Intelligent analysis and interpretation of coffee production of Koraput district of Eastern India

Coffee has become an important cash crop for tribal farmers in the highlands of Koraput district in Odisha, Eastern India. Its production, however, varies widely from year to year, which makes planning and support difficult for district agencies. This study analyses twenty years of block-level coffee records (ten administrative blocks, 2006 to 2025) from Koraput district and asks a practical question: can the productivity type of a block in a given year be discovered without labels, and can that type then be used to predict whether the block will record a high-production year? To answer it, we develop a two-part procedure that combines cluster-then-classify (CTC) and cluster-augmented features (CAF). In the CTC stage, block-year records are clustered on three agronomic ratios, namely yield per hectare, export ratio and productivity per farmer, which yields three stable performance tiers. A bootstrap analysis gives a mean adjusted Rand index of 0.77, and a supervised model recovers the tier of an unseen record with a macro F1 of 0.79. In the CAF stage, the tier label is fed back as an input to a classifier that predicts high-production block-years. Adding this single categorical feature raises F1 from about 0.27 to 0.60 on average across eight classifiers and lifts the area under the receiver operating characteristic curve from roughly 0.57 to 0.83; a McNemar test confirms that the improvement is significant ( p = 0.039). SHapley Additive exPlanations (SHAP) values show that tier membership ranks among the strongest contributors to the prediction, alongside lagged production. The same records also yield a four-group spatial map of the ten blocks and a six-group grouping of the twenty seasons, which together identify the district’s persistently weak blocks and its two crisis years. A cross-resolution check shows that the block- and year-level groupings agree moderately with the panel tiers in time but diverge in space, indicating that block-year panels carry information that spatial averages lose. The framework is simple, reproducible and interpretable, and it turns an unsupervised typology into an operational signal for targeting.