We propose to develop and commercialize a deep learning-based image classification capability that detects fine-scale and rapidly changing land surface features, using relatively low resolution and low-cost imagery and an architecture that is simple and fast to train. The proposed system promises to substantially improve the study of high frequency land cover dynamics in heterogeneous landscapes by addressing two principal roadblocks to higher spatial resolution and more frequent land cover classification: 1) the high cost of acquiring high resolution multispectral imagery on a frequent basis, and 2) the general complexity of using machine learning techniques to improve classification capabilities. Our innovation involves using time series of multispectral imagery with relatively rich spectral content as a trade-off with spatial resolution, and applying it on a pixel by pixel basis. Our Phase II focus will be on agricultural areas that frequently change on a small scale. Annual vegetable crops are a key set of relevant land cover classes. But our methodology is extensible to other land cover types, such as urban settlements and their change, and other data inputs in addition to imagery, such as time series of weather data.
Related follow-on opportunities for NASA program infusion include integration with the TOPS-SIMs irrigation management program at the Ecological Forecasting Lab at NASA Ames, and NASA Goddard’s Harvest Consortium led by the University of Maryland to enhance the use of satellite data in decision making related to food security and agriculture, and the Surface Biology and Geology (SBG) Decadal Designated Observable Study.
Related commercialization opportunities include monitoring and forecasting for industrial agriculture, particularly for fresh vegetable crops, improved cropland classification for USDA’s Cropland Data Layer, and food waste and sustainability applications addressing prioritized actions of the EPA, USDA and FDA.