Plant Vigor Assessment by Regression from Drone to Sentinel-2 Imagery

Abstract

Plant vigor assessment is an important issue in modern precision agriculture. The availability of Unmanned Arial Vehicles (UAV) and miniatured remote sensing sensors have made it possible to get precise vigor assessments. Till date, only high-resolution images are considered useful in this regard. These images are subject to cost and human effort. Naturally, it would be of much practical importance to be able to achieve precise vigor assessment from openly available images, for example satellite images. The challenge here is the low resolution of such images, for instance, images acquired with the sentinel-2A instrument from ESA’s sentinel 2 mission has a resolution of 10 m. In this research we try to tap the benefit of these freely available images while considering the accuracy issues of plant vigor assessment. Current state of the art shows the usefulness of Normalized Vegetation Index (NDVI) in relation to plant vigor assessment. It is easy to compute, and not very time-consuming for a large area, However, as the low resolution of sentinel 2 images is concerned, there is a need for rectification of NDVI values. We work around this problem with the help of some high-resolution images and regression techniques. In other words, NDVI computed from high resolution images are used to guide the vigor assessment algorithm by transfer learning. As a case study, we used UAV images acquired in Vineyards in Spain, as part of the AI4Agriculture project. Sentinel-2 images were acquired from ESA sentinel hub in the same week of acquisition as the UAV images. As there are soil tracks between the vineyard plants, we removed the soil tracks with an unsupervised classification algorithm. The transfer learning from UAV to sentinel-2 images was achieved by means of regression techniques. After visualizing and verifying the relation between NDVI computed from Sentinel 2 images and UAV images for both soil-segmented and unsegmented sentinel-2 images, we trained several regression algorithms with these two NDVI values. A comparison between the algorithm proved the boosted regression tree to be the best to model the relationship. This activity was done as part of the AI4Agriculture project. This regression model is delivered to the users who can use it to rectify NDVI computation for similar cases. The software is also available as a platform-independent service and also as an executable in Python programming language.