Otimização do mapeamento de micronutrientes do solo com base em macronutrientes e técnicas de aprendizado estatístico
Arquivos
Data
2024-02-22
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal de Viçosa
Resumo
O constante crescimento da população mundial acarreta diretamente no setor agronô- mico, resultando em um aumento na demanda por produção de alimentos, além de gerar preocupações relacionadas a limitações de áreas de cultivo e escassez de mão de obra. Surgem então a agricultura de precisão e a agricultura digital, que são res- ponsáveis por processar um grande volume de informações com o objetivo de promo- ver retorno econômico, vantagem competitiva para o produtor e minimizar os efeitos ao meio ambiente. Nota-se, portanto, a necessidade intrínseca de lidar de forma mais eficiente com os recursos e a variabilidade dos atributos do solo. Um dos ferramentais utilizados para a descrição da variabilidade espacial e mapeamento de atributos é conhecido como geoestatística. Contudo, um dos grandes desafios do método está relacionado com um número mínimo de amostras para realizar as interpolações, o que pode aumentar consideravelmente os gastos e necessidade de mão de obra para um projeto, pois a amostragem envolve a coleta e análise de atributos de todos os pontos previamente estipulados. Com o intuito de contornar a problemática relacionada a amostragem de dados em campo, este trabalho tem como objetivo reduzir o número de amostras analisadas quimicamente para micronutrientes ao predizer suas concen- trações com base nos macronutrientes, utilizando uma combinação de krigagem e métodos de machine learning (KNN). A área experimental é referente a uma parcela da fazenda “Sozinha” localizada em Goianápolis. As 150 amostras foram recolhidas nas profundidades de 0 a 0,2 𝑚, sendo cada uma composta por dez subamostras co- letadas a uma distância de até 5 𝑚 do ponto. Posteriormente foram realizadas análises físicas e químicas para quantificar os atributos presentes. Em seguida foram selecio- nadas grades modificadas (através dos métodos de amostragem aleatória simples (𝐴𝐴𝑆) e Conditioned Latin Hypercube Sampling (𝑐𝐿𝐻𝑆)) com redução de 15, 25 e 35% dos pontos originais, os quais resultaram em conjuntos de treinamento para o KNN. Posteriormente, o algoritmo KNN foi utilizado para predizer esses 23, 38 e 53 pontos amostrados e esses valores preditos foram então substituídos no conjunto de dados original. A seguir os mapas interpolados por malha e tipo de amostragem de cada um dos métodos empregados (krigagem ordinária (OK) e da diferença entre a OK e a krigagem ordinária combinada com KNN) foram obtidos. Todo o processo, desde a amostragem até as interpolações por krigagem, foi repetido por 50 vezes. Para com- parar as interpolações da krigagem ordinária no banco de dados original e nas grades modificadas foi analisada a razão entre a média da raiz quadrada do erro quadrático médio (𝑅𝑀𝑆𝐸) e do erro absoluto médio (𝑀𝐴𝐸) de ambas amostragens e o 𝑅𝑀𝑆𝐸 e 𝑀𝐴𝐸 da krigagem dos dados originais. A amostragem 𝑐𝐿𝐻𝑆 se mostrou melhor em manter as características espaciais do solo (com perda da variabilidade espacial) para os atributos estudados frente a todas as reduções de dimensionalidade quando com- parada a 𝐴𝐴𝑆. Sugere-se para trabalhos futuros, que sejam estudadas novas meto- dologias de machine learning combinadas à krigagem ordinária, além de tipos de amostragem diferentes como forma a avaliar seu comportamento frente a redução do adensamento amostral. Palavras-chave: Redução do adensamento amostral; Krigagem; KNN, Random Forest.
A continuous growth in the world population directly impacts the agronomic sector, resulting in an increased demand for food production and raising concerns related to limitations in cultivation areas and a shortage of labor. Precision agriculture and digi- tal farming emerge as solutions responsible for processing a large volume of infor- mation aimed at promoting economic returns, providing a competitive advantage for producers, and minimizing environmental effects. Therefore, an intrinsic need arises to handle resources and soil attribute variability more efficiently. One of the tools used for describing spatial variability and mapping attributes is known as geostatis- tics. However, a significant challenge in this method is associated with a minimum number of samples required for interpolations, which can considerably increase ex- penses and the need for labor in a project. This is because the sampling involves col- lecting and analyzing attributes from all predetermined points. To address the issues related to field data sampling, this study aims to reduce the number of chemically an- alyzed samples for micronutrients by predicting their concentrations based on macro- nutrients. This is achieved using a combination of kriging and machine learning meth- ods (KNN). The experimental area pertains to a section of the "Sozinha" farm located in Goianápolis. One hundred and fifty samples were collected at depths of 0 to 0.2 meters, with each composed of ten subsamples collected within a distance of up to 5 meters from the point. Subsequently, physical and chemical analyses were con- ducted to quantify the present attributes. Modified grids were then selected (using the methods of random sampling (𝐴𝐴𝑆) and Conditioned Latin Hypercube Sampling (𝑐𝐿𝐻𝑆)) with a reduction of 15, 25, and 35% of the original points, resulting in training sets for KNN. The KNN algorithm was used to predict these 23, 38, and 53 sampled points, and these predicted values were then replaced in the original dataset. Maps interpolated by mesh and sampling type for each of the employed methods (ordinary kriging (OK) and the difference between OK and ordinary kriging combined with KNN) were obtained. The entire process, from sampling to kriging interpolations, was repeated 50 times. To compare the interpolations of ordinary kriging in the original database and modified grids, the ratio between the mean square root of the mean er- ror (𝑅𝑀𝑆𝐸) and the mean absolute error (𝑀𝐴𝐸) of both samplings and the 𝑅𝑀𝑆𝐸 and 𝑀𝐴𝐸 of kriging of the original data was analyzed. The 𝑐𝐿𝐻𝑆 sampling proved to be more effective in preserving the spatial characteristics of the soil (with loss of spatial variability) for the studied attributes compared to all dimensionality reductions when compared to 𝐴𝐴𝑆. It is suggested for future work to explore new machine learning methodologies combined with ordinary kriging, as well as different sampling tech- niques, to assess their behavior in the face of sample density reduction. Keywords: Sample density reduction; Kriging; KNN; Random Forest.
A continuous growth in the world population directly impacts the agronomic sector, resulting in an increased demand for food production and raising concerns related to limitations in cultivation areas and a shortage of labor. Precision agriculture and digi- tal farming emerge as solutions responsible for processing a large volume of infor- mation aimed at promoting economic returns, providing a competitive advantage for producers, and minimizing environmental effects. Therefore, an intrinsic need arises to handle resources and soil attribute variability more efficiently. One of the tools used for describing spatial variability and mapping attributes is known as geostatis- tics. However, a significant challenge in this method is associated with a minimum number of samples required for interpolations, which can considerably increase ex- penses and the need for labor in a project. This is because the sampling involves col- lecting and analyzing attributes from all predetermined points. To address the issues related to field data sampling, this study aims to reduce the number of chemically an- alyzed samples for micronutrients by predicting their concentrations based on macro- nutrients. This is achieved using a combination of kriging and machine learning meth- ods (KNN). The experimental area pertains to a section of the "Sozinha" farm located in Goianápolis. One hundred and fifty samples were collected at depths of 0 to 0.2 meters, with each composed of ten subsamples collected within a distance of up to 5 meters from the point. Subsequently, physical and chemical analyses were con- ducted to quantify the present attributes. Modified grids were then selected (using the methods of random sampling (𝐴𝐴𝑆) and Conditioned Latin Hypercube Sampling (𝑐𝐿𝐻𝑆)) with a reduction of 15, 25, and 35% of the original points, resulting in training sets for KNN. The KNN algorithm was used to predict these 23, 38, and 53 sampled points, and these predicted values were then replaced in the original dataset. Maps interpolated by mesh and sampling type for each of the employed methods (ordinary kriging (OK) and the difference between OK and ordinary kriging combined with KNN) were obtained. The entire process, from sampling to kriging interpolations, was repeated 50 times. To compare the interpolations of ordinary kriging in the original database and modified grids, the ratio between the mean square root of the mean er- ror (𝑅𝑀𝑆𝐸) and the mean absolute error (𝑀𝐴𝐸) of both samplings and the 𝑅𝑀𝑆𝐸 and 𝑀𝐴𝐸 of kriging of the original data was analyzed. The 𝑐𝐿𝐻𝑆 sampling proved to be more effective in preserving the spatial characteristics of the soil (with loss of spatial variability) for the studied attributes compared to all dimensionality reductions when compared to 𝐴𝐴𝑆. It is suggested for future work to explore new machine learning methodologies combined with ordinary kriging, as well as different sampling tech- niques, to assess their behavior in the face of sample density reduction. Keywords: Sample density reduction; Kriging; KNN; Random Forest.
Descrição
Palavras-chave
Correlação (Estatística), Aprendizado do computador, Plantas - Nutrição - Estatística, Micronutrientes - Amostragem
Citação
OLIVEIRA, Samantha Gouvêa. Otimização do mapeamento de micronutrientes do solo com base em macronutrientes e técnicas de aprendizado estatístico. 2024. 55 f. Dissertação (Mestrado em Estatística Aplicada e Biometria) - Universidade Federal de Viçosa, Viçosa. 2024.