rahmansunbeam Posted October 30, 2019 Report Share Posted October 30, 2019 This is a very interesting mapping platform for the agriculture community. The Belarus-based startup platform uses Sentinel-2 data and AI to instantly delineate thousands of crop fields and status of 20 plus crops in USA and Europe. Quote It is not too surprising that a “farming by satellite” tool such as OneSoil originated in Belarus as agriculture is an important economic sector in the country. OneSoil is a precision farming platform which monitors fields according to criteria crucial for efficient and fruitful agricultural activities. It can detect and mark field boundaries, detect crop types, determine required amounts of fertilisers, show weather forecasts, etc. As a source of satellite imagery, it relies exclusively on Copernicus Sentinel-1 and Sentinel-2 data which it processes with sophisticated AI algorithms for user-friendly results. More specifically, OneSoil uses Copernicus Sentinel-2 multispectral images to automatically determine the crop type that grows in any given field. Then, to reduce uncertainty, it uses Sentinel-1 radar satellite images. This enables OneSoil to spot 20 types of crop with, according to its estimates, high accuracy. The OneSoil technologies can potentially determine the sowing date and the stage of plant development. However, to improve its accuracy, the model needs in situ data sets with high quality data, which are collected with help of the platform users. For instance, this feature would help farmers choose the best time for applying fertilizers and pesticides. They also have smartphone-based apps which you can use to find these solutions for your field as well. The platform applies Machine Learning, which constantly improves the service as more data and feedback is collected. Considering that a mind-boggling 376,835,301 hectares of fields across Europe and the USA have already been analyzed and catalogued, the system has reached a remarkable level of maturity. Quote To detect fields all over the world, we need to solve two main Machine Learning problems. The first one includes training a model in an efficient way. The second problem is more complex — it entails developing a cloud pipeline prediction (inference) of the model for large areas. Training a model in the cloud (x2 Graphics Processing Units (GPU) k80 with Tensorflow/Keras and multi-GPU training, for those who are in the know) takes around one day. It takes us approximately 1−2 days to predict fields in Europe and the USA for a 3 year-period. To build the map, we used images acquired by the Sentinel-2 satellites of the European Union’s Copernicus programme. A total of 250 Tb of data was processed, covering Europe and the USA. We pre-processed the images by removing the clouds, shadows and snow. We then compressed the data down to 50 Tb. In a quick calculation, approximately 140Tb of raw data (we used float32 data at the inference stage) was passed through the GPU. Following this, we searched for field boundaries and classified cultures with our Machine Learning algorithms. This produced approximately 250Gb of vector maps with field geometries and cultures. OneSoil — a Copernicus-enabled start-up from Belarus Check out their interactive map. Onesoil homepage 1 Quote Link to comment Share on other sites More sharing options...
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.