What question did this study set out to answer?

The study aims to evaluate the efficiency of random forest versus deep learning algorithms in digital soil mapping.

February 26, 2026

Digital Soil Mapping of Soil Depth and Textural Classes Using Random Forest and CNN Algorithms for Theri Soils of Tamil Nadu, India

MRM. Paul RajTamil Nadu Agricultural University RKR. KumaraperumalTamil Nadu Agricultural University SPS. PazhanivelanTamil Nadu Agricultural University

Key Points

The study aims to evaluate the efficiency of random forest versus deep learning algorithms in digital soil mapping.
Used 2000 topsoil samples for analysis
Applied random forest and 1D-resnet algorithms for soil prediction
Employed conditional hypercube sampling for data extraction
Utilized 30 environmental covariates for model calibration
Validated models using stratified data partitioning (70% calibration, 30% validation)
Random forest yielded a higher R2 (0.46) and lower RMSE (4.77 cm) compared to 1D-resnet (R2 0.16, RMSE 6.58 cm)
Overall accuracy for soil textural class prediction was higher for random forest (71%) than for 1D-resnet (54%)
Key contributing factors identified included mean air temperature, elevation, and geomorphology.

Abstract

With the data mining algorithms being optimized for digital soil mapping (DSM) procedures, determining the utility of the deep learning algorithms can help to increase the generalizability and transferability of the calibrated models. The present study aims at comparing the efficiency of the predominant random forest algorithm with that of the 1—Dimensional Residual Network (1D—ResNet 50) with bottleneck blocks variant. The pixel wise predictions and classifications were derived at 30 m spatial resolution for the soils of Tamil Nadu, occurring predominantly in the Thoothukudi district. Considering the voluminous data requirement nature of the deep learning algorithms, a total of 2000 Nos. topsoil sample observations (i.e., Soil textural classes and Depth) were extracted from the legacy soil database using the conditional hypercube sampling technique. The environmental covariates of different modalities, related to the SCORPAN factors (30 Nos.) were then utilized for calibrating the machine and deep learning models. Further, the validation of the models was performed based on the stratified data partitioning procedure (Calibration: 70%; Validation: 30%) with validation metrics for both soil textural classes and depth information. The coefficient of determination (R2) and Root Mean Square Error (RMSE) computed for the RF (0.46 and 4.77cm) and 1D ResNet 50 (0.16 and 6.58 cm) model varied greatly with deep learning algorithm having the marginal efficiency. Similarly, the overall accuracy (OA) computed for the textural class prediction was also found to be higher for RF (71%) than the 1D—ResNet50 algorithm (54%). Further, the permutation feature importance (PFI) measure resulted in Mean air temperature, Elevation, Geomorphology, Physiography, Ferrous Minerals Difference Ratio (FMDR) and Multiresolution Index of Valley Bottom Flatness (MRVBF) as the highly contributing covariates to the spatial soil predictions. Though RF depicted an increased versatility over DSM approaches, the lower efficiency of the deep learning model can be attributed to the large-scale implementation, besides the lack of integrating more specialized hyperparameter tuning mechanisms.

KI fragen

Bookmark

View Full Paper