Latest Tweets


KP Labs Sp. z o.o. (PL)


Farmers and field owners need information about the soil parameters to optimize the fertilization process. This may ultimately lead to selecting a better mix of fertilizers, and to reducing the overall amount of them. The current approach toward quantifying the soil parameters (e.g., macroelements) is very user-dependent, laborious, time-consuming, vastly manual – we have to gather and mix soil samples in the field and pass them to the lab for further chemical analysis. Also, this process does not allow us to accurately capture the information concerning the macroelements, and the number of sampling points in the field is commonly limited.

KP Labs and QZ Solutions intend to use the Intuition-1 satellite to remotely detect soil parameters (specifically: potassium – K2O, phosphate – P2O5, magnesium – Mg and pH) using on-board machine learning techniques employed to analyze the acquired hyperspectral data. Hyperspectral data captures very detailed information about the observed area; however, its volume makes the data acquisition and transfer back to Earth very costly and time expensive (due to the data transmission constraints). Hence, the whole process of detecting the soil parameters will be automated on-board the Intuition-1 satellite. To make the process efficient for large-scale imaging, additional optimizations in the on-board pre-processing chain are foreseen, including filtering of too cloudy scenes based on L0 data, so heavy processing to L1 data will be avoided. Another important pre-processing step assumes determination of the bare soil area, so the soil analysis algorithms can be used only in the right context focusing on the regions of interest.

Selected algorithms which will be used for pre-processing and estimating the soil parameters, will be deployed on the data processing unit called Leopard, verified during the on-ground testing and benchmarking campaign. As an output, an image with the soil parameters should be returned (see the example rendered in Figure below) – it will allow us to dramatically reduce the volume of the data that is to be transferred back to Earth, as we would be sending just the parameter maps.
Moreover, the project will focus on the incremental/continuous learning which directly corresponds to the real-world scenario, where the new data may be captured (and manually or semi-automatically analysed in order to obtain the ground truth) while a machine learning model is deployed on-board spacecraft. Here, gathering new data may help us improve the models in time following the lifelong learning paradigm, but, at the same time, may easily lead to the catastrophic forgetting and interference, as the newly acquired data may be of different characteristics.


Soil parameters maps



Scientific Papers


Prime contractor
KP Labs Sp. z o.o. (PL)
  • QZ Solutions Sp. z o.o. (PL)