A deep-learning approach to forest disturbance detection using Sentinel-1 imagery in the Lower Mekong Basin

Abstract:

To manage forests effectively, tropical and subtropical countries require robust programs to responsively map forest disturbance with earth observation (EO) data. For long-term stability of such programs, the tools must be cost-efficient, transparent, and able to detect change year-round. In the Lower Mekong Basin, frequent cloud cover in rainy periods of the year reduces the effectiveness of some global-scale forest alert systems based on optical data alone. Radar-based mapping of forest disturbance is promising because of radar's all-weather applicability, and the availability of Sentinel-1 radar imagery (S1-imagery) on the Google Earth Engine (GEE), and new functionality which allows GEE and TensorFlow to interface holds promise for making radar-based mapping tractable for resource-limited forest management agencies. Here, we report on a new approach to utilize deep-learning approaches to improve change detection using S1-imagery. In a project funded under the NASA SERVIR programs Lower Mekong Basin Hub, we used a two-phase learning process to train a fully convolutional deep-learning network to segment (i.e., recognize) forest disturbance. In the first phase, we sampled from the full record of existing forest disturbance alerts produced by the Global Land Analysis and Discovery Lab’s (GLAD) forest alerts project (glad.umd.edu/dataset/glad-forest-alerts) to perform the initial training of the network. We then used transfer learning to hone that model on a locally collected reference dataset of forest disturbance data to complete the model training. The network architecture is similar to UNet and uses an encoder-decoder structure to perform semantic segmentation. However, our network uses 3D convolutions in the encoder to learn space-time patterns and has added residual connections between layers to improve gradient flow. The procedure for the two-phase training methodology can be readily adapted to other remote sensing contexts where existing model outputs are abundant but reference data are scarce and/or difficult to collect. Additionally, the two-phase strategy provides flexibility for our local collaborators to augment their training dataset readily without losing the strength of the full time-series of optical-based alerts.