Predicting soil EC using spectroscopy and smartphone-based digital images

GÖZÜKARA G., ANAGÜN Y., IŞIK Ş., Zhang Y., Hartemink A. E.

CATENA, vol.231, 2023 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 231
  • Publication Date: 2023
  • Doi Number: 10.1016/j.catena.2023.107319
  • Journal Name: CATENA
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aquatic Science & Fisheries Abstracts (ASFA), Artic & Antarctic Regions, BIOSIS, CAB Abstracts, Environment Index, Veterinary Science Database, DIALNET, Civil Engineering Abstracts
  • Keywords: Digital image processing, Soil color, Deep learning, Vis-NIR, pXRF, REFLECTANCE SPECTROSCOPY, ORGANIC-MATTER, COLOR SENSOR, CAMERA, SALINITY, FIELD, PXRF, CARBON, CLAY, TOOL
  • Eskisehir Osmangazi University Affiliated: Yes

Abstract

The soil EC was predicted using smartphone-based (iPhone 11) color coordinates (RGB, HSV, and CIE L*a*b*) and individual and combined Vis-NIR and pXRF spectra. Prediction models used were: one-dimension (1D) datadriven machine learning and recurrent neural network (RNN), and two-dimension (2D) convolutional neural network (CNN). A total of 240 soil samples were collected from 0 to 20 cm depth and air-dried samples were used for smartphone-based digital images and spectra (Vis-NIR and pXRF). Most smartphone-based and Vis-based color coordinates could be used to predict EC and salinity classes. Combined Vis-NIR and pXRF spectra had the highest prediction accuracy (R2 = 0.93) for predicting EC compared to individual Vis-NIR or pXRF spectra and smartphone-based and Vis-based color coordinates. We conclude that smartphone-based digital images based on the 2D-CNN model can be used to predict EC, but we recommend using combined Vis-NIR + pXRF spectra with gated recurrent unit (GRU) model for the highest prediction accuracy.