Refined algorithm for forest early warning system with ALOS-2/PALSAR-2 ScanSAR data in tropical forest regions

In this study, an automatic change detection method for near real-time (NRT) forest monitoring in tropical re-gions is described. L-band ALOS-2/PALSAR-2 ScanSAR HH, HV, and HH/HV ratio were used to detect various deforestation stages based on their different radar scattering characteristics. The three main stages considered in this approach were as follows: (1) forest was cut, and felled trees were left on the ground, (2) felled trees were burned, and (3) felled (and burned) trees were removed. Two types of forest fires occurred in tropical rain forests and dry and open forests were considered. Multi-temporal data and image normalization techniques are used to suppress commission errors induced by the effects of seasonality and rainfall. Detection accuracies were eval-uated at 11 validation sites distributed over various forest types in Latin America, Africa, and Southeast Asia. The user's and producer's accuracies of the proposed forest monitoring algorithm were estimated to be 85.0% and 63.8%, respectively. An additional broader performance assessment using data from 191 sites showed an esti-mated total user's accuracy of 71.1%. The results indicate that detection accuracies depend on temporal se-quences, with slower, gradual transitions from forest to other land cover hampering detection performance. The developed algorithm is used in the JICA-JAXA Forest Early Warning System in the tropics (JJ-FAST), which provides forest change information for 77 countries every 42 days under all weather conditions. The NRT alert products are freely available from the JJ-FAST website 3-4 days after PALSAR-2 observation.