Change detection identifies differences in the state of a landscape between two or more dates by comparing remote sensing images acquired at different times. Techniques range from simple image differencing (subtracting one date from another) to post-classification comparison (classifying each date independently and comparing the maps) to advanced time-series analysis that tracks continuous change trajectories. Reliable change detection requires that differences between images reflect actual surface change rather than artifacts from atmospheric conditions, sensor calibration, illumination geometry, or phenological cycles. This makes preprocessing and careful image selection critical.
Remote sensing's unique power is not just mapping what is on Earth's surface at one moment, but tracking how it changes over time. With archives stretching back to 1972 (Landsat), analysts can reconstruct decades of landscape transformation -- deforestation, urban expansion, glacial retreat, coastal erosion -- at scales from individual parcels to entire continents.
The simplest approach is image differencing: subtract one date's spectral values (or derived index like NDVI) from another. Pixels with large differences are flagged as changed. This is fast and intuitive but sensitive to noise and requires very careful preprocessing to ensure that differences reflect actual change. Post-classification comparison independently classifies each date and compares the resulting maps, producing a from-to change matrix (e.g., forest-to-agriculture, agriculture-to-urban). This provides thematic change information but accumulates classification errors from both dates.
Modern time-series approaches exploit the growing density of satellite observations. Algorithms like LandTrendr fit piecewise linear models to each pixel's spectral trajectory over decades, identifying break points that correspond to disturbance events. BFAST decomposes time series into trend, seasonal, and residual components, detecting both abrupt breaks and gradual trends. Google Earth Engine and similar cloud platforms make it feasible to process thousands of images per pixel, transforming change detection from a bi-temporal exercise into continuous monitoring.
The persistent challenge is separating real change from confounding factors: phenological cycles, atmospheric variability, sensor degradation, and registration errors. Successful change detection demands not just algorithms but careful experimental design -- selecting appropriate dates, ensuring comparable preprocessing, understanding the landscape's natural variability, and validating results against independent data. A change map without accuracy assessment is an assertion, not evidence.