The NFMD is a website application that has a query system that allows users to access live- and dead-fuel moisture information used in wildland fire management (Delgado, 2009). The application uses a database that is updated by fuel specialists who observe, sample, collect, and input the fuel moisture data into the NFMD. To optimize data acquisition, and querying needed for the long-term live fuel moisture analysis, the Fuel Moisture Repository (FMR) has been created. The FMR creates a repository on a local machine and optimizes the storage of the data acquired from the NFMD. The local database is controlled by the dedicated script FMR.py, which is available on a public GitHub repository at https://github.com/wirc-sjsu/FMR. The script subsets the desired data by regions (e.g., States, GACC, or every region), acquires data from the NFMD, organizes the data by year, and creates a local repository. This data is then readily available to the user whenever they may need it. From here, the user can query the data by timeframe, stations ID, fuel species [fuel types (Chamise, Manzanita, Sagebrush, etc.) and fuel variations (new growth, old growth, whitethorn, etc.)], and geolocation (bounding box) using the Example_FMR_Code.py script. Based on the requested criteria, the script packages all the requested data into a CSV file with the name of the file being the timestamp indicating when it was created. The workflow outlined above allows for easy selection and organization of the LFMC data and was used in this study.

California is the third largest state in the United States by land mass and has a Mediterranean climate. Some regions are heavily forested while other regions experience significant aridity leading to extended grasslands and desert areas. These conditions provide diverse ecosystems across the state. The California State Water Resources Control Board divided California into 10 different bioregions based on topography and local ecosystems. Modifying these divisions, we refined spatial classification by creating a total of 9 different regions: Klamath (KLA), Modoc (MOD), Sacramento (SAC), Bay Area (BA), Sierra Nevada (SN), Central Coast (CC), San Joaquin Valley (SJV), South Coast (SC), and the Mojave Desert (MD), as seen in Figure 1. The collected data for this study has been categorized according to each respective region. During the initial analysis of LFMC in each region, a few areas (Mojave Desert, Sacramento, and San Joaquin Valley) were identified with minimal or no observations and were subsequently excluded. Thus, this study focuses solely on the six regions (KLA, MOD, BA, SN, CC, and SC) that possess a substantial amount of LFMC data.

The FMR was used to gather and subset the LFMC data from sites within the regions’ boundaries described above (Section “2.2. Regions”). Most of the regions had a diverse range of live fuels and the start times of the data records for those fuels varied from region to region (Table 1). Using one of the FMR’s built-in plotting functions, a bar plot displaying regional variability in the types of fuels as well as the number of observations was used to find the most abundant and representative fuel type in each region. To examine the spatial variability of the live fuel moisture, we separately analyzed the climatology across each of the geographical regions. The climatology of the fuels was broken into two sections, seasonality, and long-term trends. For the seasonality figures (Figures 2, 3), all the fuel data were grouped by month and then averaged across the analyzed years to account for all the observations each month. The long-term trend figures depicted in Figures 4–7 showcase the annual maxima, annual minima, and general trends (annually averaged values). Figures 4, 5 provide an overview of the long-term trends spanning the entire research period from 2000 to 2021, while Figures 6, 7 focus specifically on the trends observed in the last 5 years (2017-2021). The corresponding numerical representation of these trends can be found in Supplementary Tables 1–3 in the Appendix. To find the annual maximum trend, we identified the maximum monthly averaged fuel moisture values for each year and applied a simple linear regression model to those values. A similar process was applied to the annual minimum and general fuel moisture trends. To corroborate the results seen from the linear regression model, we also ran the Mann-Kendall Test (Hamed and Ramachandra Rao, 1998). After examining the LFMC data alone, we analyzed the data against two other datasets; first, we looked at how drought affects the annual maximum and minimum values of LFMC (Section “2.4. Drought impact on LFMC”), and second, we compared the number of large wildfires to LFMC values (Section “2.5. Burned area”).

To examine the impact of water deficit on LFMC variability, data from the Drought Monitor were used. The Drought Monitor, a product jointly generated by the National Drought Mitigation Center (NDMC) at the University of Nebraska-Lincoln, the National Oceanic and Atmospheric Administration (NOAA), and the U.S. Department of Agriculture (USDA), provides weekly drought estimates across the United States. There are 5 drought variables, each representing different drought levels on a given week: D0, D1, D2, D3, and D4. These levels are based on varying outputs from Palmer Drought Severity Index, CPC Soil Moisture Model, USGS Weekly Streamflow, Standardized Precipitation Index, and Objective Drought Indicator Blends. Further explanations of how these drought variables were created can be found at the Drought Monitor (n.d.).

For this study, California county drought data were used. The counties were grouped into the 6 regions listed above (Section “2.2. Regions”) and were used to extract and average the data from the drought dataset. The Drought Severity Metric (DSM) was created to determine the overall weekly drought severity for each region (Eq. 1), which can be correlated with LFMC observations. The metric was created as a weighted average, giving each drought level equal weight, to reduce the weekly 5 drought variables into a single value. Each drought variable [shown in Eq. 1 as Data (D0-D4)] is valued between 0–100%. With that in mind, the metric value ranges from 0–5, with 0 being no/slight drought conditions and 5 being the worst drought conditions. The metric was then applied to the drought dataset for each region. The integrated weekly drought values were then averaged across each year to create an annual drought severity value.

Once the annual drought severity value was found for each region from 2000–2021, the annual maximum and minimum LFMC values were gathered using the FMR. Using the drought severity metric as the colormap color reference, a heatmap was created to show the severity of drought each year, and the maximum, as well as minimum LFMC values, were superimposed over the drought conditions.

We leverage the data from fuel moisture repository and the burned area derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) to investigate the relationship between burnt area and the live fuel moisture. The MODIS Burned Area Product (MCD64A1) provides a gridded 500 m monthly dataset which was split into 6 subsets corresponding to the regions of interest. After sub-setting the data, our initial examination of the burned area data revealed a temporal lag in the data between the drying LFMC and the burned area. It has been documented that the MODIS Burned Area Product suffers from omission errors (non-reported burned pixels), commission errors (false alarms), overestimation (each 500 m cell was considered “burned” even if it didn’t completely burn), and burn date uncertainty that results in errors in burn date of over 15 days (Giglio et al., 2009; Ying et al., 2019). It has to be noted satellite fire observations are available at much higher frequency than the live fuel moisture data. The fuel moisture sampling is generally performed monthly, but the observations may be collected anytime within the month. As a consequence, when the fuel moisture observations are averaged monthly, observations at the very beginning of the month (representative of the previous month), or observations at the very end of the month (more representative of the following month) may skew the analysis by creating a time lag between the LFMC and fire activity.

To alleviate this time shift and the detection timing error, here we used the Last Day variable in the MODIS dataset, which provides the last day a pixel was detected as burning, and we averaged the fire data over a 2-month period. This averaging window smooths the burned area and makes it more consistent with low-frequency LFMC data. Once we organized the burned area data for each region, we utilized this data and the LFMC data in the contingency table presented in Table 2.

In the analyzed dataset, some regions had periods with burned area data but no LFMC data, so only periods where both were present were considered. We then ran the burned area and LFMC data through the contingency table, using an LFMC threshold of 79%, to obtain the TP, FN, FP, and TN values for each region. After that, we applied the output from the contingency table to Matthew’s Correlation Coefficient (MCC). The MCC was used to create a baseline correlation score as it combines all of the contingency table results into a single value, which ranges from −1 to 1, with −1 indicating a negative correlation between the predicted and observed values and 1 indicating a positive correlation between the predicted and observed values (Eq 2.). In the context of this study, a positive MCC supports the hypothesis of larger wildfires occurring when the LFMC drops below a certain threshold, while a negative MCC score rejects it, supporting an opposite hypothesis that large fires are associated with LFMC values above the threshold. A MCC value of 0 indicates a lack of relationship between the LFMC and burned area. The values of TP, FN, FP, TN, and MCC for each of the regions’ baselines are shown in Table 3.

To test the adequacy of the region specific baseline threshold, we repeated the process across each of the regions, but instead of using the predefined threshold of 79%, we examined different threshold values in the range between 30–100%. By investigating these possible thresholds, we identified optimal region-specific thresholds useful for indicating the potential for large fires by maximizing MCC. The results can be seen in Table 4.

We also examined the climatology of the fuels in each region along with the average burned area per month as well as the calculated LFMC threshold for each region. To achieve this, we found the monthly averaged burned area from our MODIS data for each region, similar to how we found the LFMC climatology, and plotted it together with the LFMC climatology and the calculated thresholds (Figures 2, 3).

The analysis of the LFMC data indicated the variability in the dominant fuel types in the analyzed bioregions. For instance, Chamise, (Adenostoma Fasciculatum) was the dominant species for three regions (Bay Area, Central Coast, and the South Coast), in 2 regions the most sampled fuel was Manzanita, (Arctostaphylos manzanita), (Klamath, and the Sierra Nevada), while Sagebrush, Artemisia tridentata, was the predominant species in the Modoc region. It is clear by looking at Table 1 that shrubland plants are the dominant species sampled across all California regions, with Chamise and Manzanita being the two most common species. It must be noted that out of all the 6 analyzed regions, 2 (BA and MOD) had a relatively small number of observations given the sampling period and the size of the regions. Not only that, but the overall number of LFMC observations is relatively small given the number of sites as well as the temporal range of the data.

To streamline the analysis, the regions containing Chamise (Bay Area, Central Coast, and South Coast) were grouped together, while the regions characterized by Manzanita (Klamath and Sierra Nevada) were grouped separately. Additionally, the region featuring Sagebrush (Modoc) was included within the manzanita group. Starting with the Chamise regions, the typical monthly values and trends were analyzed and plotted in Figures 2, 4, respectively. These three regions span the California coastline, reaching from the far north to the southernmost tip. Being in relatively similar conditions, one would assume that the LFMC of the Chamise in these regions would also look similar to each other. This is not necessarily the case. The average LFMC maxima have an extreme range across the regions, ranging from about 105–125% with these maxima occurring around April. The average LFMC minima for these regions have a much smaller range, ranging from 58–60% and occur between September and October. Although there is a large range between the LFMC maxima across the regions, the drying period between when the maxima are reached to the time when the minima are reached is similar. Here we see that although the typical timing of the LFMC maxima and minima in chamise regions is similar, the maximum LFMC varies greatly, indicating variability in the duration of the drying periods.

Further examination of the LFMC trends in the Chamise regions shows a dire picture. Each of the regions has shown a downtrend in LFMC from 2000–2021. Specifically in the general trend and maximum trends (see Figure 4). This downtrend seen over the entire study period for these regions is minor relative to the downtrends over the last 5 years (2017–2021) as presented in Figure 6. While all the regions exhibit a downtrend in LFMC, the rates at which these downtrends are occurring are different, showing again the variation in LFMC even among the same fuel type. Before discussing these results in more detail, it is important to also analyze the Manzanita regions.

The Manzanita regions share less similarities between each other compared to the regions with Chamise, and have generally higher LFMC. To start, the maxima LFMC ranges from around 128–138% which is generally higher than chamise (105–125%). As shown in Figure 3, these maxima occur in May (for Sierra Nevada) and June (for Klamath) which is later than in Chamise. The minima, however, have a smaller spread, ranging only from 83–84%, and are reached around September and October similarly like in Chamise. Due to the Klamath region reaching its maximum LFMC in June, this region experiences a rapid drying period as it reaches its minimum LFMC in September/October compared to the Sierra Nevada which reaches its maximum in May. Variations in LFMC are significant between these two regions even though they contain the same fuel type. Quickly looking at the Modoc region, with Sagebrush as the dominant fuel, the maximum LFMC reaches around 177% in May and drops to its minimum of around 74% in September. In Figure 3B it can be seen that during December, the LFMC drops to around 50%, but observations during wet periods are sparse and this may be a single observation for that month.

Moving forward to the trends, we see a different picture in Manzanita than in the Chamise regions. Aside from the Sierra Nevada region which also has a downtrend in LFMC, similar to the Chamise regions, the Klamath and Modoc regions experienced an uptrend. Both are located in mountainous, heavily forested areas in the northern part of California. Due to their location and typical climatological trends in Northern California, these regions may experience more moisture. This likely causes these regions to have an uptrend in LFMC as compared to the more Southern ones (Bay Area, Sierra Nevada, Central Coast, and South Coast) which may experience more aridity. Although these regions have an uptrend, all of the Manzanita regions have a significant downtrend in LFMC over the last 5 years (2017–2021) (Figure 7), similar to the Chamise regions.

When looking at the overall trends of the regions, two of the six regions saw an uptrend and the rest saw a downtrend in LFMC. Although there were two regions with an uptrend, on average, altogether the regions experienced a general decrease of 9.58% in LFMC. The annual maxima decreased by 17.02% and the annual minima decreased by 9.18%. This downtrend is even more prevalent when looking at the last 5 years (2017–2021) (Figure 7). The average decrease across all of the regions in the last 5 years was 17.19%, with the annual maxima decrease of 13.93% and the annual minima decrease of 7.39%.

To extend our trend analysis we also ran the Mann-Kendall Test on the monthly averaged LFMC for each region from 2000–2021 (see Supplementary Table 4 in the Appendix). The Tau values presented there indicate trends consistent with the results from the linear regression test discussed above. The only difference was observed in Modoc region where linear regression showed a slightly positive trend, while Mann-Kendall test indicated a trend that is slightly negative.

Given the correlation between LFMC and wildfire development, this massive reduction in LFMC over the last 5 years appears to be a large factor in the larger fire developments recently seen throughout California. Considering the influential role of drought on live fuel moisture, the subsequent sections examine its significance in establishing possible associations between drought conditions, LFMC, and fire activity.

Most of the observed regions experienced similar drought periods from 2000–2021 as shown in Figure 8. During this time there are three distinct drought periods across California: between 2007–2009, 2012–2016, and 2020–2021. Even though there are evident drought periods across the state, some of the regions experienced more severe/less severe drought conditions than others. An example of this can be seen between the Klamath and Sierra Nevada regions in 2007. Manzanita is the predominant fuel type in both of these regions. The Klamath region entered slight drought conditions whereas the Sierra Nevada region saw moderate drought conditions compared to the prior year. These variations in the drought conditions across all of the regions shows strong evidence that drought does play a key role in the variability of LFMC, even among the same fuel types, and confirms why studies like Nolan et al. (2020) looked into the connections between drought and LFMC.

Before delving into the drought and LFMC comparison, a few limitations of this study have to be pointed out. The first limitation is that even though drought data were available from the year 2000 to 2014, a limited amount of LFMC observations during that time made the presented analysis less robust in the early years. The second limitation is that there is only one non-drought period following 2014, which made it difficult to find any climatological patterns.

As drought conditions worsened, it was originally hypothesized that the maximum and minimum LFMC would decrease and vice versa when drought conditions lessened. As suspected, it was found that whenever a region entered a drought period, the LFMC usually dropped relative to the previous year and when drought conditions eased, the LFMC increased from the prior year. However, one interesting observation was how the LFMC behaved during extended drought periods, typically of two or more years. When a given drought period in a region extended for a few years, the maximum and minimum LFMC would stabilize and sometimes even rise during the following years. This can be observed in Figure 8 in the thirteenth row (South Coast Max) from 2013–2017. In 2013, the region entered a drought period and the max LFMC was 103%, down from the previous year of 113%. The following year, the drought conditions continued but the max LFMC rose to 108%. This trend of increasing and stabilization of the LFMC occurred until 2017 when drought conditions lessened and the LFMC rose from 112% in 2016 to 123% in 2017. This could be due to the plant’s physiology, allowing the plant to adapt to drought conditions. Another potential reason why we may observe this stabilization of the LFMC during extended drought periods may have to do with the way LFMC was defined. The equation for FMC utilizes the wet and dry mass of a plant. During periods when the plants are under high water stress, not only do they have a limited water supply, but they can potentially reduce their dry mass to offset water loss. Therefore, during extended drought periods, a decrease in the dry mass even with identical amount of water may lead to higher moisture content.

However, the most recent drought period from 2020–2021 shows that the maximum LFMC is not responding as expected. The regions that entered drought periods in 2020 saw a drop in LFMC as expected. Going into 2021, each of those regions experienced another drop in LFMC. Based on previous drought periods, there should have been some sort of stabilization or increase in LFMC. Not only is the LFMC during this period not indicating this behavior as it did in previous drought years, but the values approached or reached their lowest levels ever recorded. This may indicate that combination of the consecutive droughts led the LFMC to a breaking point after which the LFMC recovery observed earlier was not possible.

The analysis presented above clearly indicates differences in the LFMC and drought conditions across the analyzed regions. In order to investigate the possible connection between the variability in the LFMC and fire activity, the next section compares MODIS-derived burned area to the observed fluctuations in fuel moisture.

Williams et al. (2019) found that fire activity in California had increased dramatically over the past decade, in part due to a reduction in LFMC. In order to conduct our investigation into the connection between LFMC and fire activity, we utilized the Dennison and Moritz (2009) definitions of large fires and their LFMC threshold to compare our LFMC observations and MODIS data. Tables 3, 4 show the Dennison and Moritz (2009) baseline threshold and the calculated thresholds for each region, respectively. The baseline MCC scores are relatively high for the Central Coast and South Coast and low for the other regions (low MCC values denote a weak relationship between the burned area and LFMC threshold and while high values indicate a stronger relationship). Upon iterating through the various potential thresholds for each region, more optimal thresholds were found for the regions resulting in a significant increase in regional MCC. All of the new thresholds ranged between 70 and 98%. The thresholds for regions that contained Chamise (Bay Area, Central Coast, and South Coast) were within 9% of the ones published by Dennison and Moritz (2009). The regions with Manzanita (Sierra Nevada and Klamath) had thresholds that differed by up to 13%, and the regions containing other fuel types (Modoc) saw up to a 19% difference compared to Dennison and Moritz (2009).

Although most of the locally optimized thresholds differed from the 79% threshold, the regions with Chamise fuels came out to be very close to it, and our estimates for the South Coast (which included the region the original study observed) matched exactly 79% building confidence in our methodology for finding optimal LFMC thresholds. It also appears that the same fuel types share similar LFMC thresholds across different regions, even with varying LFMC.

The climatology of LFMC in each region and its monthly average burned area relative to our LFMC thresholds are shown in Figures 2, 3. In these regions, on average, when the LFMC drops below the LFMC threshold, larger burned areas occur. However, there are outliers such as in the Sierra Nevada region during August when the LFMC is higher than the threshold and larger fires were occurring. One possible reason for this could be the impact of tree mortality due to drought and bark beetles (Restaino et al., 2019) which could lead to a high number of fire occurrences and reduce the importance of LFMC in the context of fire occurrences compared to dead fuels.

There were also some differences in the thresholds between the regions with Chamise and Manzanita. One of these differences can be observed in the period when the LFMC in each region dropped below their optimal threshold (Figures 2, 3). When the LFMC drops below the threshold, there tends to be a spike (around 15,000+ acres) in the burned area in each region. Looking at the regions with Chamise, the Bay Area experienced 4 months where the LFMC dropped below its optimal threshold and saw a burned area of around 15,000+ acres, 8 months in the Central Coast, and 7 months in the South Coast. Then, in the regions with Manzanita, we see the Klamath region below the threshold for about 3 months and 2 months in the Sierra Nevada region. The annual LFMC maxima and minima data may show why this has occurred. The regions with Chamise all have minima around 60% a year while their LFMC thresholds are about 10–22% higher whereas the regions with Manzanita have minima around 84% with their LFMC thresholds being about 4–8% higher. This increasing gap between the minimum LFMC and the optimal threshold indicates a longer period where we can see a potential increase in fire occurrences and an increase in the likelihood for larger fires to occur. The gap also indicates that the differences in the fuels between the regions lead to differences in the LFMC thresholds and therefore, varying burned area activity as shown in Figures 2, 3.