Adaptive Neural Network Calibration for Hyperspectral Plant Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining substance concentrations in plants using hyperspectral image data are limited in accuracy and reliability, particularly for low concentrations of nutrients, and are not suitable for precise agricultural applications.
Innovation Solution
A training method for an adaptive evaluation algorithm that calibrates an artificial neural network using a variety of test plants in controlled physiological states, allowing for the generation of training cases with high variability in substance concentrations, enabling the algorithm to accurately determine a large number of substance concentrations, including those below the sensitivity threshold of the camera, without requiring knowledge of hyperspectral reflection properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional training methods using randomly selected plants are used, then the training process is simple, but the determination accuracy of substance concentrations (especially low concentrations) is insufficient
Solution Approach 1:
The patent applies preliminary action by bringing test plants into predetermined physiological states through controlled treatment before measurement. This pre-conditioning of plants with specific nutrient concentrations creates structured training data that enables the neural network to learn accurate relationships between hyperspectral signatures and substance concentrations, particularly for low concentrations that would otherwise be indistinguishable from noise.
Solution Approach 2:
The patent changes parameters by systematically varying nutrient concentrations in test plants to create diverse training samples. By controlling physiological states and generating training cases with known substance concentrations across a wide range, the training dataset covers extreme values and low concentrations, enabling the evaluation algorithm to accurately determine substance concentrations that were previously undetectable.
2Measurement precision
If the number of substance concentrations to be determined is increased, then more precise statements about plant health are possible, but the calibration of the neural network becomes difficult
Solution Approach 1:
The patent applies universality by creating a comprehensive training dataset that covers multiple substance types and concentration ranges simultaneously. The controlled physiological states approach generates training cases that can be used to calibrate the neural network for determining various substance concentrations (nitrogen, phosphorus, potassium, and other nutrients) with a single unified training process, rather than requiring separate calibration for each substance.
Solution Approach 2:
The patent uses feedback by providing the neural network with training cases where both input vectors (hyperspectral image data) and output vectors (substance concentrations from chemical analysis) are known. This supervised learning approach with error feedback enables the network to iteratively optimize its parameters and achieve accurate determination of multiple substance concentrations simultaneously.
3Measurement precision
If conventional vegetation indices are used, then the evaluation process is simple, but the accuracy and certainty of plant health determination are insufficient
Solution Approach 1:
The patent replaces the simple but inaccurate conventional vegetation index calculations with an artificial neural network-based adaptive evaluation algorithm. This substitution enables the system to process hyperspectral image data more effectively, identifying subtle spectral patterns associated with low substance concentrations and providing accurate, certain determinations of plant health status that go beyond the capabilities of traditional indices.
4Measurement precision
If the camera sensitivity is increased to detect low substance concentrations, then the detection capability improves, but the device cost and complexity increase
Solution Approach 1:
The patent introduces an intermediary approach by using controlled test plants with predetermined physiological states as mediators between the hyperspectral measurement system and the substance concentration determination. These pre-conditioned plants create distinct spectral signatures that amplify the signal from low substance concentrations, enabling the existing camera sensitivity to detect and quantify substances without requiring hardware modifications or increased camera sensitivity.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The method achieves high accuracy and reliability in determining substance concentrations, enabling precise statements about plant health and facilitating automated resource application, such as fertilizers and water, tailored to the current physiological state of plants.
Implementation Method 1
capturing a hyperspectral image of the at least one test plant and generating hyperspectral image data from the captured image, in which image data of a hyperspectral image point belonging to the at least one test plant is contained
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a training method for an adaptive evaluation algorithm, in particular for an artificial neural network, for determining substance concentrations (20) in at least one plant (25) based on hyperspectral image data (17) of the at least one plant (25), wherein parameters of the adaptive evaluation algorithm are calibrated based on a plurality of training cases (23), each of these training cases (23) being given by an input vector (11) comprising image data (17) of a hyperspectral image (1) of a test plant (2) belonging to the training case (23), and an output vector (21) comprising substance concentrations (20) within this test plant (2) belonging to the training case (23) determined by means of a chemical analysis (19), wherein each of the training cases (23) is generated by first bringing the test plant (23) belonging to the training case into a predetermined physiological state.The invention further relates to a hyperspectral measuring device (24) with an evaluation unit (18) and a device (27, 31, 36) for dispensing an operating medium comprising such a measuring device.