<p>Maps of (<b>a</b>) mean GFED CO, (<b>b</b>) mean Naus CO, and (<b>c</b>) their difference, and (<b>d</b>) concept diagram of the form permutations for candidate regression equations. The legend between panels a and b applies to both. The average difference in cell means is 7006 g C km<sup>−2</sup> land d<sup>−1</sup>.</p> Full article ">Figure 2
<p>Histogram of CO emissions fluxes in log scale. Bins are 0.1 units wide on a log<sub>10</sub> scale. The 48% of Naus monthly means and 50% of GFED’s whose value is zero are not shown.</p> Full article ">Figure 3
<p>Accuracy metrics for the recommended equations and FireMIP CO emissions. Each recommended equation has a different color and shape. FireMIP models included are CLM4.5, CTEM, JBSpitfire, Jules, LGSimfire, LGSpitfire, and Orchidee. All <span class="html-italic">x</span>-axes show r<sup>2</sup> explanatory power. In the left column (panels (<b>a</b>,<b>c</b>,<b>e</b>,<b>g</b>)), ratio of the mean prediction to the mean benchmark is on the <span class="html-italic">y</span>-axis. In the graphs on the right (panels (<b>b</b>,<b>d</b>,<b>f</b>,<b>h</b>)), ratio of explained variances is on the <span class="html-italic">y</span>-axis. Black arrows mark the optimal score of one for both ratio of the mean prediction and ratio of explained variances, with the arrow emphasizing that higher r<sup>2</sup> is better. Each row of panels shows the same sets of predictions, scaled as linear monthly means (panels (<b>a</b>,<b>b</b>)), as linear annual means (<b>c</b>,<b>d</b>), log monthly means (<b>e</b>,<b>f</b>), or log annual means (<b>g</b>,<b>h</b>). Ratios of means higher than 2 and ratios of explained variances higher than 5 are not plotted. FireMIP emissions accuracy is for only 2003–2013 and is addressed in <a href="#sec4dot2-fire-07-00477" class="html-sec">Section 4.2</a> below. Which recommended equation is most accurate varies markedly by the scale at which prediction accuracy is judged.</p> Full article ">Figure 4
<p>Binned distribution of monthly emission predictions from each recommended equation, compared to the merger of both CO inventories. The panels display the same underlying data but at different scales. The legend applies to both panels. Black dots describe CO fluxes for test cells and are the benchmark. Dashed gray vertical lines mark test cell benchmark means. Panel A shows the distribution of the linear-scale values, binned as the nearest multiple of 125,000. Frequencies in panel a are graphed on a log scale. In panel (<b>a</b>), the truncated <span class="html-italic">x</span>-axis omits the 0.019% (n = 40) of input fluxes larger than 1,600,000 g C km<sup>−2</sup> land d<sup>−1</sup>. For panel (<b>b</b>), the <span class="html-italic">x</span>-axis shows predicted values that have been transformed to a log scale, then binned as the nearest whole number. The spikes at log (0.32 g C km<sup>−2</sup> land d<sup>−1</sup>) are the log-scale replacement for zero emission instances. Inventory fluxes with log value of less than about 2.5 are likely to have large errors of both detection and relative magnitude. The graphs display the large differences in distributions of predictions compared to benchmark data for log versus linear forms of predictions.</p> Full article ">Figure 5
<p>Each predictor’s relative contribution to total r<sup>2</sup> in selected equations. Contributions can be calculated only for training data, and at an equation’s native outcome transformation and time scale. Because the native scales differ across equations, the stacks describe, respectively: contributions when evaluated in monthly linear space for the Linear equation, contributions in monthly log space for the Log equation, and contributions in annual linear space for the LinearPMet equation. Predictors related to dryness and deforestation are the heart of each recommended equation.</p> Full article ">