Delimitação de zonas de manejo com base em dados temporais do sentinel-1
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
A caracterização da variabilidade dos atributos do solo geralmente demanda grades de amostragem com alta densidade, o que pode resultar em custos significativos de coleta e análise laboratorial. Uma possível solução é delimitar Zonas de Manejo (ZMs) com base em dados previamente coletados para direcionar o processo de amostragem. O objetivo desta pesquisa foi avaliar o potencial de aplicação dos dados de radar na delimitação de ZMs. Para isso, foram estabelecidos os seguintes objetivos específicos: (1) desenvolver um método para mapear atributos do solo por meio da delimitação de ZMs utilizando dados de radar do Sentinel-1 (2) comparar o método desenvolvido com métodos tradicionais de amostragem na estimativa dos atributos do solo (3) avaliar a acurácia das ZMs definidas com base no índice VV/VH do Sentinel-1 na estimativa de atributos do solo, em relação a métodos convencionais que tem por base mapas de produtividade e índice NDVI (4) avaliar a acurácia das ZMs delimitadas pelos índices NDVI e VV/VH na previsão da produtividade média padronizada, utilizando mapas de produtividade coletados ao longo dos anos. Imagens do Sentinel-1 foram usadas para gerar o perfil temporal de seis índices com base nas bandas de retroespalhamento VV (vertical-vertical) e VH (vertical-horizontal) entre janeiro de 2018 e março de 2023 em 3 campos agrícolas. Para delimitar as ZMs, os índices e as bandas foram analisados individualmente por duas abordagens: (1) agrupamento fuzzy k-means aplicado às séries temporais dos índices e (2) redução de dimensionalidade via autoencoders, seguida de um agrupamento fuzzy k-means. A melhor combinação de índice e abordagem foi comparada com quatro métodos tradicionais de mapeamento de atributos do solo: amostra única composta, grade uniforme (1 amostra/ha), células retangulares (5-10 ha) e células geradas aleatoriamente. A validação cruzada deixe-um-de-fora foi utilizada para avaliar o desempenho dos métodos. Os resultados obtidos mostraram que a combinação do índice VV/VH com o algoritmo autoencoders para a delimitação de ZMs proporcionou estimativas mais precisas dos atributos do solo e superou o método convencional de amostragem do solo (uma única amostra composta) e, em muitos casos, o método de amostragem do solo por células. Em comparação com as abordagens tradicionais de delimitação de ZMs, que utilizam mapas de produtividade e NDVI, o método desenvolvido apresentou acurácia semelhante na maioria dos atributos do solo, com destaque para o atributo argila, onde seu desempenho foi significativamente superior. Ambos os métodos NDVI e VV/VH identificaram zonas de alta e baixa produtividade com comportamento similares. Os resultados deste estudo destacam o índice VV/VH do Sentinel-1 como uma alternativa viável para a delimitação de ZMs. Entretanto, mais estudos são necessários para validar sua aplicação em diferentes condições ambientais, avaliar sua estabilidade ao longo do tempo e explorar possíveis melhorias na estratégia de processamento das sereis temporais do índice VV/VH. Palavras-chave: Agricultura de precisão, Sensoriamento Remoto, Amostragem de solo, Radar de Abertura Sintética, Aprendizado Profundo
The characterization of soil attribute variability often requires high-density sampling grids, which can lead to significant costs in data collection and laboratory analysis. A possible solution is to delineate Management Zones (MZs) based on previously collected data to guide the sampling process. This study aimed to evaluate the potential application of radar data in MZs delineation. To achieve this, the following specific objectives were established: (1) develop a method to map soil attributes by delineating MZs using radar data from Sentinel-1 (2) compare the developed method with traditional soil sampling methods in estimating soil attributes (3) evaluate the accuracy of MZs defined based on the Sentinel-1 VV/VH index in estimating soil attributes, in comparison to conventional methods based on productivity maps and the NDVI index (4) assess the accuracy of MZs delineated using NDVI and VV/VH indices in predicting standardized average productivity, using productivity maps collected over the years. Sentinel-1 images were used to generate the temporal profile of six indices based on VV (vertical-vertical) and VH (vertical-horizontal) backscatter bands between January 2018 and March 2023 in three agricultural fields. To delineate MZs, the indices and bands were analyzed individually using two approaches: (1) fuzzy k-means clustering applied to the temporal series of indices and (2) dimensionality reduction via autoencoders, followed by fuzzy k-means clustering. The best combination of index and approach was compared with four traditional soil attribute mapping methods: composite single sample, uniform grid (1 sample/ha), rectangular cells (5-10 ha), and randomly generated cells. Leave-one-out cross-validation was used to assess the performance of the methods. The results showed that the combination of the VV/VH index with the autoencoder algorithm for MZs delineation provided more accurate soil attribute estimates, outperforming the conventional soil sampling method (composite single sample) and, in many cases, the soil sampling method based on cells. In comparison with traditional approaches for delineating MZs, which use yield maps and NDVI, the developed method demonstrated similar accuracy for most soil attributes, with a notable improvement in the estimation of clay content, where its performance was significantly superior. Both NDVI and VV/VH methods identified high- and low-productivity zones with similar behavior. The findings of this study highlight the Sentinel-1 VV/VH index as a viable alternative for MZs delineation. However, further studies are needed to validate its application under different environmental conditions, assess its stability over time, and explore potential improvements in the processing strategy of the VV/VH index time series. Keywords: Precision Agriculture, Remote Sensing, Soil Sampling, Synthetic Aperture Radar, Deep Learning
The characterization of soil attribute variability often requires high-density sampling grids, which can lead to significant costs in data collection and laboratory analysis. A possible solution is to delineate Management Zones (MZs) based on previously collected data to guide the sampling process. This study aimed to evaluate the potential application of radar data in MZs delineation. To achieve this, the following specific objectives were established: (1) develop a method to map soil attributes by delineating MZs using radar data from Sentinel-1 (2) compare the developed method with traditional soil sampling methods in estimating soil attributes (3) evaluate the accuracy of MZs defined based on the Sentinel-1 VV/VH index in estimating soil attributes, in comparison to conventional methods based on productivity maps and the NDVI index (4) assess the accuracy of MZs delineated using NDVI and VV/VH indices in predicting standardized average productivity, using productivity maps collected over the years. Sentinel-1 images were used to generate the temporal profile of six indices based on VV (vertical-vertical) and VH (vertical-horizontal) backscatter bands between January 2018 and March 2023 in three agricultural fields. To delineate MZs, the indices and bands were analyzed individually using two approaches: (1) fuzzy k-means clustering applied to the temporal series of indices and (2) dimensionality reduction via autoencoders, followed by fuzzy k-means clustering. The best combination of index and approach was compared with four traditional soil attribute mapping methods: composite single sample, uniform grid (1 sample/ha), rectangular cells (5-10 ha), and randomly generated cells. Leave-one-out cross-validation was used to assess the performance of the methods. The results showed that the combination of the VV/VH index with the autoencoder algorithm for MZs delineation provided more accurate soil attribute estimates, outperforming the conventional soil sampling method (composite single sample) and, in many cases, the soil sampling method based on cells. In comparison with traditional approaches for delineating MZs, which use yield maps and NDVI, the developed method demonstrated similar accuracy for most soil attributes, with a notable improvement in the estimation of clay content, where its performance was significantly superior. Both NDVI and VV/VH methods identified high- and low-productivity zones with similar behavior. The findings of this study highlight the Sentinel-1 VV/VH index as a viable alternative for MZs delineation. However, further studies are needed to validate its application under different environmental conditions, assess its stability over time, and explore potential improvements in the processing strategy of the VV/VH index time series. Keywords: Precision Agriculture, Remote Sensing, Soil Sampling, Synthetic Aperture Radar, Deep Learning
Description
Citation
GONÇALVES, Juliano de Paula. Delimitação de zonas de manejo com base em dados temporais do sentinel-1. 2025. 90 f. Tese (Doutorado em Engenharia Agrícola) - Universidade Federal de Viçosa, Viçosa. 2025.
