Using Sentinel-1/2 data , supplemented with Landsat, for detection and classification of cover crops in the Netherlands, and to estimate their growth duration

By Harm van Dinther

Abstract
In my thesis, I researched monitoring cover crops in the Netherland using satellite imagery. The research questions originate from Cosun Beet Company (CBC), the largest beet processor in the Netherlands. This cooperation is interested in knowing more about cover crop growth before the beet production of their farmers. Cover crops are non-commercial crops seeded between two main crops. The crops cover the soil during winter, which gives multiple advantages to the soil and the environment. The first part of the study contains a time-series analysis using an algorithm created by Wageningen Environmental Research (WENR). During this thesis, this code is improved and expanded. Using the NDVI of Sentinel-2 and Landsat-7/8, the algorithm detects cover crops using a threshold during a set period in the winter. The results are validated with the use of data from CBC, which resulted in an overall accuracy of 80.3, 83.5, and 85% for three winters (2017-18, 2018-19, and 2019-20). Next, the start and end times of cover crops are estimated by the algorithm, wherefrom the cover crop’s duration can be derived. The start time is estimated by the NDVI, and the end time is estimated by both the NDVI and coherence. The coherence, derived from Sentinel-1, is the amount of similarity between two consecutive radar images and can therefore give insight into the field's structure changes. Due to a lack of ground-truth data, the resulting time estimations could not be validated with actual start and end times. Instead, validation is applied by using visual analysis of the time series, wherefrom we could conclude that the algorithm performs reasonably well in estimating the beginning and end dates of cover crops. Besides the time-series analysis, also machine learning is applied for the winter of 2018-19 in the form of Random Forest. Binary classification into fields with cover crops and bare soil resulted in an overall accuracy of 81.5%. Multiclass classification into the cover crop types fodder radish, yellow mustard, grasses, Japanese oats, rye, and the class ‘other’ is executed and achieved a maximum overall accuracy of 52.5%.

One of the main difficulties during the research was the limited amount of optical imagery. In winter, cloud formation is generally high. This results in less datapoint on the time series and limited input data for the machine learning. For follow-up research, the estimations of the start and end times should be compared with collected ground-truth data to get a better conclusion about the algorithm's performance in estimating the start and end of cover crops.