Targeted land management strategies could halve peatland fire occurrences in Central Kalimantan, Indonesia

Large areas of Indonesian tropical swamp forests have been transformed into degraded peatlands through logging, agricultural expansion, and drainage, leading to recurrent fires with major climate, economic, and public health impacts. Despite recognition that fires are largely anthropogenic, existing restoration and management programs in Central Kalimantan face governance and compliance challenges, and quantitative evidence linking specific interventions to future fire reduction is limited. Prior studies often use hydrological models of canal blocking or retrospective statistical models that rely on fire-season data, limiting predictive scenario analysis. This study develops a predictive convolutional neural network (CNN) using only pre-fire season variables to reproduce fire-season hotspot distributions and to evaluate targeted land management and restoration strategies (canal blocking, reforestation, plantation establishment) and the impacts of continued deforestation in Central Kalimantan. The objective is to quantify potential reductions in fire occurrence under realistic management scenarios.

Previous work highlights the role of peatland degradation, drainage canals, and anthropogenic ignition in driving fires across Indonesian peatlands. Hydrological restoration studies argue for canal blocking to raise water tables and mitigate fires but typically lack spatially explicit fire distribution forecasts under restoration (e.g., Ritzema et al. 2014). Other studies identify key drivers of fires and ignition sources using retrospective models and contemporaneous variables (e.g., NDVI), which can inadvertently encode information about actual fires, limiting their predictive utility for scenario testing. Recent deep learning approaches, including CNNs, have improved fire susceptibility mapping in other regions but generally rely on fire-season predictors, making them descriptive rather than predictive for future scenarios. This study addresses these gaps by using only pre-fire season and static variables to enable unbiased scenario simulations.

Study area: Central Kalimantan, Indonesia, focusing on the ex-Mega Rice Project region, characterized by extensive peatlands, a pronounced dry season (May–October), and highly variable interannual fire activity. Data and preprocessing: Dependent variable comprised MODIS (1 km) and VIIRS (375 m) active fire hotspots during Aug–Oct, buffered by 500 m around grid-cell centers and filtered to confidence >50%. Predictor variables (31 total) include: land cover (reclassified to 8 classes) and a forest clearance index; pre-fire season (May–July) vegetation indices (NDVI, EVI) and ET/PET; drought indices (SPEI 3- and 12-month, measured backwards from July); number of cloud days (May–July); distances to canals, roads, rivers, and settlements (from OpenStreetMap, circa 2015); topography (SRTM elevation/slope/aspect); peat depth; Oceanic Niño Index (July–September). All predictors were normalized to [0,1] using cross-year min–max, resampled to common 0.002° resolution, and stacked annually. Each yearly stack was tiled into 32×32 patches (with 31 channels); corresponding hotspot rasters were tiled similarly (1 channel). Tiles across all years (2002–2019) were assembled into 4D arrays and randomized for training/validation. Model architecture (FireCNN): A fully convolutional binary classifier with five 3×3 convolution layers preserving spatial dimensions. K1: input 31 channels → 128 channels; K2–K4: 128→128; K5: 128→1 (probability). Activations: ReLU for K1–K4 and sigmoid for K5. Output is a per-pixel probability of fire occurrence (0–1). Training and validation: Optimizer Adam with binary cross-entropy loss; 70/30 train/validation split; 20 epochs. Metrics: accuracy, precision, recall. After inference, a 3×3 moving average smoothing was applied to reduce tile edge effects. Performance evaluated per year (2002–2019) by comparing predicted probabilities >0.5 to buffered hotspot rasters to compute TP, TN, FP, FN and derive accuracy, precision, recall. Scenario simulations: Seven scenarios were created by altering input layers while keeping others unchanged: (1) canal blocking (retain only two major canals in proximity layer); (2) convert Swamp shrubland and Scrubland to Swamp forest; (3) convert Swamp shrubland and Scrubland to Plantation; (4) convert Swamp forest to Swamp shrubland; (5) convert Swamp forest to Plantation; (6) combined: canal blocking + reforest degraded shrub/scrub to Swamp forest; (7) combined: canal blocking + convert degraded shrub/scrub to Plantation. For each year, the ratio of predicted fire pixels (>0.5) under each scenario to the corresponding baseline (observed land cover and canals) was computed to quantify percentage change. Early warning test: A second model was trained with data from 2002–2018 only and applied to 2019 pre-fire predictors to assess out-of-sample predictive capability, comparing accuracy, precision, and recall to both observed hotspots and the in-sample FireCNN results for 2019.

Model performance: During training over 20 epochs, accuracy, precision, and recall improved as loss decreased; final overall accuracy ≈93%, precision ≈75%, recall ≈50%. Across years (2002–2019), median accuracy was ≈95%, median precision ≈78%, and median recall ≈46%, with recall higher in active fire years (e.g., 2014, 2015, 2018, 2019) and lower in years with few fires (e.g., 2008, 2010). The model captured large fire clusters well but missed many isolated fires. Scenario impacts (relative to baseline median fire counts across 2002–2019): - Convert degraded Swamp shrubland and Scrubland to Plantation: −55% median fire occurrences. - Convert degraded Swamp shrubland and Scrubland to Swamp forest: ≈−40%. - Canal blocking (retain only two major canals): ≈−40%. - Combined canal blocking + reforest degraded areas to Swamp forest: −70%. - Combined canal blocking + convert degraded areas to Plantation: −76%. - Deforestation scenarios: Swamp forest → Swamp shrubland: +10%; Swamp forest → Plantation: −20% (with caveat that actual practice would likely include new canals, potentially increasing fire risk beyond modelled results). Early warning (2019 excluded from training): accuracy 81%, precision 67%, recall 10% (recall decreased from 76% when 2019 was included in training), indicating limited out-of-sample recall despite good precision/accuracy. Overall, targeted land management, especially widespread canal blocking combined with restoring degraded areas, could reduce median fire occurrences by up to about 70%, with substantial implications for carbon emission reductions.

Using pre-fire season and static predictors only, the CNN produced spatially explicit fire distributions with strong precision and accuracy, enabling unbiased scenario testing. The results reinforce the anthropogenic nature of peatland fires and provide quantitative evidence that land management can substantially alter fire regimes. While converting degraded areas to plantations yielded larger modeled reductions than reforestation, practical constraints make plantations incompatible with broad canal blocking because plantations typically require drainage canals. Hence, widespread canal blocking coupled with re-establishing swamp forest is likely the most pragmatic and effective approach. Deforestation scenarios were conservative because they did not add new canals or roads that typically accompany forest conversion, implying actual fire increases from deforestation could be higher. Although the early warning experiment showed low out-of-sample recall for 2019, the scenario analyses remain valid because they perturb a small subset of inputs within the model’s verified operating domain. Implementing effective fire mitigation policies will require balancing environmental benefits with economic and social feasibility and fostering community and institutional support.

A CNN driven by pre-fire season variables accurately mapped fire-season hotspots in Central Kalimantan and enabled robust evaluation of land management scenarios. Blocking most drainage canals and restoring degraded areas can reduce median fire occurrences by up to about 70%, offering a compelling pathway to cut peatland fire emissions and support sustainable ecosystem management. The approach demonstrates the value of predictive CNNs for policy-relevant scenario analysis and could be extended to other regions. Future work should improve input data quality and temporal specificity (e.g., dynamic infrastructure and land cover), simulate co-occurring changes such as new canals during plantation expansion, and develop enhanced early warning capabilities with better recall outside the training period.

Key limitations include: (1) hotspot detection errors (false positives/negatives) and occlusion by clouds/smoke; (2) anthropogenic proximity layers (roads, settlements, canals, rivers) sourced circa 2015 applied to all years, potentially misrepresenting earlier periods; (3) land cover maps do not distinguish plantation types with different management; (4) inability to represent timing, location, and hydrological effects of numerous small-scale canal restoration structures; (5) model recall is modest overall and poor for isolated fires; (6) scenario simplifications (e.g., plantation expansion without adding new canals/roads) likely underestimate fire risk increases from deforestation; (7) limited out-of-sample recall in the early warning test (2019). These factors may affect absolute predictions but are less likely to bias proportional changes across scenarios compared to the baseline.