PSXII-9 Application of Machine Learning Algorithms to Estimate Tropical Pasture Biomass Based on Satellite Images. (21st September 2022)

• Record Type: Journal Article
• Authors: Fernandes, Marcia H; Fernandes, Jalme; Reis, Ricardo Andrade; Tedeschi, Luis O
• Is Part Of: Journal of animal science. Volume 100(2022)Supplement 3
• Page Start: 284
• Page End: 285
• Publication Date: 2022-09-21
• Subjects: grassland -- remote sensing -- vegetation index; Livestock -- Periodicals; Livestock; Electronic journals; Periodicals; 636.005
• DOI: 10.1093/jas/skac247.518
• Languages: English
• ISSNs: 0021-8812

Abstract

Abstract: The proper quantification of forage allowance for ruminants in grass-based production systems has always been challenging. At the field scale, pasture management based on ground-level measurements, such as clipping or plate meter, is labor intensive and hampers the assessment of spatial and temporal variability. On the lookout for a solution, this study aimed to estimate marandu palisade grass forage mass based on satellite images using two machine learning (ML) algorithms, Multiple Linear Regression (MLR) and Artificial Neural Network (ANN). The experimental area comprised Marandu palisade grass pasture (Brachiaria brizantha 'Marandu'), summing 33 paddocks (42 ha), receiving or not N fertilization. Pastures were managed under continuous stocking to maintain grazing height fixed at 25 cm during the growing season, using the put-and-take methodology with young beef bulls. Field dataset collection (total forage mass and morphological composition) and satellite images were assessed from Dec/2015 to Mar/2019 during the growing or wet season. The satellite images (Landsat-8 and Sentinel-2) were downloaded from US Geological Survey (USGS, http://earthexplorer.usgs.gov ). Six spectral bands (Bd) and five vegetation indices (VI, Table 1) were used as input variables to MLR and ANN models. Datasets were randomly divided into a training set (80%) and a testing set (20%). Analyses were run in Python 3 (version 3.7). ML models are generally data-driven and require a large amount of data for better performance, and the best accuracy was achieved by using all Bd and VIs as input variables for both models (Table 1). In general, ANN produced better estimates than MLR models. Bd+VI better predicted leaf mass than total forage mass. Our results show that remotely sensed observations, based on satellite images, are a promising and effective tool for tropical grassland monitoring and management under continuous stocking rates. São Paulo Research Foundation (FAPESP) (grant ≠15/16631-5; 17/18750-7; 20/14367-7).