How it performs
TideWatch against flexiclass.
The same held-out flood images, sent to both: TideWatch, trained for this question, and flexiclass, Accelerant’s general decision service, asked it zero-shot. Every metric, with its uncertainty.
Measured 2026-11-20. How this was measured ↓
Aerial · 3,034 test images
ROC-AUC · accuracy 98.1% against 96.3%
Satellite · 594 test images
ROC-AUC · accuracy 98.0% against 83.3%
Aerial
Aerial: 3,034 test images.
Held-out drone and helicopter photos, never used to train TideWatch; the same files sent to both.
TideWatch flexiclass
Every metric
| Metric | TideWatch | flexiclass | Difference | p |
|---|---|---|---|---|
| ROC-AUCRanking: chance a flooded image scores above a dry one | 0.9940.989–0.997 | 0.9700.957–0.981 | 0.0240.015–0.035 | <0.001 |
| PR-AUCAverage precision across recall | 0.9710.958–0.981 | 0.9020.875–0.929 | 0.0690.048–0.091 | <0.001 |
| AccuracyShare of images answered right | 0.9810.976–0.985 | 0.9630.956–0.970 | 0.0180.011–0.025 | <0.001 |
| Balanced accuracyMean of recall and specificity | 0.9580.943–0.971 | 0.9010.880–0.924 | 0.0560.038–0.074 | <0.001 |
| PrecisionOf the images called flooded, the share that are | 0.8880.855–0.919 | 0.8140.768–0.857 | 0.0740.027–0.123 | 0.008 |
| RecallOf the flooded images, the share found | 0.9280.898–0.955 | 0.8240.781–0.868 | 0.1040.069–0.139 | <0.001 |
| SpecificityOf the dry images, the share called dry | 0.9870.983–0.991 | 0.9790.973–0.984 | 0.0080.002–0.015 | 0.020 |
| F1Precision and recall, equally weighted | 0.9080.884–0.928 | 0.8190.784–0.851 | 0.0890.056–0.119 | <0.001 |
| F2Recall weighted 4 times precision: a missed flood costs more | 0.9200.895–0.941 | 0.8220.786–0.859 | 0.0980.065–0.128 | <0.001 |
| MCCCorrelation of answer and truth, −1 to 1 | 0.8970.871–0.920 | 0.7980.760–0.834 | 0.0990.063–0.132 | <0.001 |
| Brier scoreMean squared error of the probability; lower is better | 0.0150.013–0.019 | 0.0300.025–0.035 | -0.015-0.019–-0.010 | <0.001 |
| Calibration errorGap between stated probability and flood rate; lower is better | 0.0160.013–0.021 | 0.0370.031–0.043 | -0.021-0.026–-0.015 | <0.001 |
95% intervals from 1,000 bootstrap resamples of whole groups (a tile’s before and after together); difference is TideWatch minus flexiclass, its p-value 2-sided. A shaded cell is the better model where the difference is significant at 5%. Decisions at a probability of 0.5.
McNemar’s exact test on the decisions: TideWatch alone right on 86 images, flexiclass alone on 32; p = <0.001. flexiclass’s own reliability, read as a probability, has a calibration error of 0.066.
TideWatch
| Called flooded | Called dry | |
|---|---|---|
| Flooded | 285 | 22 |
| Dry | 36 | 2,691 |
flexiclass
| Called flooded | Called dry | |
|---|---|---|
| Flooded | 253 | 54 |
| Dry | 58 | 2,669 |
Accuracy by subset
| Level | Images | Flooded | TideWatch | flexiclass |
|---|---|---|---|---|
| AIDER | 958 | 78 | 100.0% | 99.6% |
| The AIFloodSense dataset | 95 | 95 | 100.0% | 100.0% |
| FloodNet | 475 | 52 | 99.2% | 97.5% |
| LADI | 1,049 | 82 | 94.9% | 94.3% |
| RescueNet | 457 | 0 | 100.0% | 92.1% |
| Level | Images | Flooded | TideWatch | flexiclass |
|---|---|---|---|---|
| collapsed building | 76 | 0 | 100.0% | 100.0% |
| fire | 78 | 0 | 100.0% | 98.7% |
| flooded | 229 | 229 | 90.4% | 76.9% |
| flooded areas | 78 | 78 | 100.0% | 98.7% |
| normal | 654 | 0 | 100.0% | 100.0% |
| not flooded | 1,847 | 0 | 98.1% | 97.0% |
| traffic incident | 72 | 0 | 100.0% | 97.2% |
On data TideWatch never saw
| Left out | Images | TideWatch accuracy | flexiclass accuracy | TideWatch ROC-AUC | flexiclass ROC-AUC |
|---|---|---|---|---|---|
| The AIFloodSense dataset | 470 | 98.7% | 96.4% | — | — |
| AIDER | 6,433 | 95.8% | 99.5% | 0.999 | 1.000 |
| FloodNet | 2,343 | 96.8% | 96.6% | 0.990 | 0.988 |
| LADI | 9,969 | 92.3% | 90.5% | 0.928 | 0.898 |
| RescueNet | 4,494 | 99.8% | 91.0% | — | — |
All held-out events together
| Metric | TideWatch | flexiclass | Difference | p |
|---|---|---|---|---|
| ROC-AUCRanking: chance a flooded image scores above a dry one | 0.9470.942–0.953 | 0.9380.933–0.943 | 0.0090.005–0.014 | <0.001 |
| PR-AUCAverage precision across recall | 0.8720.862–0.884 | 0.8170.805–0.829 | 0.0560.047–0.064 | <0.001 |
| AccuracyShare of images answered right | 0.9530.950–0.955 | 0.9370.934–0.940 | 0.0150.012–0.018 | <0.001 |
| Balanced accuracyMean of recall and specificity | 0.8820.874–0.890 | 0.8380.831–0.847 | 0.0430.037–0.050 | <0.001 |
| PrecisionOf the images called flooded, the share that are | 0.8260.812–0.840 | 0.7740.759–0.790 | 0.0510.034–0.069 | <0.001 |
| RecallOf the flooded images, the share found | 0.7870.773–0.803 | 0.7060.692–0.723 | 0.0810.068–0.094 | <0.001 |
| SpecificityOf the dry images, the share called dry | 0.9760.974–0.978 | 0.9700.968–0.973 | 0.0060.003–0.009 | <0.001 |
| F1Precision and recall, equally weighted | 0.8060.795–0.817 | 0.7390.727–0.751 | 0.0670.056–0.078 | <0.001 |
| F2Recall weighted 4 times precision: a missed flood costs more | 0.7950.782–0.808 | 0.7190.706–0.734 | 0.0760.064–0.087 | <0.001 |
| MCCCorrelation of answer and truth, −1 to 1 | 0.7790.767–0.792 | 0.7040.691–0.718 | 0.0750.062–0.088 | <0.001 |
| Brier scoreMean squared error of the probability; lower is better | 0.0380.036–0.040 | 0.0500.048–0.053 | -0.012-0.014–-0.010 | <0.001 |
| Calibration errorGap between stated probability and flood rate; lower is better | 0.0250.022–0.027 | 0.0270.025–0.030 | -0.002-0.006–0.001 | 0.188 |
Augmentation, in training and at test time
| TideWatch | Accuracy | ROC-AUC | F1 | Brier | Never seen: accuracy | Never seen: ROC-AUC |
|---|---|---|---|---|---|---|
| TideWatch, shipped | 0.9810.976–0.985 | 0.9940.989–0.997 | 0.9080.884–0.928 | 0.015 | 0.953 | 0.947 |
| Shipped + test-time augmentation | 0.9830.978–0.987 | 0.9950.990–0.997 | 0.9150.892–0.935 | 0.014 | 0.962 | 0.949 |
Test-time augmentation averages the original with 3 label-safe views (flips, a small rotation or a crop), so it costs 4 encodes per image. Augmented training adds 3 transformed copies of every training image. Shaded: the best on the clean test split.
On data never seen, test-time augmentation against none, the same head and images: accuracy +0.9 points (95% interval +0.8 to +1.1, p = <0.001), ROC-AUC +0.002 (p = <0.001); McNemar: with it alone right on 280 images, without it on 61, p = <0.001. The service scores aerial images this way by default (TIDEWATCH_TTA_MODELS).
Robustness: the same images, transformed
TideWatch flexiclass
| Transform | Images | TideWatch | flexiclass |
|---|---|---|---|
| Clean images | 554 | 96.9%AUC 0.997 · flipped 0.0% | 92.4%AUC 0.976 · flipped 0.0% |
| Darker (60%) | 554 | 96.2%AUC 0.996 · flipped 1.1% | 91.5%AUC 0.981 · flipped 2.7% |
| Brighter (140%) | 554 | 96.2%AUC 0.996 · flipped 1.8% | 90.6%AUC 0.972 · flipped 3.2% |
| Lower contrast | 554 | 95.8%AUC 0.996 · flipped 1.4% | 91.7%AUC 0.979 · flipped 2.9% |
| Desaturated | 554 | 95.5%AUC 0.996 · flipped 2.2% | 91.3%AUC 0.974 · flipped 2.2% |
| Hue shift ⚠ | 554 | 90.4%AUC 0.991 · flipped 6.9% | 87.7%AUC 0.962 · flipped 5.1% |
| Colour jitter | 554 | 95.5%AUC 0.996 · flipped 1.4% | 91.9%AUC 0.977 · flipped 2.0% |
| Grayscale | 554 | 90.8%AUC 0.991 · flipped 6.9% | 86.3%AUC 0.933 · flipped 7.9% |
| Equalized | 554 | 93.5%AUC 0.994 · flipped 4.2% | 91.2%AUC 0.959 · flipped 4.2% |
| Autocontrast | 554 | 97.1%AUC 0.997 · flipped 0.5% | 93.0%AUC 0.978 · flipped 1.3% |
| Posterized (3 bits) | 554 | 93.5%AUC 0.994 · flipped 3.8% | 87.7%AUC 0.941 · flipped 7.2% |
| Solarized ⚠ | 554 | 93.7%AUC 0.993 · flipped 3.6% | 89.4%AUC 0.952 · flipped 6.0% |
| Inverted ⚠ | 554 | 83.6%AUC 0.977 · flipped 14.1% | 83.4%AUC 0.911 · flipped 12.6% |
| Gaussian blur | 554 | 95.3%AUC 0.996 · flipped 2.0% | 91.5%AUC 0.968 · flipped 3.8% |
| Sharpened | 554 | 96.2%AUC 0.997 · flipped 1.1% | 91.7%AUC 0.969 · flipped 2.2% |
| Gaussian noise | 554 | 94.6%AUC 0.993 · flipped 2.7% | 90.6%AUC 0.969 · flipped 2.9% |
| JPEG quality 20 | 554 | 95.3%AUC 0.996 · flipped 2.0% | 91.7%AUC 0.971 · flipped 2.5% |
| Mirrored | 554 | 97.3%AUC 0.997 · flipped 0.7% | 92.2%AUC 0.976 · flipped 2.3% |
| Rotated 20° | 554 | 94.8%AUC 0.996 · flipped 2.9% | 91.2%AUC 0.969 · flipped 2.3% |
| Cropped to half | 554 | 96.0%AUC 0.993 · flipped 2.7% | 90.6%AUC 0.965 · flipped 4.3% |
| Shifted | 554 | 96.2%AUC 0.997 · flipped 2.5% | 91.9%AUC 0.977 · flipped 2.7% |
| Sheared | 554 | 95.7%AUC 0.997 · flipped 1.6% | 91.3%AUC 0.975 · flipped 1.8% |
| Zoomed out | 554 | 93.1%AUC 0.996 · flipped 4.2% | 91.2%AUC 0.970 · flipped 3.4% |
| Perspective | 554 | 94.6%AUC 0.997 · flipped 2.7% | 91.0%AUC 0.970 · flipped 4.0% |
| Elastic warp | 554 | 95.8%AUC 0.996 · flipped 1.4% | 92.6%AUC 0.969 · flipped 3.1% |
| Patch erased | 554 | 93.3%AUC 0.997 · flipped 4.0% | 91.3%AUC 0.971 · flipped 3.2% |
| RandAugment | 554 | 95.5%AUC 0.996 · flipped 2.9% | 90.3%AUC 0.967 · flipped 3.2% |
| AugMix | 554 | 95.7%AUC 0.994 · flipped 1.6% | 91.7%AUC 0.972 · flipped 2.5% |
150 flooded and 150 dry test images, each under 27 torchvision transforms at a fixed strength. Flipped: the share of answers that changed from the clean image’s. ⚠ marks a transform that can change what the picture shows: brown floodwater stops looking like water.
Satellite
Satellite: 594 test images.
Held-out satellite tiles, never used to train TideWatch; the same files sent to both.
TideWatch flexiclass
Every metric
| Metric | TideWatch | flexiclass | Difference | p |
|---|---|---|---|---|
| ROC-AUCRanking: chance a flooded image scores above a dry one | 0.9990.998–1.000 | 0.8940.843–0.938 | 0.1050.062–0.157 | <0.001 |
| PR-AUCAverage precision across recall | 0.9990.997–1.000 | 0.8980.859–0.932 | 0.1010.067–0.140 | <0.001 |
| AccuracyShare of images answered right | 0.9800.966–0.992 | 0.8330.792–0.867 | 0.1460.113–0.184 | <0.001 |
| Balanced accuracyMean of recall and specificity | 0.9740.957–0.990 | 0.7920.753–0.828 | 0.1830.149–0.221 | <0.001 |
| PrecisionOf the images called flooded, the share that are | 1.0001.000–1.000 | 0.9790.951–1.000 | 0.0210.000–0.049 | 0.104 |
| RecallOf the flooded images, the share found | 0.9490.915–0.979 | 0.5910.515–0.664 | 0.3570.288–0.433 | <0.001 |
| SpecificityOf the dry images, the share called dry | 1.0001.000–1.000 | 0.9920.980–1.000 | 0.0080.000–0.020 | 0.104 |
| F1Precision and recall, equally weighted | 0.9740.955–0.989 | 0.7370.675–0.792 | 0.2360.184–0.298 | <0.001 |
| F2Recall weighted 4 times precision: a missed flood costs more | 0.9590.930–0.983 | 0.6420.569–0.710 | 0.3160.253–0.390 | <0.001 |
| MCCCorrelation of answer and truth, −1 to 1 | 0.9580.930–0.982 | 0.6690.603–0.729 | 0.2900.231–0.352 | <0.001 |
| Brier scoreMean squared error of the probability; lower is better | 0.0170.008–0.029 | 0.1320.103–0.166 | -0.114-0.148–-0.086 | <0.001 |
| Calibration errorGap between stated probability and flood rate; lower is better | 0.0220.011–0.036 | 0.1250.094–0.165 | -0.103-0.140–-0.074 | <0.001 |
95% intervals from 1,000 bootstrap resamples of whole groups (a tile’s before and after together); difference is TideWatch minus flexiclass, its p-value 2-sided. A shaded cell is the better model where the difference is significant at 5%. Decisions at a probability of 0.5.
McNemar’s exact test on the decisions: TideWatch alone right on 87 images, flexiclass alone on 0; p = <0.001. flexiclass’s own reliability, read as a probability, has a calibration error of 0.101.
TideWatch
| Called flooded | Called dry | |
|---|---|---|
| Flooded | 223 | 12 |
| Dry | 0 | 359 |
flexiclass
| Called flooded | Called dry | |
|---|---|---|
| Flooded | 139 | 96 |
| Dry | 3 | 356 |
Accuracy by subset
| Level | Images | Flooded | TideWatch | flexiclass |
|---|---|---|---|---|
| SpaceNet8 | 202 | 25 | 95.5% | 91.1% |
| xBD | 392 | 210 | 99.2% | 79.3% |
| Level | Images | Flooded | TideWatch | flexiclass |
|---|---|---|---|---|
| post flood | 265 | 207 | 96.2% | 74.0% |
| post flood crops | 28 | 28 | 92.9% | 3.6% |
| pre flood | 301 | 0 | 100.0% | 99.0% |
| Level | Images | Flooded | TideWatch | flexiclass |
|---|---|---|---|---|
| germany-flooding | 54 | 13 | 90.7% | 79.6% |
| Hurricane Florence, 2018 | 82 | 45 | 100.0% | 61.0% |
| Hurricane Harvey, 2017 | 112 | 64 | 97.3% | 78.6% |
| louisiana-flooding | 148 | 12 | 97.3% | 95.3% |
| Midwest floods, 2019 | 46 | 25 | 100.0% | 84.8% |
| Nepal floods, 2017 | 152 | 76 | 100.0% | 88.2% |
On data TideWatch never saw
| Left out | Images | TideWatch accuracy | flexiclass accuracy | TideWatch ROC-AUC | flexiclass ROC-AUC |
|---|---|---|---|---|---|
| xBD | 2,630 | 81.2% | 79.8% | 0.875 | 0.901 |
| Hurricane Florence, 2018 | 574 | 78.6% | 54.9% | 0.993 | 0.806 |
| Hurricane Harvey, 2017 | 726 | 95.0% | 87.1% | 0.999 | 0.926 |
| Midwest floods, 2019 | 306 | 92.5% | 80.4% | 0.981 | 0.884 |
| Nepal floods, 2017 | 1,024 | 96.3% | 88.5% | 0.999 | 0.981 |
| SpaceNet8 | 1,377 | 70.4% | 90.0% | 0.908 | 0.807 |
| germany-flooding | 371 | 86.3% | 81.4% | 0.949 | 0.721 |
| louisiana-flooding | 1,006 | 78.7% | 93.1% | 0.945 | 0.888 |
All held-out events together
| Metric | TideWatch | flexiclass | Difference | p |
|---|---|---|---|---|
| ROC-AUCRanking: chance a flooded image scores above a dry one | 0.8970.882–0.911 | 0.8880.870–0.906 | 0.009-0.002–0.018 | 0.088 |
| PR-AUCAverage precision across recall | 0.8690.855–0.885 | 0.8920.878–0.906 | -0.022-0.035–-0.011 | <0.001 |
| AccuracyShare of images answered right | 0.8270.814–0.839 | 0.8330.819–0.848 | -0.006-0.020–0.008 | 0.402 |
| Balanced accuracyMean of recall and specificity | 0.8270.813–0.841 | 0.7910.777–0.808 | 0.0350.022–0.048 | <0.001 |
| PrecisionOf the images called flooded, the share that are | 0.7580.740–0.776 | 0.9740.963–0.984 | -0.216-0.236–-0.196 | <0.001 |
| RecallOf the flooded images, the share found | 0.8250.799–0.850 | 0.5930.565–0.625 | 0.2320.210–0.254 | <0.001 |
| SpecificityOf the dry images, the share called dry | 0.8280.816–0.841 | 0.9900.985–0.994 | -0.162-0.174–-0.149 | <0.001 |
| F1Precision and recall, equally weighted | 0.7900.773–0.806 | 0.7370.715–0.762 | 0.0530.031–0.073 | <0.001 |
| F2Recall weighted 4 times precision: a missed flood costs more | 0.8110.790–0.832 | 0.6430.617–0.673 | 0.1680.145–0.188 | <0.001 |
| MCCCorrelation of answer and truth, −1 to 1 | 0.6450.618–0.671 | 0.6670.642–0.693 | -0.022-0.047–0.004 | 0.092 |
| Brier scoreMean squared error of the probability; lower is better | 0.1490.137–0.161 | 0.1340.122–0.145 | 0.0150.004–0.026 | 0.006 |
| Calibration errorGap between stated probability and flood rate; lower is better | 0.1370.126–0.149 | 0.1230.109–0.135 | 0.0150.001–0.028 | 0.042 |
Augmentation, in training and at test time
| TideWatch | Accuracy | ROC-AUC | F1 | Brier | Never seen: accuracy | Never seen: ROC-AUC |
|---|---|---|---|---|---|---|
| TideWatch, shipped | 0.9800.966–0.992 | 0.9990.998–1.000 | 0.9740.955–0.989 | 0.017 | 0.827 | 0.897 |
| Shipped + test-time augmentation | 0.9800.965–0.992 | 0.9990.998–1.000 | 0.9740.954–0.990 | 0.016 | 0.835 | 0.918 |
Test-time augmentation averages the original with 3 label-safe views (flips, a small rotation or a crop), so it costs 4 encodes per image. Augmented training adds 3 transformed copies of every training image. Shaded: the best on the clean test split.
On data never seen, test-time augmentation against none, the same head and images: accuracy +0.8 points (95% interval +0.1 to +1.4, p = 0.016), ROC-AUC +0.021 (p = <0.001); McNemar: with it alone right on 700 images, without it on 576, p = <0.001. The service scores satellite images this way by default (TIDEWATCH_TTA_MODELS).
Robustness: the same images, transformed
TideWatch flexiclass
| Transform | Images | TideWatch | flexiclass |
|---|---|---|---|
| Clean images | 375 | 97.1%AUC 0.999 · flipped 0.0% | 79.7%AUC 0.904 · flipped 0.0% |
| Darker (60%) | 375 | 96.5%AUC 0.999 · flipped 1.1% | 75.5%AUC 0.875 · flipped 6.9% |
| Brighter (140%) | 375 | 97.3%AUC 0.998 · flipped 1.3% | 80.8%AUC 0.902 · flipped 5.3% |
| Lower contrast | 375 | 96.0%AUC 0.999 · flipped 1.6% | 77.6%AUC 0.896 · flipped 8.0% |
| Desaturated | 375 | 95.5%AUC 0.998 · flipped 2.1% | 76.3%AUC 0.892 · flipped 7.7% |
| Hue shift ⚠ | 375 | 94.9%AUC 0.986 · flipped 2.7% | 72.3%AUC 0.877 · flipped 14.9% |
| Colour jitter | 375 | 96.5%AUC 0.998 · flipped 1.1% | 79.7%AUC 0.897 · flipped 4.3% |
| Grayscale | 375 | 93.9%AUC 0.989 · flipped 3.7% | 67.7%AUC 0.845 · flipped 15.2% |
| Equalized | 375 | 96.8%AUC 0.994 · flipped 3.5% | 71.5%AUC 0.837 · flipped 18.4% |
| Autocontrast | 375 | 97.1%AUC 0.999 · flipped 1.6% | 80.0%AUC 0.908 · flipped 5.1% |
| Posterized (3 bits) | 375 | 95.7%AUC 0.995 · flipped 4.5% | 66.7%AUC 0.753 · flipped 18.4% |
| Solarized ⚠ | 375 | 97.1%AUC 0.994 · flipped 3.2% | 79.2%AUC 0.889 · flipped 5.3% |
| Inverted ⚠ | 375 | 84.5%AUC 0.920 · flipped 14.1% | 58.1%AUC 0.713 · flipped 26.9% |
| Gaussian blur | 375 | 95.2%AUC 0.993 · flipped 5.6% | 77.9%AUC 0.889 · flipped 8.8% |
| Sharpened | 375 | 97.1%AUC 0.999 · flipped 0.5% | 80.0%AUC 0.903 · flipped 6.1% |
| Gaussian noise | 375 | 95.2%AUC 0.997 · flipped 1.9% | 76.0%AUC 0.881 · flipped 8.0% |
| JPEG quality 20 | 375 | 93.9%AUC 0.997 · flipped 3.7% | 76.5%AUC 0.891 · flipped 9.1% |
| Mirrored | 375 | 96.8%AUC 0.999 · flipped 0.8% | 79.5%AUC 0.902 · flipped 7.7% |
| Rotated 20° | 375 | 96.0%AUC 0.997 · flipped 2.1% | 78.4%AUC 0.902 · flipped 9.3% |
| Cropped to half | 375 | 95.5%AUC 0.983 · flipped 4.8% | 76.3%AUC 0.882 · flipped 9.3% |
| Shifted | 375 | 97.1%AUC 0.997 · flipped 1.6% | 77.9%AUC 0.896 · flipped 6.7% |
| Sheared | 375 | 95.7%AUC 0.998 · flipped 1.9% | 78.1%AUC 0.885 · flipped 6.4% |
| Zoomed out | 375 | 96.5%AUC 0.998 · flipped 2.7% | 76.3%AUC 0.880 · flipped 7.7% |
| Perspective | 375 | 96.0%AUC 0.998 · flipped 2.1% | 78.1%AUC 0.899 · flipped 8.0% |
| Elastic warp | 375 | 97.1%AUC 0.997 · flipped 1.6% | 78.4%AUC 0.906 · flipped 6.1% |
| Patch erased | 375 | 95.7%AUC 0.998 · flipped 1.9% | 75.5%AUC 0.891 · flipped 6.4% |
| RandAugment | 375 | 95.5%AUC 0.994 · flipped 3.7% | 76.5%AUC 0.888 · flipped 9.1% |
| AugMix | 375 | 96.5%AUC 0.997 · flipped 2.7% | 75.7%AUC 0.872 · flipped 9.3% |
150 flooded and 150 dry test images, each under 27 torchvision transforms at a fixed strength. Flipped: the share of answers that changed from the clean image’s. ⚠ marks a transform that can change what the picture shows: brown floodwater stops looking like water.
Method
How this was measured.
- Same images, never trained on. The test splits held back before TideWatch was trained, split by group so near-duplicates and a tile’s before and after never straddle train and test.
- TideWatch: the shipped heads on frozen Perception Encoder (PE-Core-G) vectors, each threshold calibrated on datasets left out, exactly as the API answers. flexiclass: its image endpoint at low effort through its own client, Gemma 4 26B, zero-shot.
- flexiclass’s question, the condition and its negation as its guide asks: “The area is flooded” / “The area is not flooded”. Aerial: “Look at this photo, taken from a drone or helicopter at an angle. Based only on what is visible, is the area in it flooded?” Satellite: “Look at this satellite image, taken looking straight down at about 0.5 m per pixel. Based only on what is visible, is the area in it flooded?”
- Probabilities. TideWatch’s probability of a flood; flexiclass’s score for the flooded label. Both decide at 0.5.
- Never seen. TideWatch refitted without each dataset or flood event in turn, against flexiclass on the same images: the fair test against a model that never saw any of them.
- Uncertainty. 95% intervals and paired p-values from bootstrap resamples of whole groups; McNemar’s exact test on the decisions.
flexiclass is tuned for many questions, not this one; a trained model is expected to lead on its own task. The test deployment of flexiclass records submitted images for its evaluation.