?? AI for Soil: Prediction of Soil Water Content
Aerial and ground-truth imagery provide spatially extensive data on soil moisture

?? AI for Soil: Prediction of Soil Water Content

Predicting soil water content is important for effective agricultural management, irrigation planning, and environmental monitoring.

The application of artificial intelligence in this domain offers significant advantages due to its ability to handle large datasets, identify complex patterns, and provide accurate predictions. Today we'll focus on predicting soil water content based on images captured with a digital camera and drone.

We'll examine which machine learning models for soil analysis best utilize the dependency between the color of the soil and soil water content.

How practically useful are the findings of these soil science studies?


Rapid Prediction of Soil Water Content Using Colour Information and Machine Learning

Country: China ????

Published: 2 March 2023

This study aims to develop a non-invasive, efficient method for predicting soil water content (SWC) by analyzing soil images using Gaussian-fitting of grey histograms and machine learning models.

The methods involve capturing soil images under controlled illumination, extracting characteristic colour parameters from multiple colour spaces, and using machine learning models (PLSR, RF, SVMR, GPR) to predict SWC. The process includes image segmentation, optimal area sampling, and statistical analysis to identify the most relevant predictors.

Key findings show that the Gaussian-fitting method effectively reduces interference from complex soil surfaces. The GPR model achieved the highest prediction accuracy (R2 = 0.95, RMSE = 2.01%, RPD = 4.95, RPIQ = 6.37).


Colour Parameters for Soil Water Content Prediction

Ten key colour parameters were identified as optimal predictors without significant loss of accuracy compared to 32 parameters.

The ten key color parameters identified as optimal predictors for soil water content (SWC) in the study are:

  1. mRc: Gaussian-fitting expectation of the red channel
  2. mGc: Gaussian-fitting expectation of the green channel
  3. mBc: Gaussian-fitting expectation of the blue channel
  4. HSV_V: Value in the HSV color space
  5. HSL_S: Saturation in the HSL color space
  6. HSL_L: Lightness in the HSL color space
  7. HI: Hue index
  8. LCH_H: Hue in the CIELCH color space
  9. CIE_Z: Virtual component in the CIEXYZ color space
  10. RI: Redness index

These parameters were selected based on their significance and contribution to the SWC prediction models

The results can be practically applied by agricultural engineers, geotechnical professionals, and environmental scientists for real-time soil moisture monitoring.

Main tools/technologies

- Digital Image Analysis

- Gaussian-fitting Gray Histogram

- Partial Least Squares Regression (PLSR)

- Random Forest (RF)

- Support Vector Machines Regression (SVMR)

- Gaussian Process Regression (GPR)

- MATLAB

Soil texture triangle representing all samples (
Testing apparatus: (a) Equipment of photography platform; and (b) A photograph under the experimental conditions with a soil specimen. Source: Liu et al., 2023
Soil image processing: (a) Image pre-processing; (b) Gaussian-fitting RGB histogram; and (c) Extracting the identity matrix of soil image from multi-colour spaces. Source: Liu et al., 2023
Fused colour characteristic parameters of colour spaces of 336 soil samples. Source: Liu et al., 2023
Steps used in the proposed soil image-based method for SWC prediction. Source: Liu et al., 2023
Summary statistics for the SWC in Soils 1–4. Source: Liu et al., 2023
(a) A typical soil image; (b) Spatial distribution of soil colour; and (c) and (d) Fluctuation of grey level of soil image surface along the
The performance of the Gaussian-fitting in acquiring characteristic parameters from soil images with different surface conditions: (a) Soil images with same water content but different holes and shadows; (b) and (c) Surface three-dimensional (3D) gray levels; (d) Histogram fitting; (e) Image with water film derived from
Mean, variance and range of RGB values for different sampling areas from 1124 pixels to 3,021,374 pixels. Source: Liu et al., 2023
Relationship curves between characteristic parameters in RGB color space and SWC of Soils 1–4. Source: Liu et al., 2023
The importance of characteristic parameters in RF model. Source: Liu et al., 2023

Soil Colour Analysis Based on a RGB Camera and an Artificial Neural Network Towards Smart Irrigation

Country: Iraq ????, Australia ????

Published: 20 January 2021

The focus of this study is targeted on a non-contact vision system using an RGB camera and a feed-forward back propagation neural network to predict irrigation requirements for loam soils by analyzing soil colour variations.

The methods involved collecting soil images under different conditions (distance, time, and illumination), using an RGB camera mounted on a tripod, and processing these images to extract RGB colour values. These values were used as input to an artificial neural network (ANN) to determine soil hydration status. The ANN was configured with two hidden layers, each with ten neurons, to achieve optimal performance.

Key findings include the system's high accuracy, with mean square errors (MSE) of 1.616 × 10?? for training, 1.004 × 10?? for testing, and 1.809 × 10?? for validation.

The ANN effectively classified soil types under various conditions, achieving a high coefficient of determination (R2 = 1 for training and testing).

Agricultural engineers, farmers, and irrigation managers can practically apply the results to improve irrigation efficiency and water management.

Main tools/technologies

  • RGB Camera
  • Feed-Forward Back Propagation Neural Network
  • MATLAB
  • Image Processing Techniques


Data collection of the loam soil under four scenarios (a) S1: sunny-dry, (b) S2: sunny-wet, (c) S3: shadow-dry, and (d) S4: shadow-wet. Source: Al-Naji et al., 2021
Data collection of soil surface colors (a) red, (b) green, and (c) blue. Source: Al-Naji et al., 2021
The block diagram illustrating the process of capturing soil images at different scenarios using a digital camera and a neural network system. Source: Al-Naji et al., 2021
The adopted ANN structure of the feed-forward back propagation neural network. Source: Al-Naji et al., 2021
Adopted ANN with related parameters. Source: Al-Naji et al., 2021
MSE of training, validation, and testing for the adopted ANN. Source: Al-Naji et al., 2021
Comparison of the current study with state-of-the-art. Source: Al-Naji et al., 2021

Smart Soil Moisture Prediction with UAV Multispectral and Thermal Imaging

Country: Hungary ????

Published: 1 August 2022

The final study for today aims to evaluate the efficacy of UAV-based thermal and multispectral cameras combined with machine learning algorithms in predicting soil water content (SWC) in agricultural fields.

The methods involved UAV flights equipped with thermal and multispectral cameras to capture soil data. Soil samples were collected and analyzed in the laboratory for SWC, which served as reference data. Machine learning algorithms (Random Forest [RF], Elastic Net Regression [ENR], General Linear Model [GLM], Robust Linear Model [RLM]) were tested using three pixel value extraction methods (single pixel, mean of 20 cm radius buffer, mean of 30 cm radius buffer).

Key findings showed that multispectral cameras provided better data for SWC prediction than thermal cameras, with the RF model achieving the highest accuracy (R2 = 0.97, nRMSE = 10%).

Thermal data were less accurate but still useful, especially with the RF model (nRMSE = 24.4%). Single pixel extraction proved to be the most effective method for data input.

Agricultural engineers and farmers can practically apply the results to enhance precision irrigation and water management.

Main tools/technologies

  • UAV (DJI Matrice M210)
  • Multispectral Camera (MicaSense RedEdge-MX)
  • Thermal Camera (Zenmuse XT2)
  • Machine Learning Algorithms
  • Pix4D Mapper
  • R Programming (caret, MASS, glmnet packages)

Location of the study area and parts of the survey equipments: (a) location of Hajdúb?sz?rmény in a European map; (b) location of the sampled agricultural area at Hajdúb?sz?rmény; (c) the red rectangle represents the sampled area; (d) a DJI Matrice M210 UAV?+?Zenmuse XT2 camera; (e) the specific design of our Ground Control Points (GCP) for mosaicking the thermal imagery; and (f) a MicaSense CRP (Calibrated Reflectance Panel). Source: Bertalan et al., 2022
Root Mean Square Errors (RMSE) and R2-values (Rsquared) of prediction models based on the spectral data of surveys conducted at 04.06.2019 (a) and at 14.09.2020 (b) based on 30 cross-validated models (GLM: General Linear Model, RLM: Robust Linear Model, RF: Random Forest Regression, ENR: Elastic Net Regression; simple abbreviations: pixel-based data; 0.20: mean of pixels of 20?cm buffer, 0.30: mean of pixels of 30?cm buffer; left and right end of boxes: lower and upper quartiles, ?: medians, dashed lines: 1.5-time interquartile ranges). Source: Bertalan et al., 2022


Root Mean Square Errors (RMSE) and R2-values (Rsquared) of prediction models using the merged datasets of surveys conducted in the morning and at noon, based on 30 cross-validated models (GLM: General Linear Model, RLM: Robust Linear Model, RF: Random Forest Regression, ENR: Elastic Net Regression; simple abbreviations: pixel based data; 0.20: mean of pixels of 20?cm buffer, 0.30: mean of pixels of 30?cm buffer; left and right end of boxes: lower and upper quartiles, ?: medians, dashed lines: 1.5 time interquartile ranges). Source: Bertalan et al., 2022


Estimation of soil water content (SWC) with multispectral (a) and thermal (b) cameras with the difference of two estimations (c) and the UAV-based RGB orthomosaic of the plot (d). CRS: UTM34N. Source: Bertalan et al., 2022
Taylor diagram of soil water content (%) predicted by Elastic Net Regression (ENR), General Linear Model (GLM), Random Forest Regression (RF) and Robust Linear Regression (RLM) (merged datasets of surveys at 04.06.2019 and 14.09.2020). Source: Bertalan et al., 2022
Nash-Sutcliffe model efficiencies (NSE) of different predictions using General Linear Models (GLM), Robust Linear Models (RLM), Random Forest Regression (RF) and Elastic Net Regression (ENR) with multispectral (m.) and thermal (t.) data (models without numbers: pixel-based sampling, 0.20: mean values of 20?cm buffers around the sampling points, 0.30: mean values of 30?cm buffers around the sampling points). Source: Bertalan et al., 2022



What's next in "AI for Soil"?

Next time we hope to find more interesting resources about applying machine learning for soil analysis. Particularly, we will take a look at estimating soil organic carbon and visual assessment of other important soil characteristics.

Would you like to share your opinion about this choice of the topic?

We'll be grateful for a few words in the comments below ??

Stay in touch! ??


Wishes of accurate predictions of soil water content,

Maryna Kuzmenko, Ph.D ????, Chief Inspiration Officer at Petiole Pro Community

#soilanalysis

Photo credit for the cover image: Al-Naji, A., Fakhri, A. B., Gharghan, S. K., & Chahl, J. (2021). Soil color analysis based on a RGB camera and an artificial neural network towards smart irrigation: A pilot study. Heliyon, 7(1), e06078. ISSN 2405-8440. https://doi.org/10.1016/j.heliyon.2021.e06078.


References for today's "AI for Soil" edition


Abdul Manan

Engineer || AgTech || Precision Crop Protection Researcher || UAV's

8 个月

Interesting!

Ishu Bansal

Optimizing logistics and transportation with a passion for excellence | Building Ecosystem for Logistics Industry | Analytics-driven Logistics

8 个月

Innovative How can we use AI and color parameters to accurately predict soil water content in agriculture?

Avinash Chandra Pandey

Crop Improvement Researcher

8 个月

Maryna Kuzmenko, Ph.D ???? Already there are submergence tolerance genes (sub1,2,3) identified and successfully introgressed in many paddy mega varieties for submergence tolerance that can help plants even up to 30 days under the water. On the other hand, there is a huge problem in wheat and maize when late monsoon or pre-winter rainfall increases moisture content that either affects on seed germination (by creating anaerobic conditions in soil) or heavy moisture in the soil leads to seedling death. Contrarily in the same field due to drought during panicle emergence or grain filling leads the yield loss to farmers. In those cases up to the discovery of tolerant genes. Maryna Kuzmenko, Ph.D ???? work can significantly help in research work and even local farmers to reduce the cost of cultivation. No words besides "excellent work Maryna Kuzmenko, Ph.D ????"

要查看或添加评论,请登录

Maryna Kuzmenko的更多文章

社区洞察

其他会员也浏览了