| 09 Apr 2026
FRAME v1.0: Advancing Fire Risk Assessment in Tropical Fragmented Forests with a Machine Learning Environment
Abstract. In this study, we develop a comprehensive Fire Risk Assessment with a Machine Learning Environment (FRAME v1.0) for tropical fragmented forest systems by adding fuel availability and anthropogenic ignition factors to a well-known climate-driven fire hazard assessment model. Our work focuses on the forests of India, a representative example of tropical fragmented forest systems in a densely populated country where fire behavior is complex and influenced strongly by natural and human factors. In this work, we first developed a Fire Danger Rating System (FDRS) based on the Fire Weather Index (FWI) module of the Canadian Forest Fire Danger Rating System (CFFDRS) and machine learning (ML) techniques. The integration of ML techniques increased the FDRS's ability to estimate fire probability by 30–50 %. While the FDRS forms the core meteorological component of FRAME v1.0, it does not account for other critical drivers. Hence, we extended this FDRS to a comprehensive fire risk assessment framework by incorporating fuel availability and anthropogenic ignition factors using machine learning predictive algorithms with fire count as the target variable. We observed that the neural network-based model performed best among all algorithms across different forest zones of India. Maximum relevance minimum redundancy analysis revealed spatial heterogeneity in dominant fire drivers, although weather remained a consistently critical factor. FRAME v1.0 provides a scalable operational foundation for fire risk assessment in tropical fragmented forests and demonstrates how machine learning can enhance physically grounded fire danger systems.
The study presents a pragmatic method for combining available variables to estimate fire danger under typical conditions in India. The methods and results show significant promise for consideration by the relevant agencies and highlight future research areas, including data gaps in the country.
However, the following points need to be addressed.
1. Fire risk computation based on several inputs, including the FWI, needs further elaboration in the paper. The relevance of hazard component i.e FWI and other vulnerability and exposure indicators, such as altitude, slope, aspect, population density, fuel availability, is well factored by the authors but the manner in which these have been integrated across the regions is not clear. This is an important challenge which needs to be dealt in more detail, including its theoretical basis.
2. The linearity between the FWI values and fire occurrence is established by the authors in 5.1 (lines 330- 357). However, the association of FWI and fires is expected to significantly vary during the fire season based on my personal experience. Towards the end of the fire season during an average fire season (avoiding both extremes), the subsequent peaks in the FWI value towards the end of the fire season are not accompanied by fires due to fuel exhaustion. This is generally expected across the regions, with some exceptions where fires trigger increased needle shedding in Pinus roxburghii forests in the western Himalayas. Fuel exhaustion is an important factor which can explain some of the exceptions noted in the paper.
3. FWI and FDRS appear to be used interchangeably across the paper, which creates some confusion in the reader's mind. Some examples are 5.2 (lines 363- 396); 356-357; Figure 7. Also, please check lines 343-344 ; 556-557 for accuracy.
4. Please check the effect of elevation in lines 512-513, as elevation is not expected to significantly influence fires in the Central indian region, which is mostly a level plateau without much ruggedness, relatively speaking.
5. I suggest looking into the explanation of the results discussed in the last para of 5.4. The interpretation can be improved by looking into the data and correlating it to the ground conditions.