Classification algorithms for selecting crop types in Asian countries based on soil and climate parameters
Classifying crops from environmental conditions can support agricultural data exploration, but an observed crop is not necessarily the most suitable crop for that land. We compare multinomial logistic regression, k-nearest neighbors (kNN), and random forest on a transparent seven-country South Asian grid benchmark. Crop-area and surface-soil layers came from the open ACASA 50 km collection; annual mean temperature and precipitation were sampled from WorldClim 2.1 climate normals. For each grid cell, the largest of 12 crop-area layers defined the observed dominant-crop label. Cells without mapped crops, incomplete inputs, or less than 50% dominance were removed, and classes with fewer than 30 records were excluded, leaving 1,070 cells in seven crop classes. No crop-area layer was used as a predictor. Disjoint 2° spatial blocks supplied 642 training, 215 validation, and 213 test cells. On the held-out test blocks, kNN achieved 73.7% accuracy and 0.461 macro-F1, versus 70.9% and 0.441 for random forest and 70.0% and 0.398 for logistic regression. Its advantage over random forest in accuracy was 2.8 percentage points, but a paired block-bootstrap interval crossed zero. Leave-one-country-out tests were markedly less reliable, especially in Afghanistan and Pakistan. The results identify a modestly stronger within-region classifier for mapped crop dominance under this protocol; they do not validate agronomic suitability, yield benefit, or transfer to new countries.
Authors
- Jahongir Keldibekov
Publication Details
- Journal
- InterConf
- Published
- 2026-11-20
- DOI
- https://doi.org/10.51582/interconf.19-20.11.2026.034
- Primary Topic
- Remote Sensing in Agriculture
- Type
- article
- Field-Weighted Citation Impact
- 0.00