Image analysis is the stage where corrected imagery becomes usable information: vegetation condition, water extent, urban texture, land cover class and the amount of change over time. The work starts with preprocessing, continues through spectral indices and classification algorithms, and ends with accuracy assessment; without that last step the outputs are merely attractive pictures rather than a basis for decisions. The method is chosen from the type of feature, the resolution of the imagery, the availability of ground reference data and the final question the project has to answer. Outputs are delivered as spatial layers, statistical tables and thematic maps, prepared for entry into GIS systems.
Preprocessing and Analysis Readiness
Before any analysis the scenes must be co-registered, meaning that corresponding pixels in different dates align to within a fraction of a pixel. Clipping to the study boundary, masking cloud and shadow, removing saturated pixels and applying consistent scale factors all belong to this step as well.
For large areas several scenes are mosaicked and the radiometric differences between them are adjusted; without that adjustment the scene boundaries appear as false lines in the final map. In multi-date work, converting every date to surface reflectance and resampling to one common grid is the precondition for a valid comparison.
- Accurate co-registration of multitemporal imagery and a check of residual shift.
- Clipping to the study boundary and one consistent pixel size across all layers.
- Cloud, shadow, snow and invalid-pixel masking applied the same way on every date.
- Radiometric adjustment of mosaic seams and conversion to surface reflectance.
Spectral Indices and Their Interpretation
Spectral indices are simple band ratios that reduce the effect of overall surface brightness and highlight differences in spectral behaviour. The Normalized Difference Vegetation Index is built from the difference of near-infrared and red, the Normalized Difference Water Index from green and near-infrared, and the Normalized Difference Built-up Index from shortwave-infrared and near-infrared.
Soil-adjusted indices, which introduce a soil adjustment factor, behave better over sparse vegetation and bright soils, while the enhanced vegetation index adds a blue band term and is less sensitive to atmospheric and background effects. For fire monitoring, the normalized burn ratio contrasts shortwave-infrared and near-infrared behaviour in burnt areas.
- Vegetation index: near-infrared and red difference for green biomass.
- Water index: separating water bodies from soil and vegetation using green and near-infrared.
- Built-up index: distinguishing sealed surfaces with the shortwave-infrared band.
- Soil-adjusted index: use in arid regions and sparse cover.
- Burn ratio: fire monitoring and burned-area mapping.
Supervised and Unsupervised Classification
In supervised classification, training samples are collected from field visits, higher-resolution imagery or visual interpretation, and the algorithm learns class boundaries from them. Random forest and support vector machines tend to give more stable results on multispectral imagery, while classical approaches such as maximum likelihood are simpler and easier to explain.
In unsupervised classification, algorithms such as k-means or ISODATA derive spectral groups from the data itself, and the classes are labelled semantically after the run. This is useful for exploring the structure of the cover types, but the number and boundaries of classes cannot be fixed in advance.
In both cases training and test samples must be kept separate, so that accuracy is measured on data the algorithm has not seen. The number of samples per class should match the extent and internal variety of that class; small classes with few samples usually produce the weakest results.
- Balanced training samples kept independent of the test samples.
- Using original bands, indices and auxiliary layers together as input features.
- Avoiding overfitting by limiting the number of features and tuning parameters.
Object-Based Analysis
In high-resolution imagery, pixel-by-pixel classification produces fragmented and noisy results because the internal variation of a single feature exceeds the difference between features. Object-based image analysis first segments the image into regions that are homogeneous in spectrum and space, then classifies each segment using its spectral, geometric, textural and contextual properties.
Segmentation scale, shape weight and compactness are the key parameters and must be set from the size of the target features rather than left at a default. The approach suits building units, agricultural parcels, linear features such as roads and channels, and irregular natural boundaries better than pixel-based classification does.
- Set segmentation scale from the smallest target feature.
- Use geometric properties such as area, compactness and length-to-width ratio.
- Use texture and neighbourhood relations to separate spectrally similar classes.
Accuracy Assessment
The confusion matrix is the main assessment tool and yields overall accuracy, the kappa coefficient, producer’s accuracy, user’s accuracy and per-class balance measures. User’s accuracy answers how many pixels labelled as a class really belong to it, and producer’s accuracy answers the reverse question; reporting the two separately is essential.
Validation samples must come from an independent reference database and be drawn by stratified sampling so that small classes are represented as well. If accuracy falls short of the project requirement, the correction path runs through additional training data, revised class definitions, a different feature set or splitting classes that were merged.
- Report the confusion matrix with per-class measures, not only overall accuracy.
- Use stratified sampling independent of the training samples.
- Record ground control points with satellite positioning for field validation.
Change Detection and Time Series
In post-classification comparison, land cover maps from two dates are compared through a transition matrix, which makes the type and direction of change explicit; classification error, however, is counted twice in the result. Image-based methods, such as index differencing or change vector analysis in band space, are more sensitive to small changes but make the type of change harder to interpret.
In time-series monitoring, indices are smoothed to remove short-term fluctuation before phenological metrics such as start, peak and end of season and long-term trend are extracted. Non-parametric statistical tests are better suited to declaring a trend significant, and separating gradual change from abrupt events is part of the interpretation.
- Consistent preprocessing and acquisition geometry across all dates.
- Smoothing the time series and removing outliers before trend extraction.
- Significance testing to separate real trend from natural fluctuation.
- Reading the result alongside auxiliary data to tell real change from processing error.
Integration with GIS Layers
Image analysis reaches its full value when combined with other spatial layers. Zonal statistics, meaning the mean and spread of indices inside each polygon, allow comparison across management units, agricultural parcels or catchment areas.
Raster-vector integration, joining attribute tables to spatial features and overlaying layers form the basis of spatial analysis and suitability modelling. The resulting layers must be delivered with a consistent coordinate reference system, attribute structure and metadata so they can be published in a Web GIS without rework.
Summary
Image analysis is a chain in which the correctness of each step determines the credibility of the next; precise indices on poorly co-registered imagery, or classification without validation, do not produce dependable results. Establishing the question, the classes, the accuracy acceptance threshold and the final delivery form before processing begins avoids rework.
The output of this stage is validated spatial layers, statistical tables and a report covering method and limitations, which can serve as the basis for periodic monitoring or as input to machine-learning models in later steps.