Hayfa Zayani , Youssef Fouad , Didier Michot , Zeineb Kassouk , Nicolas Baghdadi , Maria José Marquez Perez , Emmanuelle Vaudour , Zohra Lili-Chabaane , Christian Walter
{"title":"利用遥感图像和实验室光谱数据集的深度学习方法绘制两个不同土壤气候区的土壤碳含量图","authors":"Hayfa Zayani , Youssef Fouad , Didier Michot , Zeineb Kassouk , Nicolas Baghdadi , Maria José Marquez Perez , Emmanuelle Vaudour , Zohra Lili-Chabaane , Christian Walter","doi":"10.1016/j.geoderma.2025.117513","DOIUrl":null,"url":null,"abstract":"<div><div>Mapping soil properties can help in monitoring spatial and temporal variability in soils. However, the accuracy of predictions based on optical remote sensing data depends on the availability of cloud-free images and the presence of bare soil under suitable surface conditions. The objective of this study was to produce maps of soil organic carbon (SOC) content or soil total carbon (STC) content for two study areas with contrasting pedoclimatic conditions: Naizin (1.5 km<sup>2</sup>), which has a temperate climate and mainly intensive mixed crop-livestock farms on Luvic Cambisols and Haplic Albeluvisols, and Merguellil (40 km<sup>2</sup>), which has a semi-arid climate and mainly intensive farms on Fluvisols. We used a deep learning approach that combined remote sensing and laboratory visible, near-infrared and short-wave infrared (Vis-NIR-SWIR) datasets. We developed deep neural network models using all available bare soil pixels from Sentinel-2 (S2) images over one farming year and by measuring soil properties at 58 and 73 soil sampling points in the Naizin and Merguellil study areas, respectively. We used two approaches: (1) using only S2 bands to calibrate models and (2) adding laboratory spectral indices incrementally to the S2 bands in decreasing order of their correlation with SOC or STC content. To compare and analyse the spatial patterns in the SOC and STC maps produced by the two approaches, we applied one model from approach 1 and one model from approach 2. The results showed that this method was able to adapt to the two pedoclimatic contexts. Adding the laboratory indices to the models increased prediction accuracy for both study areas. Although the two approaches yielded similar spatial patterns of SOC or STC content, some differences were observed. Adding the laboratory indices increased intra-field variability in the distribution of SOC content for Naizin but decreased inter- and intra-field variability in the distribution of STC content for Merguellil. The method accurately mapped SOC content in 70 % of the Naizin area and STC content in more than 90 % of the Merguellil area. Future research could assess effects of using different methods for kriging, the potential of using hyper-spectral images for calculating spectral indices which would avoid effects of kriging on the spatial patterns in the maps of soil properties produced and the transferability of the models across different pedoclimatic regions.</div></div>","PeriodicalId":12511,"journal":{"name":"Geoderma","volume":"462 ","pages":"Article 117513"},"PeriodicalIF":6.6000,"publicationDate":"2025-09-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Mapping soil carbon content in two contrasting pedoclimatic regions using a deep learning approach with remote sensing imagery and laboratory spectral datasets\",\"authors\":\"Hayfa Zayani , Youssef Fouad , Didier Michot , Zeineb Kassouk , Nicolas Baghdadi , Maria José Marquez Perez , Emmanuelle Vaudour , Zohra Lili-Chabaane , Christian Walter\",\"doi\":\"10.1016/j.geoderma.2025.117513\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Mapping soil properties can help in monitoring spatial and temporal variability in soils. However, the accuracy of predictions based on optical remote sensing data depends on the availability of cloud-free images and the presence of bare soil under suitable surface conditions. The objective of this study was to produce maps of soil organic carbon (SOC) content or soil total carbon (STC) content for two study areas with contrasting pedoclimatic conditions: Naizin (1.5 km<sup>2</sup>), which has a temperate climate and mainly intensive mixed crop-livestock farms on Luvic Cambisols and Haplic Albeluvisols, and Merguellil (40 km<sup>2</sup>), which has a semi-arid climate and mainly intensive farms on Fluvisols. We used a deep learning approach that combined remote sensing and laboratory visible, near-infrared and short-wave infrared (Vis-NIR-SWIR) datasets. We developed deep neural network models using all available bare soil pixels from Sentinel-2 (S2) images over one farming year and by measuring soil properties at 58 and 73 soil sampling points in the Naizin and Merguellil study areas, respectively. We used two approaches: (1) using only S2 bands to calibrate models and (2) adding laboratory spectral indices incrementally to the S2 bands in decreasing order of their correlation with SOC or STC content. To compare and analyse the spatial patterns in the SOC and STC maps produced by the two approaches, we applied one model from approach 1 and one model from approach 2. The results showed that this method was able to adapt to the two pedoclimatic contexts. Adding the laboratory indices to the models increased prediction accuracy for both study areas. Although the two approaches yielded similar spatial patterns of SOC or STC content, some differences were observed. Adding the laboratory indices increased intra-field variability in the distribution of SOC content for Naizin but decreased inter- and intra-field variability in the distribution of STC content for Merguellil. The method accurately mapped SOC content in 70 % of the Naizin area and STC content in more than 90 % of the Merguellil area. Future research could assess effects of using different methods for kriging, the potential of using hyper-spectral images for calculating spectral indices which would avoid effects of kriging on the spatial patterns in the maps of soil properties produced and the transferability of the models across different pedoclimatic regions.</div></div>\",\"PeriodicalId\":12511,\"journal\":{\"name\":\"Geoderma\",\"volume\":\"462 \",\"pages\":\"Article 117513\"},\"PeriodicalIF\":6.6000,\"publicationDate\":\"2025-09-27\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Geoderma\",\"FirstCategoryId\":\"97\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0016706125003544\",\"RegionNum\":1,\"RegionCategory\":\"农林科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"SOIL SCIENCE\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Geoderma","FirstCategoryId":"97","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0016706125003544","RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"SOIL SCIENCE","Score":null,"Total":0}
Mapping soil carbon content in two contrasting pedoclimatic regions using a deep learning approach with remote sensing imagery and laboratory spectral datasets
Mapping soil properties can help in monitoring spatial and temporal variability in soils. However, the accuracy of predictions based on optical remote sensing data depends on the availability of cloud-free images and the presence of bare soil under suitable surface conditions. The objective of this study was to produce maps of soil organic carbon (SOC) content or soil total carbon (STC) content for two study areas with contrasting pedoclimatic conditions: Naizin (1.5 km2), which has a temperate climate and mainly intensive mixed crop-livestock farms on Luvic Cambisols and Haplic Albeluvisols, and Merguellil (40 km2), which has a semi-arid climate and mainly intensive farms on Fluvisols. We used a deep learning approach that combined remote sensing and laboratory visible, near-infrared and short-wave infrared (Vis-NIR-SWIR) datasets. We developed deep neural network models using all available bare soil pixels from Sentinel-2 (S2) images over one farming year and by measuring soil properties at 58 and 73 soil sampling points in the Naizin and Merguellil study areas, respectively. We used two approaches: (1) using only S2 bands to calibrate models and (2) adding laboratory spectral indices incrementally to the S2 bands in decreasing order of their correlation with SOC or STC content. To compare and analyse the spatial patterns in the SOC and STC maps produced by the two approaches, we applied one model from approach 1 and one model from approach 2. The results showed that this method was able to adapt to the two pedoclimatic contexts. Adding the laboratory indices to the models increased prediction accuracy for both study areas. Although the two approaches yielded similar spatial patterns of SOC or STC content, some differences were observed. Adding the laboratory indices increased intra-field variability in the distribution of SOC content for Naizin but decreased inter- and intra-field variability in the distribution of STC content for Merguellil. The method accurately mapped SOC content in 70 % of the Naizin area and STC content in more than 90 % of the Merguellil area. Future research could assess effects of using different methods for kriging, the potential of using hyper-spectral images for calculating spectral indices which would avoid effects of kriging on the spatial patterns in the maps of soil properties produced and the transferability of the models across different pedoclimatic regions.
期刊介绍:
Geoderma - the global journal of soil science - welcomes authors, readers and soil research from all parts of the world, encourages worldwide soil studies, and embraces all aspects of soil science and its associated pedagogy. The journal particularly welcomes interdisciplinary work focusing on dynamic soil processes and functions across space and time.