Two-Stage Soil Parameter Estimation From Multispectral Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to accurately and efficiently combine historical and modern data, such as satellite imagery, for estimating dynamic soil parameters like organic carbon and soil moisture, due to the limitations of one-stage machine learning architectures and the mismatch between data types.
Innovation Solution
A two-stage machine learning architecture is introduced, comprising a first stage estimator with a deep neural network for low-dynamic parameter estimation and a second stage estimator using classic algorithms, connected by a fully connected layer for latent feature extraction, to effectively process both historical and modern data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a one-stage machine learning architecture is used for soil parameter estimation, then the system structure is simple, but the accuracy of estimating dynamic soil parameters is insufficient
Solution Approach 1:
The patent divides the estimation system into two distinct stages: a first stage estimator using deep neural networks for low-dynamic parameters (soil texture, organic carbon) and a second stage estimator using classic machine learning algorithms for high-dynamic parameters (soil moisture, nitrogen, phosphorus, potassium). This segmentation allows each stage to be optimized for its specific parameter type, improving overall estimation accuracy while maintaining manageable system complexity through modular architecture.
2Reliability
If historical data and modern satellite imagery are combined, then the training data coverage is improved, but the data integration complexity increases
Solution Approach 1:
The patent performs preliminary processing of historical soil survey data and modern satellite imagery separately before integration. The first stage estimator is trained on pre-processed historical data for low-dynamic parameters, while the second stage estimator uses pre-processed satellite imagery for high-dynamic parameters. This preliminary action approach simplifies the integration process by preparing data in advance according to specific requirements, reducing the complexity of real-time data fusion.
3Measurement precision
If different machine learning methods are used for different parameter types, then the estimation accuracy for each parameter type is improved, but the system complexity increases
Solution Approach 1:
The patent applies the principle of local quality by matching different machine learning algorithms to specific parameter characteristics. Deep neural networks are used for low-dynamic parameters (soil texture, organic carbon) that require capturing complex spatial patterns, while classic machine learning algorithms are used for high-dynamic parameters (soil moisture, nutrients) that require processing temporal variations. This localized algorithm selection optimizes estimation accuracy for each parameter type while maintaining clear functional boundaries in the system.
Data Source
Figure 1a
Figure 1b
Figure 2
AI summary
The invention proposes a new system and method for estimating dynamic soil parameters based on multispectral or hyperspectral images. The system (105) has two-stage estimation architecture and a fully connected layer (108) that connects the two estimators. The system estimates dynamic parameters such as organic carbon in the soil based on multispectral or hyperspectral images (102) and also uses geocoordinates (100) and soil elevation (101) as inputs. Multispectral or hyperspectral images (102) are further processed in the pre-processing module (103) and the data (104) is transmitted to the training neural network (107) of the 1st degree estimation module (106). It further processes them in submodule (108) for the extraction of latent features, estimates the low-dynamic parameters of the soil and forwards them to the output. Also, the output data of submodule (108) is the input data of module (2) for the 2nd degree estimation which estimates the highly dynamic parameters. The low-dynamic parameters (112) are e.g. soil texture, and highly dynamic parameters (113) are e.g. the concentration of organic carbon in the soil. Submodule (108) connects said modules (106) and (110).