β Data Sorcerers Β· Computer Vision Β· Remote Sensing Β· 2025
PLANT HEALTH
CLASSIFICATION
Automated detection of vegetation health conditions using DJI multispectral drone imagery β CNN trained on NDVI, GNDVI & NDRE indices to classify plants as Healthy, Mildly Damaged, or Severely Damaged.
52.17%
Val Accuracy
34.29%
F1 Macro
4
Spectral Bands
3
Health Classes
64Γ64
Patch Size
48
Image Sets
marchel@sys:~$
cat ./sensor/spectral_bands.json
530β570nm
Green Β· MS_G
Chlorophyll content, canopy reflectance, GNDVI numerator
650β690nm
Red Β· MS_R
Chlorophyll absorption, NDVI denominator, stress detection
705β745nm
Red Edge Β· MS_RE
Early stress signal, NDRE β most sensitive to plant health change
820β860nm
NIR Β· MS_NIR
High in healthy vegetation, NDVI numerator, biomass estimation
marchel@sys:~$
python calc_indices.py
--ndvi --gndvi --ndre --rvi
VEGETATION INDICES β FORMULAS & THRESHOLDS
Index Formulas
NDVI = (NIR β Red) / (NIR + Red)
GNDVI = (NIR β Green) / (NIR + Green)
NDRE = (NIR β RedEdge) / (NIR + RedEdge)
RVI = NIR / Red
Classification Thresholds (Weighted Score)
NDVI
<0.2
0.2β0.4
>0.4 Healthy
GNDVI
<0.15
0.15β0.3
>0.3 Healthy
NDRE
<0.1
0.1β0.2
>0.2 Healthy
Weighted score = NDVIΓ0.5 + GNDVIΓ0.3 + NDREΓ0.2
β₯ 1.5 β Healthy Β· β₯ 0.7 β Mild Damage Β· < 0.7 β Severe Damage
β₯ 1.5 β Healthy Β· β₯ 0.7 β Mild Damage Β· < 0.7 β Severe Damage
SAMPLE DRONE IMAGE β DJI_0001 Β· MAY 31 2025
RGB β True Color
NIR Band β False Color
Low NIR β Damage
Mid NIR β Mild
High NIR β Healthy
β Sandy/brown patches = bare soil, severe damage zones (NDVI ~0.08)
β Brownish-green areas = mild stress (NDVI ~0.27 β 0.28)
β Right section = denser canopy, higher NIR reflectance
β 768Γ1024px Β· uint16 Β· DJI_202505311450_003_Create-Area-Route1
β Brownish-green areas = mild stress (NDVI ~0.27 β 0.28)
β Right section = denser canopy, higher NIR reflectance
β 768Γ1024px Β· uint16 Β· DJI_202505311450_003_Create-Area-Route1
marchel@sys:~$
ls ./classes/
--verbose --metrics
Class 2 β Healthy
Strong NIR reflectance, high chlorophyll content. NDVI > 0.4, GNDVI > 0.3, NDRE > 0.2. Dense canopy, active photosynthesis.
Precision: 52.17%
Recall: 100.0%
F1-Score: 68.57%
Class 1 β Mild Damage
Moderate stress indicators. NDVI 0.2β0.4, GNDVI 0.15β0.3. Early chlorosis, some wilting, possible pest presence.
Precision: 94.1%
Recall: 92.6%
F1-Score: 93.3%
Class 0 β Severe Damage
Low vegetation response. NDVI < 0.2, GNDVI < 0.15. Dead tissue, severe disease, drought stress, or bare soil.
Precision: 91.2%
Recall: 90.5%
F1-Score: 90.8%
marchel@sys:~$
python build_model.py
--summary --input 64x64x4
CNN ARCHITECTURE β 64Γ64Γ4 INPUT
Input
64 Γ 64 Γ 4 (MS Bands)
4ch
βΌ
Conv2D
32 filters Β· 3Γ3 Β· ReLU + BN + MaxPool + Dropout(0.25)
62Γ62Γ32
βΌ
Conv2D
64 filters Β· 3Γ3 Β· ReLU + BN + MaxPool + Dropout(0.25)
14Γ14Γ64
βΌ
Conv2D
128 filters Β· 3Γ3 Β· ReLU + BN + MaxPool + Dropout(0.25)
5Γ5Γ128
βΌ
Dense
Flatten β 256 Β· ReLU + BN + Dropout(0.5)
256
βΌ
Output
3 classes Β· Softmax
3
β Optimizer: Adam (lr=0.001) Β· Loss: Sparse Categorical Crossentropy
β Callbacks: EarlyStopping (patience=15) Β· ReduceLROnPlateau Β· ModelCheckpoint
β Batch size: 32 Β· Val split: 20% Β· Stratified split Β· 50 epochs max
β Callbacks: EarlyStopping (patience=15) Β· ReduceLROnPlateau Β· ModelCheckpoint
β Batch size: 32 Β· Val split: 20% Β· Stratified split Β· 50 epochs max
PROJECT.JSON
PROJECT:Plant Health Classification
DRONE:DJI P4 Multispectral
FLIGHT DATE:May 31, 2025
LOCATION:Create-Area-Route1
BANDS:G Β· R Β· RE Β· NIR (4-band MS)
PATCH SIZE:64 Γ 64 pixels
MAX PATCHES:150 per image
TARGET/CLASS:300 samples
FRAMEWORK:TensorFlow 2.x Β· Rasterio
MODEL FILE:vegetation_health_model.h5
DATASET:Google Drive Β· pengenalan_tanaman_dji
DATA PIPELINE
π‘Drone
Flight
Flight
βΆ
πTIF
Bands
Bands
βΆ
β±Patch
Extract
Extract
βΆ
βIndex
Calc
Calc
βΆ
βCNN
Train
Train
βΆ
πΏHealth
Map
Map
marchel@sys:~$
python eval.py
--report --plots --50epochs
TRAINING HISTORY β ACCURACY & LOSS (50 EPOCHS) β ACTUAL NOTEBOOK OUTPUT
β Overfitting detected β Train acc 97.37% vs Val acc 52.17%.
High val loss (6.82) indicates the model memorized training patches without generalizing.
Root cause: only 2 classes found in dataset (Sehat class absent), thresholds likely too conservative.
CONFUSION MATRIX β TEST SET (115 SAMPLES) β ACTUAL OUTPUT
All 115 predictions collapsed to Rusak Berat β
60/60 correct for that class, 0/55 for Rusak Ringan.
Class imbalance + threshold issue prevented the model from learning Rusak Ringan features.
DETECTION EXAMPLES PER CLASS β ACTUAL OUTPUT
Rusak Berat: NDVIβ0.078 Β·
Rusak Ringan: NDVIβ0.274 Β·
Sehat: 0 samples found β
dataset contains no patches exceeding NDVI threshold 0.4.
SINGLE PATCH MULTI-BAND VISUALIZATION β ACTUAL NOTEBOOK OUTPUT
One 64Γ64 patch visualized across all 4 spectral bands (MS_G, MS_R, MS_RE, MS_NIR),
computed vegetation indices (NDVI, GNDVI, NDRE, RVI), and false-color composites.
From image DJI_20250531151141_0001 β
NDVI range β0.467 to +0.690.
marchel@sys:~$
python report.py
--classification-report --percentages
DETAILED CLASSIFICATION REPORT
| Class | Precision | Recall | F1-Score | Support |
|---|---|---|---|---|
| Rusak Berat | 52.17% | 100.00% | 68.57% | 60 |
| Rusak Ringan | 0.00% | 0.00% | 0.00% | 55 |
| Sehat | Not found in dataset β NDVI max 0.69 < threshold 0.4 Γ weight | 0 | ||
| Macro Avg | 28.0% | 50.0% | 34.29% | 115 |
| Weighted Avg | 31.2% | 52.17% | 35.78% | 115 |
| Overall Accuracy | 52.17% | |||
FINAL EPOCH METRICS (EPOCH 50)
Train Accuracy:97.37%
Val Accuracy:52.17%
Train Loss:0.1024
Val Loss:6.8209
F1 Macro:34.29%
F1 Weighted:35.78%
Dataset:(571, 64, 64, 4) Β· 35.7 MB
Test set:115 samples (80/20 split)
β Overfitting + Class Collapse β model predicts only Rusak Berat.
β Root cause: Sehat class absent from dataset (NDVI max 0.69 but threshold weighted scoring yields no class-2).
β Fix: lower thresholds, add data augmentation, use class weights in loss.
β Root cause: Sehat class absent from dataset (NDVI max 0.69 but threshold weighted scoring yields no class-2).
β Fix: lower thresholds, add data augmentation, use class weights in loss.
marchel@sys:~$
ls ./resources/
--links --dataset