Reconstructing historical forest fire risk in the non-satellite era using the improved forest fire danger index and long short-term memory deep learning-a case study in Sichuan Province, southwestern China

Historical forest fire risk databases are vital for evaluating the effectiveness of past forest management approaches, enhancing forest fire warnings and emergency response capabilities, and accurately budgeting potential carbon emissions resulting from fires. However, due to the unavailability of spatial information technology, such databases are extremely difficult to build reliably and completely in the non-satellite era. This study presented an improved forest fire risk reconstruction framework that integrates a deep learning-based time series prediction model and spatial interpolation to address the challenge in Sichuan Province, southwestern China. First, the forest fire danger index (FFDI) was improved by supplementing slope and aspect information. We compared the performances of three time series models, namely, the autoregressive integrated moving average (ARIMA), Prophet and long short-term memory (LSTM) in predicting the modified forest fire danger index (MFFDI). The best-performing model was used to retrace the MFFDI of individual stations from 1941 to 1970. Following this, the Anusplin spatial interpolation method was used to map the distributions of the MFFDI at five-year intervals, which were then subjected to weighted overlay with the distance-to-river layer to generate forest fire risk maps for reconstructing the forest fire danger database. The results revealed LSTM as the most accurate in fitting and predicting the historical MFFDI, with a fitting determination coefficient (R2) of 0.709, mean square error (MSE) of 0.047, and validation R2 and MSE of 0.508 and 0.11, respectively. Independent validation of the predicted forest fire risk maps indicated that 5 out of 7 historical forest fire events were located in forest fire-prone areas, which is higher than the results determined from the original FFDI (2 out of 7). This proves the effectiveness of the improved MFFDI and indicates a high level of reliability of the historical forest fire risk reconstruction method proposed in this study.