Climate Projection and LULC Changes
Climate Projection at the Future
The climate data (temperature and precipitation) projections in future within the watershed have been studied using CanESM2 climate model for RCP2.6, RCP4.5, and RCP8.5 climate scenarios from the coupled model inter-comparison project-5 (CMIP5) experiments which have been downscaled by statistical downscaling model (SDSM). After trial and error to get the highest model performance by changing the values of bias correction and variance inflation in the SDSM model for precipitation, maximum temperature, and minimum temperature, the statistical results were estimated (Table 1- Table 3), and the mean values of the graphical results were drawn (Fig.3-Fig.5). The selected potential predictors for calibrating the model were ncepp8_ugl, ncepp8_thgl, nceps500gl, ncepshumgl, and nceptempgl, with 1.356 bias correction and 12 values of variance inflation for precipitation, ncepp1_ugl, ncepp1thgl, nceps500gl, nceps850gl, ncepshumgl, and nceptempgl predictors with the values of bias correction and variance inflation of 1 and 12 respectively, were used for the model calibrated for minimum temperature, and nceps500gl, nceps500gl. ncepp1_zgl, ncepp5_fgl, ncepp5_vgl, ncepp500gl, ncepp5thgl, ncepp8_vgl, ncepp8_zgl, and nceptempgl predictors with the values of bias correction and variance inflation of 1 and 12 respectively were used for model calibrating, validating, and testing for maximum temperature.
Table 1: The performance results of SDS model for downscaled precipitation after taking different trial and error
Period
|
RMSE
|
NSE
|
R
|
Calibration
|
3.795
|
0.319
|
0.584
|
Validation
|
1.529
|
0.309
|
0.597
|
Testing
|
RCP8.5
|
3.446
|
0.324
|
0.625
|
RCP4.5
|
3.429
|
0.331
|
0.613
|
RCP2.6
|
3.371
|
0.353
|
0.601
|
After calibrating and validating the model, the statistical evaluated values of RMSE, NSE, and R were found to be 3.446, 0.324 and 0.625, 3.429, 0.331 and 0.613, 3.371, 0.353 and 0.601 respectively for model performance downscaled precipitation during testing period under RCP8.5, RCP4.5 and RCP2.6 climate scenarios respectively. The SDS model super performed during the testing period under RCP2.6 climate scenario with downscaled precipitation, however, this scenario used to project the precipitation data at future time horizons during 2022-2050, 2051-2075 and 2076-2100 in the watershed varies a lot.
Table 2: The performance results of SDS model for downscaled minimum temperature after taking different trial and error
Period
|
RMSE
|
NSE
|
R
|
Calibration
|
1.586
|
0.509
|
0.722
|
Validation
|
1.514
|
0.582
|
0.779
|
Testing
|
RCP8.5
|
1.722
|
0.312
|
0.662
|
RCP4.5
|
1.697
|
0.332
|
0.670
|
RCP2.6
|
1.678
|
0.346
|
0.676
|
The statistical evaluated values of RMSE, NSE, and R were 1.722, 0.312 and 0.662, respectively, under RCP8.5, 1.697, 0.332, and 0.670 respectively, under RCP4.5, and 1.678, 0.346, and 0.676 respectively under RCP2.6 climate scenario (Table 2) during testing period. From this analysis, the SDS model has super performed during the testing period under RCP2.6 climate scenario with downscale minimum temperature, however, this scenario is used to project the minimum temperature data at future time horizons from 2022-2050, 2051-2075 and 2076-2100.
Table 3: The performance results of SDS model for downscaled maximum temperature after taking different trial and error
Period
|
RMSE
|
NSE
|
R
|
Calibration
|
1.265
|
0.489
|
0.715
|
Validation
|
1.179
|
0.522
|
0.780
|
Testing
|
RCP 8.5
|
1.443
|
0.114
|
0.624
|
RCP 4.5
|
1.418
|
0.144
|
0.646
|
RCP 2.6
|
1.429
|
0.129
|
0.628
|
The evaluated values of RMSE, NSE, and R were 1.443, 0.114, and 0.624, respectively, under RCP8.5, 1.418, 0.144, and 0.646 respectively under RCP4.5, and 1.429, 0.129, and 0.628 respectively under RCP2.6 climate scenario (Table 3) for the model performance of downscaled maximum temperature during testing period. From the analysis, the SDS model has super performed during the testing period under RCP4.5 climate scenario with downscale maximum temperature, therefore, this scenario is used to project the maximum temperature data at future time horizons from 2022-2050, 2051-2075 and 2076-2100.
3.1.2 LULC Changes and Scenario of the Future
Analysis LULC map in ArcGIS 10.5 by using the Landsat 8 and Landsat 7 images were downloaded from USGS for path 168, row 54 with different bands at different acquired years (2000, 2010, and 2020). The download images by using supervised classification method with seven different LULC types namely, Agricultural land, Bare land, Forest Area, Grass land, Settlement’s Area, Shrub land, and Water Body for each acquired year were declassified (Fig.6).
Based on the collected sample data confusion matrix (Table 4), the total sample points (TS) are 71, and the total corrected classified (TCS) value is 57, the sum of the product values in the total ground truth column and in the total user row is 1101. Substituting those values in to Equation1 and equation 2, the overall accuracy and kappa coefficient are 80.3% and 0.75, respectively. This implies 80.3% of land use and land cover classes are correctly classified.
Table 4: Creating confusion matrix based on the collected sample of ground truth and user classified
Class Name
|
AG
|
BL
|
F
|
GL
|
S
|
SL
|
WB
|
TGT
|
AG
|
25
|
0
|
1
|
0
|
0
|
1
|
0
|
27
|
BL
|
0
|
1
|
0
|
3
|
0
|
1
|
0
|
5
|
F
|
0
|
0
|
7
|
0
|
0
|
1
|
0
|
8
|
GL
|
2
|
0
|
2
|
3
|
1
|
0
|
0
|
8
|
S
|
0
|
0
|
0
|
0
|
6
|
0
|
0
|
6
|
SL
|
0
|
0
|
1
|
1
|
0
|
10
|
0
|
12
|
WB
|
0
|
0
|
0
|
0
|
0
|
0
|
5
|
5
|
TUC
|
27
|
1
|
11
|
7
|
7
|
13
|
5
|
71
|
Total Samples (TS)
|
71
|
Total Corrected Classified (TCS)
|
57
|
Overall Accuracy (%)
|
80.3
|
Kappa Coefficient (K)
|
0.75
|
Note: TGT Total ground truth and TUC- Total user classified
LULC change detection study was performed by the supervised classification method using the maximum likelihood classifier algorithm in ArcGIS 10.5 software during the period 2000-2020. Table 5 shows the changing area covering of each LULC class for the past 20 years.
Table 5: Changes in LULC from 2000 to 2020
Class Name
|
Areas Covered in 2000 (km2)
|
Areas Covered in 2010 (km2)
|
Areas Covered in 2020 (km2)
|
Change in %/year 2000-2010
|
Change in %/year 2010-2020
|
Change in %/year 2000-2020
|
AG
|
422.3
|
541.23
|
795.97
|
2.82
|
4.71
|
4.42
|
BL
|
16.46
|
15.681
|
12.414
|
-0.47
|
-2.08
|
-1.23
|
F
|
613.6
|
455.23
|
540.33
|
-2.58
|
1.87
|
-0.6
|
GL
|
241.6
|
219.87
|
55.768
|
-0.9
|
-7.46
|
-3.85
|
S
|
68.97
|
80.38
|
93.197
|
1.65
|
1.59
|
1.76
|
SL
|
1614.6
|
1664.6
|
1432.9
|
0.31
|
-1.39
|
-0.56
|
WB
|
0.004
|
0.516
|
46.926
|
14.24
|
8.99
|
6.52
|
Quantitative analysis of the overall LULC changes, decreases and increases in each class between 2000 and 2020. There is a considerable decrease in Forest (0.6%), grass land (3.85%), and bare land area (1.23%), and shrub lands (0.56%) per year were observed during this period (Table 5). On the other hand, there is an increase in agriculture land (4.42%), settlement area (1.76%), and surface water bodies (6.52%) for the same period. Based on the analysis, the future LULC change scenarios in the Kessem watershed for each class were decided as follows:
Scenario 1: Forest, bare land and shrub land area have been reduced and all grass land areas were covered by agriculture lands, settlement area, and surface water bodies during the period 2022–2050.
Scenario 2: Under this scenario, further reductions have been made in forest, bare land and shrub land area for the period 2051–2075. These will be afterwards covered by agriculture lands, settlement areas, and surface water bodies.
Scenario 3: Reduction has been predicted on the area covered by agriculture lands resulting in the formation of bare land during the period 2076-2100 and in addition forest and shrub lands will get reduced, increasing the settlement area and surface water bodies.
3.2 Flow/ Peak Flow Prediction at the Future Time Horizons
Effective rainfall, potential evapotranspiration [7] and Stream flow data are the main input datasets that were used in ML model. All these datasets used are observation data and estimated data based on observations to calibrate and validate the models.
The effective rainfall was estimated by using SCS-CN model considering the characteristics of sub-watershed of Kessem River. The PET was also computed using Hargreaves method based on the observed maximum and minimum temperature data at each available station. The complete stream flow data for Kessem River at Kessem Dam during the observed period of 1990-2013 were then transformed by using area ratio transform techniques from Aware Melka station to Kessem dam during the period of 1990-2009 and the observed stream flow at Kessem dam during the period of 2010-2013. Based on the climate projection data and LULC scenario, the future flow was predicted by using the performed ML models with the hybrid SCS-CN model. Using the daily time series data, the model was constructed using Kemel Tensor Flow package in Python 3. Training and testing were performed for the period 1990 to 2013, for which observed discharge data are available. In the network modeling, out of the total data, 70% (January 1990-October 2006) were selected for training and 30% (November, 2006 - December, 2013) for testing. Three different deep learning methods (LSTM, Bi-LSTM, GRU) for flow prediction during the historical period were implemented. Prediction models were subsequently applied to predict the flow for calibration and validation periods, and their performance was measured. Daily stream flow to the Kessem Dam reservoir in Kessem watershed was simulated using various deep learning models. The historical observation stream flow data were compared with the computed stream flow from RNN models (LSTM, Bi-LSTM, and GRU) using thirty lag days. A network was attempted to predict outcomes as accurately as possible. The value of this precision in the network is obtained by the cost function, which tries to penalize the network when it fails. The optimal output is the one with the lowest cost. For the applied networks of Mean Square Error [29], the cost function is used. A repetition step in training generally works with a division of training data named as a batch size. The number of samples for each batch is a hyper parameter, which is normally obtained by trial and error. The value of this parameter in all models is 128 in the best mode. In each repetition step, the cost function is computed as the mean MSE of these 128 samples of observed and predicted stream flow. The number of iteration steps for neural networks is named as an epoch and in each epoch, the stream flow time series is simulated by the network like other networks, neurons or network layers can be selected arbitrarily in recurrent networks. For the comparison of models with each other, the structures of all recurrent network models are created identically. In each network, a double hidden layer is used so that there are 12 units in each for the first layer and the second layer. The last layer output of the network at the final time step is linked to a dense layer with a single output neuron. Between the layers, a dropout equal to 10% is used. The structure of the neural network is also used in two hidden layers. The first and second layers have 12 neurons each. In all networks, the sigmoid activation function is applied for the hidden layer. The main advantage of using sigmoid is that, for all inputs greater than 0, there is a fixed derivative. This constant derivative speeds up network learning. Each method is run with different epoch numbers. After taken different trials and errors, the optimal hyper-parameter networks provide the details (Table 6).
Table 6: Optimal hyper-parameter network
Hyper-parameter
|
Values
|
Neuron
|
12
|
Optimization
|
Adam
|
Learning rate
|
0.001
|
Activation function
|
Sigmoid and Tanh
|
Max Epoch
|
4000
|
Batch size
|
128
|
The optimized model results were evaluated using Hydrostat packages with statistical error assessment techniques. In Hydrostats, statistical as well as graphical evaluations are made using error metrics function between observed and simulated flow. Graphically by plotting the predicted and observed flow (Fig.7) several descriptive statistics can be used for the evaluation of predictive models. In this study, RMSE, NSE, R2 have been purposively used, and the results of different methods based on the evaluation criteria are presented in Table 7. Among the RNN methods, Bi-LSTM performed the best.
Table 7: Statistical evaluation of the performance of ML models
ML Models
|
Training Period (1990-2006)
|
Testing (2006-2013)
|
RMSE
|
NSE
|
R2
|
KGE
|
RMSE
|
NSE
|
R2
|
KGE
|
BiLSTM
|
3.873
|
0.968
|
0.994
|
0.702
|
17.547
|
0.744
|
0.749
|
0.832
|
LSTM
|
4.276
|
0.962
|
0.974
|
0.777
|
19.878
|
0.672
|
0.727
|
0.690
|
GRU
|
4.003
|
0.966
|
0.982
|
0.743
|
20.703
|
0.644
|
0.681
|
0.712
|
The calculated discharges match well with the observed, as indicated by the high NSE and small RMSE values for the overall evaluation of the three ML models, revealing that the Bi-LSTM models outperform LSTM and GRU. Therefore, in this study, the result of Bi-LSTM model to predict the flow of Kessem River at Kessem dam within three time horizon is used.
3.3 Hydrologic Hazard Analysis for Kessem Dam
3.3.1 Inflow Hydrograph Shape
During 2022-2100, future period of inflow to Kessem Dam, the selected three events on September 2035 from 2022–2050-time horizon, September 2061 from 2051–2075-time horizon and September 2090 from the time horizon of 2076-2100 are the largest peak flow events that may occur in the Kessem dam watershed within the next 100 years.
Inflow hydrograph shapes for the three time horizons shown in Fig.8 were derived using the results from the predicted ML model. September 2035, September 2061, and September 2090 events were used for rescaling the sampled inflow flood events.
The PMF hydrograph was developed based on 9237m3/s (Qp) of the design of the peak inflow PMF for a 10,000-year return period of Kessem dam by using SCS dimensionless methods with a time to peak (Tp) of 33.39 hours, including the watershed characteristics (L = 136.64km = 448,294ft, Sl = 1.238%, CN = 79.41 and Tc = 51.016 hours) to compute discharge Q and the corresponding time t (Fig.9), which depicts the PMF hydrograph that represent the computed value of discharge Q versus time t. This hydrograph was used to compare with the results of peak discharge and stage frequency from hydrologic hazard analysis and to determine whether future flood events on the Kessem dam are at risk or not.
3.3.2 Inflow-Volume Frequency Curve
The developed volume-frequency curve of Kessem dam is based on the Log Pearson Type III distribution with mean, standard deviation, skew coefficient, and effective record length values for the future three time horizons. For the volume frequency analysis, the Bulletin 17C with EMA analysis was performed using HEC-SSP (Table 8).
Table 8: The result of statistics from volume frequency analysis for each future time horizon
Period/Statistics
|
mean (of log)
|
standard deviation (of log)
|
skew (of log)
|
effective record length
|
2022-2050
|
2.446
|
0.125
|
0.654
|
29
|
2051-2075
|
2.375
|
0.107
|
0.450
|
25
|
2076-2100
|
2.461
|
0.174
|
1.087
|
25
|
Based on the results of volume frequency analysis, the volume frequency curves within 90% uncertainty bounds were computed for each of the corresponding future time horizons (Fig.10).
3.3.3 Flood Seasonality Analysis
For the threshold value of 206.4m3/s, the frequency sample size during the period 2022-2050, 2051-2075, and 2076-2100 is 35, 31, and 37 respectively. Those are adequate sample size to analyse the flood seasonality for each time horizon. The flood seasonality histogram developed for this analysis for each time frame is presented in Fig.11. According to the result analysis, the annual flows are normal from October through May, with June-September as the wettest months, but the flood seasonal month is August during the period of 2022-2050, while the annual flows normally flow from November through June, with July-October as the wettest month but the flood seasonal month is September during both periods 2051-2075 and 2076-2100.
3.3.4 Reservoir Starting-Stage Duration Analysis
Initial reservoir levels and associated exceedance probabilities were estimated from daily reservoir elevation estimates for the period of record. The duration curve results indicate that the median reservoir elevation for the June through October period is approximately 926m, with a quartile range (25 to 75 percentiles) from about 922 to 928m (Fig.12). This reservoir elevation range was considered as the initial reservoir water surface elevation for routing the hydrographs. From the results, August produces the lowest pool duration curve. From the flood seasonality analysis section, it has been inferred that floods are most likely to occur in August and September. However, the dam is operated with consideration of this flood seasonality. Therefore, large events are most likely to occur in August and September, but they are also most likely to have low reservoir starting pools, mitigating some of the risk for large peak stage events in summer.
3.3.5 Hydrologic Hazard Curve of Kessem Dam
Once RMC-RFA is computed, it automatically creates the Stage-Frequency Curve and Hydrologic Hazard Curve plots (Figs.13 and 14). The median curve represents the uncertainty in stage frequency and peak discharge frequency due to natural variability. The 95% uncertainty bounds represent the uncertainty in stage and peak discharge frequency due to knowledge uncertainty, whereas the expected curve represents the combined uncertainty due to both natural variability and knowledge uncertainty. Those curves are used for semi quantitative risk analysis for Kessem dam.
HHA produces the expected peak discharge and the corresponding peak stage for 100 to 1,000,000 years of return period for each time horizon (Tables 9 and 10), respectively.
Table 9: Expected probable peak discharge (m3/s) for each future time horizons
Return period
|
2022-2050
|
2051-2075
|
2076-2100
|
Expected
|
95 % Bounds
|
Expected
|
95 % Bounds
|
Expected
|
95 % Bounds
|
100
|
867.67
|
669.3-1168
|
809.4
|
621-1082
|
1,230.9
|
759-2,308
|
1000
|
1,383.95
|
835-2294.8
|
1,178.9
|
705-1796
|
2,770.7
|
1,003-8,013
|
10,000
|
2,823.57
|
987-5430.6
|
2,126.3
|
790-3468
|
11,491.1
|
1,357-17,217
|
100,000
|
5,686.2
|
1,115-10,429.3
|
3,738.5
|
851-5811
|
17,292.3
|
1,636-17,726
|
1,000,000
|
12,916.72
|
1,147-15,231.7
|
7,104.67
|
887-9625
|
17,733.7
|
1,900-17,777
|
Table 10: Expected probable peak stage (m) for each future time horizon
Return period
|
2022-2050
|
2051-2075
|
2076-2100
|
Expected
|
95 % Bounds
|
Expected
|
95 % Bounds
|
Expected
|
95 % Bounds
|
100
|
931.92
|
931.4-932.6
|
931.77
|
931.2-932.4
|
932.7
|
931.6-934.5
|
1000
|
932.94
|
931.7-934.5
|
932.54
|
931.5-933.7
|
935.13
|
932.2-940.9
|
10,000
|
935.21
|
932.1-938.4
|
934.18
|
931.7-936.1
|
942.11
|
932.8-943
|
100,000
|
938.59
|
932.2-941.8
|
936.37
|
931.8-938.7
|
943
|
933.3-943
|
1,000,000
|
942.42
|
932.4-943
|
940.1
|
931.9-941.6
|
943
|
933.7-943
|
Kessem Dam has a spillway discharge capacity of 6180m3/s at the maximum water surface elevation of 939.5m. Comparing this value with the stage and peak discharge frequency curve, it indicates that the spillway is capable of passing a flood with a return period of 100 -100,000 years for the future time horizon (2022-2075).
During the period of 2022–2050, the expected peak discharge for 1/10000 AEP was equal to 2,823.57m3/s. The 10,000-year peak discharge at 95% confidence upper and lower limits is 987m3/s and 5,430.6m3/s, respectively, from the hydrological hazard analysis. The corresponding expected peak stage for 0.0001 APE is 935.21m, and the lower and upper 95% of the bounds values are 932.1m and 938.4m, respectively. It has not exceeded the PMF discharge of 6180m3/s and the maximum water surface elevation of 939.5m.
During the period of 2051–2075, the expected peak discharge for 1/10000 AEP is equal to 2,126.3m3/s. The peak discharge at 95% confidence upper and lower limits is 790m3/s and 3468m3/s, respectively. The corresponding peak stage for the expected value and 95% lower and upper bounds values of 0.0001 APE is 934.18m, 931.7m, and 936.1m, respectively. The PMF discharge of 6180m3/s and maximum water surface elevation of 939.5m are also not exceeded by those values.
During the period of 2076–2100, the expected peak discharge for a 1/10000 AEP was equal to 11,491.1m3/s. The 10,000-year peak discharge at 95% confidence upper and lower limits is 1,357m3/s and 17,217m3/s, respectively, from the hydrological hazard analysis. The corresponding expected peak stage for 0.0001APE is 942.11m, and the lower and upper 95% of the bounds values are 932.8m and 943m, respectively. It exceeds that from the PMF discharge of 6180m3/s and the maximum water surface elevation of 939.5m.
The results from this initial hydrologic hazard curve characterization and flood hydrograph (Fig.14) routing indicate that Kessem Dam may potentially be overtopped by a flood with a return period of about 10,000 years during the period of 2076-2100. However, this indicates that Kessem Dam does not meet Reclamation hydrologic hazard criteria for overtopping because it does not pass through a PMF for 2076-2100 future time horizons. Therefore, the dam requires further risk analysis study and dam safety modification to control this probable failure mode during the period of 2076 - 2100.