PS 8-108 - Estimating landscape-level plant diversity with spectral reflectance, remote sensing, and deep learning

Abstract: As land use and land cover shift rapidly due to climate change and human development, estimating plant biodiversity with reduced sampling effort is urgently needed. Sampling plant biodiversity is time-intensive, expensive, and requires specialized personnel. Even with intense investment, field sampling retains significant sources of error, further discouraging broad sampling. These challenges leave the present plant biodiversity of many ecosystems unquantified. This is particularly true in areas like the state of Florida in the southeast United States, where intense agriculture, diverse climates, and high non-native plant incidence can all rapidly change local biodiversity. The goals of our study are to use remote sensing and machine learning to rapidly quantify plant species diversity across the state of Florida.

Our approach started with collecting available data across Florida, including multispectral satellite imagery, land cover and soil maps, plant surveys conducted by four land management entities, and field-sampled plant and hyperspectral data collected by our group on agricultural lands. Hyperspectral data captures electromagnetic radiation signatures related to plants’ chemical composition and physical structure, allowing for accurate identification. We are training a U-net-based deep learning model on these data to predict plant species diversity, which we will then validate with further field sampling.

Here we show our initial synthesis of species observations and field-collected spectra for model development. We have already collected leaf spectral signatures for over 150 individual plants across the state. Additionally, we have synthesized georeferenced plant identifications for over 3,200 species in 238 families in Florida. We are using these data for training our U-net to quantify plant species diversity of a given area. These deep neural network algorithms and associated statistical models will allow us to make inferences about plant species diversity from patterns within our data. We will evaluate the effectiveness of this model using independent field data. Our model will generate a statewide map of plant species diversity across Florida, enabling frequent reassessment of plant biodiversity in an area, allowing identification of biodiversity hotspots, which can be prioritized for conservation.

In conclusion, expanding this model could help more rapidly detect where plant diversity is being lost, even across short timescales. Developing novel surveying methodology will advance the study of large-scale ecosystem services by reducing survey costs and increasing survey accuracy. The availability of a plant species diversity quantifier opens the door for further study of ecosystem diversity and function across landscapes that can inform conservation policy and prioritization.