I'm a community ecologist with a PhD in plant-pollinator interactions. Recently, I've become interested in using the vast amounts of data available from satellites to develop tools to monitor biodiversity in supply chains and hopefully incentivize it's protection.
Monitoring biodiversity in agricultural supply chains is a key metric to assess progress on the EU's Biodiversity Strategy and the Farm to Fork Policy. 10% of farms should be composed of 'high diversity landscape features' - habitats such as treelines, hedgerows, semi-natural grassland, forests, and wetlands. However, most land cover mapping benchmark datasets, such as the recent OpenEarthMap, fail to distinguish between agricultural land (pasture, arable, forestry) and the semi-natural vegetation that counts towards the 10% target. Thus, there is a lack of high-quality labelled data to develop models to measure progress towards important policy goals.
I tested commercial high (SPOT 6/7) and very high resolution (Pleiades) satellite images for their ability to pick up linear features in Irish farmland, and their ability to detect 10 landcover classes: Pasture, Semi-Natural Woodland, Conifer, Scrub, Hedgerow, Semi-Natural Grassland, Artificial, Bare Ground, Shadow and Other. A minimum resolution of 0.5m was required to accurately detect the linear features common to Irish Farmland. Due to a lack of high-quality masks that distinguish farmed from semi-natural vegetation, deep learning methods were unsuitable and so an object-based image analysis workflow was developed. The choice of segmentation parameters and number of segments were important to ensure objects captured the shape of landscape features while minimising the speckle effect and loss of resolution due to segment size. Cloudless Pleiades images for 40 farms distributed throughout the Republic of Ireland’s biogeographic regions were obtained for the summer of 2022. Each image underwent segmentation and segments were labelled according to the 10 classes above, with ~1000 points per image to ensure data were obtained from each biogeographic regions in Ireland, resulting in 80,000 data points for model development.
Various indices (NDVI, EVI 1-3, NDWI, GRVI, CVI, CCI, CIGreen) and textures (Grey-level Co-Occurrence Matrices and Local Binary Patterns) applied to indices and the pan chromatic band were added as additional features. A model comparison procedure was carried out, optimising for balanced accuracy, comparing random forests, support vector machines, multi-layer perceptron, kNN, and multi-class logistic regression. A minimum balanced accuracy of 80% was considered acceptable for monitoring purposes. Random forests performed best, with an out-of-sample balanced accuracy of 82%.
The modelling framework and dataset can be used to monitor progress towards the Green Deal targets, and pilots monitoring the biodiversity in large dairy-processors with significant supply chains will be presented. Further improvements such as annual change detection and extending to other European countries will be discussed.