Estimation of Cropping Regimes using High Temporal Frequency Moderate Resolution SITS Analysis

Abstract

Horticultural crops are economically, environmentally and nutritionally beneficial and are more intensively cultivated than field crops. However, there are many varieties, they have shorter growing periods and are cultivated on small parcels of land. The type and variety of crops, while limited by an area’s weather and soil conditions, is also at the discretion of the farmer and is influenced by market forces. This
spatially and temporally dynamic nature means that annual crop type mapping does not adequately provide the information necessary for forecasting. To plan and allocate resources, a higher frequency of crop inventorying involving identification of what crops are grown, and where and when they are grown is necessary. Remote Sensing and Geographical Information Systems provide excellent tools for the collection and analysis of vast amounts of data. Further, high temporal frequency data lends itself to monitoring operations which are pertinent to
precision agriculture. In this study, data mining operations and analysis are tested towards differentiation between vegetation and non-vegetation as the first step towards estimation of cropping regimes or patterns using Moderate Resolution Imaging Spectroradiometer (MODIS) data. MODIS daily NDVI Satellite Image Time Series (SITS) undergoes a process of cloud filtering via a Temporal Window Operation and smoothing using a Fast Fourier Transform (FFT). A process of discretization and dimensionality reduction using symbolic representation of time series data method, Symbolic Aggregate approXimation (SAX) is implemented and the resulting strings are clustered in order to reveal information inherent in the temporal profiles. Vegetation and non-vegetation land cover types were classified using clustering of state sequences and state sequence transitions with accuracies of 59% and 62% respectively.