Assessing Algorithm Performance in Satellite-Based Banana Area Mapping in Rayalaseema Zone of Andhra Pradesh, India

Abstract

In this study, a simple methodology used for estimation banana growing area in Andhra Pradesh. The methodology consists algorithms process the satellite data by using Landsat imagery from the 2024 growing season and four classification techniques such as Random Forest, K-Nearest Neighbours, its optimized KD-tree KNN, and Maximum Likelihood classification were tested. The results revealed that among the different techniques huge differences was observed, the KNN method estimated 53,239 hectares under cultivation, while Random Forest produced a more conservative 45,689 hectares. This substantial variation (about 16%) highlights how methodological choices can dramatically impact area estimates. Several important patterns emerged from our analysis. The KNN approaches, while effective at capturing spectral variability in mixed agricultural landscapes, consistently generated higher estimates. Random Forest, though more restrained, may better reflect actual planting areas. The current study suggests that combining multiple approaches with ground verification may offer the most reliable estimates, especially for smallholder plots that dominate the region's horticulture landscape.