Publicação
Integration of Optical and Radar Image Time Series with Supervised Learning and Geospatial Intelligence for Early Mapping of Rice and Maize Crops in the Baixo Mondego Region.
| Resumo: | The availability of information on the spatial distribution of agricultural crops is a critical factor in supporting the decision-making process. However, providing this information in the post-agricultural season compromises its utility in addressing technical issues related to agricultural production itself, as well as in administrative and policy decisions whose timing occurs before the crop harvest. Despite this need, studies indicate that while the distinction between agricultural crops is most effective at the end of the season, rapid and automated early mapping techniques using satellite imagery and advanced methods like machine learning are enabling high classification accuracies and can significantly improve agricultural planning and optimization of practices, policies and resource allocation.The objective of this study was to automate the early mapping of maize and rice agricultural plots using satellite imagery, with the utilization of time series of polarimetric and multispectral data, and the combination of geospatial intelligence and supervised learning in a cloud computing environment.The study is situated within the context of Portugal, where the mapping of agricultural crops from satellite images has been in the experimental phase for less than five years. The study was applied to the agricultural plots of the Baixo-Mondego valley, which is located in the central region of Portugal.To establish a robust and effective methodology for crop mapping from satellite data, a thorough review of 70 relevant articles was conducted. This review revealed that, in recent years, the most effective approach to agricultural mapping from satellite images involves the fusion of time series data from different sensors and the use of supervised computational learning.The applied methodology consisted of the incremental classification of time series of satellite images throughout the summer agricultural season, using SVM and RF algorithms. The methodology combines geospatial intelligence and supervised learning in a approach that fuses SAR and multispectral data, and another in parallel that uses only multispectral data. This methodology enables the mapping of the evolution of classification, the assessment of the significance of characteristics derived from satellites, and the evaluation of the contribution of radar data in optimizing early mapping.The findings indicated that the methodology employing fused data can enhance the efficacy of classification models by up to 6.26\%. Moreover, the magnitude of classification improvement was observed to increase with the proximity to spring. This is supported by the paucity of high-quality optical images and the incorporation of radar data with geometric information about crops at their early stages.The classification line with models trained using the RF algorithm proved to be more efficient in minimizing the time required to achieve acceptable performance (equal to or greater than 90\%), as evidenced by an accuracy of 91.45\% on June 20th and an F1-Score of 90.88\% on June 30th. In contrast, the classification line with models trained using the SVM algorithm only reached an accuracy of 92.33\% on July 10th and an F1-Score of 91.86\% on July 30th, resulting in a delay of approximately 20-30 days in achieving acceptable performance.The proposed approach proved to be relevant, as the precise and timely classification of agricultural crops not only allows for the optimization of agricultural production processes and resources but also facilitates inclusive technological advancement. This advancement maximizes the economy and minimizes the environmental impact, thereby contributing to food security. Future research should continue to refine the use of data obtained through space technology and processed through cloud computing environments, with a view to improving inclusive support for precision agriculture. |
|---|---|
| Autores principais: | Francisco, Adalberto Manuel Yambi |
| Assunto: | Early Crop Mapping Supervised Learning Time Series Satellite Data Fusion Incremental Classification Mapeamento Prococe de Culturas Agrícolas Aprendizagem Supervisionada Séries Temporais Fusão de Dados de Satélite Classificação Incremental |
| Ano: | 2024 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Coimbra |
| Idioma: | inglês |
| Origem: | Estudo Geral - Universidade de Coimbra |
| Resumo: | The availability of information on the spatial distribution of agricultural crops is a critical factor in supporting the decision-making process. However, providing this information in the post-agricultural season compromises its utility in addressing technical issues related to agricultural production itself, as well as in administrative and policy decisions whose timing occurs before the crop harvest. Despite this need, studies indicate that while the distinction between agricultural crops is most effective at the end of the season, rapid and automated early mapping techniques using satellite imagery and advanced methods like machine learning are enabling high classification accuracies and can significantly improve agricultural planning and optimization of practices, policies and resource allocation.The objective of this study was to automate the early mapping of maize and rice agricultural plots using satellite imagery, with the utilization of time series of polarimetric and multispectral data, and the combination of geospatial intelligence and supervised learning in a cloud computing environment.The study is situated within the context of Portugal, where the mapping of agricultural crops from satellite images has been in the experimental phase for less than five years. The study was applied to the agricultural plots of the Baixo-Mondego valley, which is located in the central region of Portugal.To establish a robust and effective methodology for crop mapping from satellite data, a thorough review of 70 relevant articles was conducted. This review revealed that, in recent years, the most effective approach to agricultural mapping from satellite images involves the fusion of time series data from different sensors and the use of supervised computational learning.The applied methodology consisted of the incremental classification of time series of satellite images throughout the summer agricultural season, using SVM and RF algorithms. The methodology combines geospatial intelligence and supervised learning in a approach that fuses SAR and multispectral data, and another in parallel that uses only multispectral data. This methodology enables the mapping of the evolution of classification, the assessment of the significance of characteristics derived from satellites, and the evaluation of the contribution of radar data in optimizing early mapping.The findings indicated that the methodology employing fused data can enhance the efficacy of classification models by up to 6.26\%. Moreover, the magnitude of classification improvement was observed to increase with the proximity to spring. This is supported by the paucity of high-quality optical images and the incorporation of radar data with geometric information about crops at their early stages.The classification line with models trained using the RF algorithm proved to be more efficient in minimizing the time required to achieve acceptable performance (equal to or greater than 90\%), as evidenced by an accuracy of 91.45\% on June 20th and an F1-Score of 90.88\% on June 30th. In contrast, the classification line with models trained using the SVM algorithm only reached an accuracy of 92.33\% on July 10th and an F1-Score of 91.86\% on July 30th, resulting in a delay of approximately 20-30 days in achieving acceptable performance.The proposed approach proved to be relevant, as the precise and timely classification of agricultural crops not only allows for the optimization of agricultural production processes and resources but also facilitates inclusive technological advancement. This advancement maximizes the economy and minimizes the environmental impact, thereby contributing to food security. Future research should continue to refine the use of data obtained through space technology and processed through cloud computing environments, with a view to improving inclusive support for precision agriculture. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.