An integrated feature selection approach to high water stress yield prediction DOI Creative Commons
Zongpeng Li, Xinguo Zhou,

Qian Cheng

et al.

Frontiers in Plant Science, Journal Year: 2023, Volume and Issue: 14

Published: Dec. 4, 2023

The timely and precise prediction of winter wheat yield plays a critical role in understanding food supply dynamics ensuring global security. In recent years, the application unmanned aerial remote sensing has significantly advanced agricultural research. This led to emergence numerous vegetation indices that are sensitive variations. However, not all these universally suitable for predicting yields across different environments crop types. Consequently, process feature selection index sets becomes essential enhance performance models. study aims develop an integrated method known as PCRF-RFE, with focus on selection. Initially, building upon prior research, we acquired multispectral images during flowering grain filling stages identified 35 yield-sensitive indices. We then applied Pearson correlation coefficient (PC) random forest importance (RF) methods select relevant features set. Feature filtering thresholds were set at 0.53 1.9 respective methods. union selected by both was used recursive elimination (RFE), ultimately yielding optimal subset constructing Cubist Recurrent Neural Network (RNN) results this demonstrate model, constructed using obtained through (PCRF-RFE), consistently outperformed RNN model. It exhibited highest accuracy stages, surpassing models or subsets derived from single method. confirms efficacy PCRF-RFE offers valuable insights references future research realms studies.

Language: Английский

SHAP values accurately explain the difference in modeling accuracy of convolution neural network between soil full-spectrum and feature-spectrum DOI
Liang Zhong, Guo Xi, Meng Ding

et al.

Computers and Electronics in Agriculture, Journal Year: 2024, Volume and Issue: 217, P. 108627 - 108627

Published: Jan. 13, 2024

Language: Английский

Citations

21
Evaluating the Soil Quality Index Using Three Methods to Assess Soil Fertility DOI Creative Commons
Hiba Chaudhry, Hitesh B. Vasava, Songchao Chen

et al.

Sensors, Journal Year: 2024, Volume and Issue: 24(3), P. 864 - 864

Published: Jan. 29, 2024

Soil health plays a crucial role in crop production, both terms of quality and quantity, highlighting the importance effective methods for preserving soil to ensure global food security. indices (SQIs) have been widely utilized as comprehensive measures function by integrating multiple physical, chemical, biological properties. Traditional SQI analysis involves laborious costly laboratory analyses, which limits its practicality. To overcome this limitation, our study explores use visible near-infrared (vis-NIR) spectroscopy rapid non-destructive alternative predicting properties SQIs. This specifically focused on seven indicators that contribute fertility, including pH, organic matter (OM), potassium (K), calcium (Ca), magnesium (Mg), available phosphorous (P), total nitrogen (TN). These play key roles nutrient availability, pH regulation, structure, influencing fertility overall health. By utilizing vis-NIR spectroscopy, we were able accurately predict with good accuracy using Cubist model (R2 = 0.35–0.93), offering cost-effective environmentally friendly traditional analyses. Using indicators, looked at three different approaches calculating SQI, including: (1) measured (SQI_m), is derived from laboratory-measured properties; (2) predicted (SQI_p), calculated spectral data; (3) direct prediction (SQI_dp), The findings demonstrated SQI_dp exhibited higher 0.90) compared SQI_p 0.23).

Language: Английский

Citations

20
Combination of feature selection and geographical stratification increases the soil total nitrogen estimation accuracy based on vis-NIR and pXRF spectral fusion DOI
Jianghui Song, Xiaoyan Shi, Haijiang Wang

et al.

Computers and Electronics in Agriculture, Journal Year: 2024, Volume and Issue: 218, P. 108636 - 108636

Published: Jan. 31, 2024

Language: Английский

Citations

11
Machine Learning and Feature Selection for soil spectroscopy. An evaluation of Random Forest wrappers to predict soil organic matter, clay, and carbonates DOI Creative Commons
Francisco M. Canero, Víctor Rodríguez‐Galiano, David Aragonés

et al.

Heliyon, Journal Year: 2024, Volume and Issue: 10(9), P. e30228 - e30228

Published: April 25, 2024

Soil spectroscopy estimates soil properties using the absorption features in spectra. However, modelling with is challenging due to high dimensionality of spectral data. Feature Selection wrapper methods are promising approaches reduce but barely used spectroscopy. The aim this study evaluate performance two feature selection methods, Sequential Forward (SFS) and Flotant (SFFS) built Random Forest (RF) algorithm, for reduction data predictive organic matter (SOM), clay carbonates. reflectance 100 samples, acquired from Sierra de las Nieves (Spain), was measured under laboratory conditions ASD FieldSpec Pro JR. Four different datasets were obtained after applying preprocessing raw spectra: spectra, Continuum Removal (CR), Multiplicative Scatter Correction (MSC), a so-called "Global" dataset composed raw, CR MSC features. RF models compared that Partial Least Squares Regression (PLSR) (alone). SFS SFFS outperformed PLSR alone models: best had respective ratio interquartile distance 1.93, 0.38 2.56. an accuracy 1.41, 0.29 1.81 SOM, carbonates, clay, respectively. 1.29, 1.81. application reduced number less than 1 % starting Features selected across all spectra SOM around 900 nm, 1900 2350 nm highlighted 1100 modelling, as well other 2200 which considered main clay. very important improving accuracy, reducing redundant avoiding curse or Hughes effect. Thus, research showed alternative have been applied date model paves way further scientific investigation based on machine learning.

Language: Английский

Citations

8
Non-linear memory-based learning for predicting soil properties using a regional vis-NIR spectral library DOI Creative Commons
Zheng Wang, Songchao Chen, Rui Lu

et al.

Geoderma, Journal Year: 2023, Volume and Issue: 441, P. 116752 - 116752

Published: Dec. 11, 2023

Visible near-infrared (vis-NIR) spectroscopy has gained widespread recognition as an efficient and reliable approach for the rapid monitoring of soil properties. This technique relies on robust machine learning models that convert spectra information to In particular, memory-based (MBL) emerged a powerful local modeling spectral analysis. However, conventional MBL algorithms use linear models, disregarding non-linear relationship between properties vis-NIR spectra. Therefore, we hypothesize (N-MBL) can enhance prediction. study develops evaluates N-MBL algorithm using Lateritic Red library (LRSSL) from Guangdong province in China. consists 742 samples corresponding properties, including pH, organic matter (SOM), total nitrogen (TN), phosphorus (TP), potassium (TK). As comparison, several commonly used supervised methods, such Partial least squares regression (PLSR), Cubist, Random Forest (RF), Super vector (SVM), Convolution neural network (CNN), (MBL), were compared proposed N-MBL. The results showed generally outperformed particularly when applied large with substantial number (over 500). When comparing two had more fluctuation model performance selected nearest neighbors (k) varied 30 250. k increased, higher R2 values SOM TN prediction than but lower pH TK addition, predicting TP. conclusion, is new It high potential improve accuracy multiple

Language: Английский

Citations

21
National-scale mapping of soil organic carbon stock in France: New insights and lessons learned by direct and indirect approaches DOI Creative Commons
Zhongxing Chen, Shuai Qi, Zhou Shi

et al.

Soil & Environmental Health, Journal Year: 2023, Volume and Issue: 1(4), P. 100049 - 100049

Published: Nov. 11, 2023

Soil organic carbon (SOC) plays a crucial role in soil health and global cycling, therefore accurate estimates of its spatial distribution are important for managing mitigating climate change. Digital mapping shows potential to provide high-resolution SOC across scales. To convert content density (SOCD), two inference trajectories exist predicting SOCD digital mapping: the direct approach (calculate-then-model) indirect (model-then-calculate). However, there is lack comprehensive exploration regarding differences their performance estimates, particularly regions characterized by diverse pedoclimatic conditions. bridge this knowledge gap, we evaluated approaches based on model France. Using 916 topsoils (0−20 cm) from LUCAS 2018 24 environmental covariates, random forest forward recursive feature selection were used build predictive models using approaches. The results show that, full both showed moderate (R2 = 0.28−0.32). By utilizing model, number predictors was reduced 9, enhancing 0.35) with no improvement 0.28). mean French topsoil 5.29 6.14 kg m-2 approaches, resulting stock (SOCS) 2.8 3.3 Pg, respectively. We found that clearly underestimated high (>9 m-2), while performed much better SOCD. Our findings serve as valuable reference mapping, thereby providing scientific basis maintaining health.

Language: Английский

Citations

17
The validity domain of sensor fusion in sensing soil quality indicators DOI Creative Commons
Jie Xue, Xianglin Zhang, Songchao Chen

et al.

Geoderma, Journal Year: 2023, Volume and Issue: 438, P. 116657 - 116657

Published: Sept. 4, 2023

Soil health has gained increasing attention under the rapid development of industrialization and requirement for green agriculture. Therefore, up-to-date soil information related to is urgently needed ensure food security biodiversity protection. Previous studies have shown potential proximal sensing in measuring information, while it remains challenging get cost-efficient robust estimates multiple indicators simultaneously via sensor fusion. In this study, we investigated visible near-infrared (vis-NIR), mid-infrared (MIR) spectroscopy as well three model averaging methods predicting properties, including organic matter (SOM), pH, cation exchange capacity (CEC). The are not only used fusion but also high-level fusion, which include Granger-Ramanathan (GR), Bayesian Model Averaging Spectral-Guided Ensemble Modelling (S-GEM). Here, S-GEM a recently proposed algorithm that can improve spectroscopic prediction by spectral ensemble modelling. Four widely models were evaluated, partial least square regression, Cubist, memory based learning convolutional neural network. For SOM, on algorithms was comparable Sensorsingle + Modelmultiple (MIR singly S-GEM) with R2 0.86. However, MIR performed best among all (LCCC 0.92, RMSE 3.66 g kg−1 RPIQ 3.68). 10-fold cross-validation results indicated 0.84, LCCC 0.90, 0.45 3.65. CEC, Sensormultiple GR 0.66, 0.80, 3.48 cmol 2.22. Our showed failed when performance sensors differed lot (△R2 > 0.2), use single therefore suggested case. When close < recommended. outcome study provide reference determining validity domain improving accuracy prediction.

Language: Английский

Citations

13
Improving model performance in mapping cropland soil organic matter using time-series remote sensing data DOI Creative Commons
Xianglin Zhang, Jie Xue, Songchao Chen

et al.

Journal of Integrative Agriculture, Journal Year: 2024, Volume and Issue: 23(8), P. 2820 - 2841

Published: Jan. 9, 2024

Faced with increasing global soil degradation, spatially explicit data on cropland organic matter (SOM) provides crucial for carbon pool accounting, quality assessment and the formulation of effective management policies. As a spatial information prediction technique, digital mapping (DSM) has been widely used to map at different scales. However, accuracy SOM maps is typically lower than other land cover types due inherent difficulty in precisely quantifying human disturbance. To overcome this limitation, study systematically assessed framework "information extraction-feature selection-model averaging" improving model performance using 462 samples collected Guangzhou, China 2021. The results showed that dynamic extraction, feature selection averaging could efficiently improve final predictions (R2: 0.48 0.53) without having obviously negative impacts uncertainty. Quantifying environment was an efficient way generate covariates are linearly nonlinearly related SOM, which improved R2 random forest from 0.44 extreme gradient boosting 0.37 0.43. FRFS recommended when there relatively few environmental (<200), whereas Boruta many (>500). granger-ramanathan approach average When structures initial models similar, number did not have significantly positive effects predictions. Given advantages these selected strategies over great potential high-accuracy any scales, so can provide more reliable references conservation policy-making.

Language: Английский

Citations

5
Ground-Based Hyperspectral Estimation of Maize Leaf Chlorophyll Content Considering Phenological Characteristics DOI Creative Commons
Yiming Guo, Shiyu Jiang,

Huiling Miao

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(12), P. 2133 - 2133

Published: June 13, 2024

Accurately measuring leaf chlorophyll content (LCC) is crucial for monitoring maize growth. This study aims to rapidly and non-destructively estimate the LCC during four critical growth stages investigate ability of phenological parameters (PPs) LCC. First, spectra were obtained by spectral denoising followed transformation. Next, sensitive bands (Rλ), indices (SIs), PPs extracted from all at each stage. Then, univariate models constructed determine their potential independent estimation. The multivariate regression (LCC-MR) built based on SIs, SIs + Rλ, Rλ after feature variable selection. results indicate that our machine-learning-based LCC-MR demonstrated high overall accuracy. Notably, 83.33% 58.33% these showed improved accuracy when successively introduced SIs. Additionally, model accuracies milk-ripe tasseling outperformed those flare–opening jointing under identical conditions. optimal was created using XGBoost, incorporating SI, PP variables R3 These findings will provide guidance support management.

Language: Английский

Citations

4
Using visible-near infrared spectroscopy to estimate whole-profile soil organic carbon and its fractions DOI Creative Commons

Mingxuan Qi,

Songchao Chen, Yu‐Chen Wei

et al.

Soil & Environmental Health, Journal Year: 2024, Volume and Issue: 2(3), P. 100100 - 100100

Published: July 18, 2024

Soil organic carbon (SOC) is crucial for soil health and quality, its sequestration has been suggested as a natural solution to climate change. Accurate cost-efficient determination of SOC functional fractions essential effective management. Visible near-infrared spectroscopy (vis-NIR) emerged approach. However, ability predict whole-profile content rarely assessed. Here, we measured two fractions, particulate (POC) mineral-associated (MAOC), down depth 200 ​cm in seven sequential layers across 183 dryland cropping fields northwest, southwest, south China. Then, vis-NIR spectra the samples were collected train machine learning model (partial least squares regression) SOC, POC, MAOC, ratio MAOC (MAOC/SOC – an index vulnerability). We found that accuracy indicated by coefficient validation (Rval2) 0.39, 0.30, 0.49, 0.48 MAOC/SOC, respectively. Incorporating mean annual temperature improved performance, Rval2 was increased 0.64, 0.31, 0.63, 0.51 four variables, Further incorporating into 0.82, 0.59, These results suggest combining with readily-available data total measurements enables fast accurate estimation POC diverse environmental conditions, facilitating reliable prediction dynamics over large spatial extents.

Language: Английский

Citations

4