Pi School, a new learning reality born inside Pi Campus. Pi School is the next-generation kind of school, where machine intelligence meets human creativity bridging across various disciplines within the fields of technology, design and consulting, among others. It is a place for people who believe their ideas have the power to change the world and are willing to build an amazing future. Marco will share the main counterintuitive lessons he learned along the way, bringing values and not technicalities to light, in order to create the magic of doing startup.

Agrimetrics is funded by the UK Government to catalyse the agri-food sector by inspiring and empowering researchers and businesses to use data to solve global food challenges. We have pioneered the use of new web technologies specifically for the agri-food sector, enabling our Data Platform to connect disparate data and convert it into valuable insights. Our Field Explorer APIs allow users to explore and access environmental and agricultural data at the field level. Within this framework, we are developing a national in-season crop map for the UK to provide timely information for decision-makers within the food supply chain. The in-season crop map will complement the current field boundaries already available from Agrimetrics.

A deep learning segmentation algorithm has been developed for the automated delineation of fields over the UK. The application of this algorithm mapped over 1.45 million fields for the cropping season of 2016-2017. This field map provides an ideal framework over which Earth Observation data can be analysed.

Sentinel-1 2016-2017 SAR SLC/IW imagery were pre-processed to obtain terrain-corrected backscattering coefficients, interferometric coherence and polarimetric decomposition derived parameters like entropy and anisotropy. The pre-processing pipeline was implemented in Python and deployed as Kubernetes jobs on the Azure platform for scalability and reproducibility.

The pre-processed data were extracted and aggregated on a field basis to produce SAR time series for every field. Ground truth records were used to train machine learning algorithms to recognise crop types from SAR time series at four stages in the 2016-2017 cropping season. The overall accuracy increased throughout the growing season from approximately 65% to 80%. Current efforts are devoted to training the classification algorithms with data more representative of the spatial and temporal variability of the crop types.

A similar approach is currently being tested in Senegal, where cropping cycles are shorter (sowing to harvest within 100 days) compared to the UK. Pre-processed SAR data for 2016-2018 was attributed to more than 300 digitised fields for which ground truth data was available. Early results indicate that, despite short growing cycles and SAR imagery available only every 12 days in this area, it is possible to distinguish between 5 crop species, including corn, green bean and sweet potato, with 75% accuracy.

We are working on expanding the range of classification algorithm used from Random Forest and Dynamic Time Warping to more advanced AI algorithms including recurrent neural network. In both scenarios, we are working on methods to improve classification accuracy by incorporating ancillary data from non-Earth Observation sources.

By adding within-season crop type mapping to our Field Explorer product, we aim to offer users with the most accessible and useful data resource for decision making throughout the growing season.

In the last decade we witnessed a massive technological breakthrough in Computer Science, particularly the Cloud gained increased attention both inside and outside of the IT community. Also in the last few years there was a massive shift towards Machine Learning, with many IT companies investing significant resources in developing new techniques and products aimed at bringing Machine Learning closer to the developer communities, and finally to the end-user

Unfortunately most of these tools and techniques are targeted at the Computer Vision and NLP communities and do not aim to tackle the issues faced by the Earth Observation community: geo-referenced data, hyperspectral data, coregistration, vectorial data, various eccentric data types, etc

In this paper we introduce Hugin, a tool aimed in bridging the gap between the latest Machine Learning tools and the Earth Observation ecosystem, facilitating the adoption of the latest ML tools for geospatial applications. Hugin brings together Machine Learning tools like Keras, TensorFlow, SciKit-Learn and tools specific to the Earth Observation community: GDAL, rasterio, Fiona. It aims to provide solutions for problems like coregistration, resampling, mask generation, ensembling and augmentation allowing the EO researcher to focus on their algorithms and models.

Hugin simplifies access to various types of machine learning infrastructures: it provides support for distributed training, support for cloud based ML services (Google Cloud ML), support for Kubeflow, etc. By building on top of GDAL and rasterio, Hugin supports the most common EO data (including native access for Sentinel-1 and Sentinel-2 data).

This work was initially motivated by the Urban 3D and SpaceNet RoadDetector challenges, particularly by the need to provide a reusable and extensible Machine Learning framework. The tool was further developed as part of the ESA ML4EO project, providing the core training and classification facilities required by project use cases. We also consider that Hugin represents a tool that can facilitate EO researchers experimentation with state of the art machine learning techniques.

Sensing4Farming is a global platform for digital and smart farm management by integrating data from satellite imagery, IoT (Internet of Things) ground sensors, weather forecasting and field-operation registers into a powerful management and planning system.
The platforms integrates data from:
 Copernicus Hub platform regarding weekly free available open data from Sentinel high resolution satellites;
 DigitalGlobe's very high resolution base map, elevation model and very high resolution satellite images at different critical phenological stages;
 Daily open agroclimatic stations for monitoring weather conditions;
 Qampo’ in-situ sensors for monitoring soil-water-crop balance;
 Ancillary open georeferenced data such as digital terrain model, land use map, habitat of EC interest areas, etc.;
 Daily practices performed by the farmer using ABACO farm management software;
in order to provide high added value information suitable for support farm decision-making in a cloud-based platform. The integration, processing and analysis of the different databases allow the generation of georeferenced information with great precision oriented towards smart farm management.
A first stage is designed to standardize data from multiple sources. This is essential in order to associate different information’ layers to a specific site.
Singular data points such as those collected through the in-situ sensors are geodistributed in order to generate geospatial maps of weather-related data based on interpolation methods which provides an estimation for a specific location based on a weighted average of the neighbouring observed data. In general, the application of interpolation methods to a dataset can be broken into several stages: exploratory statistical analysis of the data, structural or variographic analysis (i.e. estimation and modelling of the function which quantifies the spatial correlation), use of the spatial correlation function to determine the set of equations that determine the weights in the kriging estimator, and, finally, production of the estimated values and the associated estimation standard deviations. Therefore, geostatistical estimation procedures are potentially very useful in Earth Observation.
The geospatial weather data together with multi-temporal spectral data derived from remote sensing and with other geospatial databases (such as long time series of yearly statistics and digital model terrain) are also integrated and analyzed through machine learning techniques for:
- Phenological state analysis:
Same crops in a region may show different phenological phases depending of planting dates, seed variety and weather behavior. To monitor in near real time fields’ phenological state is critical to avoid wrong health status reports and unnecessary recommended treatments. Using satellite and weather data, at plot level with 5-days periodicity crop phenological status is detected by means of machine learning linear regression-algorithm.
- Crop health status assessment:
Intra-plot analysis starts after crop state at plot level is fixed. In this step the goal is to detect patches which show health value far from mean expected. Random forest, trained by a wide catalogue of eo-based standard and advanced products as LAI, FPAR, is the machine learning algorithm used to assess the crop health status.
- Recommendations and precision treatments:
The combination of the phenological state information and the crop health status assessment with static data, such as the crop type characteristics, leads to the provision of recommendations to the end user. These recommendations can cover many issues related to the crop. Some examples of these issues are fertilization, irrigation or diseases treatment. Thanks to machine learning techniques, data can be combined to infer conclusions by using classification methods, mainly decision trees. This approach links essential variables (climate), observed phonomena (health, indices) and prior crop-dependent knowledge.
- Zoning or site-specific management;
While some analyses and output products make sense locally (where the term local depends on the product resolution), some others are regional or site dependent. In order to obtain information at this level, it is necessary to segment the area of interest. The field must be classified according to local analyses already performed and segments must be delimited. The definition of field boundaries easies the management of the site, allowing the subsequent definitiaon of site policies or measures and the reasoning of insights. Deep learning models can be used to classify the area from prior knowledge and local products.
- Yield prediction;
Since main concerns of the end user are economy-related, one of the main points within this scope is the yield prediction. In the short and medium terms, yield can be predicted from previously mentioned products (mainly phenological state, crop health status). The information provided by these products is crossed with historical data (site dependent) and crop characteristics. Each data has its own nature, so they must be correctly integrated before processing. Then, regression methods can be applied to obtain yield prediction.
Sensing4Farming is an innovative service based on applied technology, designed and adapted to the agricultural sector. In this way, the digitization of agriculture through an agronomic management system connected to IoT4.0, geospatial technologies and smart-automatized data analysis, is made available to support decision making and improve sustainability and competitiveness in the agricultural sector through the integration of knowledge and experience in different technological areas.

Today´s decisions in the digital world are black boxes. Data is siloed and the algorithms that now dominate our daily lives are not subject to our control, favouring monopolistic structures and extraction of value by a few individuals. Our future must be different. In an age of increasing complexity, global challenges and increasing inequality, data needs to be openly shared in a secure and tamper-proof manner, and algorithms must be audible, transparent and serve the public good as a new form of commons. This talk will introduce the infrastructural requirements for the new age of collective sharing and will explain how Distributed Ledgers can contribute to making this vision a reality.