Assessing diversity in winter cover crop systems using high resolution imagery

By Natan Kramer

Abstract
The global population is expected to exceed 10 billion by 2050, increasing the demand for food production. However, agriculture practices has led to negative environmental impacts, such as deforestation, leaching of harmful substances, and greenhouse gas emissions. Intercropping, with mixed cropping systems, could be a sustainable solution to intensify agriculture. Mixed cropping systems are more productive than mono-culture alternatives, while also being more resistant to drought and disease through plant-soil feedback mechanisms. But the understanding of these mechanisms is limited. Therefore, more research into monitoring mixed cropping systems is necessary. However, monitoring mixed cropping systems is a labour intensive task, which often limits the scope of researches. This research will investigate whether high resolution remote sensing imagery in combination with feature extraction and regression techniques could be a cost-effective solution to perform more measurements for assessing specie coverage ratios in mixed cropping systems. The winter cover crop at the experimental plots of Wageningen University and research will be used as a case study. These plots contained various mixtures of vetch, oats, and radish winter cover crops.

Usually, diversity studies with high resolution imagery are performed with pixel level annotation, however this research will use plot level annotation with assigned coverage ratios. The segment-anything model was used to generate masks for the feature extract. These were compared simplified manually drawn mask. The generated masks on average had a Jaccard index of around 0.91, but the outliers occurred down to 0.45. Furthermore, the research found that the segmented mask had negligible effect on the model performance and concluded that simplified manual mask was the better option.

Three regression models, Ridge, Poisson, and MLP regression, were evaluated, and classical (colour and texture) and deep-learning features were compared as inputs. The study found that Ridge with deep-learning features had the best performance (R2=0.92, RMSE=0.07). However, upon inspection of the validation results of the best model, it was found that although the results of the feature extraction with regression were promising, that there is limitation in its practical use for monitoring winter cover crop coverage ratio change over time. The model occasionally outputs negative number and the models error range is too large to measure minor changes over time.

The finding of the research show that there is potential for feature extraction and regression models to assess coverage ratios in mixed winter cover crop systems, but there is still need for more research about the capabilities of monitoring diversity with high resolution imagery.