1. KappaMask for Sentinel-2
We are happy to announce that KappaMask is supporting now both Sentinel-2 Level-2A and Level-1C inputs. KappaMask predictor can be used for different levels and obtain results that outperformed all open-source existing methods. It generates 10 m georeferenced output .tif mask with clear, cloud shadow, semi-transparent clouds, clouds and missing classes. For validation, we make an extremely challenging test set that is covering the whole Northern European terrestrial conditions. The result is that KappaMask L1C reached 79% dice coefficient score, while Sen2Cor on the same data resulted in 55%, Fmask – 63%, S2cloudless – 64% and MAJA – 43%. KappaMask L2A model is performing even better than KappaMask L1C on cloud shadow class. However semi-transparent clouds are better recognized by L1C model, since L1C has a specific cirrus band which is not available for L2A. To make cloud mask this accurate we labelled more than 4000 sub-tiles of 512 x 512 pixels size. The predictor can be installed and run using this link: https://github.com/kappazeta/cm_predict. We share the open-source approach and appreciate community input, try it out and give us your feedback!
Marharyta Domnich, machine learning engineer
2. HaTRY model development and first result
Our harvesting time recommendation model developments have reached a point, where we can share some insights into model architecture and preliminary results.
The model consists of two parts: a regression model (M1) which predicts future time series and a binary classification model (M2) which picks up the harvesting signatures from the regression model output. Instead of a single black-box model approach, we opted for a modular design to make it easier to troubleshoot and evaluate the system as well as to re-use parts of the pipeline across projects.
To make the time series easier to predict, Sentinel-1 parameters are separated into different features by their relative orbit number (RON). With resampled and linearly interpolated time series as input, M1 smoothens the signatures further. The following figure shows the predicted vs. actual VH coherence median for RON 153, the median of backscatter VH/VV ratio for RON 58, vegetation index (NDVI) median and active temperature sum time series for a single winter rapeseed parcel.
M2 has been trained to function on time series input with weather forecast and M1 output stitched together. The following figure illustrates harvesting probabilities for another parcel (spring barley) from M2 output, in comparison to a labelled time series of Sentinel-1 coherences, backscatter VH/VV ratio and Sentinel-2 NDVI. Sentinel-1 parameters from all RONs are shown on the image.
Due to each crop type having different growth signatures and harvesting times, dedicated M1 and M2 are used for each crop type. In the scope of the HaTRY project, we have three crop types (winter wheat, winter rapeseed and spring barley) with a pair of models for each. This is illustrated in the figure below.
We expect to have a live prototype of the harvesting time recommendation service integrated with the crop monitoring system in July and will validate our model predictions with field surveys and farmers’ feedback.
Mihkel Järveoja, HaTRY project leader
3. Weekly field research for the grazing detection
It’s 8 o’clock in the morning on Friday and we are already heading northwest to perform our new routine – meet with the farmers, fly a drone over the field and record all the changes in the grassland condition.
Fieldwork is a significant part of ground reference data collection required for the development of grazing detection methodology. Our main goal is to collect enough continuous data to calculate the grazing intensity (i.e., number of animals per one hectare) of the test parcels in Estonia. Continuity in the data collection is important and hence we are doing the fieldwork on a weekly basis throughout the whole active grazing period.
Our partner for the fieldwork is the University of Tartu Geography department. Together we focus on three main activities: counting animals on the parcel with a drone, measuring grass height and monitoring soil moisture. In addition, we keep track of animal movements between the different parcels, grassland condition, mowing activities and supplementary fodder.
