Crop Stage Estimation: A Multi-Satellite Historical Model and a Scalable Neural Network Forecaster

Abstract:

The ability to predict main stages of an annual crop life cycle enables growers to make more-informed decisions concerning labor and sales planning as well as fertilization and pesticide use. During their life cycles, crops exhibit unique growth curves that can be described by the following chronologically-ordered key dates: emergence (when most seeds have sprouted), row closure (when vegetation completely covers soil), senescence (when the crop begins to brown), and harvest (when the field is harvested). We work on two tasks: (1) extracting key dates from historical satellite data and (2) forecasting key dates. Current methods of extracting key dates involve mathematical analysis of a normalized difference vegetation index (NDVI) time series; typically, key dates are determined from first, second, or third derivatives. These methods rely on heavy and sometimes error-prone interpolation and smoothing, especially when clouds obscure the satellites’ line of sight. Instead, to extract the key dates, we propose a machine-learning based model that analyzes data from multiple satellite sources and fits functions to the extracted growth curves. This method allows for uniform data gathering at scale, across many years. We use data from NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) and Copernicus' Sentinel-1 and Sentinel-2 satellites. The cloud obstruction problem is addressed through the incorporation of Sentinel-2's backscatter data, which clouds do not affect. On average, the model’s determinations of the key dates are within 1-2 weeks of their true values. In addition to designing a model to extract the key dates from a range of satellite data, we also present a neural network that forecasts key dates. This neural network is trained on temperature inputs (growing degree days) and outperforms a baseline algorithm (a historical average of key dates) by about 20% on average.