Method for prediction of soil and/or plant condition
A method and system for predicting soil and/or plant condition in precision agriculture with a classification of measurement data for providing an assignment of a measurement parcel to classes of interest. The assignment is used for providing action recommendations, particularly in real time or close to real time, to a farmer and/or to an agricultural device based on acquired measurement data, particularly remote sensing data, and wherein a classification model is trained by a machine learning algorithm, e.g. relying on deep learning for supervised and/or unsupervised learning, and is potentially continuously refined and adapted thanks to a feedback procedure.
1. A method for predicting a soil or plant condition in precision agriculture, the method comprising:
acquiring measurement data for at least one measurement parcel, which is at least partly captured by the measurement data, the measurement data comprising spectral sensor data; and
classifying the measurement data using a classification algorithm for providing an assignment of the at least one measurement parcel to at least one class of interest, based on:
the measurement data, and
a classification model comprising at least the at least one class of interest, wherein the assignment is performed using a feedback procedure providing feedback data comprising at least one of:
verification information for the assignment of the measurement parcel to the at least one class of interest,
change information of the assignment of the measurement parcel to the at least one class of interest,
a definition of a new class of interest,
an instruction for removal of a class of interest from the classification model,
a first selection of the measurement data to be used for the classification, and
identification of a second selection of the measurement data to be ignored for further processing,
wherein the feedback data is provided to a training procedure, the training procedure being based on a machine learning algorithm and providing update information for the classification model, wherein the feedback data is provided based on implicit feedback, wherein the implicit feedback is based on at least one of:
an interpretation of a lack of explicit feedback as a confirmation of the assignment, or
a comparison of an actually executed action on the at least one measurement parcel with a proposal for an action being based on the assignment of the at least one measurement parcel.
2. The method according to claim 1 , wherein the feedback data is provided based on explicit feedback by means of the feedback procedure.
3. The method according to claim 1 , wherein training procedure is adapted for processing a multitude of measurement data, feedback data and auxiliary data.
4. The method according to claim 1 , wherein the training procedure provides a set of different classification models having specific classification parameters or specific classes of interest for at least one of:
a region-specific classification,
a time-specific classification,
a soil and/or plant type specific classification, and
a history specific classification taking into account at least one of a cultivation history, a farming action history, and a harvesting history,
wherein the classification model from the set of classification models for the classification of the measurement data being provided based on at least one of:
a location information,
a time information,
a soil and/or plant type information, and
a field history information,
corresponding with the acquisition of the measurement data.
5. The method according to claim 1 , wherein the update information for the classification model is made available to a multitude of classification algorithms.
6. The method of claim 1 , wherein the feedback procedure is supported by a notification functionality providing a status information of the classification,
wherein the notification functionality being based on at least one of:
a visual representation of the status information, and
an acoustic notification of the status information.
7. The method according to claim 1 , wherein the classification is based on at least one of:
a class of interest based on soil content or soil humidity,
a class of interest based on at least one of plant status, plant health, and plant maturity,
a class of interest based on bio mass,
linear classification,
a support vector machine,
a quadratic classifier,
Kernel estimation,
boosting,
a decision tree,
deep learning,
learning vector quantization.
8. A system for predicting a soil or plant condition in precision agriculture, the comprising:
a measurement device with a data acquisition unit for an acquiring measurement data for at least one measurement parcel, the measurement data comprising
spectral sensor data;
a classification unit for executing a classification of the measurement data using a classification algorithm for providing an assignment of the at least one measurement parcel to at least one class of interest, based on the measurement data and a classification model made available to the classification unit comprising at least the at least one class of interest;
a feedback functionality for processing the assignment by a feedback procedure providing feedback data comprising at least one of:
verification information for the assignment of the measurement parcel to the at least one class of interest,
change information of the assignment of the measurement parcel to the at least one class of interest,
a definition of a new class of interest,
an instruction for removal of a class of interest from the classification model,
a first selection of the measurement data to be used for the classification, and
identification of a second selection of the measurement data to be ignored for further processing, and
wherein the feedback data is provided to a training unit comprising a training procedure, the training procedure being based on a machine learning algorithm and providing update information for the classification model, wherein the feedback data is provided based on implicit feedback, wherein the implicit feedback is based on at least one of:
an interpretation of a lack of explicit feedback as a confirmation of the assignment, or
a comparison of an actually executed action on the at least one measurement parcel with a proposal for an action being based on the assignment of the at least one measurement parcel.
9. The system according to claim 8 , wherein the system is built such that the feedback data can be provided based on
explicit feedback by means of the feedback functionality.
10. The system according to claim 8 , wherein the training unit is built for processing a multitude of measurement data, feedback data and auxiliary data.
11. The system according to claim 8 , wherein the training unit is adapted such that the training procedure provides a set of different classification models having specific classification parameters or specific classes of interest for at least one of:
a region-specific classification,
a time-specific classification,
a soil or plant type specific classification, and
a history specific classification taking into account at least one of a cultivation history, a farming action history, and a harvesting history,
wherein the classification model of the set of different classification models for the classification of the measurement data being provided based on at least one of:
a location information,
a time information,
a soil or plant type information, and
a field history information,
corresponding with the acquisition of the measurement data.
12. The system according to claim 8 , wherein the update information for the classification model is made available to a multitude of classification algorithms.
13. The system according to claim 12 , wherein the feedback procedure is supported by a notification functionality providing a status information of the classification.
14. The system according to claim 8 , wherein the classification is based on at least one of:
a class of interest based on soil content or soil humidity,
a class of interest based on at least one of plant status, plant health, and plant maturity,
a class of interest based on bio mass,
linear classification,
a support vector machine,
a quadratic classifier,
Kernel estimation,
boosting,
a decision tree, deep learning, and
learning vector quantization.