Home » Machine-Deep Learning Ensembles for Tropical Forest Biomass Carbon Predictions

Number of Visitors

Machine-Deep Learning Ensembles for Tropical Forest Biomass Carbon Predictions

Stacked machine-deep learning from optimal Sentinel-1/2 and fused predictors for generalizable mapping of aboveground biomass-carbon and uncertainty in tropical dry dipterocarp forests

This guideline and accompanying executable package are based on the following reference:
Huy, B.,  Temesgen, H., Poudel, K.P., Truong, N.Q., Khiem, N.Q. (2026). Stacked machine-deep learning from optimal Sentinel-1/2 and fused predictors for generalizable mapping of aboveground biomass–carbon and uncertainty in tropical dry dipterocarp forests.

The package implements a novel stacked ensemble framework combining Deep Learning, Random Forest, and XGBoost (DL–RF–XG), utilizing dual DL–RF base learners and an XGBoost meta-learner. Paired with seven optimal Sentinel-1 and Sentinel-2 fused predictors (comprising backscatter, polarization indices, and texture features from both radar and optical imagery), it enables spatiotemporally generalizable mapping of aboveground biomass (AGB) and associated uncertainties in tropical dry dipterocarp forests (DDFs).

Step-by-step instructions:

1. Download and extract the package

Download and extract  ‘DL-RF-XG Maps Executable Package.rar‘. This archive contains the excecutable file ‘Maps_AGB_SD.exe’

2. Prepare input data

Place your seven preprocessed Sentinel-1/2 fused raster files corresponding to your tropical DDF study area into the same directory as the executable file. The required input files are: V_VV.tif, VH.tif, VVminusVH.tif, M_NDI45.tif, M_SAPI.tif, H_SAPI.tif, and V_RED_EDGE3.tif

3. Run the prediction

Double-click ‘Maps_AGB_SD.exe’. This excecutes the trained DL-RF-XG ensemble model using your provided raster predictors to generate the final prediction maps.

⏳ Note: Execution may take several minutes depending on your system specifications and the spatial extent/resolution of the input rasters.

4. Output files

Upon successful completion, the following raster files will be generated in the working directory:

  • AGB_map.tif – Aboveground Biomass Raster Map
  • SD_uncertainty_map.tif – Standard Deviation (SD) Uncertainty Raster Map
  • Relative_SD_uncertainty_map.tif – Relative SD Uncertainty Raster Map

These output rasters are ready for spatiotemporal analysis and biomass–carbon accounting with uncertainty assessment in tropical DDF ecosystems.
————————————–
Note: Predictor Glossary:
Sentinel-1 (Radar): V_VV: Variance of Vertical transmit–vertical receive polarization, VH: Vertical transmit–horizontal receive polarization, VVminusVH: VV – VH.
Sentinel-2 (Optical): M_NDI45: Mean of Normalized Difference Index 4 and 5 Red Edge, M_SAPI: Mean of Scatter-Adjusted Pigment Index, H_SAPI: Homogeneity of SAPI, V_RED_EDGE3: Variance of RED_EDGE3
_____________________________________________________________________________________________________________

Pantropical tree aboveground biomass predictions through generalizable machine–deep learning ensembles

Guidelines for applying the executable package:
This package enable the application of the overall best-performing residual stacked Deep Learning – Random Forest (DL-RF) ensemble model DL-RF: AGB = fDL-RF(D².H.WD) for pantropical tree aboveground biomass (AGB) predictions.

The executable package was developed based on the following reference:
Huy, B., Temesgen, H., Poudel, K.P., Truong, N.Q.  (2026). Advancing pantropical tree aboveground biomass predictions for large-scale carbon quantification through generalizable machine–deep learning ensembles.

Download the article.

For detailed methodology and application, please refer to this publication.

Please follow the steps below to apply the executable package:

  1. Please download the folder “ DL-RF Ensemble Executable Package.rar” then save and extract it to your computer. This folder contains the overall best-performing residual stacked DL-RF ensemble model.
  2. Enter your three new predictor values (D, H, WD) into the ‘New_data.csv’ file located in this folder. Note: D: diameter at breast height (cm); H: tree height (m); WD: wood density (g cm⁻³).
  3. Double-click on the file named ‘prediction.exe’ in this folder to start the prediction process. Please allow a few minutes for the program to complete its process.
  4. As a result, a new file named ‘AGB_predictions.csv’ will be created in this folder, containing pantropical tree AGB (kg tree-1) predictions.

_____________________________________________________________________________________________________________

Deep Learning–eXtreme Gradient Boosting ensemble approach for reliable additive mapping of above- and belowground carbon pools in tropical dry dipterocarp forests from Sentinel-2 imagery

This guideline, which includes an executable package, was developed based on the following reference:
Huy, B., Poudel, K.P., Temesgen, H., Braun, A.C., Truong, N.Q., Tri, P.C., Hien, N.T., Khiem, N.Q. (2026). Deep Learning–eXtreme Gradient Boosting ensemble approach for reliable additive mapping of above- and belowground carbon pools in tropical dry dipterocarp forests from Sentinel-2 imagery. Remote Sensing Applications: Society and Environment, 42(2026) 102075.  https://doi.org/10.1016/j.rsase.2026.102075
https://authors.elsevier.com/a/1n8fw8M-mn1Kkf
Download the article

The provided executable package enables the application of the best-performing ensemble Deep Learning–eXtreme Gradient Boosting (DLXG) model to predict and map the additive aboveground carbon pool (AGCP), belowground carbon pool (BGCP), and their total (TGCP) using Sentinel-2 imagery in tropical dry dipterocarp forests.

By leveraging newly processed Sentinel-2 spectral bands, vegetation indices, and texture features, the DLXG model employs six optimized predictive raster layers: SAPI, NDVI, V_SR45, SWIR2, V_RED.EDGE3, and V_GREEN.

Step-by-step instructions:

1. Download and extract the package

Download and extract the folder DLXG Executable Package.rar, which contains the following files:

  • DL_model_final.h5 – Trained deep learning model
  • Final_Residual_Stacking_XGB.pkl – Trained XGBoost model
  • prediction_rasters_DLXG.exe – Executable package for automated prediction.

2. Prepare input data

Place your six updated Sentinel-2 raster files in the same folder as the executable files. The required input files are: SAPI.tif, NDVI.tif, V_SR45.tif, SWIR2.tif, V_RED.EDGE3.tif, and V_GREEN.tif

3. Run the prediction

Double-click the executable package named ‘prediction_rasters_DLXG.exe’. This will run the best- performing DLXG model to generate the final prediction maps.

⏳ Note: The entire process may take several minutes depending on your system specifications and the size of the input rasters.

4. Output files

Upon successful completion, the following output raster files will be saved in the same folder:

  • AGCP_DLXG.tif – Aboveground Carbon Pool
  • BGCP_DLXG.tif – Belowground Carbon Pool
  • TGCP_DLXG.tif – Total Ground Carbon Pool (AGCP + BGCP)

These raster outputs (AGCP_DLXG.tif, BGCP_DLXG.tif, and TGCP_DLXG.tif) are ready to be used for spatial analysis, carbon accounting, and decision-making in tropical dry dipterocarp forest ecosystems. They represent an additive predictive mapping approach that integrates Deep learning and XGBoost models into a novel ensemble learning framework using remote sensing inputs.
————————————–
Note: SAPI (Scatter-Adjusted Pigment Index), NDVI (Normalized Difference Vegetation Index), V_SR45 (Texture variance derived from Simple Ratio 45), SWIR2 (Band 12), V_RED.EDGE3 (Texture variance derived from RED.EDGE3, Band 7), and V_GREEN (Texture variance derived from GREEN, Band 3).

_____________________________________________________________________________________________________________

Deep learning additive models for simultaneous prediction of stand level above- and belowground biomass in tropical forests

Guidelines for Applying the Executable Package:
This package enables the application of the best-developed Deep Learning Additive Models (DLAMs) for simultaneously predicting stand-level above- and belowground biomass and total (AGB, BGB and ABGB, respectively), while ensuring additivity in tropical forests.

This executable package was developed based on the following reference:
Huy, B., Poudel, K.P., Temesgen, H., Salas-Eljatib, C., Truong, N.Q., Khiem, N.Q. (2025). Comparing statistical and deep learning approaches for simultaneous prediction of stand-level above- and belowground biomass in tropical forests. Science of the Total Environment 958 (2025) 177869. DOI: https://doi.org/10.1016/j.scitotenv.2024.177869
Download the article. 
For detailed methodology and application, please refer to this publication.

Please follow the steps below to apply the executable package:

  1. Please download the folder DLAMs executable packages.rar then save and extract it to your computer.
  2. Choose one of the following subfolders based on your preferred predictive covariates:
    • DLAMs 1: Includes seven optimal predictive covariates (stand basal area (G), stand volume (V), mean annual temperature (T), elevation (EL), forest type (FT), average height (Hg), and soil group (SG)).
    • DLAMs 5: Includes five predictive covariates (stand basal area (G), stand volume (V), mean annual temperature (T), elevation (EL), average height (Hg)).
    • DLAMs 9: Uses only one predictor: stand basal area (G).
    • DLAMs 10: Uses only one predictor: stand volume (V).
  3. Enter your new predictor values into the file named ‘predictors data.csv’ located in this subfolder.
    • Units of numeric variables: G (m2/ha), V (m3/ha), T (oC), EL (m), Hg (m)
    • Using acronyms for nominal categorical variables: FT (DDF for Dry Dipterocarp Forest or EBLF for Evergreen Broadleaf Forest), and SG (FA for Ferric Acrisols, OA for Orthic Acrisols, and RF for Rhodic Ferrasols).
  4. Double-click on the file named ‘prediction.exe’ in this subfolder to start the prediction process. Please allow a few minutes for the program to complete its process.
  5. As a result, a new file named ‘prediction.csv’ will be created in this subfolder, containing simultaneous predictions of stand-level aboveground biomass (AGB, tons/ha), belowground biomass (BGB, tons/ha), and total (ABGB, tons/ha), ensuring additivity.