Earth Observation for Sustainable Sugarcane Production

Abstract

Sugarcane is a high impact crop and the source of most of the world’s sugar, with India being the second largest producer and consumer of sugarcane in the world. The large resource requirements of sugarcane plantations, coupled with the demand for increased sugarcane production calls for the implementation of sustainable practices in sugarcane production over large areas quickly. Earth observation (EO) datasets and machine learning tools have been used in the past to monitor crop cultivation over large areas through spectral changes in surface reflectance that occur over time. Existing approaches that use EO for monitoring sugarcane cultivation over large areas are limited in scale by their reliance on crop calendars and phenological data derived from field surveys. Hence, this PhD aims to determine how new forms of EO imagery and machine learning tools can help to increase resource efficiency (land, labour, nutrients, water), in sugarcane production. The current state of the art for monitoring sugarcane production for improved sustainability was reviewed. Research gaps formed around using EO time-series data for monitoring sugarcane growth and management over time and the classification of sugarcane over large areas to overcome the limitations of field surveys. To overcome these limitations, new approaches were investigated to derive sugarcane growth stages using EO data from Landsat 8 and Sentinel-2 satellite data for sugarcane fields located in two study sites, Unit 1 and Unit 2, situated in the India States of Andhra Pradesh and Telangana. Automated methods for decomposing time-series data into individual growing seasons were compared to visual observations which resulted in R2 values of 0.72 and 0.84 for Unit 1 and R2 values of 0.78 and 0.82 for Unit 2, when regressing the automated start and end of the season against the visually interpreted start and end of the season. Also, land surface phenology was related to agronomic practices using the FAO growth model resulting in R2 values of 0.56 and 0.72 for Unit 1 and R2 values of 0.36 and 0.79 for Unit 2, for the relationship between the automated and visually interpreted start and end of the mid-season growth stage. These results demonstrate that the EO time-series can automatically determine the growth stages of sugarcane in India over large areas, without the need for prior knowledge of planting and harvest dates, as a tool for improving sustainable production. Synthetic aperture radar images were investigated to overcome challenges relating to unfiltered clouds and the frequency of observations. Growth information around tillering was present in the SAR data which was not present in the optical data and harvest could be more accurately determined compared to optical imagery. The results highlight the benefits of combining optical and SAR time-series information for monitoring sugarcane growth and management over time. Combining Sentinel-1 and NDVI time-series data allows managers of sugarcane factories to derive the timing of key sugarcane growing stages, harvests, and possible irrigation events. Seasonal decomposition was dependent on information regarding the year of the survey. To further overcome the dependency on survey data, an approach for automatic classification of sugarcane from 6-year long time-series data using Convolution Neural Networks was evaluated. The results include optimising parameters for the model, by analysing the effect of balancing classes in the training data, kernel size, and number of convolutional layers. The two best models for classifying sugarcane when trained using data from Unit 1, could generalise well and classify sugarcane in Unit 2. However, the performance of the two models improved when both were trained and tested using all data. These results demonstrated that a classification model, designed for the medical sector, could be successfully modified for agricultural applications. Individually the algorithms derived in these studies provide decision support over large areas and automate information retrieval for sustainable resource management, yield, and biodiversity. Together the developed algorithms can be used as a strong tool to be implemented into existing sustainability frameworks for improving the sustainability of sugarcane production.