Predicción de propiedades físicas/químicas del suelo mediante firmas espectrales y clasificación de pH.
Fecha
2021-12
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad EARTH
Resumen
Dentro de la agricultura se realizan prácticas de manejo, entre ellas está la determinación de nutrientes en el sistema suelo con análisis químicos en laboratorios especializados. Mediante el aprendizaje automático ha surgido el uso del espectro Visible-Infrarrojo cercano (Vis-NIR) para la predicción nutricional en el suelo, disminuyendo los costos, tiempos en la obtención de resultados y la contaminación. Para la predicción de 29 propiedades físicas/químicas del suelo se tomaron firmas dentro del espectro de 400 nm a 2500 nm, analizadas con el método estadístico de Mínimos Cuadrados Parciales (PLSR) con previo preprocesamiento y eliminación de datos atípicos (outliers). Las muestras fueron proporcionadas por el Laboratorio de Suelos, Aguas y Foliares de la Universidad EARTH. Se usó el coeficiente de determinación (R2) de mayor importancia de selección. Se comparó la calidad de la predicción del modelo utilizando 2318 y 3381 firmas de muestras de suelo. Se obtuvo que solamente tres variables presentaron mayor R2, a mayor cantidad de muestras, siendo: N con R2=0.73, Na con R2=0.60, B con R2=0.25 y dos relaciones fueron encontradas; absorción de sodio (SAR) con R2=0.57 y Ca/Mg con R2=0.51. Por lo que la calidad de predicción del modelo en cuanto al número de muestras no mostró diferencias en la mayoría de las variables, diciendo que a mayor sea la cantidad de firmas, menor puede ser la calidad de predicción, dependiendo de la agrupación de datos que se realicen. También se realizó la comparación entre la clasificación de datos con un rango de pH de 5.5 a 6.5. Se observó una diferencia positiva en 17 variables, Entre las que mayor predicción se obtuvo están: Si con R2=0.97, Ca con R2=0.93, Capacidad de intercambio catiónico (CICe) con R2=0.92, Mg con R2=0.92, C con R2=0.89, Fe con R2=0.85, K con R2=0.79, también variables como arcilla con R2=0.75, y arena con R2=0.74. La clasificación del suelo según sus características físicas o químicas hace una mejor predicción, también el lugar de procedencia puede ser una forma de clasificación, de allí la importancia de la georreferenciación de las muestras.
In agriculture, there are diverse management practices, one of these is the determination of the nutrients in the soil system using chemical analysis in specialized laboratories. Through machine learning, the use of the Vis-NIR spectrum for the soil nutritional prediction has been started, decreasing the cost and time for obtaining the results, and protecting the environment. For the prediction of 29 soil physical/chemical properties, signatures within the spectrum of 400nm and 2500nm were taken and analyzed with the Partial Leas Square (PLSR) statistics method, these were preprocessed, and the outliers were eliminated. The soil samples were taken from the Soil, Water and Foliar Laboratory at EARTH University. The determination coefficient (R2) was selected as the one with the biggest importance. The model prediction quality of 2318 samples were compared to 3381 soil samples. Only three variables presented higher value in R2 in more samples, these are: N with R2=0.73 R2, Na with R2=0.60 R2, and B with R2=0.25. Also, two ratio in the soil’s content were found: Sodium absorption (SAR) with R2=0.57 and Ca/Mag with R2=0.51. The prediction quality of the statistical model and the increasing of the soil samples did not show difference in most variables, given that the more samples of signatures does not necessarily mean greater quality of the prediction; however, it depends on the data aggrupation done. Also, a comparison of the data classified in pH ranges of 5.5 to 6.5 was done. A positive difference could be observed in 17 variables, among the ones that obtained higher values in the prediction are S with R2=0.97, Ca with R2=0.93, Cation exchange capacity (CICe) with R2=0.92, Mg with R2=0.92, C with R2=0.89, Fe with R2=0.85, K with R2=0.79, as well the variables like clay with R2=0.75 and sand with R2=0.74. The classification of soil based on its physical or chemical characteristics makes a better prediction, even the place of origin could be considered. This highlights the importance that georeferencing of samples has for obtaining knowledge of soil properties and its analysis.
In agriculture, there are diverse management practices, one of these is the determination of the nutrients in the soil system using chemical analysis in specialized laboratories. Through machine learning, the use of the Vis-NIR spectrum for the soil nutritional prediction has been started, decreasing the cost and time for obtaining the results, and protecting the environment. For the prediction of 29 soil physical/chemical properties, signatures within the spectrum of 400nm and 2500nm were taken and analyzed with the Partial Leas Square (PLSR) statistics method, these were preprocessed, and the outliers were eliminated. The soil samples were taken from the Soil, Water and Foliar Laboratory at EARTH University. The determination coefficient (R2) was selected as the one with the biggest importance. The model prediction quality of 2318 samples were compared to 3381 soil samples. Only three variables presented higher value in R2 in more samples, these are: N with R2=0.73 R2, Na with R2=0.60 R2, and B with R2=0.25. Also, two ratio in the soil’s content were found: Sodium absorption (SAR) with R2=0.57 and Ca/Mag with R2=0.51. The prediction quality of the statistical model and the increasing of the soil samples did not show difference in most variables, given that the more samples of signatures does not necessarily mean greater quality of the prediction; however, it depends on the data aggrupation done. Also, a comparison of the data classified in pH ranges of 5.5 to 6.5 was done. A positive difference could be observed in 17 variables, among the ones that obtained higher values in the prediction are S with R2=0.97, Ca with R2=0.93, Cation exchange capacity (CICe) with R2=0.92, Mg with R2=0.92, C with R2=0.89, Fe with R2=0.85, K with R2=0.79, as well the variables like clay with R2=0.75 and sand with R2=0.74. The classification of soil based on its physical or chemical characteristics makes a better prediction, even the place of origin could be considered. This highlights the importance that georeferencing of samples has for obtaining knowledge of soil properties and its analysis.
Descripción
Palabras clave
PROPIEDADES FISICAS DEL SUELO, ANALISIS ESPECTRAL, CLASIFICACION DE SUELOS, FERTILIDAD DEL SUELO