MAPPING RICE ECOSYSTEMS IN ILOCOS REGION, PHILIPPINES USING SENTINEL-1A SAR TIME SERIES DATA AND RANDOM FOREST

摘要

Boosting rice production in the Philippines is necessary to feed the growing population. To optimize production, information about the extent and location of irrigated and rainfed rice areas is important for effective targeting of appropriate ecosystem-specific technologies and interventions. This study is a first attempt to develop a methodology to classify irrigated and rainfed ecosystems using multi-temporal Sentinel-1A Synthetic Aperture Radar (SAR) images, and other derived layers from ancillary sources. Random Forest (RF), an ensemble machine learning classifier, was used for rice ecosystem classification. Sentinel-1A VH-polarized Interferometric Wide (IW) swath mode images acquired during 2016 wet season cropping (between May and December) were used to derive a rice mask and 21 predictor variables for the RF classification. Additional predictor variables that were derived from ancillary sources include elevation, flow accumulation, proximity to water bodies, and cumulative rainfall. Ground observations were used for training the random forest. In addition to field data, government statistics and an irrigation network map were used to validate the results. The rice mask had an overall accuracy of 93% with kappa index of 0.86, and the ecosystem classification yielded an overall accuracy of 68.3% with kappa index of 0.37. We found that the mean backscatter difference between the ecosystems was most evident at the start of the season (SoS), i.e., agronomic flooding, wherein irrigated rice had lower mean backscatter (-19.5 dB) than rainfed rice (-17.6 dB). The key factors in the RF classification were the backscatter at the SoS. the span from the SoS to the peak of SAR (PoS). and the sum of absolute gradients within the temporal series. However, significant discrepancies from government statistics were observed, with underestimation of irrigated areas (R2 = 0.79 and RMSE = 998 ha) and overestimation of rainfed areas (R2 = 0.54 and RMSE = 971 ha). Comparison with existing maps showed correct classification of irrigated rice areas near the stream and irrigation network. This suggests good agreement in areas with surface water irrigation. The results suggest that further improvement in the methodology is necessary to achieve better accuracy. Nevertheless, findings from this initial study contribute to future research on the development of a robust method for rice ecosystem classification.