Statistical Machine Learning for Corn Yield Prediction Based High-Resolution Satellite Imagery: Comparison Between Raw Data and a Multimodality Approach

Abstract

Agriculture is an essential source which ensures a balance on different fields. It contributes in particular to meet the needs of human beings. To do this, this domain is based on several indices and measures, especially the agricultural yield. It constitutes a very significant element allowing to know the quality and the production of a crop. The majority of studies are based on the use of vegetation indices to predict crop yields. In this chapter, we propose a novel approach combining various spectral information to predict corn yield using statistical machine learning techniques. We have tested nine different algorithms using different methodologies based on high resolution satellite data. The first one is based on raw image Sentinel-2 data, namely Red–Green–Blue (RGB). For the second one, we have used only multispectral bands. We have also established four Vegetation Indices (VIs), Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Enhanced Vegetation Index (EVI), and Triangular Greenness Index (TGI) to predict the same yield. The last used technique to predict this measure was a multimodality approach combining various categories of spectral information. The results highlight that the three RGB bands present better scores using every machine learning algorithm, and the multispectral data and VIs present similar values of statistical metrics. The AdaBoost regressor algorithm provided the best scores for the four different inputs. Our study has focused on the advantage of statistical machine learning techniques to enhance the prediction of corn yield. In general, the results prove the effectiveness of machine learning in capturing the different relationships and explaining the variability of agricultural yield over the years.