Spatial Machine Learning

Audit ecological observations for spatial clustering, sampling bias, and mismatch with the intended prediction domain.

Align raster predictor layers to a common grid while selecting an appropriate resampling method for each data type.

Extract raster predictor values at point or polygon observations while preserving observation identifiers and geometry.

Merge tabular field records with extracted spatial predictors and handle missing, out-of-bounds, or censored observations in Python.

Assemble a modeling table that retains coordinates and spatial group identifiers separately from the predictor features.

Explain how model complexity, bias, variance, and sample size affect generalization in ecological prediction.

Select regression or classification metrics that match a stated ecological prediction objective and its error costs.

Construct a scikit-learn preprocessing pipeline in which imputation, scaling, and categorical encoding are learned only from training data.

Fit a regularized linear or logistic model as an interpretable non-spatial baseline.

Explain how decision-tree split rules and stopping controls influence underfitting and overfitting.

Implement a random forest regression or classification model in a reproducible scikit-learn pipeline.

Implement a gradient-boosted tree model and identify its principal complexity controls.

Handle an imbalanced ecological classification problem using class weighting or resampling confined to each training fold.

Distinguish spatial interpolation, environmental-covariate mapping, and extrapolation to new geographic domains.

Estimate the scale of spatial dependence in observations or residuals using a variogram or spatial autocorrelation statistic.

Explain why spatial dependence and nearby duplicate information can make random cross-validation optimistic.

Choose spatial block size and orientation using ecological scale, sampling geometry, and estimated dependence range.

Implement reproducible spatially grouped cross-validation splits without allowing observations from the same spatial unit, site, or buffer to cross folds.

Fit an ordinary-kriging benchmark for a continuous response using training data only.

Compare random and spatial cross-validation estimates and interpret the gap between them.

Construct an outer spatial assessment loop and an inner spatial tuning loop so that final test regions remain isolated.

Derive ecologically meaningful terrain, distance, and neighborhood predictors without using information from the response.

Choose neighborhood and focal-window sizes that correspond to the ecological process and raster resolution.

Audit a predictor set for spatial, target, temporal, and sampling-design leakage before model fitting.

Perform feature selection entirely within spatial cross-validation folds.

Tune random-forest or gradient-boosting hyperparameters against a spatially valid objective.

Evaluate how prediction resolution or support affects model performance and map interpretation.

Generate a continuous prediction raster in memory-safe windows while preserving grid metadata and NoData behavior.

Compare machine-learning and kriging predictions on the same spatial holdout data.

Calculate held-out permutation importance without allowing spatial test data to influence model fitting.

Use partial-dependence or accumulated-local-effects plots to describe a fitted ecological response relationship.

Map spatially held-out residuals and diagnose remaining geographic structure in model errors.

Calculate environmental similarity or an area-of-applicability measure for a prediction domain.

Generate predictive intervals for continuous outcomes using a quantile-capable tree model fitted only on training data.

Assess the calibration and spatial coverage of predictive intervals on held-out regions.

Communicate model accuracy, extrapolation risk, uncertainty, and ecological limitations together on a prediction map.

Explain how spectral, spatial, radiometric, and temporal resolution constrain an ecological satellite-imagery application.

Select a satellite sensor, bands, and acquisition period that match a stated ecological phenomenon.

Distinguish top-of-atmosphere measurements from surface reflectance and identify the preprocessing needed for modeling.

Query a spatiotemporal asset catalog and load cloud-optimized satellite imagery in Python.

Apply cloud, cloud-shadow, and invalid-pixel masks using quality-assurance information.

Create a cloud-reduced temporal composite without treating missing observations as valid reflectance.

Compute spectral indices whose band combinations represent a stated ecological property.

Derive simple seasonal or phenological features from a satellite time series.

Design a class taxonomy and reference-label protocol that are spatially and temporally consistent with the imagery.

Diagnose mixed-pixel and label-support problems at class boundaries and heterogeneous ecological sites.

Choose between pixel-based and object-based classification for a stated mapping objective.

Derive texture and multiscale contextual features at scales justified by the ecological classes.

Create spatial train, validation, and test partitions by site or scene before sampling pixels or image objects.

Train a class-weighted random-forest or boosted-tree image classifier using spatial validation.

Diagnose class-specific errors and probability calibration on spatially held-out imagery.

Produce class and probability rasters and use mapped errors to identify systematic failure modes.

Represent multispectral image chips and labels as tensors with explicit channel, spatial, and batch dimensions.

Explain how convolution, receptive field, activation, pooling, and feature maps create spatial representations.

Assign source areas to spatial partitions before extracting chips so that adjacent or overlapping chips cannot cross data splits.

Implement a PyTorch dataset and data loader for georeferenced image chips.

Normalize and augment ecological imagery without introducing physically invalid values or corrupting labels.

Implement a compact convolutional neural network for image-chip classification.

Train a convolutional network with an optimizer while monitoring loss and spatially held-out performance across epochs.

Diagnose overfitting from learning curves and apply early stopping, weight decay, or dropout.

Choose between whole-chip classification and semantic segmentation based on the required ecological map output.

Select a pretrained convolutional backbone and adapt its input channels and normalization when multispectral imagery differs from RGB pretraining data.

Fine-tune a pretrained convolutional backbone for an ecological image-classification task.

Convert polygon or raster reference data into aligned semantic-segmentation target chips.

Select loss functions and evaluation metrics for an imbalanced semantic-segmentation problem.

Fine-tune an encoder-decoder segmentation model on spatially partitioned training data.

Run tiled segmentation inference and reduce seams or edge artifacts when mosaicking predictions.

Compare classical and deep-learning image classifiers in terms of accuracy, transferability, data needs, and computation.

Refactor a completed workflow into reusable, configuration-driven functions and test them on a small spatial subset.

Validate final raster outputs for coordinate system, transform, alignment, NoData handling, value ranges, and file metadata before release.

Write a concise model card describing intended use, validation geography, limitations, and uncertainty.

Assemble a leakage-resistant continuous-mapping portfolio project with a kriging baseline, spatial validation, uncertainty outputs, and explicit acceptance criteria.

Assemble a satellite or aerial-image classification portfolio project with spatial holdouts, class-specific evaluation, and explicit acceptance criteria.

Present predictions, uncertainty, applicability, and decision-relevant limitations in a professional project report.