Parameter-Efficient Multilingual Vision–Language Adaptation for Spatio-Temporal Climate-Risk Interpretation from Satellite Imagery

Climate-risk monitoring requires models that connect satellite observations with natural-language questions while supporting users working in different languages. We present ClimateRS-VL, a parameter-efficient adaptation of Qwen2.5-VL-3B-Instruct for English, Gujarati, and Hindi climate-risk interpretation. The benchmark combines 7182 regional optical–mask pairs from six districts of Gujarat, 1026 nominal regional temporal pairs, 740 reservoir-year metadata records, 409 reservoir-year image samples, and 178,503 programmatically generated multilingual references. A prompt audit confirms that explicit answer-bearing numerical and categorical variables were absent from all 126,810 training, 25,854 validation, and 25,839 test prompts. On the full 25,839-response test evaluation, English, Gujarati, and Hindi VQA obtain BLEU scores of 83.9230, 80.9125, and 83.2524, with BERTScore-F1 values of 0.9891, 0.9961, and 0.9953. A same-backbone study on a common 3000-response subset yields an average VQA BLEU of 0.5625 for zero-shot Qwen2.5-VL, 1.1138 with contextual prompting but no adaptation, and 85.5191 with contextual prompting and LoRA. Question-aligned cross-lingual consistency is 0.9673, although this metric measures agreement rather than correctness. Class-balanced analysis further shows that raw accuracies above 0.98 on four regional risk questions arise from severe low-risk class prevalence: the model predicts the low-risk class for every sample, yielding a balanced accuracy of 0.5000. The protocol is a temporal holdout over recurring monitoring locations rather than an evaluation of unseen-location generalization. The results support parameter-efficient multilingual adaptation for structured in-domain generation while identifying limitations in numerical visual reasoning, minority-risk discrimination, reference diversity, and spatial generalization.

Authors

Institutions

Publication Details

Journal
Remote Sensing
Published
2026-11-17
DOI
https://doi.org/10.3390/rs18223945
Primary Topic
Multimodal Machine Learning Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Parameter-Efficient Multilingual Vision–Language Adaptation for Spatio-Temporal Climate-Risk Interpretation from Satellite Imagery

Ritesh Vyas, Akash Kumar Bhoi, Santosh Kumar Satapathy, Paolo Barsocchi et al.
2 links
Remote Sensing
Multimodal Machine Learning Applications
article

Parameter-Efficient Multilingual Vision–Language Adaptation for Spatio-Temporal Climate-Risk Interpretation from Satellite Imagery

Ritesh Vyas, Akash Kumar Bhoi, Santosh Kumar Satapathy, Paolo Barsocchi, Parita Vishal Shah
article en

Abstract

Climate-risk monitoring requires models that connect satellite observations with natural-language questions while supporting users working in different languages. We present ClimateRS-VL, a parameter-efficient adaptation of Qwen2.5-VL-3B-Instruct for English, Gujarati, and Hindi climate-risk interpretation. The benchmark combines 7182 regional optical–mask pairs from six districts of Gujarat, 1026 nominal regional temporal pairs, 740 reservoir-year metadata records, 409 reservoir-year image samples, and 178,503 programmatically generated multilingual references. A prompt audit confirms that explicit answer-bearing numerical and categorical variables were absent from all 126,810 training, 25,854 validation, and 25,839 test prompts. On the full 25,839-response test evaluation, English, Gujarati, and Hindi VQA obtain BLEU scores of 83.9230, 80.9125, and 83.2524, with BERTScore-F1 values of 0.9891, 0.9961, and 0.9953. A same-backbone study on a common 3000-response subset yields an average VQA BLEU of 0.5625 for zero-shot Qwen2.5-VL, 1.1138 with contextual prompting but no adaptation, and 85.5191 with contextual prompting and LoRA. Question-aligned cross-lingual consistency is 0.9673, although this metric measures agreement rather than correctness. Class-balanced analysis further shows that raw accuracies above 0.98 on four regional risk questions arise from severe low-risk class prevalence: the model predicts the low-risk class for every sample, yielding a balanced accuracy of 0.5000. The protocol is a temporal holdout over recurring monitoring locations rather than an evaluation of unseen-location generalization. The results support parameter-efficient multilingual adaptation for structured in-domain generation while identifying limitations in numerical visual reasoning, minority-risk discrimination, reference diversity, and spatial generalization.

Remote SensingVol. 18(22)
Istituto di Scienza e Tecnologie dell'Informazione "Alessandro Faedo" (IT), Symbiosis International University (IN), Pandit Deendayal Energy University (IN), National Research Council (IT), Birla Institute of Technology and Science, Pilani (IN)
Climate action
Multimodal Machine Learning Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon