Detection of Forest Burned Area Using a Spatiotemporal Algorithm Based on Spectral Index Time Series Data

Timely and accurate wildfire detection is critical for ecological monitoring and disaster response. Using a model to handle both long-term time series and multi-scale spatial features is challenging. In this study, we propose a novel burned area detection algorithm based on the CL-UNet (ConvLstm-U-Net) model. CL-UNet integrates convolutional long short-term memory (ConvLSTM) units with a multi-scale encoder-decoder architecture to capture both spatial and temporal dependencies inherent in satellite image time series. The model is trained to forecast spectral index values for the target year, and discrepancies between the predicted and observed values are used to identify anomalous changes. Burned areas are detected using a root mean square error (RMSE)-based thresholding method, which provides a robust mathematical foundation for selecting the optimal threshold. Using MODIS products with a temporal coverage spanning up to 9 years, we constructed spectral index time series (SITS) based on multiple spectral indices. Experimental results from two large-scale forest fire demonstrate that Burn Area Index (BAI) exhibits the most qualified and stable performance. Under various conditions, the Recall, Precision and F1 score of the model using the BAI time series as input data all exceed 80%. The results further indicate that CL-UNet achieves outstanding performance in both study areas. When utilizing the BAI time series as input, the Recall, Precision and F1 score of CL-UNet are 1% to 10% higher than those of other models. The proposed algorithm shows strong potential for general spatiotemporal disturbance detection tasks in remote sensing applications beyond wildfire monitoring.