earthbench-general-skill

name	earthbench-general-skill
description	Global static EarthBench skill for raster, image, geospatial, numeric, and choice-matching workflows.
allowed-tools	["ATI","ChangeOS","InstructSAM","MSCN","RemoteSAM","SAM2","SM3Det","analyze_hotspot_direction","apply_cloud_mask","average_ratio_exceeding_threshold","band_ratio","bboxes2centroids","calc_batch_fire_pixels","calc_batch_image_hotspot_percentage","calc_batch_image_hotspot_tif","calc_batch_image_max","calc_batch_image_mean","calc_batch_image_mean_mean","calc_batch_image_mean_threshold","calc_batch_image_sum","calc_threshold_value_mean","calculate_area","calculate_band_mean_by_condition","calculate_batch_nbr","calculate_batch_ndsi","calculate_batch_ndti","calculate_batch_ndvi","calculate_batch_ndwi","calculate_bbox_area","calculate_max_lst_by_ndvi","calculate_mean_lst_by_ndvi","calculate_multi_band_threshold_ratio","[Truncated]"]
metadata	{"benchmark":"EarthBench","generator":"Manus","baseline":"manus_plain_skill","generation_constraints":"earthbench_v2_folded_gold_no_instance_memory_v4","allowed_gold_input":"train124_firstarg_folded_v2","gold_trim_policy":"20260520_firstarg_shortnotice_repeatfold_v2","skill_injection_mode":"global_static","runtime_state_transitions":"disabled","date":"2026-05-20"}

EarthBench General Skill

File Listing and Modality Triage

Every task begins with discovering available data files:

1. List the data directory using get_filelist(dir_path="benchmark/data/question<ID>").
2. Classify files by naming convention:
   • Files containing _LST → Land Surface Temperature brightness temperature rasters.
   • Files containing _NDVI or _b4/_b5/_Red/_NIR → Vegetation reflectance bands.
   • Files containing _BT_31/_BT_32 → MODIS thermal bands for split-window LST.
   • Files containing _NBR/_NDSI/_NDTI/_NDWI → Pre-computed spectral indices.
   • Files with RGB or multi-band indicators → Detection/segmentation inputs.
   • Files with fire/hotspot indicators → Fire analysis inputs.
3. Determine the workflow family from the question keywords and available file types.

Raster Grouping

When multiple TIF files exist for a time series:

• Group files by date extracted from filenames (pattern: Region_YYYY-MM-DD_Band.tif).
• Process each temporal group with the same tool sequence.
• Collect per-date outputs before aggregation.
• When the gold shows repeated tool calls (e.g., multiple compute_tvdi or lst_single_channel), apply the same tool to each date's file set.

RGB/Image Detection and Segmentation

For tasks requiring object detection, segmentation, or change detection:

Object Detection

SM3Det(image_path="<path>", prompt="<object_category>")
RemoteSAM(image_path="<path>", prompt="<object_category>")

• Use SM3Det for small/medium object detection in remote sensing imagery.
• Use RemoteSAM for prompted segmentation of specific objects.
• Use SAM2 for general-purpose segmentation.
• Use InstructSAM when natural language instructions guide segmentation.

Change Detection

ChangeOS(image1_path="<before>", image2_path="<after>")

• Requires two temporally distinct images of the same region.
• Returns change mask or change statistics.

Post-Detection Processing

bboxes2centroids(bboxes=<detection_output>)
calculate_bbox_area(bboxes=<detection_output>)
centroid_distance_extremes(centroids=<centroid_list>)
count_connected_components(mask_path="<segmentation_output>")
calculate_area(mask_path="<segmentation_output>")

Spectral Index Workflows

Vegetation Index (NDVI)

calculate_batch_ndvi(red_paths=[<red_band_files>], nir_paths=[<nir_band_files>], output_dir="benchmark/out/question<ID>")

Water Index (NDWI)

calculate_batch_ndwi(green_paths=[<green_files>], nir_paths=[<nir_files>], output_dir="benchmark/out/question<ID>")

Snow Index (NDSI)

calculate_batch_ndsi(green_paths=[<green_files>], swir_paths=[<swir_files>], output_dir="benchmark/out/question<ID>")

Burn Ratio (NBR)

calculate_batch_nbr(nir_paths=[<nir_files>], swir_paths=[<swir_files>], output_dir="benchmark/out/question<ID>")

Tillage Index (NDTI)

calculate_batch_ndti(swir1_paths=[<swir1_files>], swir2_paths=[<swir2_files>], output_dir="benchmark/out/question<ID>")

Band Ratio (Generic)

band_ratio(numerator_path="<band_a>", denominator_path="<band_b>")

Drought Index (TVDI)

compute_tvdi(lst_path="<LST_file>", ndvi_path="<NDVI_file>", output_dir="benchmark/out/question<ID>")

• Requires paired LST and NDVI files for the same date.
• Process each date independently, then aggregate.

Land Surface Temperature (LST) Workflows

Single-Channel Method (Landsat 8 Band 10)

lst_single_channel(bt_path="<BT10_file>", red_path="<Red_file>", nir_path="<NIR_file>", output_dir="benchmark/out/question<ID>")

• Requires brightness temperature (Band 10) plus Red and NIR bands for NDVI-based emissivity.
• Produces LST in Kelvin.

Multi-Channel / Split-Window Method (MODIS Bands 31, 32)

lst_multi_channel(bt31_path="<BT31_file>", bt32_path="<BT32_file>", output_dir="benchmark/out/question<ID>")
split_window(bt31_path="<BT31_file>", bt32_path="<BT32_file>", output_dir="benchmark/out/question<ID>")

MODIS Day/Night LST

modis_day_night_lst(day_path="<day_file>", night_path="<night_file>")

Temperature-Emissivity Separation

temperature_emissivity_separation(thermal_paths=[<thermal_files>], output_dir="benchmark/out/question<ID>")

TTM LST

ttm_lst(thermal_path="<file>", output_dir="benchmark/out/question<ID>")

Kelvin to Celsius Conversion

kelvin_to_celsius(value=<kelvin_value>)

• Apply only when the question or choice explicitly uses Celsius.

Product Raster Aggregation

After generating per-date product rasters (TVDI, LST, NDVI, etc.), aggregate:

Spatial Averaging Per Image

calculate_tif_average(file_list=[<output_tifs>], output_dir="benchmark/out/question<ID>")
calc_batch_image_mean(file_list=[<output_tifs>])

• calculate_tif_average produces a single averaged raster from multiple inputs.
• calc_batch_image_mean returns a list of per-image spatial mean values.

Batch Statistics

calc_batch_image_max(file_list=[<tifs>])
calc_batch_image_sum(file_list=[<tifs>])
calc_batch_image_mean_mean(file_list=[<tifs>])
calc_batch_image_mean_threshold(file_list=[<tifs>], threshold=<value>)
calc_batch_image_hotspot_percentage(file_list=[<tifs>], threshold=<value>)
calc_batch_image_hotspot_tif(file_list=[<tifs>], threshold=<value>)
calc_batch_fire_pixels(file_list=[<tifs>])

Threshold and Exceedance Counting

calculate_threshold_ratio(image_paths=[<tifs>], threshold=<value>)
count_images_exceeding_threshold_ratio(image_paths=[<tifs>], threshold=<value>, ratio=<percentage>)
count_images_exceeding_mean_multiplier(image_paths=[<tifs>], multiplier=<value>)
count_above_threshold(values=[<list>], threshold=<value>)
average_ratio_exceeding_threshold(image_paths=[<tifs>], threshold=<value>)

Trend and Statistical Analysis

compute_linear_trend(y=[<annual_values>])
sens_slope(values=[<time_series>])
coefficient_of_variation(values=[<list>])
kurtosis(values=[<list>])
skewness(values=[<list>])
count_spikes_from_values(values=[<list>])

Arithmetic Helpers

difference(a=<value1>, b=<value2>)
subtract(a=<value1>, b=<value2>)
division(a=<numerator>, b=<denominator>)
multiply(a=<value1>, b=<value2>)
percentage_change(old=<value1>, new=<value2>)
mean(values=[<list>])
max_value_and_index(values=[<list>])
min_value_and_index(values=[<list>])
ceil_number(value=<float>)
image_division_mean(image_path="<tif>", divisor=<value>)
calculate_tif_difference(file1="<tif1>", file2="<tif2>")

Conditional Band Analysis

calculate_band_mean_by_condition(band_path="<tif>", condition_path="<mask_or_index>", threshold=<value>, above=<bool>)
calculate_mean_lst_by_ndvi(red_paths="<red>", nir_paths="<nir>", lst_path="<lst>", ndvi_threshold=<value>, above=<bool>)
calculate_max_lst_by_ndvi(red_paths="<red>", nir_paths="<nir>", lst_path="<lst>", ndvi_threshold=<value>, above=<bool>)
calculate_multi_band_threshold_ratio(image_paths=[<tifs>], thresholds=[<values>])

Fire and Hotspot Analysis

calc_batch_fire_pixels(file_list=[<tifs>])
calc_batch_image_hotspot_percentage(file_list=[<tifs>], threshold=<value>)
analyze_hotspot_direction(hotspot_map="<tif>")
create_fire_increase_map(before_path="<tif1>", after_path="<tif2>")

Cloud Masking

apply_cloud_mask(image_path="<tif>", cloud_mask_path="<mask>")

• Apply before any spectral index or LST computation if cloud mask files are present.

Water Turbidity

calculate_water_turbidity_ntu(red_path="<red_band>", nir_path="<nir_band>")

ATI (Apparent Thermal Inertia)

ATI(day_lst_path="<day>", night_lst_path="<night>", albedo_path="<albedo>")

Unit and Sign Guardrails

• LST outputs are in Kelvin unless explicitly converted. Check whether choices use K or C.
• TVDI is dimensionless [0, 1]. Higher values indicate drier conditions.
• NDVI is dimensionless [-1, 1]. Vegetation thresholds are typically 0.2–0.7.
• Percentage ratios from threshold tools are in percent (0–100), not fractions.
• difference(a, b) computes a - b. Verify operand order matches the question's phrasing (e.g., "how much cooler" = warmer - cooler).
• percentage_change(old, new) computes (new - old) / old * 100.
• Linear trend slope sign: negative = decreasing, positive = increasing.
• When the question asks "proportion below X", compute 100 - proportion_above_X using difference.

Choice Matching

EarthBench tasks typically require selecting from multiple-choice options (A/B/C/D):

1. Compute the numeric result using the appropriate workflow.
2. Compare to each choice:
   • Extract the numeric value or key phrase from each option.
   • Match the computed result to the closest option.
3. Tie-breaking rules:
   • If the computed value falls between two options, choose the nearest.
   • If the question asks about a trend direction, match direction first, then magnitude.
   • If choices differ in units, ensure unit consistency before comparing.
4. Return the letter only (A, B, C, or D).

Common Choice Patterns

• Trend questions: choices differ in direction (increasing/decreasing) and rate.
• Count questions: choices are integers; match exactly.
• Percentage questions: choices are percentages; match to nearest.
• Temperature difference questions: choices include units (K or C); verify unit match.

Final Answer Normalization

• Return exactly one letter (A, B, C, or D) for multiple-choice questions.
• For numeric-answer questions, return the number in the requested format.
• Do not include explanation, units, or extra text in the final answer.
• Do not include the option text, only the letter.

Recovery When Paths Conflict

If tool outputs, filenames, or choices produce unexpected results:

1. Re-check file grouping — ensure LST/NDVI/band files are correctly paired by date and region.
2. Verify tool arguments — confirm paths point to existing output files, not input files when aggregating.
3. Check intermediate values — use read_file or run_python_script to inspect raster statistics.
4. Try alternative LST method — if lst_single_channel produces unexpected values, check if split_window or lst_multi_channel is more appropriate for the sensor.
5. Recompute with explicit paths — if batch tools fail, process files individually and aggregate manually.
6. Validate against choices — if no choice matches, re-examine whether the question asks for "above" vs "below", "exceeding" vs "below", or requires a complement calculation.
7. Use run_python_script — as a last resort, write a Python script to manually read rasters and compute the answer using numpy/rasterio.

Workflow Family Quick Reference

Question Pattern	Typical Tool Chain
TVDI trend over years	get_filelist → compute_tvdi (per date) → calculate_tif_average (per year) → calc_batch_image_mean → compute_linear_trend
LST threshold proportion	get_filelist → lst_single_channel/lst_multi_channel → calculate_threshold_ratio → optionally difference
LST by vegetation class	get_filelist → lst_single_channel → calculate_mean_lst_by_ndvi (twice) → difference
Exceedance day count	get_filelist → LST tool (per date) → count_images_exceeding_threshold_ratio
Spike detection	get_filelist → index tool → calc_batch_image_mean → count_spikes_from_values
Object detection/counting	get_filelist → SM3Det/RemoteSAM → bboxes2centroids → counting/area tools
Change detection	get_filelist → ChangeOS → count_connected_components/calculate_area
Fire analysis	get_filelist → calc_batch_fire_pixels → aggregation → trend
Water turbidity	get_filelist → calculate_water_turbidity_ntu
NDVI/NDWI/NDSI batch	get_filelist → calculate_batch_ndvi/ndwi/ndsi → aggregation