AI-driven digital soil Mapping of the Eastern Amazon: integrating pedological knowledge and pedometrics

Abstract

The lack of more detailed information on soil distribution in the Amazon is an impediment to effective soil governance in this region. Therefore, the overall objective of this study was to evaluate the performance and uncertainty of assembly algorithm approaches, legacy mapping, environmental covariates, and reference area. To this end, this thesis was divided into two chapters, titled: I) Integrating Tacit Knowledge and AI for Digital Soil Mapping in Eastern Amazonia: Ensemble Learning, Model Performance, and Uncertainty Incorporation; and II) Transferring soil–landscape relationships in data-limited regions of the Amazon: a knowledgeguided digital soil mapping approach based on reference areas. Output data from both chapters followed the common steps below. Fifteen environmental covariates based on the SCORPAN model were used. To generate the predictive models, we used the algorithms and four input datasets: Random Forest, Ranger, Xgboost, C 5.0, and Ensemble Learning. We also evaluated two levels of detail for the Mapping Units of the legacy map: taxonomic order level and large group level. Terrain covariates were derived from two distinct elevation models: digital surface model and digital terrain model. A total of 20 predictive models were analyzed. Key findings from Chapter 2: the algorithm assembly approach using a digital surface model and unit mapping at the Large Group taxonomic level achieved statistically superior performance in terms of accuracy and Kappa compared to individual models at the Large Group taxonomic level. Environmental covariates such as: 22,000-year-old Mean Annual Temperature, Channel Network Base Level, Altitude, and Land Surface Temperature showed greater importance in predictive modeling; And as highlights of Chapter 3: in the external validation in the extrapolation area, the ensemble learning predictive models using data derived from the digital surface model, in mapping units at the Oreder taxonomic level, ensemble learning using data derived from the digital terrain model, in mapping units at the Order taxonomic level and ensemble learning using data derived from the digital surface model, in mapping units at the Great group taxonomic level, and the prediction of the mapping units GLEISSOLOS and GLEISSOLOS SÁLICOS Sódicos stood out as the most accurate in the extrapolation area. The maps generated from the legacy map and for the extrapolation area, using reference area, successfully mapped mapping units related to the hydromorphic environment. The pioneering nature of this study in the Amazon revealed the potential of the ensemble learning approach for mapping soil classes under hydromorphic environments, environments that occupy a large area in the Amazon and are of significant importance.

The lack of more detailed information on soil distribution in the Amazon is an impediment to effective soil governance in this region. Therefore, the overall objective of this study was to evaluate the performance and uncertainty of assembly algorithm approaches, legacy mapping, environmental covariates, and reference area. To this end, this thesis was divided into two chapters, titled: I) Integrating Tacit Knowledge and AI for Digital Soil Mapping in Eastern Amazonia: Ensemble Learning, Model Performance, and Uncertainty Incorporation; and II) Transferring soil–landscape relationships in data-limited regions of the Amazon: a knowledgeguided digital soil mapping approach based on reference areas. Output data from both chapters followed the common steps below. Fifteen environmental covariates based on the SCORPAN model were used. To generate the predictive models, we used the algorithms and four input datasets: Random Forest, Ranger, Xgboost, C 5.0, and Ensemble Learning. We also evaluated two levels of detail for the Mapping Units of the legacy map: taxonomic order level and large group level. Terrain covariates were derived from two distinct elevation models: digital surface model and digital terrain model. A total of 20 predictive models were analyzed. Key findings from Chapter 2: the algorithm assembly approach using a digital surface model and unit mapping at the Large Group taxonomic level achieved statistically superior performance in terms of accuracy and Kappa compared to individual models at the Large Group taxonomic level. Environmental covariates such as: 22,000-year-old Mean Annual Temperature, Channel Network Base Level, Altitude, and Land Surface Temperature showed greater importance in predictive modeling; And as highlights of Chapter 3: in the external validation in the extrapolation area, the ensemble learning predictive models using data derived from the digital surface model, in mapping units at the Oreder taxonomic level, ensemble learning using data derived from the digital terrain model, in mapping units at the Order taxonomic level and ensemble learning using data derived from the digital surface model, in mapping units at the Great group taxonomic level, and the prediction of the mapping units GLEISSOLOS and GLEISSOLOS SÁLICOS Sódicos stood out as the most accurate in the extrapolation area. The maps generated from the legacy map and for the extrapolation area, using reference area, successfully mapped mapping units related to the hydromorphic environment. The pioneering nature of this study in the Amazon revealed the potential of the ensemble learning approach for mapping soil classes under hydromorphic environments, environments that occupy a large area in the Amazon and are of significant importance.

The lack of more detailed information on soil distribution in the Amazon is an impediment to effective soil governance in this region. Therefore, the overall objective of this study was to evaluate the performance and uncertainty of assembly algorithm approaches, legacy mapping, environmental covariates, and reference area. To this end, this thesis was divided into two chapters, titled: I) Integrating Tacit Knowledge and AI for Digital Soil Mapping in Eastern Amazonia: Ensemble Learning, Model Performance, and Uncertainty Incorporation; and II) Transferring soil–landscape relationships in data-limited regions of the Amazon: a knowledgeguided digital soil mapping approach based on reference areas. Output data from both chapters followed the common steps below. Fifteen environmental covariates based on the SCORPAN model were used. To generate the predictive models, we used the algorithms and four input datasets: Random Forest, Ranger, Xgboost, C 5.0, and Ensemble Learning. We also evaluated two levels of detail for the Mapping Units of the legacy map: taxonomic order level and large group level. Terrain covariates were derived from two distinct elevation models: digital surface model and digital terrain model. A total of 20 predictive models were analyzed. Key findings from Chapter 2: the algorithm assembly approach using a digital surface model and unit mapping at the Large Group taxonomic level achieved statistically superior performance in terms of accuracy and Kappa compared to individual models at the Large Group taxonomic level. Environmental covariates such as: 22,000-year-old Mean Annual Temperature, Channel Network Base Level, Altitude, and Land Surface Temperature showed greater importance in predictive modeling; And as highlights of Chapter 3: in the external validation in the extrapolation area, the ensemble learning predictive models using data derived from the digital surface model, in mapping units at the Oreder taxonomic level, ensemble learning using data derived from the digital terrain model, in mapping units at the Order taxonomic level and ensemble learning using data derived from the digital surface model, in mapping units at the Great group taxonomic level, and the prediction of the mapping units GLEISSOLOS and GLEISSOLOS SÁLICOS Sódicos stood out as the most accurate in the extrapolation area. The maps generated from the legacy map and for the extrapolation area, using reference area, successfully mapped mapping units related to the hydromorphic environment. The pioneering nature of this study in the Amazon revealed the potential of the ensemble learning approach for mapping soil classes under hydromorphic environments, environments that occupy a large area in the Amazon and are of significant importance.