Dados do Trabalho

Título

TEMPORAL SOIL CARBON SPATIALIZATION BY REMOTE SENSING AND MACHINE LEARNING: A CASE STUDY IN MARANHAO, BRAZIL

Resumo

Soil organic matter (SOM) indicates soil quality, contributing to soil stability and nutrient availability. In addition, it makes the soil relevant in carbon storage. However, spatialization of SOM is hard in continental dimensions, as in Brazil, where several factors affect SOM content, including climate, land use, soil texture, and drainage. We aimed to map the spatiotemporal distribution of SOM using a machine-learning algorithm and remotely sensed covariates in two periods, 2013-2014 and 2020-2021. The area of interest comprises the municipalities of Balsas-MA and Fortaleza dos Nogueiras-MA, covering about 30,000 hectares, where we sampled 249 sites in 2014 and 405 between 2020 and 2021. These samples contained information on the SOM content, serving as a response variable by 150 Cubist algorithm models. We randomly separated the samples into a calibration (70%) and validation (30%) set. The predictor covariates included TAGEE, sand and clay maps, soil synthetic image (SYSI) from GEOS3 method, NDVI, NDWI and TVDI indexes, Landsat 8 bands 2, 3, 4, 5, 6 and 7 medians from two periods in study. The models created SOM content maps for 2014 and 2020/2021 in a smaller study area of 1981 hectares. The R², RMSE, and RPIQ statistical parameters evaluated the models. The modeling reached values of 0.465, 5.361 (g.kg-1) and 0.634, for the 2014 and 0.248, 5.250 (g.kg-¹) and 0.747, for the 2020/2021 model. The mean of SOM was 22.79 (g.kg-¹) and 22.61 (g.kg-¹) for 2014 and 2020/2021, respectively. We observed that the average SOM content decreased over 7 years. The technique demonstrated the possibility of SOM prediction in time and space with acceptable levels of accuracy. Brazil has the most significant potential for carbon storage globally, being a player fundamental in the country's policy of agriculture decarbonization, which makes this technique an advantage since it can monitor carbon dynamics over large areas.

Palavras-chave

Digital soil mapping; Pedometrics; Soil health; Soil Quality

Instituição financiadora

Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP), no âmbito do projeto 2022/13995-0 e 2021/05129-8

Agradecimentos

Authors thanks GeoCis for support in processing models and statistical analysis.