Detección de anomalías en cultivos agrícolas en base a información obtenida de sensores remotos. Una comparación entre Redes Neuronales y modelos ARIMA
Cargando...
Fecha
Authors
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Rosario
Resumen
Este trabajo analiza la detección de anomalías en cultivos agrícolas mediante el uso de
series temporales de Normalized Difference Vegetation Index (NDVI) obtenidas por
sensores remotos, comparando el rendimiento de dos modelos predictivos: el modelo
estadístico Seasonal Autoregressive Integrated Moving Average (SARIMA) y la red
neuronal Long Short-Term Memory (LSTM). El estudio se aplicó a datos satelitales del
sensor Sentinel-2 correspondientes a cultivos de soja en el noreste de la provincia de Santa
Fe, Argentina.
Se implementaron técnicas de preprocesamiento como el suavizado con filtros
Savitzky-Golay y medias móviles, y se aplicó aumentación artificial de datos para
compensar la baja ocurrencia de eventos anómalos. El modelo SARIMA mostró un ajuste
adecuado bajo criterios de información como AIC, BIC y HQIC, pero falló en la detección
de anomalías abruptas, como la caída del NDVI atribuida a una helada. En contraste,
LSTM fue capaz de detectar eventos significativos que superaron el umbral de 3
desviaciones estándar (3σ), demostrando una mayor sensibilidad y adaptabilidad ante
patrones no lineales.
Además, se identificaron valores de NDVI mayores a 0.8 que, si bien fueron señalados
como atípicos por el modelo, no se interpretaron como errores, sino como eventos
inusuales según especialistas agrónomos. Los hallazgos confirman que el modelo LSTM
es más adecuado que SARIMA para la detección temprana de anomalías en series
temporales NDVI en regiones agrícolas con alta variabilidad climática.
This research focuses on anomaly detection in agricultural crops using NDVI time series data derived from remote sensing, comparing the performance of two predictive models: the SARIMA statistical model and the LSTM neural network. The study uses Sentinel-2 satellite imagery applied to soybean crops in northeastern Santa Fe, Argentina. Preprocessing techniques such as Savitzky-Golay smoothing and moving averages were implemented, along with artificial data augmentation to address the scarcity of labeled anomalies. While SARIMA showed reasonable model fit AIC, BIC and HQIC, it failed to detect abrupt changes, such as a NDVI drop due to a frost event. In contrast, LSTM successfully detected events exceeding the 3σ (three standard deviation) threshold, 2 demonstrating greater sensitivity to nonlinear patterns. The model also flagged NDVI values above 0.8, which agronomic experts considered unusual but not erroneous, likely caused by atmospheric effects or sensor noise. Findings support that LSTM outperforms SARIMA in detecting anomalies in NDVI time series for agricultural regions with high climatic variability
This research focuses on anomaly detection in agricultural crops using NDVI time series data derived from remote sensing, comparing the performance of two predictive models: the SARIMA statistical model and the LSTM neural network. The study uses Sentinel-2 satellite imagery applied to soybean crops in northeastern Santa Fe, Argentina. Preprocessing techniques such as Savitzky-Golay smoothing and moving averages were implemented, along with artificial data augmentation to address the scarcity of labeled anomalies. While SARIMA showed reasonable model fit AIC, BIC and HQIC, it failed to detect abrupt changes, such as a NDVI drop due to a frost event. In contrast, LSTM successfully detected events exceeding the 3σ (three standard deviation) threshold, 2 demonstrating greater sensitivity to nonlinear patterns. The model also flagged NDVI values above 0.8, which agronomic experts considered unusual but not erroneous, likely caused by atmospheric effects or sensor noise. Findings support that LSTM outperforms SARIMA in detecting anomalies in NDVI time series for agricultural regions with high climatic variability
